Daily Papers

Surgical scene understanding is critical for surgical training and robotic decision-making in robot-assisted surgery. Recent advances in Multimodal Large Language Models (MLLMs) have demonstrated great potential for advancing scene perception in the medical domain, facilitating surgeons to understand surgical scenes and procedures. However, these methods are primarily oriented towards image-based analysis or global video understanding, overlooking the fine-grained video reasoning that is crucial for analyzing specific processes and capturing detailed task execution within a surgical procedure. To bridge this gap, we propose SurgVidLM, the first video language model designed to address both full and fine-grained surgical video comprehension. To train our SurgVidLM, we construct the SVU-31K that is a large-scale dataset with over 31K video-instruction pairs, enabling both holistic understanding and detailed analysis of surgical procedures. Building on this resource, SurgVidLM incorporates a two-stage StageFocus mechanism: the first stage extracts global procedural context, while the second stage performs high-frequency local analysis guided by temporal cues. We also develop the Multi-frequency Fusion Attention to effectively integrate low- and high-frequency visual tokens, ensuring the preservation of critical task-specific details. Experimental results demonstrate that SurgVidLM significantly outperforms state-of-the-art Vid-LLMs of comparable parameter scale in both full and fine-grained video understanding tasks, showcasing its superior capability in capturing the context of complex robot-assisted surgeries. Our code and dataset will be publicly accessible soon.

Recent works have shown that the computational efficiency of video recognition can be significantly improved by reducing the spatial redundancy. As a representative work, the adaptive focus method (AdaFocus) has achieved a favorable trade-off between accuracy and inference speed by dynamically identifying and attending to the informative regions in each video frame. However, AdaFocus requires a complicated three-stage training pipeline (involving reinforcement learning), leading to slow convergence and is unfriendly to practitioners. This work reformulates the training of AdaFocus as a simple one-stage algorithm by introducing a differentiable interpolation-based patch selection operation, enabling efficient end-to-end optimization. We further present an improved training scheme to address the issues introduced by the one-stage formulation, including the lack of supervision, input diversity and training stability. Moreover, a conditional-exit technique is proposed to perform temporal adaptive computation on top of AdaFocus without additional training. Extensive experiments on six benchmark datasets (i.e., ActivityNet, FCVID, Mini-Kinetics, Something-Something V1&V2, and Jester) demonstrate that our model significantly outperforms the original AdaFocus and other competitive baselines, while being considerably more simple and efficient to train. Code is available at https://github.com/LeapLabTHU/AdaFocusV2.

Vision-language-action (VLA) models have become an important paradigm for robotic manipulation by unifying visual perception, language understanding, and action generation. Since reliable manipulation requires sufficient observation of objects, end-effectors, and their spatial relationships, multi-view inputs are widely used to provide complementary spatial information. However, our experiments show that existing VLA models do not fully exploit such multi-view complementarity. The policy tends to persistently rely on the wrist view, lacking the ability to dynamically adjust global-local view dependency across manipulation stages, which limits stage adaptability and robustness. To address this issue, we propose FocusVLA, a stage-aware multi-view routing framework. FocusVLA predicts future interaction intent from historical actions and uses it to guide dynamic view routing during action generation. By adaptively coordinating global and local observations across different task stages, FocusVLA improves multi-view utilization. On RoboTwin 2.0, FocusVLA improves the average success rate over the strongest baseline π_{0.5} by 10.10 pp. In real-world evaluation, FocusVLA improves the success rate from 71.7\% to 85.0\%, achieving a 13.3\% gain over π_{0.5}. Moreover, the intent prediction and dynamic routing modules of FocusVLA can be integrated into existing VLA frameworks without modifying their core architectures.

Focus is a cornerstone of photography, yet autofocus systems often fail to capture the intended subject, and users frequently wish to adjust focus after capture. We introduce a novel method for realistic post-capture refocusing using video diffusion models. From a single defocused image, our approach generates a perceptually accurate focal stack, represented as a video sequence, enabling interactive refocusing and unlocking a range of downstream applications. We release a large-scale focal stack dataset acquired under diverse real-world smartphone conditions to support this work and future research. Our method consistently outperforms existing approaches in both perceptual quality and robustness across challenging scenarios, paving the way for more advanced focus-editing capabilities in everyday photography. Code and data are available at www.learn2refocus.github.io

In text-to-image generation tasks, the advancements of diffusion models have facilitated the fidelity of generated results. However, these models encounter challenges when processing text prompts containing multiple entities and attributes. The uneven distribution of attention results in the issues of entity leakage and attribute misalignment. Training from scratch to address this issue requires numerous labeled data and is resource-consuming. Motivated by this, we propose an attribution-focusing mechanism, a training-free phase-wise mechanism by modulation of attention for diffusion model. One of our core ideas is to guide the model to concentrate on the corresponding syntactic components of the prompt at distinct timesteps. To achieve this, we incorporate a temperature control mechanism within the early phases of the self-attention modules to mitigate entity leakage issues. An object-focused masking scheme and a phase-wise dynamic weight control mechanism are integrated into the cross-attention modules, enabling the model to discern the affiliation of semantic information between entities more effectively. The experimental results in various alignment scenarios demonstrate that our model attain better image-text alignment with minimal additional computational cost.

DiT diffusion models have achieved great success in text-to-video generation, leveraging their scalability in model capacity and data scale. High content and motion fidelity aligned with text prompts, however, often require large model parameters and a substantial number of function evaluations (NFEs). Realistic and visually appealing details are typically reflected in high resolution outputs, further amplifying computational demands especially for single stage DiT models. To address these challenges, we propose a novel two stage framework, FlashVideo, which strategically allocates model capacity and NFEs across stages to balance generation fidelity and quality. In the first stage, prompt fidelity is prioritized through a low resolution generation process utilizing large parameters and sufficient NFEs to enhance computational efficiency. The second stage establishes flow matching between low and high resolutions, effectively generating fine details with minimal NFEs. Quantitative and visual results demonstrate that FlashVideo achieves state-of-the-art high resolution video generation with superior computational efficiency. Additionally, the two-stage design enables users to preview the initial output before committing to full resolution generation, thereby significantly reducing computational costs and wait times as well as enhancing commercial viability .

Different stages of manipulation tasks exhibit varying levels of difficulty, suggesting stage-dependent motion speeds and temporal prediction horizons. However, existing IL-based visuomotor policies typically imitate the execution speed of expert demonstrations and operate with a fixed temporal prediction horizon, limiting flexibility and overall task throughput. In this paper, we introduce AutoSpeed, a model-agnostic learning framework that enables existing visuomotor policies to predict trajectories with stage-adaptive motion speeds, without requiring speed or stage annotations. We treat future trajectories at different speeds as candidate optimization targets, evaluate each candidate using a composite cost that trades off prediction error against prediction horizon, and optimize the policy toward the minimum-cost candidate. With a fixed-length action sequence, speed modulation adjusts the effective temporal prediction horizon: simple stages are executed faster with a longer prediction horizon, whereas complex stages are executed more slowly with a shorter prediction horizon. Specifically, we implement speed modulation in the frequency domain via the discrete cosine transform (DCT), which enables smooth, non-integer speed scaling and thus preserves motion continuity. Extensive evaluations show that AutoSpeed substantially reduces task execution time while also improving success rates. Under the AutoSpeed framework, the inferred motion speeds exhibit a strong correspondence with task stages.

We introduce RegionFocus, a visual test-time scaling approach for Vision Language Model Agents. Understanding webpages is challenging due to the visual complexity of GUI images and the large number of interface elements, making accurate action selection difficult. Our approach dynamically zooms in on relevant regions, reducing background clutter and improving grounding accuracy. To support this process, we propose an image-as-map mechanism that visualizes key landmarks at each step, providing a transparent action record and enables the agent to effectively choose among action candidates. Even with a simple region selection strategy, we observe significant performance gains of 28+\% on Screenspot-pro and 24+\% on WebVoyager benchmarks on top of two state-of-the-art open vision language model agents, UI-TARS and Qwen2.5-VL, highlighting the effectiveness of visual test-time scaling in interactive settings. We achieve a new state-of-the-art grounding performance of 61.6\% on the ScreenSpot-Pro benchmark by applying RegionFocus to a Qwen2.5-VL-72B model. Our code will be released publicly at https://github.com/tiangeluo/RegionFocus.

Multi-exposure fusion (MEF) expands the luminance range beyond what a single exposure can capture. Combining images taken at different exposure levels requires handling geometric differences while naturally merging their complementary brightness information. It often demands generative completion where details are missing. Diffusion-based generative methods address these challenges, however, they are computationally expensive and struggle to preserve fine structures in saturated regions. We propose LIIFusion, a coarse-to-fine framework that balances fusion quality and efficiency in generative MEF. The coarse stage performs low resolution generative fusion, enhanced by an adaptive exposure correction that recovers structure lost in saturated over-exposed areas. The fine stage adapts a local implicit image function into a multi-exposure fusion function: conditioned on the HR OE/UE sources and the coarse output, it queries arbitrary target coordinates and fuses source evidence regardless of the HR input resolution. LIIFusion achieves up to 3.5times speed-up over existing generative methods while maintaining or improving structural fidelity and perceptual quality. We believe this framework provides an effective pathway toward making generative MEF more practical in real-world applications.

We present FlashLips, a two-stage, mask-free lip-sync system that decouples lips control from rendering and achieves real-time performance running at over 100 FPS on a single GPU, while matching the visual quality of larger state-of-the-art models. Stage 1 is a compact, one-step latent-space editor that reconstructs an image using a reference identity, a masked target frame, and a low-dimensional lips-pose vector, trained purely with reconstruction losses - no GANs or diffusion. To remove explicit masks at inference, we use self-supervision: we generate mouth-altered variants of the target image, that serve as pseudo ground truth for fine-tuning, teaching the network to localize edits to the lips while preserving the rest. Stage 2 is an audio-to-pose transformer trained with a flow-matching objective to predict lips-poses vectors from speech. Together, these stages form a simple and stable pipeline that combines deterministic reconstruction with robust audio control, delivering high perceptual quality and faster-than-real-time speed.

Graphical User Interface (GUI) agents increasingly act over streaming omni-modal observations spanning screenshots, audio, video, interaction history, and device states. However, exposing this complete context uniformly at every step ignores that relevance is transient and sparse: the signal determining one action can become clutter at the next. Because executed actions feed back into the context, these errors compound into Sequential Distraction, driving the agent into irreversible trajectory drift. We formulate the underlying requirement as Focused Perception: a representation that is step-adaptive, action-sufficient, and non-intrusive. We realize this in FocusGUI, a lightweight adapter that dynamically prioritizes multimodal channels against a step-specific decision reference. Instead of uniformly processing all inputs, it compresses the most relevant evidence into a compact latent and subtly injects it into the frozen decoder's embeddings, guiding the model's attention without altering its native parameter space. FocusGUI trains from action supervision alone, requiring neither modality-relevance labels nor backbone fine-tuning. Across four omni-modal backbones, it raises overall Type Match by +33.2 and Exact Match by 3.0 times over the frozen backbone on average, and Type Match by +30.1 over the strongest existing enhancement method.

Echocardiography is crucial for cardiovascular disease detection but relies heavily on experienced sonographers. Echocardiography probe guidance systems, which provide real-time movement instructions for acquiring standard plane images, offer a promising solution for AI-assisted or fully autonomous scanning. However, developing effective machine learning models for this task remains challenging, as they must grasp heart anatomy and the intricate interplay between probe motion and visual signals. To address this, we present EchoWorld, a motion-aware world modeling framework for probe guidance that encodes anatomical knowledge and motion-induced visual dynamics, while effectively leveraging past visual-motion sequences to enhance guidance precision. EchoWorld employs a pre-training strategy inspired by world modeling principles, where the model predicts masked anatomical regions and simulates the visual outcomes of probe adjustments. Built upon this pre-trained model, we introduce a motion-aware attention mechanism in the fine-tuning stage that effectively integrates historical visual-motion data, enabling precise and adaptive probe guidance. Trained on more than one million ultrasound images from over 200 routine scans, EchoWorld effectively captures key echocardiographic knowledge, as validated by qualitative analysis. Moreover, our method significantly reduces guidance errors compared to existing visual backbones and guidance frameworks, excelling in both single-frame and sequential evaluation protocols. Code is available at https://github.com/LeapLabTHU/EchoWorld.

Humans can effortlessly locate desired objects in cluttered environments, relying on a cognitive mechanism known as visual search to efficiently filter out irrelevant information and focus on task-related regions. Inspired by this process, we propose Dyfo (Dynamic Focus), a training-free dynamic focusing visual search method that enhances fine-grained visual understanding in large multimodal models (LMMs). Unlike existing approaches which require additional modules or data collection, Dyfo leverages a bidirectional interaction between LMMs and visual experts, using a Monte Carlo Tree Search (MCTS) algorithm to simulate human-like focus adjustments. This enables LMMs to focus on key visual regions while filtering out irrelevant content, without introducing additional training caused by vocabulary expansion or the integration of specialized localization modules. Experimental results demonstrate that Dyfo significantly improves fine-grained visual understanding and reduces hallucination issues in LMMs, achieving superior performance across both fixed and dynamic resolution models. The code is available at https://github.com/PKU-ICST-MIPL/DyFo_CVPR2025

We present MicroCinema, a straightforward yet effective framework for high-quality and coherent text-to-video generation. Unlike existing approaches that align text prompts with video directly, MicroCinema introduces a Divide-and-Conquer strategy which divides the text-to-video into a two-stage process: text-to-image generation and image\&text-to-video generation. This strategy offers two significant advantages. a) It allows us to take full advantage of the recent advances in text-to-image models, such as Stable Diffusion, Midjourney, and DALLE, to generate photorealistic and highly detailed images. b) Leveraging the generated image, the model can allocate less focus to fine-grained appearance details, prioritizing the efficient learning of motion dynamics. To implement this strategy effectively, we introduce two core designs. First, we propose the Appearance Injection Network, enhancing the preservation of the appearance of the given image. Second, we introduce the Appearance Noise Prior, a novel mechanism aimed at maintaining the capabilities of pre-trained 2D diffusion models. These design elements empower MicroCinema to generate high-quality videos with precise motion, guided by the provided text prompts. Extensive experiments demonstrate the superiority of the proposed framework. Concretely, MicroCinema achieves SOTA zero-shot FVD of 342.86 on UCF-101 and 377.40 on MSR-VTT. See https://wangyanhui666.github.io/MicroCinema.github.io/ for video samples.

We present DualFocus, a novel framework for integrating macro and micro perspectives within multi-modal large language models (MLLMs) to enhance vision-language task performance. Current MLLMs typically singularly focus on inputs at a predefined resolution, resulting in deficiencies in detailed questions involving local regions. We introduced a DualFocus mechanism where the model concentrates on the image from a macro perspective, responses to the question, and identifies suitable sub-regions to zoom in for subsequent micro perspective analysis. Via the integration of answers from both macro and micro perspectives, the model is adept at addressing tasks that encompass global, detailed, and combined considerations. To endows the DualFocus mechanism in MLLMs, we curated a tailored dataset derived from the Visual Genome (VG) and adapted it to align with the training regimen of DualFocus. Through comparative studies across different model sizes and benchmarks, we demonstrate DualFocus's superiority in balancing detailed examination with holistic insight, significantly reducing hallucination instances in MLLMs and improving their performance in various vision-language tasks.

Image retouching is commonly formulated as enhancing overall visual quality through color adjustment, but in practice, it also serves to emphasize visual focus by guiding viewers' attention toward a specific subject or region. Achieving such focus-oriented retouching is inherently challenging, as it requires well-coordinated global and local adjustments to manipulate perceptual saliency while maintaining visual naturalness. This intricate process typically demands substantial professional expertise. In this study, we propose EyeControl, a MLLM-driven agent with a diffusion-based retouching executor that enables visual focus enhancement under weak user intent. With only a few clicks or coarse strokes, EyeControl directs visual attention to the intended region, effectively "dotting the eye" of the image. The core idea is to explicitly link the weak user intention with the target editing region and the corresponding tonal adjustment operations during retouching. To achieve this, the system first interprets the intent and image content to infer the visual focus and generate structured intent guidance for the retouching executor. Second, the retouching executor is encouraged to respond more strongly to the target region, explicitly aligning its attention map with a designed pseudo-intent map. We also introduce an operation-consistency constraint to improve coordination between global and local adjustments, achieving more natural and coherent retouching. Additionally, we contribute ControlArt-Bench, a high-quality evaluation dataset for visual focus enhancement. Extensive evaluations demonstrate that EyeControl yields perceptually appealing results with stronger intent alignment. Code can be found in https://github.com/DragonisCV/EyeControl.

Pre-trained flow-based models excel at synthesizing complex scenes yet lack a direct mechanism for disentangling and customizing their underlying concepts from one-shot real-world sources. To demystify this process, we first introduce a novel differential probing technique to isolate and analyze the influence of individual concept tokens on the velocity field over time. This investigation yields a critical insight: the generative process is not monolithic but unfolds in three distinct stages. An initial Blueprint Stage establishes low-frequency structure, followed by a pivotal Instantiation Stage where content concepts emerge with peak intensity and become naturally disentangled, creating an optimal window for manipulation. A final concept-insensitive refinement stage then synthesizes fine-grained details. Guided by this discovery, we propose ConceptWeaver, a framework for one-shot concept disentanglement. ConceptWeaver learns concept-specific semantic offsets from a single reference image using a stage-aware optimization strategy that aligns with the three-stage framework. These learned offsets are then deployed during inference via our novel ConceptWeaver Guidance (CWG) mechanism, which strategically injects them at the appropriate generative stage. Extensive experiments validate that ConceptWeaver enables high-fidelity, compositional synthesis and editing, demonstrating that understanding and leveraging the intrinsic, staged nature of flow models is key to unlocking precise, multi-granularity content manipulation.

Diffusion models (DMs) have become the leading choice for generative tasks across diverse domains. However, their reliance on multiple sequential forward passes significantly limits real-time performance. Previous acceleration methods have primarily focused on reducing the number of sampling steps or reusing intermediate results, failing to leverage variations across spatial regions within the image due to the constraints of convolutional U-Net structures. By harnessing the flexibility of Diffusion Transformers (DiTs) in handling variable number of tokens, we introduce RAS, a novel, training-free sampling strategy that dynamically assigns different sampling ratios to regions within an image based on the focus of the DiT model. Our key observation is that during each sampling step, the model concentrates on semantically meaningful regions, and these areas of focus exhibit strong continuity across consecutive steps. Leveraging this insight, RAS updates only the regions currently in focus, while other regions are updated using cached noise from the previous step. The model's focus is determined based on the output from the preceding step, capitalizing on the temporal consistency we observed. We evaluate RAS on Stable Diffusion 3 and Lumina-Next-T2I, achieving speedups up to 2.36x and 2.51x, respectively, with minimal degradation in generation quality. Additionally, a user study reveals that RAS delivers comparable qualities under human evaluation while achieving a 1.6x speedup. Our approach makes a significant step towards more efficient diffusion transformers, enhancing their potential for real-time applications.

Given a task described by a few examples, how should a model be specialized to it? Four mechanisms are available -- zero-shot, in-context attention, test-time gradient adaptation, and emitting specialist weights from a hypernetwork -- yet the operating regime of the last is rarely mapped. We run the identical four-way comparison across six tasks spanning regression, generation, language modeling, reinforcement learning, and clinical and genomic classification, holding the specialist, the context, and (where we can) the training budget fixed. The clearest wins for emission are about cost at matched quality: it ties the state-of-the-art amortized tabular model (TabPFN) on clinical few-shot classification while emitting a reusable specialist instead of re-attending the support set per query, and reaches noise-floor shape generation with a 132-float per-instance program. On few-shot sinusoid regression it is 2--3 orders of magnitude below MAML at zero test-time gradient steps -- a margin that narrows to sim30times but persists once training budgets are equalized. Emission cannot match in-context attention on high-dimensional sequence modeling: under matched-budget pre-training a one-pass adapter recovers only a minority of the in-context gain (14.0pm0.9% at 5M, 11.2pm0.5% at 15M), and a LoRA-rank sweep shows this shortfall is a partial capacity limit -- capture climbs from 5% to 21% as rank grows but plateaus far below full recovery. Mechanism ablations confirm the emitted specialist is genuinely task-conditioned, not a memorized prior; and, more speculatively, emitted specialists compose in weight space -- interpolating two of them tracks the corresponding blend of their functions. We close with a falsifiable thesis, operationalized through a per-task resolution measure, bounding when each conditioning mechanism should be preferred.

Diffusion models have achieved remarkable success in image and video generation tasks. However, the high computational demands of Diffusion Transformers (DiTs) pose a significant challenge to their practical deployment. While feature caching is a promising acceleration strategy, existing methods based on simple reusing or training-free forecasting struggle to adapt to the complex, stage-dependent dynamics of the diffusion process, often resulting in quality degradation and failing to maintain consistency with the standard denoising process. To address this, we propose a LEarnable Stage-Aware (LESA) predictor framework based on two-stage training. Our approach leverages a Kolmogorov-Arnold Network (KAN) to accurately learn temporal feature mappings from data. We further introduce a multi-stage, multi-expert architecture that assigns specialized predictors to different noise-level stages, enabling more precise and robust feature forecasting. Extensive experiments show our method achieves significant acceleration while maintaining high-fidelity generation. Experiments demonstrate 5.00x acceleration on FLUX.1-dev with minimal quality degradation (1.0% drop), 6.25x speedup on Qwen-Image with a 20.2% quality improvement over the previous SOTA (TaylorSeer), and 5.00x acceleration on HunyuanVideo with a 24.7% PSNR improvement over TaylorSeer. State-of-the-art performance on both text-to-image and text-to-video synthesis validates the effectiveness and generalization capability of our training-based framework across different models. Our code is included in the supplementary materials and will be released on GitHub.

Video view synthesis, allowing for the creation of visually appealing frames from arbitrary viewpoints and times, offers immersive viewing experiences. Neural radiance fields, particularly NeRF, initially developed for static scenes, have spurred the creation of various methods for video view synthesis. However, the challenge for video view synthesis arises from motion blur, a consequence of object or camera movement during exposure, which hinders the precise synthesis of sharp spatio-temporal views. In response, we propose a novel dynamic deblurring NeRF framework for blurry monocular video, called DyBluRF, consisting of an Interleave Ray Refinement (IRR) stage and a Motion Decomposition-based Deblurring (MDD) stage. Our DyBluRF is the first that addresses and handles the novel view synthesis for blurry monocular video. The IRR stage jointly reconstructs dynamic 3D scenes and refines the inaccurate camera pose information to combat imprecise pose information extracted from the given blurry frames. The MDD stage is a novel incremental latent sharp-rays prediction (ILSP) approach for the blurry monocular video frames by decomposing the latent sharp rays into global camera motion and local object motion components. Extensive experimental results demonstrate that our DyBluRF outperforms qualitatively and quantitatively the very recent state-of-the-art methods. Our project page including source codes and pretrained model are publicly available at https://kaist-viclab.github.io/dyblurf-site/.

Multi-subject personalized image generation aims to synthesize customized images containing multiple specified subjects without requiring test-time optimization. However, achieving fine-grained independent control over multiple subjects remains challenging due to difficulties in preserving subject fidelity and preventing cross-subject attribute leakage. We present FocusDPO, a framework that adaptively identifies focus regions based on dynamic semantic correspondence and supervision image complexity. During training, our method progressively adjusts these focal areas across noise timesteps, implementing a weighted strategy that rewards information-rich patches while penalizing regions with low prediction confidence. The framework dynamically adjusts focus allocation during the DPO process according to the semantic complexity of reference images and establishes robust correspondence mappings between generated and reference subjects. Extensive experiments demonstrate that our method substantially enhances the performance of existing pre-trained personalized generation models, achieving state-of-the-art results on both single-subject and multi-subject personalized image synthesis benchmarks. Our method effectively mitigates attribute leakage while preserving superior subject fidelity across diverse generation scenarios, advancing the frontier of controllable multi-subject image synthesis.

Multimodal large language models (MLLMs) have achieved remarkable progress on various vision-language tasks, yet their visual perception remains limited. Humans, in comparison, perceive complex scenes efficiently by dynamically scanning and focusing on salient regions in a sequential "blink-like" process. Motivated by this strategy, we first investigate whether MLLMs exhibit similar behavior. Our pilot analysis reveals that MLLMs naturally attend to different visual regions across layers and that selectively allocating more computation to salient tokens can enhance visual perception. Building on this insight, we propose Blink, a dynamic visual token resolution framework that emulates the human-inspired process within a single forward pass. Specifically, Blink includes two modules: saliency-guided scanning and dynamic token resolution. It first estimates the saliency of visual tokens in each layer based on the attention map, and extends important tokens through a plug-and-play token super-resolution (TokenSR) module. In the next layer, it drops the extended tokens when they lose focus. This dynamic mechanism balances broad exploration and fine-grained focus, thereby enhancing visual perception adaptively and efficiently. Extensive experiments validate Blink, demonstrating its effectiveness in enhancing visual perception and multimodal understanding.

This research delves into the problem of interactive editing of human motion generation. Previous motion diffusion models lack explicit modeling of the word-level text-motion correspondence and good explainability, hence restricting their fine-grained editing ability. To address this issue, we propose an attention-based motion diffusion model, namely MotionCLR, with CLeaR modeling of attention mechanisms. Technically, MotionCLR models the in-modality and cross-modality interactions with self-attention and cross-attention, respectively. More specifically, the self-attention mechanism aims to measure the sequential similarity between frames and impacts the order of motion features. By contrast, the cross-attention mechanism works to find the fine-grained word-sequence correspondence and activate the corresponding timesteps in the motion sequence. Based on these key properties, we develop a versatile set of simple yet effective motion editing methods via manipulating attention maps, such as motion (de-)emphasizing, in-place motion replacement, and example-based motion generation, etc. For further verification of the explainability of the attention mechanism, we additionally explore the potential of action-counting and grounded motion generation ability via attention maps. Our experimental results show that our method enjoys good generation and editing ability with good explainability.

How a vision-language model internally solves the task of describing an image is far from obvious. We find that the model develops a specific mechanism for this: a small set of attention heads in its language-model backbone, which we call gaze heads, whose attention tracks the image region the model is currently describing. We find them with a simple correlation score from a few forward passes, using comic strips as a controlled testbed where narrative order is laid out spatially. These gaze heads do not just track the image tokens being described: redirecting their attention to a chosen region forces the VLM to describe that region instead. A single attention-mask intervention on the top-100 gaze heads, fewer than 9% of all heads, steers the model's answer to any chosen comic panel at 83.1% accuracy, while the same intervention on random heads fails to redirect the answer, and intervening on all heads destroys generation. The same lever also extends to continuous control: switching the gaze target mid-generation makes the model wrap up its current panel description and move to the new one within a few tokens. Beyond comics, the same intervention redirects answers to chosen regions in natural COCO images. The mechanism further recurs across model sizes from 2B to 32B parameters and across other VLM architectures, although some frozen-encoder families show no comparable head set. More broadly, this shows that targeted edits identified through mechanistic analysis can serve as practical inference-time levers for steering multimodal model behavior, without any retraining. Our code, interactive demo, and datasets are available at https://gaze.baulab.info/

Depth-of-field control is essential in photography, but getting the perfect focus often takes several tries or special equipment. Single-image refocusing is still difficult. It involves recovering sharp content and creating realistic bokeh. Current methods have significant drawbacks. They need all-in-focus inputs, depend on synthetic data from simulators, and have limited control over aperture. We introduce Generative Refocusing, a two-step process that uses DeblurNet to recover all-in-focus images from various inputs and BokehNet for creating controllable bokeh. Our main innovation is semi-supervised training. This method combines synthetic paired data with unpaired real bokeh images, using EXIF metadata to capture real optical characteristics beyond what simulators can provide. Our experiments show we achieve top performance in defocus deblurring, bokeh synthesis, and refocusing benchmarks. Additionally, our Generative Refocusing allows text-guided adjustments and custom aperture shapes.

Attention mechanism has emerged as a foundation module of modern deep learning models and has also empowered many milestones in various domains. Moreover, FlashAttention with IO-aware speedup resolves the efficiency issue of standard attention, further promoting its practicality. Beyond canonical attention, attention with bias also widely exists, such as relative position bias in vision and language models and pair representation bias in AlphaFold. In these works, prior knowledge is introduced as an additive bias term of attention weights to guide the learning process, which has been proven essential for model performance. Surprisingly, despite the common usage of attention with bias, its targeted efficiency optimization is still absent, which seriously hinders its wide applications in complex tasks. Diving into the computation of FlashAttention, we prove that its optimal efficiency is determined by the rank of the attention weight matrix. Inspired by this theoretical result, this paper presents FlashBias based on the low-rank compressed sensing theory, which can provide fast-exact computation for many widely used attention biases and a fast-accurate approximation for biases in general formalization. FlashBias can fully take advantage of the extremely optimized matrix multiplication operation in modern GPUs, achieving 1.5times speedup for AlphaFold, and over 2times speedup for attention with bias in vision and language models without loss of accuracy.

The text-to-video (T2V) generation models, offering convenient visual creation, have recently garnered increasing attention. Despite their substantial potential, the generated videos may present artifacts, including structural implausibility, temporal inconsistency, and a lack of motion, often resulting in near-static video. In this work, we have identified a correlation between the disparity of temporal attention maps across different blocks and the occurrence of temporal inconsistencies. Additionally, we have observed that the energy contained within the temporal attention maps is directly related to the magnitude of motion amplitude in the generated videos. Based on these observations, we present BroadWay, a training-free method to improve the quality of text-to-video generation without introducing additional parameters, augmenting memory or sampling time. Specifically, BroadWay is composed of two principal components: 1) Temporal Self-Guidance improves the structural plausibility and temporal consistency of generated videos by reducing the disparity between the temporal attention maps across various decoder blocks. 2) Fourier-based Motion Enhancement enhances the magnitude and richness of motion by amplifying the energy of the map. Extensive experiments demonstrate that BroadWay significantly improves the quality of text-to-video generation with negligible additional cost.

Smooth speaker transitions are fundamental to effective conversation and rely on an interlocutor's ability to predict when to enter the conversation. This ability depends on accurately interpreting and expressing the verbal and non-verbal cues that signal when a speaker wishes to take or relinquish the floor. The process becomes even more complex in noisy, natural, multi-party settings, with multiple interlocutors available. This study models how gaze and speech, together with perceived interpersonal closeness, signal conversational floor changes in free four-person dialogue. Using the GaMMA corpus, we trained logistic regression models using interpretable, behaviourally motivated features extracted before each turn-taking event to classify floor-transfer outcomes as gaps or overlaps. Predictors included gaze features such as transition motifs and behavioural contrasts, entropy, gaze-based addressee identity, and mutual gaze, alongside speech features derived from speaker loudness, as well as perceived interpersonal closeness (IOS) between speakers. Results show that gaze features capture predictive structure, and that combining them with loudness improves performance (ROC AUC = 0.76 +- 0.04). Loudness reflected speaker control, while gaze dispersion and addressing indexed listener readiness and competitive entry. Performance remained robust across noise conditions, indicating that gaze provides a complementary, noise-resilient cue to turn-taking dynamics.

Focused transmit beamforming is the most commonly used acquisition scheme for echocardiograms, but suffers from relatively low frame rates, and in 3D, even lower volume rates. Fast imaging based on unfocused transmits has disadvantages such as motion decorrelation and limited harmonic imaging capabilities. This work introduces a patient-adaptive focused transmit scheme that has the ability to drastically reduce the number of transmits needed to produce a high-quality ultrasound image. The method relies on posterior sampling with a temporal diffusion model to perceive and reconstruct the anatomy based on partial observations, while subsequently taking an action to acquire the most informative transmits. This active perception modality outperforms random and equispaced subsampling on the 2D EchoNet-Dynamic dataset and a 3D Philips dataset, where we actively select focused elevation planes. Furthermore, we show it achieves better performance in terms of generalized contrast-to-noise ratio when compared to the same number of diverging waves transmits on three in-house echocardiograms. Additionally, we can estimate ejection fraction using only 2% of the total transmits and show that the method is robust to outlier patients. Finally, our method can be run in real-time on GPU accelerators from 2023. The code is publicly available at https://tue-bmd.github.io/ulsa/

Recent advances in large-scale text-to-image models have revolutionized creative fields by generating visually captivating outputs from textual prompts; however, while traditional photography offers precise control over camera settings to shape visual aesthetics -- such as depth-of-field -- current diffusion models typically rely on prompt engineering to mimic such effects. This approach often results in crude approximations and inadvertently altering the scene content. In this work, we propose Bokeh Diffusion, a scene-consistent bokeh control framework that explicitly conditions a diffusion model on a physical defocus blur parameter. By grounding depth-of-field adjustments, our method preserves the underlying scene structure as the level of blur is varied. To overcome the scarcity of paired real-world images captured under different camera settings, we introduce a hybrid training pipeline that aligns in-the-wild images with synthetic blur augmentations. Extensive experiments demonstrate that our approach not only achieves flexible, lens-like blur control but also supports applications such as real image editing via inversion.

The Multi-Head Attention mechanism is central to LLM operation, and multiple works target its compute and memory efficiency during training. While most works focus on approximating the scaled dot product, the memory consumption of the linear projections that compute the Q, K, and V tensors from the input x is often overlooked. To address this, we propose Point-Approximate Matrix Multiplication (PAMM), a novel tensor compression technique that reduces memory consumption of the Q,K,V projections in attention layers by a factor of up to times 512, effectively erasing their memory footprint, while achieving similar or better final perplexity. PAMM is fully composable with efficient attention techniques such as FlashAttention, making it a practical and complementary method for memory-efficient LLM training.

We present a target-aware video diffusion model that generates videos from an input image in which an actor interacts with a specified target while performing a desired action. The target is defined by a segmentation mask and the desired action is described via a text prompt. Unlike existing controllable image-to-video diffusion models that often rely on dense structural or motion cues to guide the actor's movements toward the target, our target-aware model requires only a simple mask to indicate the target, leveraging the generalization capabilities of pretrained models to produce plausible actions. This makes our method particularly effective for human-object interaction (HOI) scenarios, where providing precise action guidance is challenging, and further enables the use of video diffusion models for high-level action planning in applications such as robotics. We build our target-aware model by extending a baseline model to incorporate the target mask as an additional input. To enforce target awareness, we introduce a special token that encodes the target's spatial information within the text prompt. We then fine-tune the model with our curated dataset using a novel cross-attention loss that aligns the cross-attention maps associated with this token with the input target mask. To further improve performance, we selectively apply this loss to the most semantically relevant transformer blocks and attention regions. Experimental results show that our target-aware model outperforms existing solutions in generating videos where actors interact accurately with the specified targets. We further demonstrate its efficacy in two downstream applications: video content creation and zero-shot 3D HOI motion synthesis.

Attention guides our gaze to fixate the proper location of the scene and holds it in that location for the deserved amount of time given current processing demands, before shifting to the next one. As such, gaze deployment crucially is a temporal process. Existing computational models have made significant strides in predicting spatial aspects of observer's visual scanpaths (where to look), while often putting on the background the temporal facet of attention dynamics (when). In this paper we present TPP-Gaze, a novel and principled approach to model scanpath dynamics based on Neural Temporal Point Process (TPP), that jointly learns the temporal dynamics of fixations position and duration, integrating deep learning methodologies with point process theory. We conduct extensive experiments across five publicly available datasets. Our results show the overall superior performance of the proposed model compared to state-of-the-art approaches. Source code and trained models are publicly available at: https://github.com/phuselab/tppgaze.

When a steering intervention changes a language model's behavior, which part of the intervention deserves the credit? We test this question for a target-centered, scaled conceptor complement by removing its fitted covariance filter while preserving the same target-specific center and gain. At the original aperture α=.05 and gain β=2, all 21 fitted maps across three models and seven targets expand every direction at the generation intervention layers. The matrix correction is 0.093%--7.815% of the matched-control displacement on held-out extraction activations. In 480 paired comparisons within a 1,920-response study, retaining the matrix changes many individual texts but does not establish a consistent additional target-suppression benefit on the four-target panel. A geometric aperture sweep shows that the operating regime changes with the filter setting. Separately, an invertible soft filter preserves what an unrestricted affine readout can recover, whereas dedicated erasure methods reduce prediction by newly fitted probes. Together, the results show why a whole-intervention comparison cannot by itself attribute generated behavior to the fitted matrix: analyze the complete transformation, hold shared components fixed, and test its output consequences directly.

Robotic foundation models have recently made substantial progress in multi-task capability, cross-embodiment transfer, and language-conditioned control. Yet robust deployment across diverse real-world settings remains difficult, in part because policies often fail to distinguish causally relevant visual structure from spurious scene-level correlations. We identify this failure mode as shortcut learning: the tendency to exploit predictive but non-causal correlations in the training distribution rather than the task-relevant visual evidence that determines successful action. Although shortcut learning has been extensively studied in computer vision and broader machine learning, its role in robotic foundation models remains comparatively underexplored. We propose Artificial Foveated Perception (AFP), a lightweight, policy-agnostic module that takes the same vision and language inputs as Vision-Language-Action and World Action Model pipelines and predicts task-conditioned masks over relevant objects, the robot, and other action-critical regions. We use these masks primarily as an auxiliary grounding signal during fine-tuning, aligning policy attention with task-relevant regions while leaving the core architecture unchanged. After fine-tuning, the policy executes on the original observation stream without requiring AFP in the control loop. We evaluate AFP across state-of-the-art robotic foundation models and show that foveated perception reduces fine-tuning time, suppresses overfitting, and improves generalization under environmental perturbations. Ablations over mask quality and grounding-loss design further show that these gains arise from directing policy learning toward task-relevant visual evidence. These results suggest that task-conditioned foveated perception is a practical mechanism for making robotic foundation models more robust, data-efficient, and scalable.

Video diffusion transformers are expensive because attention dominates long spatiotemporal token sequences. We identify the high-sparsity trap: at extreme attention sparsity, step-local training losses keep decreasing while terminal generation quality stagnates or degrades. The trap is one of supervision: the dominant terminal errors originate in the high-noise structure-generation stage, and terminal-aligned training corrects terminal errors that substantially extended step-local training cannot. This yields a simple staging principle: first adapt the sparse architecture into a coarse prior, then correct the terminal distribution. We instantiate the principle as \method, a unified acceleration framework for visual generation that combines a short sparse warm-up, few-step trajectory-mixed distillation, and FP8 quantization with fused kernels. \method sustains 97% attention sparsity with strong visual quality on long-sequence 720P generation across Wan2.1/Wan2.2 backbones and T2V/I2V tasks, and 90% sparsity on Wan2.1-T2V-1.3B-480P. With 3-step CFG-free inference, \method achieves a 265times end-to-end speedup over the 50-step CFG dense baseline for Wan2.1-T2V-14B-720P on a single RTX~5090 (220times on H100), and denoises a Wan2.1-T2V-1.3B-480P video in 1.3s.

Recent advances in training-free attention control methods have enabled flexible and efficient text-guided editing capabilities for existing generation models. However, current approaches struggle to simultaneously deliver strong editing strength while preserving consistency with the source. This limitation becomes particularly critical in multi-round and video editing, where visual errors can accumulate over time. Moreover, most existing methods enforce global consistency, which limits their ability to modify individual attributes such as texture while preserving others, thereby hindering fine-grained editing. Recently, the architectural shift from U-Net to MM-DiT has brought significant improvements in generative performance and introduced a novel mechanism for integrating text and vision modalities. These advancements pave the way for overcoming challenges that previous methods failed to resolve. Through an in-depth analysis of MM-DiT, we identify three key insights into its attention mechanisms. Building on these, we propose ConsistEdit, a novel attention control method specifically tailored for MM-DiT. ConsistEdit incorporates vision-only attention control, mask-guided pre-attention fusion, and differentiated manipulation of the query, key, and value tokens to produce consistent, prompt-aligned edits. Extensive experiments demonstrate that ConsistEdit achieves state-of-the-art performance across a wide range of image and video editing tasks, including both structure-consistent and structure-inconsistent scenarios. Unlike prior methods, it is the first approach to perform editing across all inference steps and attention layers without handcraft, significantly enhancing reliability and consistency, which enables robust multi-round and multi-region editing. Furthermore, it supports progressive adjustment of structural consistency, enabling finer control.

Deploying massive diffusion models for real-time, infinite-duration, audio-driven avatar generation presents a significant engineering challenge, primarily due to the conflict between computational load and strict latency constraints. Existing approaches often compromise visual fidelity by enforcing strictly unidirectional attention mechanisms or reducing model capacity. To address this problem, we introduce SoulX-FlashTalk, a 14B-parameter framework optimized for high-fidelity real-time streaming. Diverging from conventional unidirectional paradigms, we use a Self-correcting Bidirectional Distillation strategy that retains bidirectional attention within video chunks. This design preserves critical spatiotemporal correlations, significantly enhancing motion coherence and visual detail. To ensure stability during infinite generation, we incorporate a Multi-step Retrospective Self-Correction Mechanism, enabling the model to autonomously recover from accumulated errors and preventing collapse. Furthermore, we engineered a full-stack inference acceleration suite incorporating hybrid sequence parallelism, Parallel VAE, and kernel-level optimizations. Extensive evaluations confirm that SoulX-FlashTalk is the first 14B-scale system to achieve a sub-second start-up latency (0.87s) while reaching a real-time throughput of 32 FPS, setting a new standard for high-fidelity interactive digital human synthesis.

The development of video diffusion models unveils a significant challenge: the substantial computational demands. To mitigate this challenge, we note that the reverse process of diffusion exhibits an inherent entropy-reducing nature. Given the inter-frame redundancy in video modality, maintaining full frame rates in high-entropy stages is unnecessary. Based on this insight, we propose TPDiff, a unified framework to enhance training and inference efficiency. By dividing diffusion into several stages, our framework progressively increases frame rate along the diffusion process with only the last stage operating on full frame rate, thereby optimizing computational efficiency. To train the multi-stage diffusion model, we introduce a dedicated training framework: stage-wise diffusion. By solving the partitioned probability flow ordinary differential equations (ODE) of diffusion under aligned data and noise, our training strategy is applicable to various diffusion forms and further enhances training efficiency. Comprehensive experimental evaluations validate the generality of our method, demonstrating 50% reduction in training cost and 1.5x improvement in inference efficiency.

Tuning-free long video diffusion has been proposed to generate extended-duration videos with enriched content by reusing the knowledge from pre-trained short video diffusion model without retraining. However, most works overlook the fine-grained long-term video consistency modeling, resulting in limited scene consistency (i.e., unreasonable object or background transitions), especially with multiple text inputs. To mitigate this, we propose the Consistency Noise Injection, dubbed CoNo, which introduces the "look-back" mechanism to enhance the fine-grained scene transition between different video clips, and designs the long-term consistency regularization to eliminate the content shifts when extending video contents through noise prediction. In particular, the "look-back" mechanism breaks the noise scheduling process into three essential parts, where one internal noise prediction part is injected into two video-extending parts, intending to achieve a fine-grained transition between two video clips. The long-term consistency regularization focuses on explicitly minimizing the pixel-wise distance between the predicted noises of the extended video clip and the original one, thereby preventing abrupt scene transitions. Extensive experiments have shown the effectiveness of the above strategies by performing long-video generation under both single- and multi-text prompt conditions. The project has been available in https://wxrui182.github.io/CoNo.github.io/.

For artistic applications, video generation requires fine-grained control over both performance and cinematography, i.e., the actor's motion and the camera trajectory. We present ActCam, a zero-shot method for video generation that jointly transfers character motion from a driving video into a new scene and enables per-frame control of intrinsic and extrinsic camera parameters. ActCam builds on any pretrained image-to-video diffusion model that accepts conditioning in terms of scene depth and character pose. Given a source video with a moving character and a target camera motion, ActCam generates pose and depth conditions that remain geometrically consistent across frames. We then run a single sampling process with a two-phase conditioning schedule: early denoising steps condition on both pose and sparse depth to enforce scene structure, after which depth is dropped and pose-only guidance refines high-frequency details without over-constraining the generation. We evaluate ActCam on multiple benchmarks spanning diverse character motions and challenging viewpoint changes. We find that, compared to pose-only control and other pose and camera methods, ActCam improves camera adherence and motion fidelity, and is preferred in human evaluations, especially under large viewpoint changes. Our results highlight that careful camera-consistent conditioning and staged guidance can enable strong joint camera and motion control without training. Project page: https://elkhomar.github.io/actcam/.

Diffusion Transformers (DiT) have emerged as a widely adopted backbone for high-fidelity image and video generation, yet their iterative denoising process incurs high computational costs. Existing training-free acceleration methods rely on feature caching and reuse under the assumption of temporal stability. However, reusing features for multiple steps may lead to latent drift and visual degradation. We observe that model outputs evolve smoothly along much of the diffusion trajectory, enabling principled predictions rather than naive reuse. Based on this insight, we propose PrediT, a training-free acceleration framework that formulates feature prediction as a linear multistep problem. We employ classical linear multistep methods to forecast future model outputs from historical information, combined with a corrector that activates in high-dynamics regions to prevent error accumulation. A dynamic step modulation mechanism adaptively adjusts the prediction horizon by monitoring the feature change rate. Together, these components enable substantial acceleration while preserving generation fidelity. Extensive experiments validate that our method achieves up to 5.54times latency reduction across various DiT-based image and video generation models, while incurring negligible quality degradation.

State-of-the-art text-to-video models excel at generating isolated clips but fall short of creating the coherent, multi-shot narratives, which are the essence of storytelling. We bridge this "narrative gap" with HoloCine, a model that generates entire scenes holistically to ensure global consistency from the first shot to the last. Our architecture achieves precise directorial control through a Window Cross-Attention mechanism that localizes text prompts to specific shots, while a Sparse Inter-Shot Self-Attention pattern (dense within shots but sparse between them) ensures the efficiency required for minute-scale generation. Beyond setting a new state-of-the-art in narrative coherence, HoloCine develops remarkable emergent abilities: a persistent memory for characters and scenes, and an intuitive grasp of cinematic techniques. Our work marks a pivotal shift from clip synthesis towards automated filmmaking, making end-to-end cinematic creation a tangible future. Our code is available at: https://holo-cine.github.io/.

Bokeh and monocular depth estimation are tightly coupled through the same lens imaging geometry, yet current methods exploit this connection in incomplete ways. High-quality bokeh rendering pipelines typically depend on noisy depth maps, which amplify estimation errors into visible artifacts, while modern monocular metric depth models still struggle on weakly textured, distant and geometrically ambiguous regions where defocus cues are most informative. We introduce BokehDepth, a two-stage framework that decouples bokeh synthesis from depth prediction and treats defocus as an auxiliary supervision-free geometric cue. In Stage-1, a physically guided controllable bokeh generator, built on a powerful pretrained image editing backbone, produces depth-free bokeh stacks with calibrated bokeh strength from a single sharp input. In Stage-2, a lightweight defocus-aware aggregation module plugs into existing monocular depth encoders, fuses features along the defocus dimension, and exposes stable depth-sensitive variations while leaving downstream decoder unchanged. Across challenging benchmarks, BokehDepth improves visual fidelity over depth-map-based bokeh baselines and consistently boosts the metric accuracy and robustness of strong monocular depth foundation models.

Multimodal Diffusion Transformers (MM-DiTs) have achieved remarkable progress in text-driven visual generation. However, even state-of-the-art MM-DiT models like FLUX struggle with achieving precise alignment between text prompts and generated content. We identify two key issues in the attention mechanism of MM-DiT, namely 1) the suppression of cross-modal attention due to token imbalance between visual and textual modalities and 2) the lack of timestep-aware attention weighting, which hinder the alignment. To address these issues, we propose Temperature-Adjusted Cross-modal Attention (TACA), a parameter-efficient method that dynamically rebalances multimodal interactions through temperature scaling and timestep-dependent adjustment. When combined with LoRA fine-tuning, TACA significantly enhances text-image alignment on the T2I-CompBench benchmark with minimal computational overhead. We tested TACA on state-of-the-art models like FLUX and SD3.5, demonstrating its ability to improve image-text alignment in terms of object appearance, attribute binding, and spatial relationships. Our findings highlight the importance of balancing cross-modal attention in improving semantic fidelity in text-to-image diffusion models. Our codes are publicly available at https://github.com/Vchitect/TACA

Attention steering is an important technique for controlling model focus, enabling capabilities such as prompt highlighting, where the model prioritises user-specified text. However, existing attention steering methods require explicit storage of the full attention matrix, making them incompatible with memory-efficient implementations like FlashAttention. We introduce Spectral Editing Key Amplification (SEKA), a training-free steering method that tackles this by directly editing key embeddings before attention computation. SEKA uses spectral decomposition to steer key embeddings towards latent directions that amplify attention scores for certain tokens. We extend this to Adaptive SEKA (AdaSEKA), a query-adaptive variant that uses a training-free routing mechanism to dynamically combine multiple expert subspaces based on the prompt's semantic intent. Our experiments show both methods significantly outperform strong baselines on standard steering benchmarks while adding much lower latency and memory overhead, in compatibility with optimised attention.

Event cameras are novel vision sensors that output pixel-level brightness changes ("events") instead of traditional video frames. These asynchronous sensors offer several advantages over traditional cameras, such as, high temporal resolution, very high dynamic range, and no motion blur. To unlock the potential of such sensors, motion compensation methods have been recently proposed. We present a collection and taxonomy of twenty two objective functions to analyze event alignment in motion compensation approaches (Fig. 1). We call them Focus Loss Functions since they have strong connections with functions used in traditional shape-from-focus applications. The proposed loss functions allow bringing mature computer vision tools to the realm of event cameras. We compare the accuracy and runtime performance of all loss functions on a publicly available dataset, and conclude that the variance, the gradient and the Laplacian magnitudes are among the best loss functions. The applicability of the loss functions is shown on multiple tasks: rotational motion, depth and optical flow estimation. The proposed focus loss functions allow to unlock the outstanding properties of event cameras.

Video diffusion models have achieved remarkable progress in generating high-quality videos. However, these models struggle to represent the temporal succession of multiple events in real-world videos and lack explicit mechanisms to control when semantic concepts appear, how long they persist, and the order in which multiple events occur. Such control is especially important for movie-grade video synthesis, where coherent storytelling depends on precise timing, duration, and transitions between events. When using a single paragraph-style prompt to describe a sequence of complex events, models often exhibit semantic entanglement, where concepts intended for different moments in the video bleed into one another, resulting in poor text-video alignment. To address these limitations, we propose Prompt Relay, an inference-time, plug-and-play method to enable fine-grained temporal control in multi-event video generation, requiring no architectural modifications and no additional computational overhead. Prompt Relay introduces a penalty into the cross-attention mechanism, so that each temporal segment attends only to its assigned prompt, allowing the model to represent one semantic concept at a time and thereby improving temporal prompt alignment, reducing semantic interference, and enhancing visual quality.

Training and evaluation of state-of-the-art computer vision algorithms for reliable shallow depth of field (DoF) rendering and defocus deblurring remain constrained by a persistent lack of large-scale, full-frame, high fidelity, real-image datasets. Optical effects of shallow DoF and defocus blur depend intimately on camera optical configuration set with focal length and aperture; requiring rigorous evaluation of the models when these parameters systematically change. Further, modern applications such as AR, VR, smartphones, industrial robots, etc. deploy stereo or multi-camera systems. We present MODEST - the first ultra-high-resolution(5472x3648px, 20MP), multi-optics depth of field stereo DSLR dataset that methodically varies focal length and aperture for a series of complex, real-world scenes, capturing the optical realism and complexity of professional camera systems. With 20,000 images across 50 distinct optical configurations, focal length in 28-70mm, aperture in f/22-f/2.8 for multiple stereo viewpoints for 10 scenes; this ultra-high-resolution, full-range optics coverage enables controlled analysis of geometric and optical effects for shallow DoF rendering and defocus deblurring. Each scene is curated to have challenging visual elements: reflective surfaces, transparent glass walls, fine-grained details, point lights, and multi-scale depth illusions. In addition, we provide intrinsics and extrinsics calibration images to support ever-evolving calibration methods. We evaluate several SOTA DoF and defocus deblurring methods and demonstrate failure cases and limitations. Our comprehensive tuning analysis demonstrates how MODEST evaluates sensitivity of the SOTA DoF models for the actual focal configuration of input images. This work attempts to bridge the realism gap between synthetic, low-resolution training data and inference generalization on high-resolution real-camera optics.

Recent diffusion models achieve the state-of-the-art performance in image generation, but often suffer from semantic inconsistencies or hallucinations. While various inference-time guidance methods can enhance generation, they often operate indirectly by relying on external signals or architectural modifications, which introduces additional computational overhead. In this paper, we propose Tangential Amplifying Guidance (TAG), a more efficient and direct guidance method that operates solely on trajectory signals without modifying the underlying diffusion model. TAG leverages an intermediate sample as a projection basis and amplifies the tangential components of the estimated scores with respect to this basis to correct the sampling trajectory. We formalize this guidance process by leveraging a first-order Taylor expansion, which demonstrates that amplifying the tangential component steers the state toward higher-probability regions, thereby reducing inconsistencies and enhancing sample quality. TAG is a plug-and-play, architecture-agnostic module that improves diffusion sampling fidelity with minimal computational addition, offering a new perspective on diffusion guidance.

Cinematic camera control relies on a tight feedback loop between director and cinematographer, where camera motion and framing are continuously reviewed and refined. Recent generative camera systems can produce diverse, text-conditioned trajectories, but they lack this "director in the loop" and have no explicit supervision of whether a shot is visually desirable. This results in in-distribution camera motion but poor framing, off-screen characters, and undesirable visual aesthetics. In this paper, we introduce VERTIGO, the first framework for visual preference optimization of camera trajectory generators. Our framework leverages a real-time graphics engine (Unity) to render 2D visual previews from generated camera motion. A cinematically fine-tuned vision-language model then scores these previews using our proposed cyclic semantic similarity mechanism, which aligns renders with text prompts. This process provides the visual preference signals for Direct Preference Optimization (DPO) post-training. Both quantitative evaluations and user studies on Unity renders and diffusion-based Camera-to-Video pipelines show consistent gains in condition adherence, framing quality, and perceptual realism. Notably, VERTIGO reduces the character off-screen rate from 38% to nearly 0% while preserving the geometric fidelity of camera motion. User study participants further prefer VERTIGO over baselines across composition, consistency, prompt adherence, and aesthetic quality, confirming the perceptual benefits of our visual preference post-training.

Few-step autoregressive video diffusion generates a long video by splitting the video into temporal chunks and generating chunk-by-chunk, each through a short sequence of denoising stages. To memorize chunks that are already generated, previous methods reconstruct a clean or less-noisy key--value (KV) cache by additional forwards to build the cache without advancing an output latent. However, every denoising forward itself already computes the in-flight KV of the current chunk. We introduce FlashForward, which directly reuses this cache to avoid the heavy cache-update-only model forwards. After the current chunk completes one denoising stage, its stage-specific cache is already available for the next chunk. Assigning one GPU to each stage therefore lets different chunks occupy different stages concurrently. This early availability has a quality cost: the resulting stage-matched history is noisy, causing appearance and motion drift among chunks. To complement it, FlashForward produces sparse auxiliary clean anchor latents before the corresponding region is generated so the generation trajectories can be stabilized by this two-sided conditioning. The two memories operate at different temporal scales: sparse clean anchor KV supplies coarse, long-range two-sided structural guidance, while dense stage-matched history preserves fine, recent evolution. With up to four GPUs, FlashForward runs 1.16--1.69times faster than HiAR and 1.42--2.92times faster than Self-Forcing for 16 FPS videos of 20 seconds or longer across 1.3B and 14B backbone scales at 480p and 720p. On VBench, for the 1.3B model at 480p, it achieves higher scores and remains stable at longer durations, demonstrating that FlashForward generates high-quality and temporally consistent videos across durations of 20s, 35s and 65s at a much faster generation speed.

Audio-driven portrait animation aims to synthesize realistic and natural talking head videos from an input audio signal and a single reference image. While existing methods achieve high-quality results by leveraging high-dimensional intermediate representations and explicitly modeling motion dynamics, their computational complexity renders them unsuitable for real-time deployment. Real-time inference imposes stringent latency and memory constraints, often necessitating the use of highly compressed latent representations. However, operating in such compact spaces hinders the preservation of fine-grained spatiotemporal details, thereby complicating audio-visual synchronization RAP (Real-time Audio-driven Portrait animation), a unified framework for generating high-quality talking portraits under real-time constraints. Specifically, RAP introduces a hybrid attention mechanism for fine-grained audio control, and a static-dynamic training-inference paradigm that avoids explicit motion supervision. Through these techniques, RAP achieves precise audio-driven control, mitigates long-term temporal drift, and maintains high visual fidelity. Extensive experiments demonstrate that RAP achieves state-of-the-art performance while operating under real-time constraints.

In this work, we build a simple but strong baseline for sounding video generation. Given base diffusion models for audio and video, we integrate them with additional modules into a single model and train it to make the model jointly generate audio and video. To enhance alignment between audio-video pairs, we introduce two novel mechanisms in our model. The first one is timestep adjustment, which provides different timestep information to each base model. It is designed to align how samples are generated along with timesteps across modalities. The second one is a new design of the additional modules, termed Cross-Modal Conditioning as Positional Encoding (CMC-PE). In CMC-PE, cross-modal information is embedded as if it represents temporal position information, and the embeddings are fed into the model like positional encoding. Compared with the popular cross-attention mechanism, CMC-PE provides a better inductive bias for temporal alignment in the generated data. Experimental results validate the effectiveness of the two newly introduced mechanisms and also demonstrate that our method outperforms existing methods.

Recent advances in flow-based generative models have enabled training-free, text-guided image editing by inverting an image into its latent noise and regenerating it under a new target conditional guidance. However, existing methods struggle to preserve source fidelity: higher-order solvers incur additional model inferences, truncated inversion constrains editability, and feature injection methods lack architectural transferability. To address these limitations, we propose SteerFlow, a model-agnostic editing framework with strong theoretical guarantees on source fidelity. In the forward process, we introduce an Amortized Fixed-Point Solver that implicitly straightens the forward trajectory by enforcing velocity consistency across consecutive timesteps, yielding a high-fidelity inverted latent. In the backward process, we introduce Trajectory Interpolation, which adaptively blends target-editing and source-reconstruction velocities to keep the editing trajectory anchored to the source. To further improve background preservation, we introduce an Adaptive Masking mechanism that spatially constrains the editing signal with concept-guided segmentation and source-target velocity differences. Extensive experiments on FLUX.1-dev and Stable Diffusion 3.5 Medium demonstrate that SteerFlow consistently achieves better editing quality than existing methods. Finally, we show that SteerFlow extends naturally to a complex multi-turn editing paradigm without accumulating drift.

Attention sinks, representation collapse, and norm stratification are treated as transformer-specific pathologies. We show they are not specific to attention: they are what content-based routing does under a fixed similarity metric. We give a reframing identity: softmax attention is Boltzmann-weighted aggregation over Euclidean distances with constant key norms, so its score omits a -|k|^2 term and is blind to key magnitude. This predicts that any router whose metric is ill-matched to its representations should compensate, by concentrating its routing and collapsing the routed representations. We test it on routers that score and aggregate over different axes: softmax attention over tokens (nine pretrained transformers), graph attention over nodes, a selective state-space model and a recurrent mixer over time, and learned residuals over depth. All develop the same signature, and two within-model ablations show it is caused by the routing mechanism rather than by incidental dynamics. The form is contingent, set by the strength of the positional brake each router carries alongside its content score; we sweep that brake and move the onset across its whole range. The mechanism is not contingent, and it does not require norm stratification: a router with norm-normalized keys concentrates just the same. We do not claim these models implement Riemannian geometry; the geometric view is a diagnostic that names the inadequacy of the flat, norm-blind metric.

Human visual search is serial: the fovea must land on a candidate to confirm it, and those landings form a scanpath. Whether multimodal large language models (MLLMs), given the same foveated input, search as humans do bears on their use as models of human vision and on attention-alignment scores. We compare three general-purpose MLLMs with human eye-movement scanpaths on goal-directed search (COCO-Search18), driving each model fixation by fixation through an identical, human-matched foveated view and assessing it along three axes: the decision of target presence, the efficiency of reaching the target, and the gaze process itself. The axes dissociate. On the decision and on target acquisition the models match or exceed humans, detecting present targets near ceiling and reaching them on the first saccade more often than people do. The gaze process is not human. Under the human-matched condition, all three share one signature: low-entropy, large-amplitude, self-consistent scanpaths that agree with themselves far more closely than two humans agree with each other. That is consistent with a single-pass, non-serial architecture rather than a limit of acuity. Matched retinal input reproduces where humans look but not how the looking unfolds in time, and no degradation regime recovers human-like search at human-like success. The gap sits on a process axis that answer-alignment and saliency metrics do not measure. Because they miss it, such metrics cannot certify human-like vision, and zero-shot models suit outcome and spatial questions but not temporal, process-level ones.

Vision-Language-Action (VLA) policies that execute fixed-length action chunks can exhibit multimodal bifurcation: a cross-chunk inconsistency in which adjacent chunks generated from independent Gaussian latents can converge to incompatible trajectory modes, producing abrupt discontinuities at chunk boundaries. Existing remedies either require backpropagation through the policy at each denoising step, rely on rejection sampling, or require retraining, each trading computational cost or task reliability for smoother transitions. We propose SEAM (Smooth Execution of Action-Chunked Motion), a training-free inference-time method for flow matching VLAs. SEAM exploits a simple synchronous-execution insight: after the robot consumes the executed prefix, the previous chunk's unexecuted tail is already available as an analytic consistency reference. Its core mechanism, Velocity-guided Loss Steering (VLS), derives a time-dependent target from this tail and applies a closed-form correction after each Euler step without backpropagating through the policy network. On LIBERO-10 with pi_0.5, SEAM reduces boundary jerk by 28%, reduces chunk transition discontinuity by 27%, preserves baseline-level task success, and keeps denoising-loop cost near the unguided baseline.

Autoregressive generation underlies real-time, causal video synthesis and is the backbone of video world models. Today's dominant approach distills a bidirectional diffusion model into a causal few-step student, but existing pipelines make this conversion prohibitively expensive, requiring offline Probability Flow ODE} (PF-ODE) trajectories that cost tens of thousands of GPU-hours and terabytes of storage. We show it can be made dramatically more efficient. Pushing this efficiency to its extreme, we reduce trajectory supervision to a single sample, the strictest setting possible, and find that it causes severe quality collapse. We trace this to the teacher's trajectory geometry: the high-noise regime, governing semantic layout and motion, and the low-noise regime, governing fine texture exhibit fundamentally different curvature and optimization behavior, so a single monolithic student cannot jointly fit both, causing severe representation interference. Motivated by this, we propose Glance Forcing, a lightweight distillation framework that decouples the denoising trajectory into two complementary slow and fast experts, each guided by a tailored, stage-specific objective, while keeping the pretrained teacher entirely frozen. This decoupling further reveals a general mechanism underlying few-step distillation: PF-ODE distillation mainly straightens the few-step trajectory geometry, while the subsequent distribution-matching stage independently governs distributional quality. Using only one PF-ODE trajectory, Glance Forcing already surpasses the fully-trained state-of-the-art baseline, raising the VBench Total score from 84.04 to 84.22 while cutting trajectory-collection cost from thousands of GPU-hours to about one A100-GPU-hour, with markedly larger gains in dynamic degree, VisionReward, and instruction-following, and consistent preference from human raters over strong baselines.

Robotic ultrasound couples a high-rate contact-force loop with slower, delayed image feedback, so image-guided ultrasound probe rotation can perturb contact force before the resulting image response is observed. We derive a closed-form orientation-rate limit that bounds the modeled rotation-induced estimated-force excursion over a finite horizon while accounting for disturbance rejection by the fast force loop. The limit depends on local contact stiffness, force-loop gains, a conservative rotation-to-force gain bound, the excursion budget, and the prediction horizon. We implement this model in a dual-rate controller with timestamp-based delay reconstruction and joint-torque-based force estimation, and evaluate it on a curved gelatin phantom using paired controller comparisons and component ablations. Relative to unconstrained image guidance, the proposed rate-limited controller reduced first-second root-mean-square (RMS) estimated-force error by 0.40 N while increasing cue-convergence time by 0.94 s. A fixed rate cap near the analytically predicted ceiling produced no resolvable difference in force error and converged 0.32 s faster, indicating that the principal practical value of the model is the rate-design rule rather than online prediction. Delay reconstruction had no resolvable effect at the tested latency. A single-subject popliteal scan demonstrated feasibility, although the image cue was noise-limited on heterogeneous tissue.

We study the problem of human gaze modeling, which aims to generate the gaze patterns a viewer produces while observing a visual stimulus. Gaze is primarily captured through two modalities: continuous eye-tracking trajectories, which describe fine-grained motion dynamics, and discrete scanpaths, which describe high-level fixation structure. Because gaze varies substantially across viewers and trials, we treat this variability as a defining property rather than noise and model gaze as a stochastic generative process. Existing generative gaze models supervise on only one of these two representations in isolation. We hypothesize that trajectories and scanpaths describe gaze at complementary scales and are jointly informative during training, and test this hypothesis through ST-DiffEye, a joint trajectory-scanpath diffusion framework that couples both modalities by concatenating them as an additional raw input channel, requiring no architectural overhead beyond an input and output channel expansion. We further introduce a principled evaluation framework based on the Continuous Ranked Probability Score (CRPS), which generalizes any existing sequence similarity metric into a proper scoring rule that jointly assesses the accuracy and diversity of generated gaze. Experiments on task-driven visual search, covering both target-present and target-absent scenarios, and on free-viewing benchmarks demonstrate state-of-the-art performance. These results, along with detailed ablations, confirm the benefit of joint modeling and the value of distribution-aware evaluation in capturing the intrinsic variability of human gaze. Project webpage: https://st-diffeye.github.io/

A common recipe to improve diffusion models at test-time so that samples score highly against a user-specified reward is to introduce the gradient of the reward into the dynamics of the diffusion itself. This procedure is often ill posed, as user-specified rewards are usually only well defined on the data distribution at the end of generation. While common workarounds to this problem are to use a denoiser to estimate what a sample would have been at the end of generation, we propose a simple solution to this problem by working directly with a flow map. By exploiting a relationship between the flow map and velocity field governing the instantaneous transport, we construct an algorithm, Flow Map Trajectory Tilting (FMTT), which provably performs better ascent on the reward than standard test-time methods involving the gradient of the reward. The approach can be used to either perform exact sampling via importance weighting or principled search that identifies local maximizers of the reward-tilted distribution. We demonstrate the efficacy of our approach against other look-ahead techniques, and show how the flow map enables engagement with complicated reward functions that make possible new forms of image editing, e.g. by interfacing with vision language models.

Diffusion models produce impressive results in modalities ranging from images and video to protein design and text. However, generating samples with user-specified properties remains a challenge. Recent research proposes fine-tuning models to maximize rewards that capture desired properties, but these methods require expensive training and are prone to mode collapse. In this work, we propose Feynman Kac (FK) steering, an inference-time framework for steering diffusion models with reward functions. FK steering works by sampling a system of multiple interacting diffusion processes, called particles, and resampling particles at intermediate steps based on scores computed using functions called potentials. Potentials are defined using rewards for intermediate states and are selected such that a high value indicates that the particle will yield a high-reward sample. We explore various choices of potentials, intermediate rewards, and samplers. We evaluate FK steering on text-to-image and text diffusion models. For steering text-to-image models with a human preference reward, we find that FK steering a 0.8B parameter model outperforms a 2.6B parameter fine-tuned model on prompt fidelity, with faster sampling and no training. For steering text diffusion models with rewards for text quality and specific text attributes, we find that FK steering generates lower perplexity, more linguistically acceptable outputs and enables gradient-free control of attributes like toxicity. Our results demonstrate that inference-time scaling and steering of diffusion models, even with off-the-shelf rewards, can provide significant sample quality gains and controllability benefits. Code is available at https://github.com/zacharyhorvitz/Fk-Diffusion-Steering .

Cinematic video generation requires both temporal inference from limited inputs (i.e., sparse rather than frame-by-frame guidance) and factor-aware control of character, scene, action, and camera. Multimodal large language models (MLLMs) offer useful priors for both capabilities, but their representations are not automatically actionable for video generation: a Video Diffusion Transformer (DiT) may bypass them through native conditioning pathways, and using them does not ensure either factor-aware control or generation-compatible temporal prediction. We propose Factor-Aware Predictive Learning (FAPL), which establishes effective MLLM latent use as a shared prerequisite and then targets the two capabilities with complementary learning signals. First, latent-only generator grounding trains the DiT with a frozen MLLM and connector, strengthening the DiT's use of MLLM latents for generation. Second, for factor-aware control, paired videos specify a target-factor change and non-target preservation, and factor-span replacement trains the DiT to realize this relationship in its output. Third, for temporal inference, text-only and text-plus-complete-video observations share a video target, and the trained DiT is then frozen while the MLLM and connector are optimized through it. This final stage trains limited-input representations for generation while retaining complete-observation reconstruction as a complementary training mode. Extensive experiments demonstrate our consistently better performance on three cinematic video generation tasks (i.e., sparse-to-video, factor-aware editing, and multi-shot generation) over baselines and validate FAPL's contributions to MLLM feature use, temporal control, and factor coordination.

Video inbetweening creates smooth and natural transitions between two image frames, making it an indispensable tool for video editing and long-form video synthesis. Existing works in this domain are unable to generate large, complex, or intricate motions. In particular, they cannot accommodate the versatility of user intents and generally lack fine control over the details of intermediate frames, leading to misalignment with the creative mind. To fill these gaps, we introduce MultiCOIN, a video inbetweening framework that allows multi-modal controls, including depth transition and layering, motion trajectories, text prompts, and target regions for movement localization, while achieving a balance between flexibility, ease of use, and precision for fine-grained video interpolation. To achieve this, we adopt the Diffusion Transformer (DiT) architecture as our video generative model, due to its proven capability to generate high-quality long videos. To ensure compatibility between DiT and our multi-modal controls, we map all motion controls into a common sparse and user-friendly point-based representation as the video/noise input. Further, to respect the variety of controls which operate at varying levels of granularity and influence, we separate content controls and motion controls into two branches to encode the required features before guiding the denoising process, resulting in two generators, one for motion and the other for content. Finally, we propose a stage-wise training strategy to ensure that our model learns the multi-modal controls smoothly. Extensive qualitative and quantitative experiments demonstrate that multi-modal controls enable a more dynamic, customizable, and contextually accurate visual narrative.

Talking head synthesis is vital for virtual avatars and human-computer interaction. However, most existing methods are typically limited to accepting control from a single primary modality, restricting their practical utility. To this end, we introduce ACTalker, an end-to-end video diffusion framework that supports both multi-signals control and single-signal control for talking head video generation. For multiple control, we design a parallel mamba structure with multiple branches, each utilizing a separate driving signal to control specific facial regions. A gate mechanism is applied across all branches, providing flexible control over video generation. To ensure natural coordination of the controlled video both temporally and spatially, we employ the mamba structure, which enables driving signals to manipulate feature tokens across both dimensions in each branch. Additionally, we introduce a mask-drop strategy that allows each driving signal to independently control its corresponding facial region within the mamba structure, preventing control conflicts. Experimental results demonstrate that our method produces natural-looking facial videos driven by diverse signals and that the mamba layer seamlessly integrates multiple driving modalities without conflict.

Language-guided human following is an important capability for embodied agents, but existing benchmarks typically assume that the target person is visible at the start of an episode. This setting simplifies the problem and overlooks a more realistic requirement: an agent often needs to first find a language-described target and then persistently follow that target in a dynamic environment. While recent work has started to study human search, existing settings are typically evaluated in task-specific scenarios and often rely on stronger prior knowledge of the environment. Moreover, they usually treat searching and following as separate tasks and still lack a unified benchmark for systematic evaluation. To address these limitations, we introduce the Unified Embodied Seeking and Following Benchmark (UESF-Bench), a large-scale and diverse benchmark for embodied human seeking and following. The benchmark requires agents to handle semantic-guided exploration, reliable behavior switching and recovery, and delayed identity grounding. To this end, we propose SeekFollow-VLA, a vision-language-action framework with a task-driven routing mechanism for latent phase inference and transition modeling between seeking and following. Experimental results show that SeekFollow-VLA achieves clear improvements over both single-head and dual-head baselines across single-person and multi-person environments, establishing a baseline for unified embodied seek-and-follow.

While recent advancements in generative models have achieved remarkable visual fidelity in video synthesis, creating coherent multi-shot narratives remains a significant challenge. To address this, keyframe-based approaches have emerged as a promising alternative to computationally intensive end-to-end methods, offering the advantages of fine-grained control and greater efficiency. However, these methods often fail to maintain cross-shot consistency and capture cinematic language. In this paper, we introduce STAGE, a SToryboard-Anchored GEneration workflow to reformulate the keyframe-based multi-shot video generation task. Instead of using sparse keyframes, we propose STEP2 to predict a structural storyboard composed of start-end frame pairs for each shot. We introduce the multi-shot memory pack to ensure long-range entity consistency, the dual-encoding strategy for intra-shot coherence, and the two-stage training scheme to learn cinematic inter-shot transition. We also contribute the large-scale ConStoryBoard dataset, including high-quality movie clips with fine-grained annotations for story progression, cinematic attributes, and human preferences. Extensive experiments demonstrate that STAGE achieves superior performance in structured narrative control and cross-shot coherence. Our code will be available at this url.

Robot foundation models achieve strong in-distribution performance but often degrade under visual distribution shifts. When learning to generate actions from pretrained visual representations, models may exploit task-irrelevant visual cues that correlate with demonstrated actions within the training distribution. Such vision-action shortcuts can undermine generalization when these correlations change under distribution shifts. Mitigating these shortcuts requires constraining how visual information is used for action generation while preserving task-relevant spatial information. We propose Latent Interface Training (LIT), a framework-agnostic two-stage strategy that first establishes a spatial-goal-conditioned action prior without images, then constrains visual conditioning through a pose-supervised latent interface. Stage 1 trains the action expert to generate action chunks conditioned on language, robot state, and each demonstrated chunk's terminal SE(3) end-effector pose, learning goal-directed action generation independently of visual cues. Stage 2 introduces a latent interface that aggregates visual and semantic representations and serves as the pretrained action expert's only visual conditioning pathway. The interface is supervised to reconstruct the terminal pose previously used to condition Stage 1, encouraging it to retain the goal-relevant spatial information needed for action generation. Across four vision-language-action and world-action architectures (Pi0.5, MolmoAct2, FAST-WAM, and ImageWAM), LIT improves overall LIBERO-Plus success by 3.87-10.70 percentage points while preserving or improving average LIBERO success. Real-world evaluations show 13.30-16.70 percentage-point gains in success aggregated across three tasks under unseen camera configurations, lighting variations, and distractors.

Text-driven video editing aims to modify video content according to natural language instructions. While recent training-free approaches have made progress by leveraging pre-trained diffusion models, they typically rely on inversion-based techniques that map input videos into the latent space, which often leads to temporal inconsistencies and degraded structural fidelity. To address this, we propose FlowDirector, a novel inversion-free video editing framework. Our framework models the editing process as a direct evolution in data space, guiding the video via an Ordinary Differential Equation (ODE) to smoothly transition along its inherent spatiotemporal manifold, thereby preserving temporal coherence and structural details. To achieve localized and controllable edits, we introduce an attention-guided masking mechanism that modulates the ODE velocity field, preserving non-target regions both spatially and temporally. Furthermore, to address incomplete edits and enhance semantic alignment with editing instructions, we present a guidance-enhanced editing strategy inspired by Classifier-Free Guidance, which leverages differential signals between multiple candidate flows to steer the editing trajectory toward stronger semantic alignment without compromising structural consistency. Extensive experiments across benchmarks demonstrate that FlowDirector achieves state-of-the-art performance in instruction adherence, temporal consistency, and background preservation, establishing a new paradigm for efficient and coherent video editing without inversion.

Classifier-free guidance (CFG) typically applies a shared scale to all latent tokens, although their velocity predictions evolve differently during sampling. We introduce Trajectory-Adaptive Classifier-Free Guidance (TrajCFG), a training-free method for flow-matching models that adapts token-wise guidance using historical changes in guided velocity direction. TrajCFG maps this feedback to bounded scales, assigning stronger guidance to larger directional changes and weaker guidance to more consistent directions. It combines adjacent-step and interval-normalized longer-span feedback, reusing standard CFG predictions without additional model evaluations. Controlled diagnostics and ablations examine the roles of directional feedback, token-wise allocation, and history-span fusion. Experiments on text-to-video, image-to-video, text-to-audio-video, text-to-image, and image-to-3D generation show improvements over base samplers across multiple backbones. For text-to-video generation, TrajCFG improves VBench Total scores by 0.30-0.75 points over the strongest evaluated baseline on each of three backbones. These results support historical directional feedback as a broadly applicable signal for fine-grained guidance allocation. The project page is available at https://jyliu-98.github.io/trajcfg-project-page/.

In this work, we propose a training-free, trajectory-based controllable T2I approach, termed TraDiffusion. This novel method allows users to effortlessly guide image generation via mouse trajectories. To achieve precise control, we design a distance awareness energy function to effectively guide latent variables, ensuring that the focus of generation is within the areas defined by the trajectory. The energy function encompasses a control function to draw the generation closer to the specified trajectory and a movement function to diminish activity in areas distant from the trajectory. Through extensive experiments and qualitative assessments on the COCO dataset, the results reveal that TraDiffusion facilitates simpler, more natural image control. Moreover, it showcases the ability to manipulate salient regions, attributes, and relationships within the generated images, alongside visual input based on arbitrary or enhanced trajectories.

Large language models can now process increasingly long inputs, yet their ability to effectively use information spread across long contexts remains limited. We trace this gap to how attention budget is spent during supervised fine-tuning (SFT) on long sequences: positional biases and attention sinks cause the model to allocate most of its attention to positionally privileged tokens rather than semantically relevant content. This training-time attention dilution (the starvation of content tokens in the attention distribution) weakens the gradient signal, limiting the model's ability to learn robust long-context capabilities. We introduce FocuSFT, a bilevel optimization framework that addresses this problem at training time. An inner loop adapts lightweight fast-weight parameters on the training context to form a parametric memory that concentrates attention on relevant content, and the outer loop performs SFT conditioned on this sharpened representation. Both loops apply bidirectional attention over context tokens while preserving causal masking for responses, reducing the causal asymmetry that gives rise to attention sinks and aligning inner-outer behavior. On BABILong, FocuSFT improves accuracy by up to +14pp across 4K--32K context lengths; on RULER, it raises CWE aggregation from 72.9\% to 81.1\% at 16K; and on GPQA with agentic tool use, it yields a 24\% relative gain in pass@1. Attention analysis shows that FocuSFT reduces attention sink mass by 529times and triples context engagement during training. Code: https://github.com/JarvisPei/FocuSFT

Modern Text-to-Image (T2I) diffusion models have achieved remarkable semantic alignment, yet they often suffer from a significant lack of variety, converging on a narrow set of visual solutions for any given prompt. This typicality bias presents a challenge for creative applications that require a wide range of generative outcomes. We identify a fundamental trade-off in current approaches to diversity: modifying model inputs requires costly optimization to incorporate feedback from the generative path. In contrast, acting on spatially-committed intermediate latents tends to disrupt the forming visual structure, leading to artifacts. In this work, we propose to apply repulsion in the Contextual Space as a novel framework for achieving rich diversity in Diffusion Transformers. By intervening in the multimodal attention channels, we apply on-the-fly repulsion during the transformer's forward pass, injecting the intervention between blocks where text conditioning is enriched with emergent image structure. This allows for redirecting the guidance trajectory after it is structurally informed but before the composition is fixed. Our results demonstrate that repulsion in the Contextual Space produces significantly richer diversity without sacrificing visual fidelity or semantic adherence. Furthermore, our method is uniquely efficient, imposing a small computational overhead while remaining effective even in modern "Turbo" and distilled models where traditional trajectory-based interventions typically fail.

Vision--Language--Action (VLA) policies such as π_{0.5} and OpenVLA perform well on many manipulation tasks, but they are often reactive: the next action is predicted from the current observation, instruction, and proprioceptive state. This assumption breaks down in stage-dependent manipulation, where visually similar states may require different actions depending on latent task progress and previous interaction outcomes. We argue that such tasks require not only memory, but dynamics-aware belief updates: the policy should preserve task progress during stable or occluded phases and revise its belief near contact, release, or subgoal transitions. We introduce Temporally Conditioned Memory-Fusion Policies (TFP), a lightweight memory-action framework for VLA backbones. TFP maintains an episode-local task-progress belief with Liquid Time-Constant dynamics and injects the updated belief directly into the flow-matching action decoder through adaptive modulation. This lets temporally accumulated context shape the generated action chunk, rather than serving only as passive history context. With a 3.3B-parameter model, TFP improves the average success rate from \(96.9\%\) to \(98.75\%\) on LIBERO and from \(91.4\%\) to \(93.77\%\) on LIBERO-plus. On the memory-focused MIKASA ShellGameTouch diagnostic, TFP achieves success up to \(75.0\%\). Mechanistic analyses show that write-gain changes near manipulation events are about \(6\times\) larger than far non-event phases, and hidden-state interventions show that the belief causally modulates generated action chunks. These results suggest that compact, event-sensitive memory dynamics can improve VLA policies under occlusion, visual perturbation, and stage-dependent task structure.

Robotic manipulation integrates vision, touch, and language, whose importance shifts across stages: vision guides reaching, while touch, through its evolution over time, decides grasping, alignment, and contact. Yet existing multi-modal manipulation policies typically use fixed temporal contexts and fusion strategies, despite shifts in what each modality contributes across different skills. We study how vision and touch should be combined at the level of primitive skills, asking what each skill needs from each sensor, and propose a skill-conditioned representation in which the queried skill conditions fusion over modality-specific short-term observation tokens while attending to a sparse event memory that retains terminal observations from the last K executed skills. Evaluated by skill progress estimation on three contact-rich tasks, it reduces slip-detection delay by 87% against fine-tuned SOTA progress models, twist-completion delay by 67.5% against a vision-only ablation, and progress error on a blind search task by 92% through sparse event memory. Gains concentrate exactly where completion is defined by contact or task history. More broadly, our results suggest that observation formation not only policy architecture is a central challenge in multi-modal representation. Project Website: http://what-to-attend-what-to-keep.github.io/

LLM-based multi-agent pipelines flip from correct to incorrect answers under simulated peer disagreement at rates we term yield, a vulnerability widely attributed to RLHF-induced sycophancy. We test this attribution across four model families and find it largely wrong: pretrained base models exhibit the same substitution pattern as their Instruct variants, averaging higher yield than Instruct. Using activation patching, we localize the corruption to a narrow mid-layer window where attention carries the causal weight and MLP contribution is negligible; patching above this window restores 96% of the clean-to-pressured P(correct) gap. The attack surface decomposes into two independent factors (channel framing and consensus strength) whose interaction produces a 47.5 percentage-point yield gap at majority consensus, preserved across jury sizes N in {4, 5, 6}. Two converging activation-space interventions show that pressure suppresses clean-reasoning features rather than activating a new sycophancy circuit. A single correctly-arguing dissenter reduces yield by 54-73 percentage points across all framings tested, whereas the strongest prompt-level defense fails on attack variants outside its design surface. Mitigations should target the mechanism, structured dissent at the pipeline level, rather than prompt-level defenses.

Recent advances in text-conditioned video diffusion have greatly improved video quality. However, these methods offer limited or sometimes no control to users on camera aspects, including dynamic camera motion, zoom, distorted lens and focus shifts. These motion and optical aspects are crucial for adding controllability and cinematic elements to generation frameworks, ultimately resulting in visual content that draws focus, enhances mood, and guides emotions according to filmmakers' controls. In this paper, we aim to close the gap between controllable video generation and camera optics. To achieve this, we propose AKiRa (Augmentation Kit on Rays), a novel augmentation framework that builds and trains a camera adapter with a complex camera model over an existing video generation backbone. It enables fine-tuned control over camera motion as well as complex optical parameters (focal length, distortion, aperture) to achieve cinematic effects such as zoom, fisheye effect, and bokeh. Extensive experiments demonstrate AKiRa's effectiveness in combining and composing camera optics while outperforming all state-of-the-art methods. This work sets a new landmark in controlled and optically enhanced video generation, paving the way for future optical video generation methods.

Despite recent advances in image-to-video generation, better controllability and local animation are less explored. Most existing image-to-video methods are not locally aware and tend to move the entire scene. However, human artists may need to control the movement of different objects or regions. Additionally, current I2V methods require users not only to describe the target motion but also to provide redundant detailed descriptions of frame contents. These two issues hinder the practical utilization of current I2V tools. In this paper, we propose a practical framework, named Follow-Your-Click, to achieve image animation with a simple user click (for specifying what to move) and a short motion prompt (for specifying how to move). Technically, we propose the first-frame masking strategy, which significantly improves the video generation quality, and a motion-augmented module equipped with a short motion prompt dataset to improve the short prompt following abilities of our model. To further control the motion speed, we propose flow-based motion magnitude control to control the speed of target movement more precisely. Our framework has simpler yet precise user control and better generation performance than previous methods. Extensive experiments compared with 7 baselines, including both commercial tools and research methods on 8 metrics, suggest the superiority of our approach. Project Page: https://follow-your-click.github.io/

Unified diffusion editors often rely on a fixed, shared backbone for diverse tasks, suffering from task interference and poor adaptation to heterogeneous demands (e.g., local vs global, semantic vs photometric). In particular, prevalent ControlNet and OmniControl variants combine multiple conditioning signals (e.g., text, mask, reference) via static concatenation or additive adapters which cannot dynamically prioritize or suppress conflicting modalities, thus resulting in artifacts like color bleeding across mask boundaries, identity or style drift, and unpredictable behavior under multi-condition inputs. To address this, we propose Condition-Aware Routing of Experts (CARE-Edit) that aligns model computation with specific editing competencies. At its core, a lightweight latent-attention router assigns encoded diffusion tokens to four specialized experts--Text, Mask, Reference, and Base--based on multi-modal conditions and diffusion timesteps: (i) a Mask Repaint module first refines coarse user-defined masks for precise spatial guidance; (ii) the router applies sparse top-K selection to dynamically allocate computation to the most relevant experts; (iii) a Latent Mixture module subsequently fuses expert outputs, coherently integrating semantic, spatial, and stylistic information to the base images. Experiments validate CARE-Edit's strong performance on contextual editing tasks, including erasure, replacement, text-driven edits, and style transfer. Empirical analysis further reveals task-specific behavior of specialized experts, showcasing the importance of dynamic, condition-aware processing to mitigate multi-condition conflicts.

Point tracking in videos is a fundamental task with applications in robotics, video editing, and more. While many vision tasks benefit from pre-trained feature backbones to improve generalizability, point tracking has primarily relied on simpler backbones trained from scratch on synthetic data, which may limit robustness in real-world scenarios. Additionally, point tracking requires temporal awareness to ensure coherence across frames, but using temporally-aware features is still underexplored. Most current methods often employ a two-stage process: an initial coarse prediction followed by a refinement stage to inject temporal information and correct errors from the coarse stage. These approach, however, is computationally expensive and potentially redundant if the feature backbone itself captures sufficient temporal information. In this work, we introduce Chrono, a feature backbone specifically designed for point tracking with built-in temporal awareness. Leveraging pre-trained representations from self-supervised learner DINOv2 and enhanced with a temporal adapter, Chrono effectively captures long-term temporal context, enabling precise prediction even without the refinement stage. Experimental results demonstrate that Chrono achieves state-of-the-art performance in a refiner-free setting on the TAP-Vid-DAVIS and TAP-Vid-Kinetics datasets, among common feature backbones used in point tracking as well as DINOv2, with exceptional efficiency. Project page: https://cvlab-kaist.github.io/Chrono/

Event cameras are increasingly popular in robotics due to beneficial features such as low latency, energy efficiency, and high dynamic range. Nevertheless, their downstream task performance is greatly influenced by the optimization of bias parameters. These parameters, for instance, regulate the necessary change in light intensity to trigger an event, which in turn depends on factors such as the environment lighting and camera motion. This paper introduces feedback control algorithms that automatically tune the bias parameters through two interacting methods: 1) An immediate, on-the-fly fast adaptation of the refractory period, which sets the minimum interval between consecutive events, and 2) if the event rate exceeds the specified bounds even after changing the refractory period repeatedly, the controller adapts the pixel bandwidth and event thresholds, which stabilizes after a short period of noise events across all pixels (slow adaptation). Our evaluation focuses on the visual place recognition task, where incoming query images are compared to a given reference database. We conducted comprehensive evaluations of our algorithms' adaptive feedback control in real-time. To do so, we collected the QCR-Fast-and-Slow dataset that contains DAVIS346 event camera streams from 366 repeated traversals of a Scout Mini robot navigating through a 100 meter long indoor lab setting (totaling over 35km distance traveled) in varying brightness conditions with ground truth location information. Our proposed feedback controllers result in superior performance when compared to the standard bias settings and prior feedback control methods. Our findings also detail the impact of bias adjustments on task performance and feature ablation studies on the fast and slow adaptation mechanisms.

Imitation learning is an efficient method for teaching robots a variety of tasks. Diffusion Policy, which uses a conditional denoising diffusion process to generate actions, has demonstrated superior performance, particularly in learning from multi-modal demonstrates. However, it relies on executing multiple actions predicted from the same inference step to retain performance and prevent mode bouncing, which limits its responsiveness, as actions are not conditioned on the most recent observations. To address this, we introduce Responsive Noise-Relaying Diffusion Policy (RNR-DP), which maintains a noise-relaying buffer with progressively increasing noise levels and employs a sequential denoising mechanism that generates immediate, noise-free actions at the head of the sequence, while appending noisy actions at the tail. This ensures that actions are responsive and conditioned on the latest observations, while maintaining motion consistency through the noise-relaying buffer. This design enables the handling of tasks requiring responsive control, and accelerates action generation by reusing denoising steps. Experiments on response-sensitive tasks demonstrate that, compared to Diffusion Policy, ours achieves 18% improvement in success rate. Further evaluation on regular tasks demonstrates that RNR-DP also exceeds the best acceleration method (DDIM) by 6.9% in success rate, highlighting its computational efficiency advantage in scenarios where responsiveness is less critical. Our project page is available at https://rnr-dp.github.io

Autoregressive (AR) video diffusion models enable low-latency streaming generation by synthesizing videos chunk by chunk with cached visual context, but this chunk-wise formulation makes temporal instruction following ambiguous. A single global prompt does not specify which sub-event should be realized in each chunk, while naively switching to step-wise prompts often leads to delayed reactions, blended step semantics, and error propagation across prompt transitions. These failures are difficult to address with supervised fine-tuning or distillation alone: SFT suffers from exposure bias, while rollout-based distillation still optimizes low-level denoising or teacher-distribution matching rather than directly enforcing action ordering and prompt-transition correctness. We address these challenges with TempAct, a planner--executor reinforcement learning framework that jointly optimizes temporal decomposition and step-conditioned execution for temporally plausible AR video generation. TempAct uses an LLM planner to explore span-aware step prompts that are executable by the video model, and trains an AR diffusion executor to follow these prompts under its own generated histories. Its key mechanism is hierarchical group exploration: candidate plans form planning groups, and each plan induces an execution group of multiple continuations from a shared visual context, enabling plan-level credit assignment for long-horizon temporal outcomes and executor-level credit assignment for prompt-switch behavior. We further design hierarchical rewards that combine plan-quality and full-video temporal feedback for the planner with local transition-level step-following rewards, aesthetic regularization, and KL constraints for the executor. Experiments on Self-Forcing and LongLive show that TempAct improves temporal consistency while preserving overall visual quality.

Contact-rich policies often fail because distinct physical states look alike yet require different actions. Cameras may not reveal whether a connector is aligned or fully seated, while many normal-only tactile sensors can miss the tangential interactions perpendicular to the grasping direction that distinguish these states. We ask whether a learning policy needs calibrated shear measurements, or only a repeatable observation that separates shear-dependent contact states. We introduce TacGooseBumps (TacGB), a passive domed film that mechanically encodes tangential loading as pattern changes in an existing sensor's pressure map. Tangential loading tilts each dome and redistributes pressure across its footprint; an end-to-end policy consumes the resulting maps without added electronics, force reconstruction, or taxel-level dome alignment. Across four imitation-learning tasks and two data-collection pipelines, TacGB improves goal attainment, efficiency, and contact quality: insertion success increases by up to 36 percentage points, and successful insertions are completed faster, while fragile-object placement becomes gentler and drawing becomes more continuous and straight. Signal, stage-wise, failure-mode, and trajectory analyses link these gains to contact regimes in which task-relevant tangential interactions are poorly resolved by vision and normal pressure alone. Together, these results show that shear need not be measured metrically to benefit robot learning; it can instead be mechanically encoded without changing the underlying tactile sensor or the policy's pressure-map input format.

Robust depth estimation under dynamic and adverse lighting conditions is essential for robotic systems. Currently, depth foundation models, such as Depth Anything, achieve great success in ideal scenes but remain challenging under adverse imaging conditions such as extreme illumination and motion blur. These degradations corrupt the visual signals of frame cameras, weakening the discriminative features of frame-based depths across the spatial and temporal dimensions. Typically, existing approaches incorporate event cameras to leverage their high dynamic range and temporal resolution, aiming to compensate for corrupted frame features. However, such specialized fusion models are predominantly trained from scratch on domain-specific datasets, thereby failing to inherit the open-world knowledge and robust generalization inherent to foundation models. In this work, we propose ADAE, an event-guided spatiotemporal fusion framework for Depth Anything in degraded scenes. Our design is guided by two key insights: 1) Entropy-Aware Spatial Fusion. We adaptively merge frame-based and event-based features using an information entropy strategy to indicate illumination-induced degradation. 2) Motion-Guided Temporal Correction. We resort to the event-based motion cue to recalibrate ambiguous features in blurred regions. Under our unified framework, the two components are complementary to each other and jointly enhance Depth Anything under adverse imaging conditions. Extensive experiments have been performed to verify the superiority of the proposed method. Our code will be released upon acceptance.

Generating lifelike conversational avatars requires modeling not just isolated speakers, but the dynamic, reciprocal interaction of speaking and listening. However, modeling the listener is exceptionally challenging: direct audio-driven training fails, producing stiff, static listening motions. This failure stems from a fundamental imbalance: the speaker's motion is strongly driven by speech audio, while the listener's motion primarily follows an internal motion prior and is only loosely guided by external speech. This challenge has led most methods to focus on speak-only generation. The only prior attempt at joint generation relies on extra speaker's motion to produce the listener. This design is not end-to-end, thereby hindering the real-time applicability. To address this limitation, we present UniLS, the first end-to-end framework for generating unified speak-listen expressions, driven by only dual-track audio. Our method introduces a novel two-stage training paradigm. Stage 1 first learns the internal motion prior by training an audio-free autoregressive generator, capturing the spontaneous dynamics of natural facial motion. Stage 2 then introduces the dual-track audio, fine-tuning the generator to modulate the learned motion prior based on external speech cues. Extensive evaluations show UniLS achieves state-of-the-art speaking accuracy. More importantly, it delivers up to 44.1\% improvement in listening metrics, generating significantly more diverse and natural listening expressions. This effectively mitigates the stiffness problem and provides a practical, high-fidelity audio-driven solution for interactive digital humans.

Contact-rich manipulation requires not only vision-dominant task semantics but also closed-loop reactions to force/torque (F/T) transients. Yet, generative visuomotor policies are typically constrained to low-frequency updates due to inference latency and action chunking, underutilizing F/T for control-rate feedback. Furthermore, existing force-aware methods often inject force continuously and indiscriminately, lacking an explicit mechanism to schedule when / how much / where to apply force across different task phases. We propose PhaForce, a phase-scheduled visual--force policy that coordinates low-rate chunk-level planning and high-rate residual correction via a unified contact/phase schedule. PhaForce comprises (i) a contact-aware phase predictor (CAP) that estimates contact probability and phase belief, (ii) a Slow diffusion planner that performs dual-gated visual--force fusion with orthogonal residual injection to preserve vision semantics while conditioning on force, and (iii) a Fast corrector that applies control-rate phase-routed residuals in interpretable corrective subspaces for within-chunk micro-adjustments. Across multiple real-robot contact-rich tasks, PhaForce achieves an average success rate of 86% (+40 pp over baselines), while also substantially improving contact quality by regulating interaction forces and exhibiting robust adaptability to OOD geometric shifts.

Recent advancements in video diffusion models have significantly enhanced audio-driven portrait animation. However, current methods still suffer from flickering, identity drift, and poor audio-visual synchronization. These issues primarily stem from entangled appearance-motion representations and unstable inference strategies. In this paper, we introduce ConsistTalk, a novel intensity-controllable and temporally consistent talking head generation framework with diffusion noise search inference. First, we propose an optical flow-guided temporal module (OFT) that decouples motion features from static appearance by leveraging facial optical flow, thereby reducing visual flicker and improving temporal consistency. Second, we present an Audio-to-Intensity (A2I) model obtained through multimodal teacher-student knowledge distillation. By transforming audio and facial velocity features into a frame-wise intensity sequence, the A2I model enables joint modeling of audio and visual motion, resulting in more natural dynamics. This further enables fine-grained, frame-wise control of motion dynamics while maintaining tight audio-visual synchronization. Third, we introduce a diffusion noise initialization strategy (IC-Init). By enforcing explicit constraints on background coherence and motion continuity during inference-time noise search, we achieve better identity preservation and refine motion dynamics compared to the current autoregressive strategy. Extensive experiments demonstrate that ConsistTalk significantly outperforms prior methods in reducing flicker, preserving identity, and delivering temporally stable, high-fidelity talking head videos.

Recent advances in Diffusion Transformers (DiTs) have enabled high-quality joint audio-video generation, producing videos with synchronized audio within a single model. However, existing controllable generation frameworks are typically restricted to video-only control. This restricts comprehensive controllability and often leads to suboptimal cross-modal alignment. To bridge this gap, we present MMControl, which enables users to perform Multi-Modal Control in joint audio-video generation. MMControl introduces a dual-stream conditional injection mechanism. It incorporates both visual and acoustic control signals, including reference images, reference audio, depth maps, and pose sequences, into a joint generation process. These conditions are injected through bypass branches into a joint audio-video Diffusion Transformer, enabling the model to simultaneously generate identity-consistent video and timbre-consistent audio under structural constraints. Furthermore, we introduce modality-specific guidance scaling, which allows users to independently and dynamically adjust the influence strength of each visual and acoustic condition at inference time. Extensive experiments demonstrate that MMControl achieves fine-grained, composable control over character identity, voice timbre, body pose, and scene layout in joint audio-video generation.

Large Language Models (LLMs) often allocate disproportionate attention to specific tokens, a phenomenon commonly referred to as the attention sink. While such sinks are generally considered detrimental, prior studies have identified a notable exception: the model's consistent emphasis on the first token of the input sequence. This structural bias can influence a wide range of downstream applications and warrants careful consideration. Despite its prevalence, the precise mechanisms underlying the emergence and persistence of attention sinks remain poorly understood. In this work, we trace the formation of attention sinks around the first token of the input. We identify a simple mechanism, referred to as the P0 Sink Circuit, that enables the model to recognize token at position zero and induce an attention sink within two transformer blocks, without relying on any semantic information. This mechanism serves as the basis for the attention sink on position zero. Furthermore, by analyzing training traces from a 30B A3B MoE model trained from scratch, we find that this mechanism emerges early in training and becomes increasingly concentrated in the first two layers, suggesting a possible signal for tracking pre training convergence states.

Audio-driven streaming avatar generation requires real-time synthesis of speech-synchronized videos with dynamic and diverse motion. Self Forcing uses Distribution Matching Distillation (DMD) to distill bidirectional video diffusion models into causal, few-step generators for real-time streaming. However, DMD minimizes a reverse KL divergence, which is inherently mode-seeking: it causes the student to discard high-dynamic modes and collapse onto static outputs, compressing both dynamics and diversity of generated videos. We find that this collapse is region-heterogeneous: person regions involving pose and gesture variations suffer the largest diversity loss, the audio-driven mouth region shows a small loss, and the background remains nearly stable. Based on this observation, we propose Routed Forcing, which routes the distillation objective by semantic region and noise stage to improve dynamics and diversity while preserving visual quality. Specifically, (1) Where to Force: Semantic-Region Routing applies Data-Forcing Distillation (DFD), which supervises the student with real videos, to the person region where diversity collapse is most severe, while retaining DMD for the mouth and background to preserve lip synchronization and scene stability. (2) When to Force: Noise-Stage Routing activates DFD at high noise stages, where real video serves as effective supervision to inject diverse and dynamic motion patterns. At low noise stages, DMD is used to refine details, avoiding blur and artifacts from spatial differences between real video and student-generated video. Experiments show that Routed Forcing improves dynamics by up to 45% and diversity by 7-25% over Self Forcing, while preserving video quality and lip synchronization.

Search-augmented recommenders can use the latest search episode and earlier searches. We introduce FocalSwap to test how these inputs contribute to the next recommendation. For each recommendation, it changes the latest search slot or all search inputs while holding the target, candidates, pre-search history, and scorer fixed. It compares neutral masks with matched observed-search replacements, recording rank changes, RR gains relative to each control, and preference for the true focal search. A planted signal and GenSAR's current-query task confirm that the audit detects known input dependence. We evaluate behavioral and lexical evidence on KuaiSAR and Qilin, and recommendation checkpoints trained using the released SESRec and UniSAR implementations. Across three independently trained UniSAR checkpoints, the estimated all-source benefit depends on the control. On the same 66,139 rows, replacing all search inputs with a neutral ID yields a positive signed contribution, whereas replacing them with category/profile-matched observed search sequences gives smaller estimates, with an interval below zero for one checkpoint. The paired RR gaps range from +2.770 times 10^{-3} to +6.604 times 10^{-3}, and every adjusted interval is above zero. Neutral masks change 59.7-61.7% of target ranks, while matched replacements change 44.1-44.8%. The control contrast persists across five alternative donor assignments per checkpoint. Under the primary neutral-mask and category-matched controls, positive latest-slot contribution is not established in SESRec or UniSAR. These results show why masking evidence should be compared with observed-search replacements when attributing a benefit to broader search state or the latest episode.

Aligned language models gate behaviors such as refusal and language routing through sparse feed forward neurons, yet no theory predicts when a single neuron intervention controls a behavior coherently rather than collapsing the output. We develop a budget normalized control window framework for single neuron steering. A dose along one write direction reduces to one control coordinate: the alignment between the residual stream and the write, driven along a universal saturation curve in units of a coherence budget set by the residual norm divided by the write norm. Coherent control exists when a behavior trigger lies below the collapse ceiling. The same coordinate governs benign mode switches and refusal; the ceiling follows from weights and one generic forward pass, while triggers are measured at rollout. On fifteen held out neurons, the predicted ceiling has mean absolute error 0.14, about 0.07 in bulk layers, and the committed open or closed verdict holds on eleven against a ten of fifteen majority baseline. Closed cases expose three failure modes rather than violations: collapse before trigger, too little depth to propagate, or a normalization that caps how far one neuron can push. The law explains why local gradient attribution anti predicts control: true controllers write off the readout axis and carry a near zero first order gradient. A forward only contrastive screen made precise by the window recovers controllers that attribution misses. On refusal, the hardest case, intervention success is typed, not scalar: coherent bypass and strict actionable reach separate, so a neuron can flip refusal in fluent, on task text with no actionable content, and genuine actionable reach appears only for three of six audited Llama pivots and only at later rollout horizons. Single neuron steering is therefore a budgeted, typed audit of controllability rather than a fixed dose anecdote.

Recently, active vision has reemerged as an important concept for manipulation, since visual occlusion occurs more frequently when main cameras are mounted on the robot heads. We reflect on the visual occlusion issue and identify its essence as the absence of information useful for task completion. Inspired by this, we come up with the more fundamental problem of Exploratory and Focused Manipulation (EFM). The proposed problem is about actively collecting information to complete challenging manipulation tasks that require exploration or focus. As an initial attempt to address this problem, we establish the EFM-10 benchmark that consists of 4 categories of tasks that align with our definition (10 tasks in total). We further come up with a Bimanual Active Perception (BAP) strategy, which leverages one arm to provide active vision and another arm to provide force sensing while manipulating. Based on this idea, we collect a dataset named BAPData for the tasks in EFM-10. With the dataset, we successfully verify the effectiveness of the BAP strategy in an imitation learning manner. We hope that the EFM-10 benchmark along with the BAP strategy can become a cornerstone that facilitates future research towards this direction. Project website: EFManipulation.github.io.

Video generation is experiencing rapid growth, driven by advances in diffusion models and the development of better and larger datasets. However, producing high-quality videos remains challenging due to the high-dimensional data and the complexity of the task. Recent efforts have primarily focused on enhancing visual quality and addressing temporal inconsistencies, such as flickering. Despite progress in these areas, the generated videos often fall short in terms of motion complexity and physical plausibility, with many outputs either appearing static or exhibiting unrealistic motion. In this work, we propose a framework to improve the realism of motion in generated videos, exploring a complementary direction to much of the existing literature. Specifically, we advocate for the incorporation of a retrieval mechanism during the generation phase. The retrieved videos act as grounding signals, providing the model with demonstrations of how the objects move. Our pipeline is designed to apply to any text-to-video diffusion model, conditioning a pretrained model on the retrieved samples with minimal fine-tuning. We demonstrate the superiority of our approach through established metrics, recently proposed benchmarks, and qualitative results, and we highlight additional applications of the framework.

In real-world scene perception human observers generate sequences of fixations to move image patches into the high-acuity center of the visual field. Models of visual attention developed over the last 25 years aim to predict two-dimensional probabilities of gaze positions for a given image via saliency maps. Recently, progress has been made on models for the generation of scan paths under the constraints of saliency as well as attentional and oculomotor restrictions. Experimental research demonstrated that task constraints can have a strong impact on viewing behavior. Here we propose a scan path model for both fixation positions and fixation durations, which includes influences of task instructions and interindividual differences. Based on an eye-movement experiment with four different task conditions, we estimated model parameters for each individual observer and task condition using a fully Bayesian dynamical modeling framework using a joint spatial-temporal likelihood approach with sequential estimation. Resulting parameter values demonstrate that model properties such as the attentional span are adjusted to task requirements. Posterior predictive checks indicate that our dynamical model can reproduce task differences in scan path statistics across individual observers.

Audio-driven human animation models often suffer from identity drift during temporal autoregressive generation, where characters gradually lose their identity over time. One solution is to generate keyframes as intermediate temporal anchors that prevent degradation, but this requires an additional keyframe generation stage and can restrict natural motion dynamics. To address this, we propose Lookahead Anchoring, which leverages keyframes from future timesteps ahead of the current generation window, rather than within it. This transforms keyframes from fixed boundaries into directional beacons: the model continuously pursues these future anchors while responding to immediate audio cues, maintaining consistent identity through persistent guidance. This also enables self-keyframing, where the reference image serves as the lookahead target, eliminating the need for keyframe generation entirely. We find that the temporal lookahead distance naturally controls the balance between expressivity and consistency: larger distances allow for greater motion freedom, while smaller ones strengthen identity adherence. When applied to three recent human animation models, Lookahead Anchoring achieves superior lip synchronization, identity preservation, and visual quality, demonstrating improved temporal conditioning across several different architectures. Video results are available at the following link: https://lookahead-anchoring.github.io.

To the best of our knowledge, we first present a live system that generates personalized photorealistic talking-head animation only driven by audio signals at over 30 fps. Our system contains three stages. The first stage is a deep neural network that extracts deep audio features along with a manifold projection to project the features to the target person's speech space. In the second stage, we learn facial dynamics and motions from the projected audio features. The predicted motions include head poses and upper body motions, where the former is generated by an autoregressive probabilistic model which models the head pose distribution of the target person. Upper body motions are deduced from head poses. In the final stage, we generate conditional feature maps from previous predictions and send them with a candidate image set to an image-to-image translation network to synthesize photorealistic renderings. Our method generalizes well to wild audio and successfully synthesizes high-fidelity personalized facial details, e.g., wrinkles, teeth. Our method also allows explicit control of head poses. Extensive qualitative and quantitative evaluations, along with user studies, demonstrate the superiority of our method over state-of-the-art techniques.

Cinematic video depicts multiple subjects acting or interacting at specific moments, captured with deliberate camera movement, and stitched together by shot transitions. Together, these elements demand a level of fine-grained control beyond current text-to-video models. Existing work addresses each axis in isolation: multi-subject personalization, temporal control, multi-shot synthesis, or camera control; no prior framework jointly integrates all four. We present CineOrchestra, a unified video diffusion model that controls subjects, events, cameras, and shot transitions simultaneously. Our key insight is that these heterogeneous cinematic elements share a fundamental structure: each is an entity acting over a specific temporal interval, which can therefore all be expressed through one shared structure of entity-centric conditioning primitives, augmented with reference images for visual entities. This formulation reduces the architectural challenge to a single positional encoding problem, which we solve with two parameter-free coordinated rotary embeddings: (a) an interval-sampled temporal RoPE that yields consistent attention behavior across events of dramatically varying duration, and (b) a 2D entity-temporal cross-attention RoPE that disambiguates per-entity conditions and routes each to its corresponding spatiotemporal region. On two new benchmarks, CineOrchestra outperforms six per-axis specialists on dense caption following and shot-transition timing, with consistent gains in a pairwise user study and component ablations.

Advanced autoregressive (AR) video generation models have improved visual fidelity and interactivity, but the quadratic complexity of attention remains a primary bottleneck for efficient deployment. While existing sparse attention solutions have shown promise on bidirectional models, we identify that applying these solutions to AR models leads to considerable performance degradation for two reasons: isolated consideration of chunk generation and insufficient utilization of past informative context. Motivated by these observations, we propose Light Forcing, the first sparse attention solution tailored for AR video generation models. It incorporates a Chunk-Aware Growth mechanism to quantitatively estimate the contribution of each chunk, which determines their sparsity allocation. This progressive sparsity increase strategy enables the current chunk to inherit prior knowledge in earlier chunks during generation. Additionally, we introduce a Hierarchical Sparse Attention to capture informative historical and local context in a coarse-to-fine manner. Such two-level mask selection strategy (\ie, frame and block level) can adaptively handle diverse attention patterns. Extensive experiments demonstrate that our method outperforms existing sparse attention in quality (\eg, 84.5 on VBench) and efficiency (\eg, 1.2{sim}1.3times end-to-end speedup). Combined with FP8 quantization and LightVAE, Light Forcing further achieves a 2.3times speedup and 19.7\,FPS on an RTX~5090 GPU. Code will be released at https://github.com/chengtao-lv/LightForcing{https://github.com/chengtao-lv/LightForcing}.

Vision-Language-Action (VLA) foundation models have recently achieved strong progress in embodied intelligence. To reduce policy-call frequency while preserving temporal coherence, most generative policies adopt an action chunk mechanism, executing multiple future actions in an open-loop manner under a fixed action horizon. However, this "predict-then-blindly-execute" paradigm sacrifices closed-loop reactivity: in contact-rich physical interactions, even small local perturbations can rapidly amplify within the open-loop blind spot, leading to compounding errors and ultimately task failure. To address this limitation, we propose VLA-Corrector, a lightweight corrective inference framework for action-chunked VLA policies. Without modifying the backbone policy weights, VLA-Corrector introduces a lightweight Latent-space Vision Monitor (LVM) that continuously compares predicted and actual visual feature evolution, enabling online detection of visual dynamics deviations. Once persistent deviation is detected, the system triggers a truncation event, discards the remaining stale actions, and invokes corrective replanning via Online Gradient Guidance (OGG). The detect-and-correct mechanism of VLA-Corrector naturally induces an event-triggered adaptive action horizon: it preserves long-horizon execution when the current chunk remains reliable, and invokes short-horizon corrective replanning when execution begins to drift. In doing so, VLA-Corrector mitigates the trade-off imposed by static horizons between execution robustness and policy-call frequency. It can be integrated into different VLA models without further retraining the VLA backbone, interrupting compounding errors while preserving much of the efficiency benefit of action chunking and substantially improving robustness in long-horizon, contact-rich robotic manipulation tasks.

We demonstrate a new computational illumination technique that achieves large space-bandwidth-time product, for quantitative phase imaging of unstained live samples in vitro. Microscope lenses can have either large field of view (FOV) or high resolution, not both. Fourier ptychographic microscopy (FPM) is a new computational imaging technique that circumvents this limit by fusing information from multiple images taken with different illumination angles. The result is a gigapixel-scale image having both wide FOV and high resolution, i.e. large space-bandwidth product (SBP). FPM has enormous potential for revolutionizing microscopy and has already found application in digital pathology. However, it suffers from long acquisition times (on the order of minutes), limiting throughput. Faster capture times would not only improve imaging speed, but also allow studies of live samples, where motion artifacts degrade results. In contrast to fixed (e.g. pathology) slides, live samples are continuously evolving at various spatial and temporal scales. Here, we present a new source coding scheme, along with real-time hardware control, to achieve 0.8 NA resolution across a 4x FOV with sub-second capture times. We propose an improved algorithm and new initialization scheme, which allow robust phase reconstruction over long time-lapse experiments. We present the first FPM results for both growing and confluent in vitro cell cultures, capturing videos of subcellular dynamical phenomena in popular cell lines undergoing division and migration. Our method opens up FPM to applications with live samples, for observing rare events in both space and time.

Full-duplex spoken language models (FD-SLMs) enable seamless speech interaction by allowing models to listen and speak simultaneously, yet the internal mechanism by which they coordinate listening and speaking remains underexplored. We analyze the predictive behavior encoded in FD-SLM hidden representations and find that they exhibit stream-specific predictive patterns: during listening, they preferentially predict the incoming user stream, whereas during speaking, they preferentially predict the model output stream. Building on this observation, we show that FD-SLMs dynamically modulate their internal predictive focus between two states: a generative state aligned with model output generation and a perceptive state aligned with incoming user input. However, this modulation can lag behind abrupt changes in conversational context. During user interruptions, the model remains transiently biased toward the generative state before transitioning into the perceptive state, causing it to miss the beginning of the incoming input. We term this delayed internal transition state inertia. To quantify its downstream impact, we introduce the Zero-Buffer Benchmark (ZBB), a diagnostic benchmark for evaluating immediate interruption comprehension when user speech begins abruptly. We evaluate this setting using response correctness and initial-word occurrence rate (IWOR). Finally, we mitigate state inertia through activation steering with a perception vector, a training-free intervention with little additional computational overhead. Across multiple state-of-the-art FD-SLMs, activation steering substantially improves interruption handling; for example, on PersonaPlex, it improves correctness from 28% to 45% and IWOR from 40% to 72% without any fine-tuning.

Seamlessly moving objects within a scene is a common requirement for image editing, but it is still a challenge for existing editing methods. Especially for real-world images, the occlusion situation further increases the difficulty. The main difficulty is that the occluded portion needs to be completed before movement can proceed. To leverage the real-world knowledge embedded in the pre-trained diffusion models, we propose a Diffusion-based framework specifically designed for Occluded Object Movement, named DiffOOM. The proposed DiffOOM consists of two parallel branches that perform object de-occlusion and movement simultaneously. The de-occlusion branch utilizes a background color-fill strategy and a continuously updated object mask to focus the diffusion process on completing the obscured portion of the target object. Concurrently, the movement branch employs latent optimization to place the completed object in the target location and adopts local text-conditioned guidance to integrate the object into new surroundings appropriately. Extensive evaluations demonstrate the superior performance of our method, which is further validated by a comprehensive user study.

Sora-like video generation models have achieved remarkable progress with a Multi-Modal Diffusion Transformer MM-DiT architecture. However, the current video generation models predominantly focus on single-prompt, struggling to generate coherent scenes with multiple sequential prompts that better reflect real-world dynamic scenarios. While some pioneering works have explored multi-prompt video generation, they face significant challenges including strict training data requirements, weak prompt following, and unnatural transitions. To address these problems, we propose DiTCtrl, a training-free multi-prompt video generation method under MM-DiT architectures for the first time. Our key idea is to take the multi-prompt video generation task as temporal video editing with smooth transitions. To achieve this goal, we first analyze MM-DiT's attention mechanism, finding that the 3D full attention behaves similarly to that of the cross/self-attention blocks in the UNet-like diffusion models, enabling mask-guided precise semantic control across different prompts with attention sharing for multi-prompt video generation. Based on our careful design, the video generated by DiTCtrl achieves smooth transitions and consistent object motion given multiple sequential prompts without additional training. Besides, we also present MPVBench, a new benchmark specially designed for multi-prompt video generation to evaluate the performance of multi-prompt generation. Extensive experiments demonstrate that our method achieves state-of-the-art performance without additional training.

Subject-driven image personalization---generating new images that preserve the identity of one or several reference subjects in novel scenes---is a foundational capability for modern visual content creation. It is currently dominated by generalized methods that fine-tune a pretrained multimodal diffusion transformer (MMDiT) on hundreds of thousands to millions of paired (reference, composed-target) examples, where each composed target is a synthesized image of the subject in a novel scene. Producing such targets demands a costly multi-stage curation pipeline---LLM-based prompt generation, T2I-based composed-target synthesis, reference-subject extraction, VLM-based quality filtering, and correspondence labeling---and tightly couples each method to a particular target synthesizer and curation choice. We introduce CRAFT (Constrained Reward via Attention Fine-Tuning), a single-step ReFL framework that fine-tunes a pre-trained reference-aware MMDiT via LoRA adapters using a compact reference-only data construction---10K reference images and subject masks, with no composed-target supervision. CRAFT realizes a Where to look principle: attention-level rewards align noise- and phrase-token attention with the correct reference subject, and the resulting per-subject attention masks gate a pixel-level identity reward to keep image-space supervision consistent with the learned attention routing. Applied to FLUX.2-klein-9B, CRAFT achieves state-of-the-art performance on XVerseBench while using no composed-target supervision---only 10K reference-only samples, whereas prior generalized methods require 150K to over 2M composed-target pairs. The same recipe transfers to other reference-aware backbones, consistently improving performance. Project page: https://jihun999.github.io/projects/CRAFT/.

Motion-centric video editing remains difficult for large generative video models, which often respond well to appearance changes but struggle to produce specific, localized actions or state transitions in an existing clip. We introduce Sound Sparks Motion, a training-free framework that enables motion editing in an audio-visual video generation model by tuning its internal multimodal conditioning signals at test time. Rather than modifying model weights, our method tunes only two lightweight variables: an audio latent derived from the source video and a residual perturbation in the text-conditioning. We find that this combination can encourage motion edits that the underlying model often struggles to realize under prompt-only control. Since there is no direct way to evaluate temporal alignment between text and motion, we guide the tuning process using a vision-language model that provides feedback indicating whether the intended motion appears in the generated video. This simple supervision yields an effective semantic objective for motion editing, while regularization and perceptual-temporal constraints help preserve content and visual quality. Beyond per-video tuning, we show that the learned latent controls are transferable across videos, suggesting that they capture reusable motion-edit directions rather than overfitting to a single example. Our results highlight multimodal conditioning tuning, particularly through the audio pathway, as a promising direction for motion-aware video editing, and suggest that test-time tuning can serve as a lightweight probing mechanism that helps reveal latent motion controls embedded in the model's multimodal conditioning. Code and data are available via our project page: https://amirhossein-razlighi.github.io/Sound_Sparks_Motion/

Transformer-based diffusion models have recently superseded traditional U-Net architectures, with multimodal diffusion transformers (MM-DiT) emerging as the dominant approach in state-of-the-art models like Stable Diffusion 3 and Flux.1. Previous approaches have relied on unidirectional cross-attention mechanisms, with information flowing from text embeddings to image latents. In contrast, MMDiT introduces a unified attention mechanism that concatenates input projections from both modalities and performs a single full attention operation, allowing bidirectional information flow between text and image branches. This architectural shift presents significant challenges for existing editing techniques. In this paper, we systematically analyze MM-DiT's attention mechanism by decomposing attention matrices into four distinct blocks, revealing their inherent characteristics. Through these analyses, we propose a robust, prompt-based image editing method for MM-DiT that supports global to local edits across various MM-DiT variants, including few-step models. We believe our findings bridge the gap between existing U-Net-based methods and emerging architectures, offering deeper insights into MMDiT's behavioral patterns.

Current autoregressive video diffusion models are constrained by three core bottlenecks: (i) the finite temporal horizon imposed by the base model's 3D Rotary Positional Embedding (3D-RoPE), (ii) slow prompt responsiveness in maintaining fine-grained action control during long-form rollouts, and (iii) the inability to realize discontinuous cinematic transitions within a single generation stream. We introduce infty-RoPE, a unified inference-time framework that addresses all three limitations through three interconnected components: Block-Relativistic RoPE, KV Flush, and RoPE Cut. Block-Relativistic RoPE reformulates temporal encoding as a moving local reference frame, where each newly generated latent block is rotated relative to the base model's maximum frame horizon while earlier blocks are rotated backward to preserve relative temporal geometry. This relativistic formulation eliminates fixed temporal positions, enabling continuous video generation far beyond the base positional limits. To obtain fine-grained action control without re-encoding, KV Flush renews the KV cache by retaining only two latent frames, the global sink and the last generated latent frame, thereby ensuring immediate prompt responsiveness. Finally, RoPE Cut introduces controlled discontinuities in temporal RoPE coordinates, enabling multi-cut scene transitions within a single continuous rollout. Together, these components establish infty-RoPE as a training-free foundation for infinite-horizon, controllable, and cinematic video diffusion. Comprehensive experiments show that infty-RoPE consistently surpasses previous autoregressive models in overall VBench scores.

Recent Large Vision Language Models (LVLMs) demonstrate promising capabilities in unifying visual understanding and generative modeling, enabling both accurate content understanding and flexible editing. However, current approaches treat "what to see" and "how to edit" separately: they either perform isolated object segmentation or utilize segmentation masks merely as conditional prompts for local edit generation tasks, often relying on multiple disjointed models. To bridge these gaps, we introduce FOCUS, a unified LVLM that integrates segmentation-aware perception and controllable object-centric generation within an end-to-end framework. FOCUS employs a dual-branch visual encoder to simultaneously capture global semantic context and fine-grained spatial details. In addition, we leverage a MoVQGAN-based visual tokenizer to produce discrete visual tokens that enhance generation quality. To enable accurate and controllable image editing, we propose a progressive multi-stage training pipeline, where segmentation masks are jointly optimized and used as spatial condition prompts to guide the diffusion decoder. This strategy aligns visual encoding, segmentation, and generation modules, effectively bridging segmentation-aware perception with fine-grained visual synthesis. Extensive experiments across three core tasks, including multimodal understanding, referring segmentation accuracy, and controllable image generation, demonstrate that FOCUS achieves strong performance by jointly optimizing visual perception and generative capabilities.

Vision-Language-Action (VLA) models have become a cornerstone in robotic policy learning, leveraging large-scale multimodal data for robust and scalable control. However, existing VLA frameworks primarily address short-horizon tasks, and their effectiveness on long-horizon, multi-step robotic manipulation remains limited due to challenges in skill chaining and subtask dependencies. In this work, we introduce Long-VLA, the first end-to-end VLA model specifically designed for long-horizon robotic tasks. Our approach features a novel phase-aware input masking strategy that adaptively segments each subtask into moving and interaction phases, enabling the model to focus on phase-relevant sensory cues and enhancing subtask compatibility. This unified strategy preserves the scalability and data efficiency of VLA training, and our architecture-agnostic module can be seamlessly integrated into existing VLA models. We further propose the L-CALVIN benchmark to systematically evaluate long-horizon manipulation. Extensive experiments on both simulated and real-world tasks demonstrate that Long-VLA significantly outperforms prior state-of-the-art methods, establishing a new baseline for long-horizon robotic control.

Long exposure photography produces stunning imagery, representing moving elements in a scene with motion-blur. It is generally employed in two modalities, producing either a foreground or a background blur effect. Foreground blur images are traditionally captured on a tripod-mounted camera and portray blurred moving foreground elements, such as silky water or light trails, over a perfectly sharp background landscape. Background blur images, also called panning photography, are captured while the camera is tracking a moving subject, to produce an image of a sharp subject over a background blurred by relative motion. Both techniques are notoriously challenging and require additional equipment and advanced skills. In this paper, we describe a computational burst photography system that operates in a hand-held smartphone camera app, and achieves these effects fully automatically, at the tap of the shutter button. Our approach first detects and segments the salient subject. We track the scene motion over multiple frames and align the images in order to preserve desired sharpness and to produce aesthetically pleasing motion streaks. We capture an under-exposed burst and select the subset of input frames that will produce blur trails of controlled length, regardless of scene or camera motion velocity. We predict inter-frame motion and synthesize motion-blur to fill the temporal gaps between the input frames. Finally, we composite the blurred image with the sharp regular exposure to protect the sharpness of faces or areas of the scene that are barely moving, and produce a final high resolution and high dynamic range (HDR) photograph. Our system democratizes a capability previously reserved to professionals, and makes this creative style accessible to most casual photographers. More information and supplementary material can be found on our project webpage: https://motion-mode.github.io/

Most current visual trackers adopt a matching-based architecture trained exclusively on tracking datasets, whose performance gains depend heavily on the length of the input context, and have now reached a bottleneck. While high-performance tracking increasingly relies on foundation models, existing methods use them monolithically, adapting a foundation model into a tracker or modify a segmentation foundation model into a tracking pipeline, which fails to exploit complementary strengths. Matching-based trackers excel at instance-level correspondence but lack semantic discrimination and fine-grained foreground perception, whereas segmentation foundation models produce precise masks yet struggle with instance discrimination and multimodal extension. Both paradigms also lack error-correction capabilities for long-term tracking. To address these issues, we propose ACTrack, an agentic coordination framework that treats heterogeneous models as invocable tools under an event-triggered mechanism. ACTrack coordinates a Tracker-based Instance Matching Tool for target discrimination, a SAM3 Motion Tool for mask-derived motion priors, a SAM3 Perception Tool for detecting distractors and instance-conflict cues, and a VLM Reprompt Tool activated only under persistent conflict to mitigate error accumulation. We design a complete tool-invocation trigger mechanism and an inter-tool coordination mechanism, enabling the complementary strengths of different model tools to be fully integrated. Experiments show that ACTrack substantially surpasses the strongest and the largest trackers on eight RGB benchmarks. Furthermore, a parameter-efficient adaptation strategy enables parameter sharing and reuse across tools, achieving unified multimodal tracking with only 30\% trainable parameters while substantially outperforming prior methods on multimodal benchmarks such as LasHeR, VisEvent, TNL2K, and DepthTrack.

The first-in-first-out (FIFO) video diffusion, built on a pre-trained text-to-video model, has recently emerged as an effective approach for tuning-free long video generation. This technique maintains a queue of video frames with progressively increasing noise, continuously producing clean frames at the queue's head while Gaussian noise is enqueued at the tail. However, FIFO-Diffusion often struggles to keep long-range temporal consistency in the generated videos due to the lack of correspondence modeling across frames. In this paper, we propose Ouroboros-Diffusion, a novel video denoising framework designed to enhance structural and content (subject) consistency, enabling the generation of consistent videos of arbitrary length. Specifically, we introduce a new latent sampling technique at the queue tail to improve structural consistency, ensuring perceptually smooth transitions among frames. To enhance subject consistency, we devise a Subject-Aware Cross-Frame Attention (SACFA) mechanism, which aligns subjects across frames within short segments to achieve better visual coherence. Furthermore, we introduce self-recurrent guidance. This technique leverages information from all previous cleaner frames at the front of the queue to guide the denoising of noisier frames at the end, fostering rich and contextual global information interaction. Extensive experiments of long video generation on the VBench benchmark demonstrate the superiority of our Ouroboros-Diffusion, particularly in terms of subject consistency, motion smoothness, and temporal consistency.

Reinforcement learning (RL) post-training improves the reward alignment of flow-based generators, but often degrades perceptual quality in ways that are not captured by the reward proxy. We identify a simple structural signature of this drift: across three post-training methods (NFT, AWM, DPO), RL fine-tuning inflates the per-step velocity norm |v_θ| by 5% to 15% relative to the reference. A form of norm inflation has been studied in classifier-free guidance (CFG), where rescaling the velocity back to a reference norm at inference time can mitigate the resulting artifacts. However, this inference-time correction does not transfer cleanly to RL: rescaling v_θ to match |v_{ref}| at inference time neither improves reward nor fixes the quality degradation, because the inflation is co-adapted into the model weights. Furthermore, an adjoint sensitivity analysis shows that velocity magnitude rescaling carries no coherent first-order reward signal at the batch level, indicating that suppressing norm inflation is unlikely to remove a consistently reward-carrying component. Since inference-time renormalization fails while norm suppression carries no reward cost, training-time intervention is the appropriate strategy. Together, these findings motivate \methodname, a hinge penalty that activates only when |v_θ| exceeds |v_{ref}| and composes additively with any velocity-local base loss. Across two base models, three post-training methods, and two reward proxies, \methodname consistently improves MLLM-judged image quality and forensic realism while preserving reward, with gains that amplify under few-step inference and are not explained by early stopping.

Scaling video diffusion transformers (DiTs) is limited by their quadratic 3D attention, even though most of the attention mass concentrates on a small subset of positions. We turn this observation into VSA, a trainable, hardware-efficient sparse attention that replaces full attention at both training and inference. In VSA, a lightweight coarse stage pools tokens into tiles and identifies high-weight critical tokens; a fine stage computes token-level attention only inside those tiles subjecting to block computing layout to ensure hard efficiency. This leads to a single differentiable kernel that trains end-to-end, requires no post-hoc profiling, and sustains 85\% of FlashAttention3 MFU. We perform a large sweep of ablation studies and scaling-law experiments by pretraining DiTs from 60M to 1.4B parameters. VSA reaches a Pareto point that cuts training FLOPS by 2.53times with no drop in diffusion loss. Retrofitting the open-source Wan-2.1 model speeds up attention time by 6times and lowers end-to-end generation time from 31s to 18s with comparable quality. These results establish trainable sparse attention as a practical alternative to full attention and a key enabler for further scaling of video diffusion models.

This paper presents a novel method for exerting fine-grained lighting control during text-driven diffusion-based image generation. While existing diffusion models already have the ability to generate images under any lighting condition, without additional guidance these models tend to correlate image content and lighting. Moreover, text prompts lack the necessary expressional power to describe detailed lighting setups. To provide the content creator with fine-grained control over the lighting during image generation, we augment the text-prompt with detailed lighting information in the form of radiance hints, i.e., visualizations of the scene geometry with a homogeneous canonical material under the target lighting. However, the scene geometry needed to produce the radiance hints is unknown. Our key observation is that we only need to guide the diffusion process, hence exact radiance hints are not necessary; we only need to point the diffusion model in the right direction. Based on this observation, we introduce a three stage method for controlling the lighting during image generation. In the first stage, we leverage a standard pretrained diffusion model to generate a provisional image under uncontrolled lighting. Next, in the second stage, we resynthesize and refine the foreground object in the generated image by passing the target lighting to a refined diffusion model, named DiLightNet, using radiance hints computed on a coarse shape of the foreground object inferred from the provisional image. To retain the texture details, we multiply the radiance hints with a neural encoding of the provisional synthesized image before passing it to DiLightNet. Finally, in the third stage, we resynthesize the background to be consistent with the lighting on the foreground object. We demonstrate and validate our lighting controlled diffusion model on a variety of text prompts and lighting conditions.

In recent years, diffusion models have emerged as the most powerful approach in image synthesis. However, applying these models directly to video synthesis presents challenges, as it often leads to noticeable flickering contents. Although recently proposed zero-shot methods can alleviate flicker to some extent, we still struggle to generate coherent videos. In this paper, we propose DiffSynth, a novel approach that aims to convert image synthesis pipelines to video synthesis pipelines. DiffSynth consists of two key components: a latent in-iteration deflickering framework and a video deflickering algorithm. The latent in-iteration deflickering framework applies video deflickering to the latent space of diffusion models, effectively preventing flicker accumulation in intermediate steps. Additionally, we propose a video deflickering algorithm, named patch blending algorithm, that remaps objects in different frames and blends them together to enhance video consistency. One of the notable advantages of DiffSynth is its general applicability to various video synthesis tasks, including text-guided video stylization, fashion video synthesis, image-guided video stylization, video restoring, and 3D rendering. In the task of text-guided video stylization, we make it possible to synthesize high-quality videos without cherry-picking. The experimental results demonstrate the effectiveness of DiffSynth. All videos can be viewed on our project page. Source codes will also be released.