---
base_model:
- black-forest-labs/FLUX.1-Kontext-dev
datasets:
- handsomeWilliam/Relation252K
license: other
license_name: nvidia-license-non-commercial
license_link: LICENSE
pipeline_tag: image-to-image
---

# LoRWeB: Spanning the Visual Analogy Space with a Weight Basis of LoRAs

<div align="center">
  
[![arXiv](https://img.shields.io/badge/arXiv-2602.15727-b31b1b.svg)](https://huggingface.co/papers/2602.15727)
[![Project Website](https://img.shields.io/badge/🌐-Project%20Website-blue)](https://research.nvidia.com/labs/par/lorweb)
[![GitHub Repository](https://img.shields.io/badge/GitHub-LoRWeB-black?logo=github)](https://github.com/NVlabs/LoRWeB)
[![Evaluation Dataset](https://img.shields.io/badge/🤗-Evaluation%20Dataset-yellow)](https://huggingface.co/datasets/hilamanor/LoRWeB_evalset)

</div>

Visual analogy learning enables image manipulation through demonstration rather than textual description, allowing users to specify complex transformations difficult to articulate in words. 
Given a triplet {**a**, **a'**, **b**}, the goal is to generate **b'** such that **a** : **a'** :: **b** : **b'**.

**LoRWeB** specializes the model for each analogy task at inference time through dynamic composition of learned transformation primitives. It introduces a learnable basis of LoRA modules to span the space of different visual transformations and a lightweight encoder that dynamically selects and weighs these basis LoRAs based on the input analogy pair.

<div align="center">
  
**Hila Manor**<sup>1,2</sup>,&ensp; **Rinon Gal**<sup>2</sup>,&ensp; **Haggai Maron**<sup>1,2</sup>,&ensp; **Tomer Michaeli**<sup>1</sup>,&ensp; **Gal Chechik**<sup>2,3</sup>

<sup>1</sup>Technion - Israel Institute of Technology  &ensp;&ensp; <sup>2</sup>NVIDIA &ensp;&ensp; <sup>3</sup>Bar-Ilan University

</div>

<div align="center">
<img src="https://github.com/NVlabs/LoRWeB/raw/main/assets/teaser.jpg" alt="Teaser" width="800"/>

<i>Given a prompt and an image triplet {**a**, **a'**, **b**} that visually describe a desired transformation, LoRWeB dynamically constructs a single LoRA from a learnable basis of LoRA modules, and produces an editing result **b'** that applies the same analogy to the new image.</i>
</div>


## 🛠 Sample Usage

To perform inference using the LoRWeB weights, use the `inference.py` script from the [official GitHub repository](https://github.com/NVlabs/LoRWeB):

```bash
python inference.py \
  -w "path/to/lorweb_model.safetensors" \
  -c "config/your_config.yaml" \
  -a "data/path_to_a_img.jpg" \
  -t "data/path_to_atag_img.jpg" \
  -b "data/path_to_b_img.jpg" \
  -o "outputs/generated_btag_img_path.jpg"
```

### ℹ️ Additional Information

**This model is a reproduction of the original model from the paper. It was trained from scratch using Technion resources.** This might introduce differences from the results reported in the paper. Please see the `samples` directory for examples of this model's outputs on the {**a**, **a'**, **b**} triplets from the teaser figure.

Please see our full modelcard and further details in the [GitHub Repo](https://github.com/NVlabs/LoRWeB).

## 📚 Citation

If you use this model in your research, please cite:

```bibtex
@article{manor2026lorweb,
    title={Spanning the Visual Analogy Space with a Weight Basis of LoRAs},
    author={Manor, Hila and Gal, Rinon and Maron, Haggai and Michaeli, Tomer and Chechik, Gal},
    journal={arXiv preprint arXiv:2602.15727},
    year={2026}
}
```

## 🙏🏻 Acknowledgements

This project builds upon:
- [FLUX.1-Kontext](https://huggingface.co/black-forest-labs/FLUX.1-Kontext-dev) by Black Forest Labs
- [Diffusers](https://github.com/huggingface/diffusers) by Hugging Face
- [PEFT](https://github.com/huggingface/peft) by Hugging Face
- [AI-Toolkit](https://github.com/ostris/ai-toolkit) for training infrastructure