|
|
TL;DR: MagicNodes, it's a plug-and-play multi-pass "render-machine" for SD/SDXL models. Simple one-node start, expert-grade results. Core is ZeResFDG (Frequency-Decoupled + Rescale + Zero-Projection) and the always-on QSilk Micrograin Stabilizer, complemented by practical stabilizers (NAG, local masks, EPS, Muse Blend, Polish). Ships with a four-pass preset for robust, clean, and highly detailed outputs. Our pipeline runs through several purposeful passes: early steps assemble global shapes, mid steps refine important regions, and late steps polish without overcooking the texture. We gently stabilize the amplitudes of the "image’s internal draft" (latent) and adapt the allowed value range per region: where the model is confident we give more freedom, and where it’s uncertain we act more conservatively. The result is clean gradients, crisp edges, and photographic detail even at very high resolutions and, as a side effect on SDXL models, text becomes noticeably more stable and legible. |
| 
|
Photo Portrait
| | |
|:--:|:--:|
| 
| 
|
| | |
|:--:|:--:|
| 
| 
|
Photo Cup
| | |
|:--:|:--:|
| 
| 
|
Photo Dog
| | |
|:--:|:--:|
| 
| 
|
---
## Features
- ZeResFDG: LF/HF split, energy rescale, and zero-projection (stable early, sharp late)
- NAG (Normalized Attention Guidance): small attention variance normalization (positive branch)
- Local spatial gating: optional CLIPSeg masks for faces/hands/pose
- EPS scale: small early-step exposure bias
- QSilk Micrograin Stabilizer: gently smooths rare spikes and lets natural micro-texture (skin, fabric, tiny hairs) show through — without halos or grid patterns. Always on, zero knobs, near‑zero cost.
- Adaptive Quantile Clip (AQClip): softly adapts the allowed range per region. Confident areas keep more texture; uncertain ones get cleaner denoising. Tile‑based with seamless blending (no seams). Optional Attn mode uses attention confidence for an even smarter balance.
- MGHybrid scheduler: hybrid Karras/Beta sigma stack with smooth tail blending and tiny schedule jitter (ZeSmart-inspired) for more stable, detail-friendly denoising; used by CADE and SuperSimple by default
- Seed Latent (MG_SeedLatent): fast, deterministic latent initializer aligned to VAE stride; supports pure-noise starts or image-mixed starts (encode + noise) to gently bias content; batch-ready and resolution-agnostic, pairs well with SuperSimple recommended latent sizes for reproducible pipelines
- Muse Blend and Polish: directional post-mix and final low-frequency-preserving clean-up
- SmartSeed (CADE Easy and SuperSimple): set `seed = 0` to auto-pick a good seed from a tiny low-step probe. Uses a low-discrepancy sweep, avoids speckles/overexposure, and, if available, leverages CLIP-Vision (with `reference_image`) and CLIPSeg focus text to favor semantically aligned candidates. Logs `Smart_seed_random: Start/End`.
I highly recommend working with SmartSeed.
- CADE2.5 pipeline does not just upscale the image, it iterates and adds small details, doing it carefully, at every stage.
## Hardware
- The pipeline is designed for good hardware (tested on RTX5090 (32Gb) and RAM 92Gb), try to keep the starting latency very small, because there is an upscale at the steps and you risk getting errors if you push up the starting values.
- start latent ~ 672x944 -> final ~ 3688x5192 across 4 steps.
- Notes
- Lowering the starting latent (e.g., 512x768) or lower, reduces both VRAM and RAM.
- Disabling hi-res depth/edges (ControlFusion) reduces peaks. (not recommended!)
- Depth weights add a bit of RAM on load; models live under `depth-anything/`.
## Install (ComfyUI 0.3.60, tested on this version)
Preparing:
I recomend update pytorch version: 2.8.0+cu129.
1. PyTorch install: `pip install torch==2.8.0 torchvision==0.23.0 torchaudio==2.8.0 --index-url https://download.pytorch.org/whl/cu129`
2. CUDA manual download and install: https://developer.nvidia.com/cuda-12-9-0-download-archive?target_os=Windows&target_arch=x86_64&target_version=11&target_type=exe_local
3. Install `SageAttention 2.2.0`, manualy `https://github.com/thu-ml/SageAttention` or use script `scripts/check_sageattention.bat`. The installation takes a few minutes, wait for the installation to finish.
p.s.: To work, you definitely need to install `SageAttention v.2.2.0`, `version 1.0.6` is not suitable for pipeline.
Next:
1. Clone or download this repo into `ComfyUI/custom_nodes/`
2. Install helpers: `pip install -r requirements.txt`
3. I recomend, take my negative LoRA `mg_7lambda_negative.safetensors` in HF https://huggingface.co/DD32/mg_7lambda_negative/blob/main/mg_7lambda_negative.safetensors and place the file in ComfyUI, to `ComfyUI/models/loras`
4. download model `depth_anything_v2_vitl.pth` https://huggingface.co/depth-anything/Depth-Anything-V2-Large/tree/main and place inside in to `depth-anything/` folder.
5. Workflows
Folder `workflows/` contains ready-to-use graphs:
- `mg_SuperSimple-Workflow.json` — one-node pipeline (2/3/4 steps) with presets
- `mg_Easy-Workflow.json` — the same logic built from individual Easy nodes
You can save this workflow to ComfyUI `ComfyUI\user\default\workflows`
6. Restart ComfyUI. Nodes appear under the "MagicNodes" categories.
💥 I strongly recommend use `mg_Easy-Workflow` workflow + default settings + your model and my negative LoRA `mg_7lambda_negative.safetensors`, for best result.
## 🚀 "One-Node" Quickstart (MG_SuperSimple)
Start with `MG_SuperSimple` for the easiest path:
1. Drop `MG_SuperSimple` into the graph
2. Connect `model / positive / negative / vae / latent` and a `Load ControlNet Model` module
3. Choose `step_count` (2/3/4) and Run
or load `mg_SuperSimple-Workflow` in panel ComfyUI
Notes:
- When "Custom" is off, presets fully drive parameters
- When "Custom" is on, the visible CADE controls override the Step presets across all steps; Step 1 still enforces `denoise=1.0`
- CLIP Vision (if connected) is applied from Step 2 onward; if no reference image is provided, SuperSimple uses the previous step image as reference
## ❗Tips
(!) There are almost always artifacts in the first step, don't pay attention to them, they will be removed in the next steps. Keep your prompt clean and logical, don't duplicate details and be careful with symbols.
0) `MG_SuperSimple-Workflow` is a bit less flexible than `MG_Easy-Workflow`, but extremely simple to use. If you just want a stable, interesting result, start with SuperSimple.
1) Recommended negative LoRA: `mg_7lambda_negative.safetensors` with `strength_model = -1.0`, `strength_clip = 0.2`. Place LoRA files under `ComfyUI/models/loras` so they appear in the LoRA selector.
2) Download a CLIP Vision model and place it under `ComfyUI/models/clip_vision` (e.g., https://huggingface.co/openai/clip-vit-large-patch14; heavy alternative: https://huggingface.co/laion/CLIP-ViT-H-14-laion2B-s32B-b79K). SuperSimple/CADE will use it for reference-based polish.
3) Samplers: i recomend use `ddim` for many cases (Draw and Realism style). Scheduler: use `MGHybrid` in this pipeline.
4) Denoise: higher -> more expressive and vivid; you can go up to 1.0. The same applies to CFG: higher -> more expressive but may introduce artifacts. Suggested CFG range: ~4.5–8.5.
5) If you see unwanted artifacts on the final (4th) step, slightly lower denoise to ~0.5–0.6 or simply change the seed.
6) You can get interesting results by repeating steps (in Easy/Hard workflows), e.g., `1 -> 2 -> 3 -> 3`. Just experiment with it!
7) Recommended starting latent close to ~672x944 (other aspect ratios are fine). With that, step 4 produces ~3688x5192. Larger starting sizes are OK if the model and your hardware allow.
8) Unlucky seeds happen — just try another. (We may later add stabilization to this process.)
9) Rarely, step 3 can show a strange grid artifact (in both Easy and Hard workflows). If this happens, try changing CFG or seed. Root cause still under investigation.
10) Results depend on checkpoint/LoRA quality. The pipeline “squeezes” everything SDXL and your model can deliver, so prefer high‑quality checkpoints and non‑overtrained LoRAs.
11) Avoid using more than 3 LoRAs at once, and keep only one “lead” LoRA (one you trust is not overtrained). Too many/strong LoRAs can spoil results.
12) Try connecting reference images in either workflow — you can get unusual and interesting outcomes.
13) Very often, the image in `step 3 is of very good quality`, but it usually lacks sharpness. But if you have a `weak system`, you can `limit yourself to 3 steps`.
14) SmartSeed (auto seed pick): set `seed = 0` in Easy or SuperSimple. The node will sample several candidate seeds and do a quick low‑step probe to choose a balanced one. You’ll see logs `Smart_seed_random: Start` and `Smart_seed_random: End. Seed is: