ComfyUI Extension: ComfyUI-MagCache
official implementation of zehong-ma/MagCache for ComfyUI
Custom Nodes (0)
README
ComfyUI-MagCache
π« Introduction
Magnitude-aware Cache (MagCache) is a training-free caching approach. It estimates the fluctuating differences among model outputs across timesteps based on the robust magnitude observations, and thereby accelerating the inference using the error modeling mechanism and adaptive cache strategy. MagCache works well for Video Diffusion Models, Image Diffusion models. For more details and results, please visit our project page and code.
MagCache has now been integrated into ComfyUI and is compatible with the ComfyUI native nodes. ComfyUI-MagCache is easy to use, simply connect the MagCache node with the ComfyUI native nodes for seamless usage.
π₯ Latest News
- If you like our project, please give us a star β on GitHub for the latest update.
- [2025/6/30] π₯ Support Flux-Kontext with 2x speedup. Please see the demo here.
- [2025/6/19] π₯ Support FramePack with Gradio Demo in MagCache-FramePack.
- [2025/6/18] π We're collecting the best parameter settings from the community. <br> πOpen this discussion issue to contribute your configuration!
- [2025/6/17] π₯ Support Wan2.1-VACE-1.3B, achieving a 1.7Γ acceleration.
- [2025/6/17] π₯ MagCache is supported by ComfyUI-WanVideoWrapper. Thanks @kijai.
- [2025/6/16] π₯ Support Chroma. Thanks @kabachuha for the codebase.
- [2025/6/10] π₯ Support Wan2.1 T2V&I2V, HunyuanVideo T2V, FLUX-dev T2I
Installation
<!-- Installation via ComfyUI-Manager is preferred. Simply search for ComfyUI-MagCache in the list of nodes and click install. ### Manual installation -->- Go to comfyUI custom_nodes folder,
ComfyUI/custom_nodes/
- git clone https://github.com/zehong-ma/ComfyUI-MagCache.git
- Go to ComfyUI-MagCache folder,
cd ComfyUI-MagCache/
- pip install -r requirements.txt
- Go to the project folder
ComfyUI/
and runpython main.py
Usage
Download Model Weights
Please first to prepare the model weights in ComfyUI format by referring to the follow links:
MagCache
We're collecting the best parameter settings from the community. Open this discussion issue to contribute your configuration!
To use MagCache node, simply add MagCache
node to your workflow after Load Diffusion Model
node or Load LoRA
node (if you need LoRA). Generally, MagCache can achieve a speedup of 2x to 3x with acceptable visual quality loss. The following table gives the recommended magcache_thresh, retention_ratio and magcache_K βfor different models:
| Models | magcache_thresh | retention_ratio | magcache_K |
|:----------------------------:|:-----------------:|:-----------------:|:-----------------:|
| FLUX | 0.24 | 0.1 | 5 |
| FLUX-Kontext | 0.05 | 0.2 | 4 |
| Chroma | 0.10 | 0.25 | 2 |
| HunyuanVideo-T2V | 0.24 | 0.2 | 6 |
| Wan2.1-T2V-1.3B | 0.12 | 0.2 | 4 |
| Wan2.1-T2V-14B | 0.24 | 0.2 | 6 |
| Wan2.1-I2V-480P-14B | 0.24 | 0.2 | 6 |
| Wan2.1-I2V-720P-14B | 0.24 | 0.2 | 6 |
| Wan2.1-VACE-1.3B | 0.02 | 0.2 | 3 |
If the image/video after applying MagCache is of low quality, please decrease magcache_thresh
and magcache_K
. The default parameters are configured for extremely fast inference(2x-3x), which may lead to failures in some cases.
The demo workflows (flux, flux-kontext, chroma, hunyuanvideo, wan2.1_t2v, wan2.1_i2v, and wan2.1_vace) are placed in examples folder. The workflow chroma_calibration is used to calibrate the mag_ratios
for Chroma
when the number of inference steps differs from the pre-defined value.
In our experiments, the videos generated by Wan2.1 are not as high-quality as those produced by the original unquantized version.
Compile Model
To use Compile Model node, simply add Compile Model
node to your workflow after Load Diffusion Model
node or MagCache
node. Compile Model uses torch.compile
to enhance the model performance by compiling model into more efficient intermediate representations (IRs). This compilation process leverages backend compilers to generate optimized code, which can significantly speed up inference. The compilation may take long time when you run the workflow at first, but once it is compiled, inference is extremely fast.
Acknowledgments
Thanks to ComfyUI-TeaCache, ComfyUI, ComfyUI-MagCache, MagCache, TeaCache, HunyuanVideo, FLUX, Chroma, and Wan2.1.