ComfyUI-EffectErase
A ComfyUI extension with 2 custom nodes.
ComfyUI-EffectErase
A custom node for ComfyUI that implements Joint Video Object Removal and Insertion using the Wan2.1 1.3B Video DiT architecture.
This repository allows you to perform advanced, temporally consistent Video-to-Video inpainting and object removal directly inside your ComfyUI workflows without complicated setups.
Features
- Object Removal: Erase moving objects from videos by providing an input video and a mask.
- Auto-Dependencies: Automatically installs
modelscope,peft, and HuggingFace dependencies at startup. - Model Auto-Download: Automatically fetches the required Wan2.1 model and EffectErase LoRAs on the first generation and saves them to
models/EffectErase/. - Acceleration LoRA Support: Supports attaching native Wan2.1 acceleration LoRAs to speed up inference times.
- FlowMatch Scheduler Configuration: Exposes
sigma_shift, CFG, inference steps, and seed controls.
Usage
Simply plug the EffectEraseObjectRemoval node into your ComfyUI workflow.
- Provide a
video_fg_bg(Image Tensor representing video frames). - Provide a
video_mask(Target Object Mask). - Connect the output to a Save Video or Video Combine node.
Acknowledgements & Credits
This node is just a wrapper to bring the state-of-the-art research into the ComfyUI ecosystem. All credit for the mathematical implementation, training, and neural architecture modifications belongs entirely to the original authors.
- EffectErase: Developed by FudanCVL. Official Repository
- DiffSynth-Studio: The inference framework used to execute the custom modified WanRemovePipeline natively. DiffSynth-Studio
- Wan2.1: The underlying Video Diffusion Transformer baseline.
If you use this node in your research or projects, please consider starring and citing the original EffectErase repository.