ComfyUI-WanVideoWrapper
ComfyUI-WanVideoWrapper is a work-in-progress set of custom nodes for ComfyUI that wrap WanVideo (Wan2.1 and related video models) and provide convenience utilities, model loaders, and interoperability features to run video generation and related tasks inside ComfyUI. The project is intended as a sandbox to experiment with new models and features quickly without modifying ComfyUI core, while offering some patcher nodes to bridge functionality back into native workflows.
Key features
- Wrapper nodes for WanVideo (Wan2.1) and many related models, letting users load Transformer/Video models, VAE, text encoders, and CLIP vision components from specific locations.
- Support for multiple model formats including GGUF and fp8-scaled models (author recommends certain fp8 scaled variants).
- Memory / VRAM ergonomics guidance and runtime improvements: the repo documents changes to how LoRA weights are handled (assigning LoRA weights as buffers to modules to enable block-swapping and async offloading), and gives practical advice for compensating block sizes when using block swap.
- Workarounds and notes about torch.compile / PyTorch/Triton caches that can affect first-run VRAM usage (with instructions to clear Triton caches on Windows).
- Bundled examples and visual assets showing vid2vid, context-window tests, long-frame generation experiments, and integrations with many third-party models and techniques.
- Compatibility/adapters for a broad set of external projects and model collections (e.g., SkyReels, WanVideoFun, ReCamMaster, Phantom, ATI, Uni3C, etc.), plus training-free techniques and related tools.
Installation & models
- Install by cloning the repo into ComfyUI's
custom_nodesfolder. - Install Python dependencies via
pip install -r requirements.txt(or use the portable ComfyUI embedded Python path shown in the README). - Models are recommended to be placed into specific ComfyUI model folders (text_encoders, clip_vision, diffusion_models, vae). The README links to a Hugging Face collection maintained by the author for ready-to-use models.
Usage notes & performance tips
- The wrapper aims to be less reliant on
torch.compilefor VRAM efficiency and documents workarounds for compile-related VRAM spikes. Users are advised to re-run once to let caches populate or clear Triton caches if experiencing anomalies on Windows. - LoRA handling was changed so unmerged LoRA weights become buffers inside modules: this enables block-swapping and prefetch for async offloading, but increases VRAM if block swapping is not used. The README explains how to compute the effective VRAM change and how many additional blocks to swap to compensate.
- The author emphasizes that this project is often a fast experimental sandbox and not always intended to replace native ComfyUI implementations; some features may be temporary, WIP, or eventually upstreamed into ComfyUI proper.
Examples & scope
- The repository includes multiple example outputs (context-window tests, vid2vid with different model sizes, TeaCache examples) and demonstrates large-frame/context capabilities (e.g., thousands of frames with sliding windows and block offloading).
- Supported extras list shows wide interoperability with many open-source video-related models, utilities, and experimental projects (links provided in the README).
Caveats
- The project is labeled "WORK IN PROGRESS" and the author explicitly warns of potential instability and compatibility issues. New account issue reporting was temporarily restricted to reduce spam.
- Users should be familiar with ComfyUI and model placement conventions, and pay attention to VRAM/block-swap settings when using large models or unmerged LoRAs.
Summary
ComfyUI-WanVideoWrapper is a pragmatic toolkit for experimenting with WanVideo-based video generation inside ComfyUI: it streamlines model loading, offers memory/VRAM guidance, and connects many community models and techniques into the ComfyUI ecosystem for rapid prototyping and exploration.
