Xmarre Supercharges WAN Video with ComfyUI-Spectrum-WAN-Proper
ComfyUI-Spectrum-WAN-Proper is a custom node for ComfyUI that speeds up WAN video generation. It uses a technique called Spectrum, which forecasts denoiser features instead of running the full network at every step, reducing total processing time.
The node was created by developer Xmarre after upstream changes broke the original Spectrum implementation for WAN models. This updated version maintains backwards compatibility while adding support for the latest WAN video backends.
Model Size: requirements vary & VRAM GPU: 34.5GB+ recommended
Speed and efficiency features
- Provides up to 1.56x speedup compared to standard sampling
- Supports Wan 2.1, Wan 2.2 TI2V 5B, and Wan 2.2 14B models
- Offers a low VRAM mode that keeps memory usage near baseline levels
- Works with existing ComfyUI workflows without major changes
- Uses Chebyshev polynomial forecasting for accurate feature prediction
Video creators working with WAN models on limited hardware will find this tool helpful for cutting down render times. The node integrates directly into standard workflows, so users do not need to rebuild their pipelines from scratch.
Developer notes and considerations
The developer addressed a significant VRAM issue in early testing. Initial versions increased memory usage from 33GB to 38-40GB, which was an oversight. A new optional low_vram_exact path now limits the increase to approximately 34.5GB without sacrificing speed or output quality.
One known bug remains: the model stays patched for Spectrum after the first run, even if the node is bypassed. Users need to restart ComfyUI or reload the model fully to return to standard sampling. The developer noted that
'video output is different but I couldn't detect any raw quality degradation.'
This implementation focuses only on native ComfyUI WAN backends and does not support third-party wrappers or other video models like LTX or Hunyuan.
Get ComfyUI-Spectrum-WAN-Proper on GitHub.
