This repository provides ComfyUI nodes for ComfyUI-TeaCache, a training-free caching method designed to accelerate inference for diffusion models by leveraging timestep-specific output differences. It targets users of ComfyUI working with image, video, and audio diffusion models, offering significant speedups with minimal quality degradation.

How It Works

TeaCache estimates and caches fluctuating differences in model outputs across diffusion timesteps. This approach avoids recomputing redundant information, leading to faster inference. The integration into ComfyUI is seamless, requiring only node connections within existing workflows.

Quick Start & Requirements

• Installation: Preferred via ComfyUI-Manager. Manual installation involves cloning the repository into custom_nodes and running pip install -r requirements.txt.
• Prerequisites: ComfyUI, Python.
• Usage: Connect the TeaCache node after model loading nodes. Recommended rel_l1_thresh and max_skip_steps values are provided for various models.
• Documentation: Examples are available in the examples folder.

Highlighted Details

• Supports a wide range of models including FLUX, PuLID-FLUX, HunyuanVideo, LTX-Video, CogVideoX, and Wan2.1 variants.
• Achieves speedups ranging from 1.2x to 2.3x, with some configurations offering lossless speedups.
• Includes a Compile Model node leveraging torch.compile for further inference acceleration.
• Offers a "retention mode" for Wan2.1 models to improve generation speed and quality.

Maintenance & Community

• Active development with recent updates in March 2025.
• Acknowledgments to the original TeaCache repository.

Licensing & Compatibility

• The repository does not explicitly state a license in the provided README.

Limitations & Caveats

• The README does not specify a license, which may impact commercial use or integration into closed-source projects.
• Initial compilation for the Compile Model node can be time-consuming.