Faster ComfyUI diffusion by skipping the redundant timesteps

What it does

ComfyUI-TeaCache is essentially a well-maintained glue layer that wraps the TeaCache inference shortcut into ComfyUI-native nodes. It drops a TeaCache node between your model loader and sampler to cache redundant computations across diffusion timesteps, cutting generation time for a long list of image and video models without retraining. It also bundles a Compile Model node that runs torch.compile for an extra performance kick.

The interesting bit

TeaCache stands for Timestep Embedding Aware Cache: it estimates how much model outputs fluctuate between steps and skips the ones it considers close enough. The ComfyUI integration exposes this as a single node with per-model tuning knobs—rel_l1_thresh, start_percent, and end_percent—so you can trade a little fidelity for a lot less waiting without leaving the node graph.

Key highlights

• Claims training-free speedups around 1.5× to 3× for supported models, with a published table of exact thresholds for FLUX, HiDream, HunyuanVideo, Wan2.1, and others.
• Covers both image pipelines (FLUX, Lumina-Image-2.0) and video pipelines (LTX-Video, CogVideoX, Wan2.1), plus LoRA and ControlNet support where noted.
• Includes a cache_device option to park the cache on CPU if VRAM is scarce, or on CUDA for lower latency.
• Bundles a Compile Model node for additional backend optimization via torch.compile.
• Ships with ready-made example workflows for every supported model.

Caveats

• Aggressive rel_l1_thresh values can visibly degrade output; the README advises lowering the threshold if quality drops.
• The Compile Model node triggers a lengthy compilation on the first workflow run.
• CogVideoX support requires kijai’s separate ComfyUI-CogVideoXWrapper nodes.

Verdict

Worth a look if you live in ComfyUI and want cheaper inference without swapping hardware. Skip it if you are allergic to tuning tolerance knobs or not using ComfyUI at all.