TeaCache For Vid Gen

The video-side twin of TeaCacheForImgGen, and just as thoroughly superseded. Before the pack unified everything into one TeaCache node, image and video models each got their own caching node - this one covers HunyuanVideo and LTX-Video, added in the pack's first couple of weeks, before Wan 2.1 support even existed. If you're starting a new workflow, reach for TeaCache instead; it covers these two models plus Flux, HiDream, Lumina and Wan, all through one node.

How it works

Mechanically it's the same trick as the rest of the pack: TeaCache estimates how close a denoising step's output would be to the cached one, and skips the step outright when the difference is negligible. The README's own numbers for the two models this node covers make the tuning gap obvious - HunyuanVideo at rel_l1_thresh 0.15 nets roughly a 1.9x speedup, while LTX-Video needs a much lower 0.06 for about 1.7x. LTX's schedule is simply more sensitive to skipped steps.

The inputs and outputs that matter

model_type is a two-way dropdown - hunyuan_video or ltxv - and nothing else, because that's the entire scope of this node. rel_l1_thresh is the only real dial, defaulting to 0.15 (tuned toward HunyuanVideo; drop toward 0.06 if you're actually on LTX-Video). Same pieces missing as its image-side sibling: no start_percent/end_percent windowing, no cache_device choice. Output is a patched model into your sampler, as always.

How to install it

Same pack. Via ComfyUI Manager: search "ComfyUI-TeaCache", install, restart. Manually:

cd ComfyUI/custom_nodes
git clone https://github.com/welltop-cn/ComfyUI-TeaCache
cd ComfyUI-TeaCache
pip install -r requirements.txt

No model weights to fetch - it patches whatever HunyuanVideo or LTX-Video model you've already loaded.

Common issues & troubleshooting

LTX-Video output comes out blurry or garbled. Early LTX support in this pack had a rough patch - people reported not knowing where to place the node, or getting oddly soft results while support was still catching up. Make sure you're on a current pack version, and consider TeaCache instead of this legacy node if the issue persists - it's had more iteration since.

You want Wan or Flux support. Not here. This node's model_type genuinely only offers HunyuanVideo and LTX-Video. Use the unified TeaCache node for anything else, including Wan 2.1's retention-mode presets.

Quality dropped hard at the default threshold. 0.15 is a HunyuanVideo number. If you're running LTX-Video, that's roughly double the README's recommended 0.06 - bring it down.

Multiple TeaCache-family nodes in one workflow, and only one threshold seems to apply. That was a genuine early bug in this pack - with more than one TeaCache node present, rel_l1_thresh would silently use whichever node's value was set last rather than each node's own. It's long fixed, but if you're pinned to a very old copy of the pack, that's a real explanation before you start second-guessing your graph.