HunyuanVideo TeaCache

TeaCache is a caching trick for video diffusion: across the many denoising steps of a generation, a lot of the model's internal computation is nearly identical from one step to the next, and TeaCache detects when a step is close enough to a cached one to skip the redundant work. This wrapper's implementation predates the better-known port to Wan (the same acceleration idea shows up as WanVideoTeaCacheKJ in kijai's other wrapper, though that one's since been superseded by native alternatives) - HunyuanVideo was one of the models TeaCache was demonstrated on first.

This is a settings-only node. It doesn't do anything until you wire its output into HyVideoSampler's optional teacache_args socket.

The inputs that matter

• rel_l1_thresh (default 0.15) - the similarity threshold that decides when a step is "close enough" to skip. Higher values skip more aggressively: faster, but more likely to introduce visible artifacts. The tooltip says this outright. This default is noticeably more conservative than the equivalent setting in kijai's Wan wrapper, which suggests it's tuned to be a safer starting point here.
• cache_device (default offload_device) - where the cached tensors live. Keeping them off the GPU trades a bit of transfer overhead for headroom, which matters more on this model than most given how tight VRAM already runs.
• start_step / end_step (default 0 / -1) - the range of steps caching applies to. The defaults mean "the whole run"; narrowing this to skip caching on the earliest steps, where the model is establishing overall structure, is a common way to keep quality up while still getting most of the speedup on later, more repetitive steps.

Output: teacache_args, into the sampler.

How to install it

Ships with the pack - nothing extra to download for the node itself.

• ComfyUI Manager - search ComfyUI-HunyuanVideoWrapper, install, restart.
• Manual - cd ComfyUI/custom_nodes && git clone https://github.com/kijai/ComfyUI-HunyuanVideoWrapper, then pip install -r ComfyUI-HunyuanVideoWrapper/requirements.txt, restart.

No model download tied to this node specifically - it just changes how the sampler computes, given a model you've already loaded.

Common issues & troubleshooting

Faces or fine detail drift or wobble across the clip. This is the caching tradeoff, and it's consistent across every TeaCache-style implementation the community has used: skipping "close enough" steps means occasionally reusing something that should have actually changed, and faces are where you notice it first. If identity consistency matters more than speed for a given shot, lower rel_l1_thresh or disable the node entirely for that run.

Barely any speedup. A low rel_l1_thresh is conservative by design - that's the safety margin, not a bug. Raise it gradually and watch quality as you go, rather than jumping straight to an aggressive value.

Nothing changed at all. Confirm teacache_args is actually wired into HyVideoSampler - a disconnected settings node does nothing, silently.

Artifacts appear specifically in the first second or two of the clip. Try raising start_step so caching doesn't kick in until the model has finished laying down the early structure of the shot - that's usually where aggressive caching hurts the most.