TTP_TeaCache HunyuanVideo Sampler

Hunyuan Video sampling is slow - full-quality settings can eat minutes per clip even on decent hardware - and TeaCache is the community's standard answer to that, not something TTPlanetPig invented from scratch. The pack's own README credits its origin honestly: the technique comes from ali-vilab/TeaCache, and this specific ComfyUI implementation leans on facok/ComfyUI-TeaCacheHunyuanVideo. If you've seen TeaCache nodes for Flux or other models before, this is the same idea, wired for Hunyuan's SamplerCustomAdvanced-style graph.

How TeaCache actually works. Diffusion sampling reruns the transformer at every step, and a lot of those steps produce outputs that are barely different from the one before. TeaCache tracks how much consecutive steps' internal activations actually change (a relative L1 threshold), and when the change is small enough, it reuses the cached result instead of recomputing the full pass - skipping compute on steps that wouldn't have moved the output much anyway. That's why the trade-off is speed for a small, real quality hit rather than something for nothing: the model isn't doing less work per se, it's doing less repeated work.

Inputs and outputs. This node is a drop-in swap for ComfyUI's native SamplerCustomAdvanced - same required inputs: noise, guider, sampler, sigmas, latent_image. Where it differs is speedup, a five-step dial from Original (1x) through Fast (1.6x) (the default), Faster (2.1x), Ultra Fast (3.2x), up to Shapeless Fast (4.4x) - the author's own tooltip names these exact multipliers. If you want finer control than the five presets, flip enable_custom_speed on and set custom_speed directly (1.0–4.4). It gives you two latent outputs, output and denoised_output, matching the pattern of ComfyUI's native advanced sampler.

The number that matters: the README's own benchmark is a 720x480, 65-frame Hunyuan Video clip finishing in about 55 seconds on an RTX 4090, roughly a 2.1x speedup - that's the "Faster" tier, not the most aggressive one.

Where to set speedup. Start at the default Fast (1.6x). The community consensus on TeaCache generally (measured on the Flux implementation, but the shape of the trade-off is the same technique) is that output ends up "pretty damn close to the same, just... worse" - good enough that people run it for fast iteration and switch it off (or dial it down) for a final pass. The pack's own README is equally direct about the ceiling: quality and motion can change once acceleration gets too aggressive, so treat Ultra Fast and Shapeless Fast as preview-quality settings, not something you'd want on your final render.

Install. ComfyUI Manager, search "Comfyui_TTP_Toolset", or cd ComfyUI/custom_nodes && git clone https://github.com/TTPlanetPig/Comfyui_TTP_Toolset, then restart. This node itself needs no extra downloads, but it's not a standalone tool - it replaces one node inside a working Hunyuan Video sampling graph, so you need your checkpoint, text encoders, and the rest of that pipeline already set up.

Common issues. If a step in your graph relies on precise, deterministic per-step behavior (some advanced CFG or guidance-editing techniques don't compose cleanly with any caching sampler, TeaCache included), try speedup=Original (1x) first to rule out the caching as the cause before debugging further upstream. And if motion looks noticeably worse or waxy at higher speedup values - that's not a bug, that's the trade the author's README warns about directly; step back down a tier rather than assuming something's broken.