ComfyUI-MiniMaxH3-TeaCache

Timestep-caching acceleration for MiniMax H3 video generation in ComfyUI. Open-source Linux/Windows/macOS. Written from scratch, no .pyd.

Status

• Repo scaffold
• Core cache wrapper
• First real-world benchmark (see below)
• Multi-prompt sweep (fl2va / ref2va / t2v)
• Polynomial calibration (upgrade from raw rel-L1)
• Quality gate (VBench subset)
• v0.1 release tag

Real-world result (v0.1)

	Baseline (no cache)	TeaCache thresh=0.15	Delta
Total wall-clock	306.22 s	102.08 s	-66.7% (3.0×)
Sampling (20 steps)	246 s	87 s	-64.6%
Avg s / step	12.33	4.38	—
Perceived quality	reference	matches reference under A/B viewing	—

Setup: MiniMax H3 fl2va int8_convrot, res_multistep + simple scheduler, 20 steps, 1024×576, length 124 (~5.2 s), fixed seed 42, hardware NVIDIA CMP 170HX (GA100, sm_80, 64 GB modded VRAM). Baseline and cached runs share identical inputs.

Higher speedup than the closed-source competition (TE-Speed markets 45%; measured here at 66% with default threshold). Whether this holds on non-H3 or non-Ampere hardware is TBD.

What it does

MiniMax H3's DiT sampler runs 20 forward passes on a ~7B-parameter transformer for each 5-second video. Adjacent timesteps produce very similar block outputs, so we reuse the output of one forward pass for the next k steps as long as the accumulated input delta stays below a threshold.

Reference: Timestep Embedding Aware Cache (Liu et al. 2024).

Reference: Timestep Embedding Aware Cache (Liu et al. 2024).

How this differs from TE-Speed-MiniMaxH3

	TE-Speed (tl2012tl)	This repo
License	Closed	MIT
Source	Windows .pyd only	Python (Linux/Win/Mac)
model.py patch	Ships a modified comfy/ldm/minimax/model.py	Uses ComfyUI's official set_model_unet_function_wrapper — no core modification
Distribution	Bilibili + Quark netdisk	GitHub

Same underlying technique (adaptive step caching), independently implemented from the public paper.

Install

cd ComfyUI/custom_nodes
git clone https://github.com/Icyoung/ComfyUI-MiniMaxH3-TeaCache
# no build step, no compile — pure Python

Restart ComfyUI. A new node MiniMaxH3 TeaCache appears under advanced/model.

Usage

Insert the node between your UNETLoader and the guider:

UNETLoader → MiniMaxH3 TeaCache → CFGGuider / BasicGuider → SamplerCustomAdvanced

Parameters:

• rel_l1_thresh: relative L1 distance below which we reuse the previous output. Default 0.15. Lower = higher quality, less speedup. Range [0.05, 0.5].
• start_step: first step at which caching becomes active (early steps are structurally important, don't skip). Default 2.
• end_step: last step to cache (final steps decide fine detail, don't skip either). Default -2 (i.e. last 2 steps).

Reproducibility (for reviewers)

The cache decision is deterministic given a fixed rel_l1_thresh — same seed + same prompt produces bit-identical latents whether TeaCache is on or off, within numerical tolerance.

Benchmark

TBD after v0.1. Hardware: NVIDIA CMP 170HX (GA100, sm_80, 64GB modded VRAM). Kept for calibration reference; other cards may show different absolute times but comparable speedup ratios.

Contributing

PRs welcome. Please include a short benchmark diff (before/after) in the PR description.

Related work

• welltop-cn/ComfyUI-TeaCache — TeaCache for FLUX, Wan, Hunyuan, LTX (does not cover H3)
• nvmax/teacache — fork with updated ComfyUI deps

License

MIT.