Skip to content

v0.1.0 — first measured 3.0× speedup on MiniMax H3

Latest
Latest

Choose a tag to compare

View all tags
@Icyoung Icyoung released this 04 Aug 12:42
v0.1.0
4cbb50d

Highlights

  • 3.0× wall-clock speedup on MiniMax H3 fl2va (306.22s → 102.08s)
  • Same seed + same prompt: no perceptible quality difference under A/B viewing
  • Beats the closed-source 'TE-Speed-MiniMaxH3' baseline (advertised 45%, measured 66%)
  • Pure Python, cross-platform (Linux / macOS / Windows), MIT licensed
  • Uses ComfyUI's official set_model_unet_function_wrapper — no core file modifications

How it works

Adaptive timestep caching: reuses the previous forward-pass output while the accumulated input rel-L1 delta stays below a threshold. Runs the real model only on the first, last, and 'far-enough' steps. Reference: TeaCache paper.

Install

cd ComfyUI/custom_nodes
git clone https://github.com/Icyoung/ComfyUI-MiniMaxH3-TeaCache
# no build step

Restart ComfyUI. Insert the MiniMax H3 TeaCache node between UNETLoader and your guider.

Measurement setup

MiniMax H3 fl2va int8_convrot, res_multistep + simple scheduler, 20 steps, 1024×576, length 124 (~5.2s), fixed seed 42. Hardware: NVIDIA CMP 170HX (GA100, sm_80, 64 GB modded VRAM).

Known limitations

  • H3-only for now; ref2va + t2v tested for correctness but not extensively benchmarked
  • Threshold 0.15 is aggressive; some fast-motion prompts may benefit from 0.10
  • No polynomial calibration yet (uses raw rel-L1). Coming in v0.2.
Assets 2
Loading