Skip Hunyuan3D DiT steps by forecasting flow-matching velocity

One ComfyUI node, HiCache Accelerate (Hunyuan3D), wired between Hy3DModelLoader and the mesh sampler. On compute steps the shape DiT runs normally and its output is cached; on skipped steps the DiT is not called and the flow-matching velocity is forecast from the cached anchors.

cd ComfyUI/custom_nodes
git clone https://github.com/Archerkattri/ComfyUI-HiCache
pip install -r ComfyUI-HiCache/requirements.txt   # just: hicache-pp

Evidence

Sampling speedup, hermite interval=3 (in ComfyUI)	2.68x — 0.70 s vs 1.87 s baseline over 30 steps
Mesh F1@0.05 vs unaccelerated baseline (hermite i3)	0.825, above the 0.751 different-seed floor
DiT forwards run, hermite i3 (30 steps)	11/30 computed, 19 skipped (dmd i5 runs 7/30)
DMD interval=5 in ComfyUI (2-mini)	1.69x speedup, F1 0.719 — at the seed-noise floor
Upstream Hunyuan3D-2.1, DMD interval=5	0.860 F-score at 1.79x speedup (baseline 0.911)
Upstream Hunyuan3D-2-mini, DMD interval=5	0.794 — exactly lossless vs baseline, at 1.69 s
CPU unit tests	45 passed, no ComfyUI needed

Media

Runtime path diagram — Runtime path: scheduled compute steps anchor the velocity history; skipped steps are forecast; incompatible schedules fall back to the full model.

Project icon, also used as the README banner.

Install & verify

pytest   # expect 45 passed (CPU-only, no ComfyUI needed)

Limits

What this release does not claim (from the release notes):

In-ComfyUI GPU validation.

Family

Runs on	ComfyUI + ComfyUI-Hunyuan3DWrapper
Sibling nodes	ComfyUI-TRELLIS-HiCache · ComfyUI-TRELLIS2-HiCache
Standalone counterparts	hunyuan2-plus-plus · hunyuan2.1-plus-plus
Forecast core	hicache-plus-plus (PyPI)
External	ComfyUI docs