Forecast TRELLIS DiT velocity on skipped sampling steps

Training-free acceleration for TRELLIS image-to-3D in ComfyUI. It forecasts the flow-matching velocity on skipped DiT steps instead of running the transformer, on both TRELLIS stages (sparse-structure and SLaT), via the hicache-pp library.

cd ComfyUI/custom_nodes
git clone https://github.com/Archerkattri/ComfyUI-TRELLIS-HiCache
pip install hicache-pp

Evidence

Speedup, interval=2, both stages (historical RTX 5090 run)	2.1x on TRELLIS-image-large, demo image, 25+25 steps
Speedup, interval=3, both stages (historical RTX 5090 run)	2.5x on TRELLIS-image-large, demo image, 25+25 steps
Chamfer vs stock, interval=2 (unit-cube units)	0.0059 near-lossless; about 0.6% of the ~1.0 object extent
Chamfer vs stock, interval=3 (unit-cube units)	0.0145 more aggressive setting
Default interval	2 is the default per the README
CPU contract suite	15 tests pass on CPU with a dummy DiT (no ComfyUI, no GPU)

Media

TRELLIS HiCache runtime path — Runtime path: sparse-structure and SLaT keep independent cache state and schedules.

Project icon; the node drops between the TRELLIS loader and the sampler.

Install & verify

pytest # expect 15 passed: CPU contract suite, no ComfyUI or GPU
# tests/validate_gpu.py is not run here; it needs a real TRELLIS pipeline and CUDA

Limits

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

In-ComfyUI GPU validation (no GPU hardware).

Family

Runs on	ComfyUI + ComfyUI_TRELLIS
Sibling nodes	ComfyUI-HiCache · ComfyUI-TRELLIS2-HiCache
Standalone counterparts	faster-trellis-plus-plus
Forecast core	hicache-plus-plus (PyPI)
External	ComfyUI docs