Faster TRELLIS image-to-3D with Hermite velocity forecasting

faster-trellis is microsoft/TRELLIS image-to-3D with a training-free acceleration stack built into the flow-matching sampler. It forecasts and reuses the final CFG-combined velocity, so the sampler runs fewer, cheaper network evaluations per asset; the weights, decoders, and mesh + Gaussian + radiance-field outputs are the base model’s, untouched.

git clone --recurse-submodules https://github.com/Archerkattri/faster-trellis
cd faster-trellis
# TRELLIS deps (CUDA toolchain required); see setup.sh for the full option list.
. ./setup.sh --new-env --basic --xformers --flash-attn --spconv --nvdiffrast
# Deployment-contract runtime (budget/telemetry/manifest):
pip install "hicache-pp @ git+https://github.com/Archerkattri/hicache-plus-plus@master"

Evidence

End-to-end speedup vs base TRELLIS	2.85× — 1.22 s vs 3.49 s per asset, Toys4K, RTX 5090, 25 steps/stage
Speedup vs Fast-TRELLIS	1.35× — 1.22 s vs 1.65 s at equal-or-better quality
F1@0.05 mean	0.825 vs 0.823 (Fast-TRELLIS); base 0.841
Chamfer distance	0.0581 vs 0.0594 (Fast-TRELLIS); base 0.0556
SLaT network evals (median)	6 vs 14 (Fast-TRELLIS); base ~25
Benchmark scope	32 objects all variants align; mesh-decoder output, Go-ICP similarity alignment, identical harness every row

Media

faster-trellis demo render — Demo render from the repository assets.

faster-trellis teaser — Teaser image from the repository assets.

Sampler integration diagram — Sampler integration: token carving and velocity forecasting are controlled separately.

Install & verify

# CPU-only accelerator tests (no GPU, no TRELLIS weights) — from CI:
python tests/test_hicache.py
python tests/test_adaptive_cfg.py
python tests/test_hicache_contract.py  # 4 deployment-contract tests

Limits

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

The results card is retained as prior-run evidence from the stated benchmark setup. It is not a current acceptance result and does not establish a universal speedup, losslessness, or quality ordering.
Re-run the manifest command in example_faster.py on the target GPU before making a new claim.
The published numbers above were measured with the as-released code and remain valid as-measured; re-validation on this model is pending.

Family

Base model	TRELLIS v1, accelerated by the hicache-plus-plus forecast core
Same base	faster-trellis-plus-plus
All 13 adapters	TRELLIS: faster-trellis, faster-trellis-plus-plus · TRELLIS.2: fast-trellis2, hermit-trellis2, hermit-trellis2-plus-plus · Hunyuan3D-2: hunyuan2-plus, hunyuan2-plus-plus · Hunyuan3D-2.1: hunyuan2.1-plus, hunyuan2.1-plus-plus · SAM: sam3d-plus, sam3d-plus-plus, fastsam3d-plus, fastsam3d-plus-plus
External	Paper arXiv:2412.01506 · project page · live demo