>
hermit-trellis2 0.1.1

Hermite-cached acceleration for TRELLIS.2 image-to-3D

Training-free Hermite carved-hybrid for TRELLIS.2-4B image-to-3D: it forecasts the final CFG-combined velocity with a dual-scaled Hermite basis on the sparse-structure stage and carves structured-latent tokens. Weights, decoders and the full 1024_cascade mesh + texture are left untouched.

git clone https://github.com/Archerkattri/hermit-trellis2
cd hermit-trellis2
# TRELLIS.2 runtime deps (torch, flash-attn, spconv/flex_gemm, o-voxel, cumesh,
# nvdiffrast) per microsoft/TRELLIS.2. Place / symlink weights at ckpts/TRELLIS.2-4B.
# Deployment-contract runtime (budget/telemetry/manifest):
pip install "hicache-pp @ git+https://github.com/Archerkattri/hicache-plus-plus@master"
01

Evidence

End-to-end speedup1.88× (6.26 s vs 11.75 s base, RTX 5090, full 1024_cascade)
F1@0.05 mean0.869 vs base 0.860 (TaylorSeer sibling 0.900)
F1 median0.957 vs base 0.932
Chamfer distance0.059 vs base 0.057, within run-to-run noise
Benchmark scope35 of 40 Toys4K objects completed by all three rows, seed 42
SS forecast intervalcompute 1 step, forecast 1 (GF_HICACHE_SS_INTERVAL=2)
SLaT carve ratio0.10 of tokens cached per step (GF_CARVE_RATIO)
SS first-enhancefirst 6 steps always computed (GF_HICACHE_FIRST_ENHANCE)
02

Media

hermit-trellis2 teaser
Teaser: image-to-3D output of the accelerated pipeline.
sampler integration diagram
Sampler integration: Hermite forecast on sparse-structure, carving on SLaT.
example input image 1
Example input image from assets/example_image.
example input image 2
Example input image from assets/example_image.
03

Install & verify

# 5 CPU contract tests (CI job cpu-contract: torch CPU + hicache-pp master)
python tests/test_hicache_contract.py
# expected: 5 x [PASS]
04

Limits

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

  • The historical-results card “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.”
05

Family

Base modelTRELLIS.2, accelerated by the hicache-plus-plus forecast core
Same basefast-trellis2 · hermit-trellis2-plus-plus
All 13 adaptersTRELLIS: 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
ExternalTRELLIS.2 project page · 4B weights · background arXiv:2412.01506