Fast-SAM3D slat sampler with Hermite velocity forecasting
Image-to-3D with Fast-SAM3D plus a tree-aware HiCache cache on its slat-stage flow-matching. The dynamics network is skipped on most slat-stage solver steps and the cached velocity is forecast with a scaled-Hermite polynomial.
# Inference dependencies (weights are downloaded separately, see below) pip install -r requirements.inference.txt # Base-model weights are not committed; download from upstream Fast-SAM3D # https://github.com/wlfeng0509/Fast-SAM3D
01
Evidence
| F1@0.05 vs uncached Taylor baseline | 1.000 — single-input historical result |
|---|---|
| Hermite speed vs Taylor baseline | 0.922× — same single-input run, configuration-specific |
| Mean symmetric Chamfer distance vs baseline | 0.01623 — same historical result |
| CPU contract tests passing | Eight — no GPU required |
| Benchmark scope | 1 input, 1 seed, 30000 points per output |
| Configs compared in benchmark card | 4 — taylor/hermite basis × adaptive CFG off/on |
| Accel selftest cadence | steps 0,1 full; 2 forecast; 5,10,11 full |
02
Media
03
Install & verify
# dev test runner pip install -r requirements.dev.txt # README: eight CPU contract tests pass pytest tests/ # CPU selftest, no GPU or SAM3D model needed python -m sam3d_objects.model.backbone.generator.flow_matching.accel
04
Limits
What this release does not claim (from the release notes):
- GPU speedup validation and weight-gated demo cells.
05
Family
| Base model | Fast-SAM3D, accelerated by the hicache-plus-plus forecast core |
|---|---|
| Same base | sam3d-plus · sam3d-plus-plus · fastsam3d-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:2602.05293 · SAM 3D arXiv:2511.16624 |