Hermite velocity caching for Hunyuan3D-2 mini sampling
Hunyuan3D-2 mini with HiCache wired natively into its flow-matching denoise loop: skip the DiT on most sampling steps and forecast the cached velocity instead — no retraining, no monkey-patching.
git clone https://github.com/Archerkattri/hunyuan2-plus.git
cd hunyuan2-plus
pip install -r requirements.txt # pulls hicache-pp>=1.2.1 via requirements.txt
01
Evidence
| Base model | 0.6B parameters (Hunyuan3D-2 mini shape generator) |
|---|---|
| Uncached baseline | 0.794 F-score@0.05 on Toys4K, 10 objects |
| HiCache i3/o2 (corrected) | 0.760 F-score@0.05; 0.921 vs 0.911 vanilla over 7 aligned objects |
| HiCache i5/o3 probe (corrected) | 0.784 F-score@0.05 vs 0.794 baseline |
| HiCache i6/o3 probe | 0.404 F-score@0.05 — polynomial skip ceiling below baseline |
| Hunyuan3D-2.1 reference | 0.88 → 0.74 → 0.38 F-score at interval 3 / 5 / 6 |
| CPU contract tests | 3 passed — cache identity, schedule stability, reset contract |
02
Media
03
Install & verify
python -m pytest tests/ -v
# expected: 3 passed — CPU contract tests
04
Limits
What this release does not claim (from the release notes):
- GPU speedup validation and weight-gated demo cells.
05
Family
| Base model | Hunyuan3D-2, accelerated by the hicache-plus-plus forecast core |
|---|---|
| Same base | hunyuan2-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:2501.12202 |