Hunyuan3D-2.1 image-to-3D with exponential velocity cache

Tencent's Hunyuan3D-2.1 image/text-to-3D, accelerated by the HiCache++ exponential velocity cache on its DiT flow-matching loop. On most sampling steps the DiT forward is skipped and the flow-matching velocity is forecast from cached anchors, so the DiT runs only every interval steps.

git clone https://github.com/Archerkattri/hunyuan2.1-plus-plus.git
cd hunyuan2.1-plus-plus
pip install -r requirements.txt
# model weights are not committed; download Hunyuan3D-2.1 from upstream

Evidence

DMD F-score at interval-5	≈ 0.86 historical Toys4K result, 3-seed evidence (not current validation)
Hermite F-score at interval-5	≈ 0.74 polynomial baseline drops where DMD holds
Uncached baseline F-score	≈ 0.91 full-compute reference, excluding one Go-ICP-degenerate sphere
CPU contract tests	4 passed DMD adapter, snapshot, fallback, and CFG-index contracts
DMD snapshot window	≥ 4 snapshots uniform window required; shorter windows fall back to Hermite
Example cache schedule	interval 5 with first_enhance 2 and history 6

Media

HiCache++ integration diagram — Shape-sampler integration: DMD velocity modes at the DiT boundary

demo input image — Demo input image used by the image-to-3D pipeline

Install & verify

python -m hy3dshape.hy3dshape.hicache_dmd # CPU self-test, no GPU or model needed
python -m pytest tests/ -v # 4 passed

Limits

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

GPU speedup validation and weight-gated demo cells.

Family

Base model	Hunyuan3D-2.1, accelerated by the hicache-plus-plus forecast core
Same base	hunyuan2.1-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:2506.15442