When to use this repo

These repos are complementary accelerators, not competing solutions — each speeds up a different base generator, and the + / ++ suffix is a method choice, not a rival product. Pick by (1) which base model you run, then (2) which forecast basis you want:

base generator	+ = HiCache (Hermite)	++ = HiCache++ (DMD)
Hunyuan3D-2.1	hunyuan2.1-plus	hunyuan2.1-plus-plus
Hunyuan3D-2 mini	hunyuan2-plus	hunyuan2-plus-plus
SAM 3D Objects	sam3d-plus	sam3d-plus-plus
Fast-SAM3D	fastsam3d-plus	fastsam3d-plus-plus
TRELLIS (v1)	faster-trellis	faster-trellis-plus-plus
TRELLIS.2-4B (v2)	hermit-trellis2	hermit-trellis2-plus-plus

• + (HiCache / scaled-Hermite): the published polynomial velocity-forecast basis — conservative, reproduces the HiCache paper. Use it to deploy the established method.
• ++ (HiCache++ / DMD exponential): our Dynamic-Mode-Decomposition basis — the same near-lossless quality at wider skip intervals, where the polynomial diverges. Use it when you push the cache interval for more speed.
• standalone / model-agnostic: hicache-plus-plus — the forecaster itself, to add DMD caching to your own diffusion/flow model.
• fast-trellis2 = the TaylorSeer baseline fork (the upstream "Fast" accel) — the v2 reference point, not a HiCache variant.

This repo: hunyuan2-plus — Hunyuan3D-2 mini × HiCache (Hermite) — published cache on the deployed mini shape model.

---

What it is

This is a fork of Hunyuan3D-2 mini (Tencent's 0.6B shape generator, Hunyuan3DDiTFlowMatchingPipeline) with HiCache built directly into the diffusion sampler. The image-to-3D model is unchanged; the only addition is a feature cache on its flow-matching loop.

A flow-matching sampler integrates dx/dt = v_θ(x, t) over N steps, calling the expensive DiT at every step to get the velocity v. HiCache computes the DiT only on 1/interval of the steps and forecasts the velocity on the rest from a small cache of recent anchors — skipping (interval-1)/interval of the forward passes. The forecast is a first-class part of the denoise loop (__call__ calls the cache helpers directly — there is no runtime monkey-patching).

Method — Hermite polynomial forecast

At each compute ("full") step HiCache stores the CFG-combined velocity F_t and updates backward finite-difference derivatives Δ^i F_t. At a skipped step k steps past the last compute step, it forecasts

F_hat_{t+k} = F_t + Σ_{i≥1} (Δ^i F_t / i!) · H̃_i(k)

with the dual-scaled physicist's Hermite polynomial H̃_n(x) = σ^n · H_n(σ·x), σ ∈ (0,1). The σ contraction keeps the high-order terms bounded, giving a more stable extrapolation than the equivalent Taylor (monomial) series — TaylorSeer is the special case H̃_i(k) = k^i. With fewer than two anchors the forecast degenerates to plain reuse of the cached velocity (the correct zero-information forecast).

Use

Load the mini pipeline and turn the cache on with one call before sampling:

from PIL import Image
from hy3dgen.shapegen import Hunyuan3DDiTFlowMatchingPipeline

pipe = Hunyuan3DDiTFlowMatchingPipeline.from_pretrained(
    'tencent/Hunyuan3D-2mini', subfolder='hunyuan3d-dit-v2-mini',
)

# HiCache (Hermite): one DiT step then (interval-1) forecasts.
pipe.enable_hicache(
    interval=4,        # N: 1 compute step, then interval-1 forecasts
    max_order=1,       # highest finite-difference / Hermite order
    first_enhance=2,   # always compute the first few steps (warm-up)
    end_enhance=None,  # always compute steps with index >= end_enhance (default: none)
    sigma=0.5,         # Hermite contraction σ ∈ (0,1)
)
# pipe.disable_hicache()  # restore the uncached sampler

mesh = pipe(image=Image.open('assets/demo.png').convert('RGBA'))[0]
mesh.export('demo.glb')

enable_hicache just stores the config; Hunyuan3DDiTFlowMatchingPipeline.__call__ reads it and runs the compute/forecast schedule natively (see hy3dgen/shapegen/hicache.py and the denoise loop in hy3dgen/shapegen/pipelines.py).

Results

HiCache is the polynomial baseline: lossless at low skip intervals, then degrading as the skip grows, because the Hermite (polynomial) basis is only a local truncation of the trajectory's true exponential class and diverges under extrapolation. On Hunyuan3D-2.1 the Toys4K F-score@0.05 falls 0.88 → 0.74 → 0.38 at interval 3 / 5 / 6.

For the exponential forecaster that stays lossless at larger skips — 0.794 F-score at interval-5, exactly matching the uncached baseline on the deployed Hunyuan3D-2 mini — see the hunyuan2-plus-plus fork (HiCache++ / DMD), built on the standalone hicache-plus-plus library.

Sign-convention update (2026-06-10)

The vendored Hermite forecast in hicache.py evaluated the basis at x = -k; the corrected convention from hicache-plus-plus 1.2.0 is x = +k (the upstream TaylorSeer distance convention; -k flips every odd-order term, extrapolating backwards). This fork now ships the corrected forecast. All previously published numbers were measured with the as-released code and remain valid as-measured.

Re-validated on the deployed Hunyuan3D-2 mini (Toys4K, 10 objects, F-score@0.05, the deterministic protocol of the family benchmark; the as-released arm reproduces the published cells bit-exact):

config	as-released (x = -k)	corrected (x = +k)
vanilla (uncached)	0.794	0.794
HiCache i3/o2	0.792	0.760
HiCache i5/o3 (probe)	0.738	0.784
HiCache i6/o3 (probe)	0.578	0.404

Reading the table honestly:

• At the published interval (i3/o2) the corrected forecast matches the as-released one. The 0.792 vs 0.760 full-set gap is a single Go-ICP alignment timeout on a rotationally symmetric bowl (scored unaligned, a known protocol artifact); over the 7 objects aligned in both runs the means are 0.922 (as-released) vs 0.921 (corrected), vanilla 0.911.
• At i5/o3 the corrected sign is clearly better (0.784 vs 0.738, baseline 0.794), and it rescues a catastrophic as-released cell (bottle 0.25 to 0.97). This is the regime where the backwards extrapolation of the old sign bites.
• At i6/o3 both conventions are far below baseline (0.58 / 0.40 vs 0.79): the polynomial skip ceiling dominates regardless of sign. For wide skips use the exponential (DMD) fork, which stays at 0.793 at interval-5 with the corrected warm-up fallback (re-validated).

Attribution

• Hunyuan3D-2 / Hunyuan3D-2 mini © Tencent — see LICENSE and NOTICE (Tencent Hunyuan 3D 2.0 Community License Agreement; note its territorial limits, large-user threshold, and no-competing-model-training restrictions). This fork adds only the HiCache integration; the model, weights, and pipeline are unchanged.
• HiCache — Training-free Acceleration of Diffusion Models via Hermite Polynomial Feature Forecasting (arXiv:2508.16984). The Hermite/finite-difference core in hicache.py is a clean reimplementation; only the loop wiring is Hunyuan-specific. Built on TaylorSeer (monomial feature caching).

Weights & data

Model weights and demo/example assets are not committed to this repo — only the acceleration architecture (code + integration). Download the base-model weights from the upstream project, Tencent-Hunyuan/Hunyuan3D-2, per its instructions, and point the loader at them (see the code / upstream README). This keeps the repository lightweight and avoids redistributing third-party weights.

Citation

If you use this work, please cite the base model (Tencent Hunyuan3D-2) and the acceleration method (HiCache):

@misc{lai2025flashvdm,
      title={Unleashing Vecset Diffusion Model for Fast Shape Generation}, 
      author={Zeqiang Lai and Yunfei Zhao and Zibo Zhao and Haolin Liu and Fuyun Wang and Huiwen Shi and Xianghui Yang and Qinxiang Lin and Jinwei Huang and Yuhong Liu and Jie Jiang and Chunchao Guo and Xiangyu Yue},
      year={2025},
      eprint={2503.16302},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2503.16302}, 
}
@misc{hunyuan3d22025tencent,
    title={Hunyuan3D 2.0: Scaling Diffusion Models for High Resolution Textured 3D Assets Generation},
    author={Tencent Hunyuan3D Team},
    year={2025},
    eprint={2501.12202},
    archivePrefix={arXiv},
    primaryClass={cs.CV}
}

@misc{yang2024hunyuan3d,
    title={Hunyuan3D 1.0: A Unified Framework for Text-to-3D and Image-to-3D Generation},
    author={Tencent Hunyuan3D Team},
    year={2024},
    eprint={2411.02293},
    archivePrefix={arXiv},
    primaryClass={cs.CV}
}

@misc{hicache2025, title={HiCache: Training-free Acceleration of Diffusion Models via Hermite Polynomial Feature Forecasting}, eprint={2508.16984}, archivePrefix={arXiv}, year={2025}}

---

Family

Part of the HiCache++ acceleration family.

• Family hub: hicache-plus-plus — the basis library behind this adapter.
• Sibling: hunyuan2-plus-plus — the same base model with the HiCache++ (Dynamic Mode Decomposition / Prony) exponential-forecast variant.