🛠 fast-trellis2

Fast-TRELLIS's training-free acceleration, faithfully ported onto TRELLIS.2.

TRELLIS.2-4B · 1024_cascade (mesh + texture) · training-free · single RTX 5090 · MIT

This is a port, not a new method. It re-implements the acceleration introduced by Fast-TRELLIS (built for TRELLIS v1) on the v2 sampler stack, so you get the same speedup on TRELLIS.2. All credit for the acceleration design belongs to the Fast-TRELLIS authors; all credit for the base model belongs to microsoft/TRELLIS.2. For our own acceleration method on v2 — a Hermite (HiCache) sparse-structure forecast paired with a token-carved SLaT sampler — see the sibling repo hermit-trellis2.

fast-trellis2 wires Fast-TRELLIS's cross-step caching into TRELLIS.2's three flow-matching stages (sparse structure + shape SLaT + texture SLaT). Microsoft TRELLIS.2 model / decoder / o-voxel code is left untouched.

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:

• + (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: fast-trellis2 — TRELLIS.2-4B × TaylorSeer baseline fork — the upstream "Fast" accel, the v2 reference point.

At a glance

40 Toys4K objects, 1024_cascade, RTX 5090, seed 42; means over the 35 objects every config completed in one run. Geometry is scored on the o-voxel mesh decoder output with area-weighted surface sampling, after a globally-optimal (Go-ICP) similarity alignment to the ground-truth mesh. Latency is end-to-end generation, one object at a time, weights resident.

_{CD ↓ lower is better; F1 ↑ higher is better. The cross-step caching cuts latency ~1.9× while improving geometry — higher mean F-score (0.900 vs 0.860) and lower Chamfer distance (0.048 vs 0.057) than the unaccelerated base. The mean is deflated by a few rotationally-symmetric objects (ball, bowl…) Go-ICP cannot orient uniquely, hence the per-object median alongside. For our own v2 acceleration method — the Hermite carved hybrid (HiCache SS forecast + token-carved SLaT) — see the sibling repo hermit-trellis2.}

Quickstart

git clone --recursive https://github.com/Archerkattri/fast-trellis2
cd fast-trellis2
bash setup.sh --new-env --basic --flash-attn --o-voxel --flexgemm --nvdiffrast --nvdiffrec

Enable acceleration by swapping in the ported samplers; the pipeline auto-detects them and flips enable_faster on:

from trellis2.pipelines import Trellis2ImageTo3DPipeline, samplers

pipeline = Trellis2ImageTo3DPipeline.from_pretrained("microsoft/TRELLIS.2-4B").cuda()

pipeline.sparse_structure_sampler = samplers.FlowEulerGuidanceIntervalSampler_taylor(
    sigma_min=pipeline.sparse_structure_sampler.sigma_min)
pipeline.shape_slat_sampler = samplers.FlowEulerGuidanceIntervalSampler_faster(
    sigma_min=pipeline.shape_slat_sampler.sigma_min)
pipeline.tex_slat_sampler = samplers.FlowEulerGuidanceIntervalSampler_faster(
    sigma_min=pipeline.tex_slat_sampler.sigma_min)
# pipeline.enable_faster auto-enables once the *_faster / *_taylor samplers are set.

mesh = pipeline.run(image)[0]

See example.py (stock) and example_faster.py (accelerated).

SPARSE_CONV_BACKEND=spconv SPCONV_ALGO=native python example_faster.py

What the port replicates

Fast-TRELLIS's three components, wired into the TRELLIS.2 samplers (all credit: Fast-TRELLIS):

Credits & license

This repo reproduces the method of, and depends on, two MIT-licensed projects — both are credited because this work reproduces their contributions:

MIT. See LICENSE and NOTICE. The port wiring © 2026 Krishi Attri; the acceleration design © the Fast-TRELLIS authors; the base model © Microsoft.

Krishi Attri · krishiattriwork@gmail.com · github.com/Archerkattri

@misc{attri2026fasttrellis2,
  title  = {fast-trellis2: Fast-TRELLIS acceleration ported to TRELLIS.2},
  author = {Krishi Attri}, year = {2026},
  howpublished = {\url{https://github.com/Archerkattri/fast-trellis2}}
}
@article{trellis2,
  title   = {Native and Compact Structured Latents for 3D Generation},
  author  = {Microsoft TRELLIS.2 Team},
  journal = {arXiv preprint arXiv:2512.14692}, year = {2025}
}
@misc{fasttrellis,
  title  = {Fast-TRELLIS}, author = {wlfeng0509},
  howpublished = {\url{https://github.com/wlfeng0509/Fast-SAM3D/tree/Fast-TRELLIS}}
}

Family

Part of the HiCache++ acceleration family.

• Family hub: hicache-plus-plus — the basis library behind this adapter.
• Siblings (same base model, our HiCache-family accelerators): hermit-trellis2 — HiCache (scaled-Hermite) — and hermit-trellis2-plus-plus — HiCache++ (Dynamic Mode Decomposition / Prony).