TRELLIS.2 4B — image-to-3D with PBR materials

Microsoft's TRELLIS.2 structured-latent 3D generator. ≈100 s per asset on an RTX 5090, and unlike most shape models it returns material already baked in — no separate texturing pass.

What this repo is: the TRELLIS.2 4B pipeline at FP16 — weights only, not a retrain. The samples below were generated with this build.

Samples

Single image in, textured mesh out. Source images were rendered with Z-Image Turbo; the meshes are Blender Workbench renders of the raw .glb — no cleanup, no retopology, and no separate paint stage.


_{source — `a cute chunky retro robot toy standing upright, rounded metal body, simple friendly face, studio product photograph on a plain white background`}	_{mesh — 26.3 MB `.glb`, generated in 98 s, seed 42}

_{source — `an ornate ceramic teapot with a curved spout and handle, glazed blue and white porcelain, studio product photograph`}	_{mesh — 26.3 MB `.glb`, seed 42}

Turntable

Four views, 90° apart — the back is inferred, not copied from the source.

Note the level of preserved relief: the concentric ear knobs, the recessed chest display, and the fastener detail across the back panel all survive, and the surface reads as metal rather than flat clay because material comes out with the geometry.

Recommended settings

Parameter	Production value	Meaning
`seed`	any	Deterministic per seed
input	one RGB(A) image	Centred subject, plain background
output	`.glb`	Mesh with baked material

Supported modes: img2mesh, txt2mesh, and the _textured variants

Notes and gotchas

Material comes for free. img2mesh already returns a shaded mesh, so the separate paint stage other shape models need is unnecessary here. That is the main reason to pick TRELLIS.2 over a pure geometry model.
It is the slow tier — ≈100 s versus ≈20–30 s for a 1.5B geometry-only model and ≈6 s for a 0.6B one. Use it for finals, not for iteration.
Sparse-conv backend. The pipeline logs [SPARSE] Conv backend: flex_gemm; Attention backend: flash_attn on load and pulls a valeoai/NAF estimator from torch.hub the first time — the first run on a fresh machine needs network access and is slower.
Give the GPU to one model at a time. At ≈4B plus sparse-conv working memory this wants the card to itself; sharing a GPU with another model-serving process produces failures that look like crashes rather than clean out-of-memory errors.
Generated meshes come out facing the source image's front, so a fixed camera works across a batch.

Provenance

Upstream base model: microsoft/TRELLIS family, TRELLIS.2 4B
This build: FP16 redistribution of the TRELLIS.2 pipeline. Weights are not retrained here.
License: MIT, per upstream.

Downloads last month: 380

Inference Providers NEW

Image-to-3D

This model isn't deployed by any Inference Provider. 🙋 Ask for provider support


_{source — `a cute chunky retro robot toy standing upright, rounded metal body, simple friendly face, studio product photograph on a plain white background`}	_{mesh — 26.3 MB `.glb`, generated in 98 s, seed 42}

_{source — `an ornate ceramic teapot with a curved spout and handle, glazed blue and white porcelain, studio product photograph`}	_{mesh — 26.3 MB `.glb`, seed 42}