TRELLIS.2
Microsoft's second-generation image-to-3D model, and the current open-weights benchmark. TRELLIS.2 turns a single image into a textured mesh with full PBR materials — and, unusually for a frontier 3D model, ships the weights and the training code under a permissive licence.
If you want the best open-weights image-to-3D you can actually download and self-host, TRELLIS.2 is it. The 4-billion-parameter model produces textured meshes with real PBR materials — base colour, roughness, metallic, opacity — from one image, and Microsoft released both the weights and the training pipeline under the MIT licence. The catch is hardware: the official requirement is a 24 GB NVIDIA GPU on Linux, so "runs locally" for prosumers currently means leaning on community quantised builds, not the reference code.
What it is
TRELLIS.2 is a single-image-to-3D generative model from Microsoft Research, published in December 2025 (paper: arXiv 2512.14692, "Native and Compact Structured Latents for 3D Generation"). You give it one image of an object and it returns a full 3D asset — a mesh in GLB format with textures and physically-based-rendering material channels, ready to drop into a renderer or a game engine.
The headline idea is its representation. Where the original generation of models predicted a mesh or a radiance field directly, TRELLIS.2 generates a structured latent it calls O-Voxel — an "omni-voxel" that encodes geometry and appearance together in one sparse grid, compressed by a sparse 3D VAE. Because it is not tied to an iso-surface field, it can represent things earlier models struggled with: open surfaces, non-manifold geometry and internal structure. A 4B flow-matching model then generates that latent from the image.
Where it wins
- Genuinely open. MIT-licensed, with inference code, the pretrained 4B weights and the full training pipeline (shape VAE, texture VAE, flow models) all released. That is rare at this quality tier.
- PBR-native output. It models base colour, roughness, metallic and opacity — so assets look right under real lighting and support transparency, rather than baking everything into a flat diffuse texture.
- Handles awkward geometry. The O-Voxel representation copes with open surfaces, non-manifold parts and internal structure that field-based methods approximate poorly.
- Fast at low resolution. On an H100, a 512³ generation takes around 3 seconds; higher-fidelity 1024³ is around 17 seconds and 1536³ around 60 seconds — a clear quality/time dial.
Where it still hurts
- Hardware bar is high. The official requirement is an NVIDIA GPU with at least 24 GB of memory, on Linux, verified on A100 and H100. The reference code is not a laptop tool.
- "Runs locally" needs the community. Prosumers reach it through third-party ComfyUI wrappers with GGUF quantisation, which can fit a Q4 build in roughly 8 GB — but quantisation softens textures and these paths are unofficial, so treat them as community efforts, not supported ones. We found no official macOS build.
- Single object, single image. Like its predecessors it expects one clean subject; scenes, occlusion and busy backgrounds still degrade results.
- Print-readiness varies. Generative meshes usually want a repair and decimation pass before slicing.
The AI angle
TRELLIS.2 is the AI, and it marks a fork in the road for this category. The best 3D generation is now splitting into two camps: open weights you can download and self-host — TRELLIS.2 (MIT) and Hunyuan3D-2.1 (open weights + PBR) — versus open access, where the strongest topology work is only reachable through a hosted API. Tencent's art-grade Hunyuan3D-PolyGen, for instance, is hosted-only. That distinction — open weights vs open access — is the live story in image-to-3D right now, and it is exactly why an open, MIT-licensed frontier model like TRELLIS.2 matters: you can run it, fine-tune it, audit it and build on it. We line TRELLIS.2 up against Hunyuan3D-2.1, SAM 3D and TripoSR in our Image-to-3D in 2026 comparison. On 3d.2nth.ai it also serves as an eval subject — a strong open model to benchmark quality, latency and cost against faster baselines like TripoSR.
Start here
- Model card and weights: huggingface.co/microsoft/TRELLIS.2-4B.
- Code and training pipeline: github.com/microsoft/TRELLIS.2 — read the hardware notes first (24 GB NVIDIA, Linux).
- Project page with results and the O-Voxel explainer: microsoft.github.io/TRELLIS.2.
- On a consumer GPU? Start from a community ComfyUI + GGUF build and expect a quality trade-off.