4B-scale open image-to-3D
Official TRELLIS.2-4B weights are published for high-fidelity image-to-3D generation (Hugging Face model card; MIT license).
Microsoft Research · Open image-to-3D
TRELLIS.2
Native structured latents for high-fidelity image-to-3D (Microsoft Research)
TRELLIS.2 is Microsoft’s open large 3D generative model for high-fidelity image-to-3D. It learns native, compact structured latents from 3D data so a single image can drive detailed geometry and PBR-style appearance—documented in the technical report “Native and Compact Structured Latents for 3D Generation” (arXiv:2512.14692).
What is TRELLIS.2?
TRELLIS.2 continues Microsoft Research’s structured-latent line of work (following TRELLIS / SLAT-style 3D generation). The official model card lists TRELLIS.2-4B as an image-to-3D model under the MIT license, with paper link arXiv:2512.14692.
Project materials describe TRELLIS.2 as a state-of-the-art open image-to-3D system: roughly 4 billion parameters, native 3D VAEs with strong spatial compression, and generation of high-resolution PBR-textured assets (project page cites outputs up to 1536³ resolution grids).
On this site you can open the official Hugging Face Space embed to try single-image generation without installing CUDA stacks locally. For production pipelines, combine the demo with the public weights and code on GitHub / Hugging Face.
When to use TRELLIS.2
Choose TRELLIS.2 when you want an open, research-backed baseline for high-fidelity image-conditioned 3D assets—especially if you care about structured latents, PBR-oriented appearance, and reproducible MIT-licensed weights rather than a closed API.
How TRELLIS.2 works
STEP 1
Condition on a single image
Provide a clear object-centric photo or render. The generative pipeline conditions 3D latents on image features (see paper for architecture details).
STEP 2
Decode structured 3D latents
Native 3D VAEs and structured latents decode into geometry and appearance suitable for textured asset export in the demo workflow.
STEP 3
Inspect and export
Use the Hugging Face Gradio UI to preview results and download outputs according to the Space’s available export options.
Key capabilities
Native structured latents
The paper focuses on learning structured latent representations from native 3D data, improving scalability versus loosely injecting 2D features alone.
High-resolution PBR assets
Project documentation highlights compact latents with high spatial compression and PBR-textured outputs at high grid resolutions (up to 1536³ on the project page).
Open demo + code path
Official Space: microsoft/TRELLIS.2. Code and models are linked from the Microsoft TRELLIS.2 project page and GitHub repository.
Technical snapshot
Figures below come from the linked papers, model cards, or project pages—not from unbenchmarked third-party marketing claims.
- Task
- Image-to-3D (pipeline tag on HF model card)
- Model
- TRELLIS.2-4B (microsoft/TRELLIS.2-4B)
- Paper
- arXiv:2512.14692
- License
- MIT (model card)
- Reported scale
- ~4B parameters; high-res PBR outputs (project page)
- Demo
- huggingface.co/spaces/microsoft/TRELLIS.2
Who it is for
Concept-to-3D for games & VFX
Turn concept art or product shots into reviewable 3D proxies before full DCC sculpting or retopo.
Research baseline
Compare against other open image-to-3D systems (Hunyuan3D, Direct3D-S2, Pixal3D, etc.) with a well-documented Microsoft stack.
PBR-oriented previews
When you need textured, material-aware previews from one image rather than geometry-only reconstruction.
Limitations & practical notes
- Third-party Hugging Face Spaces may queue, sleep, or hit Zero GPU limits; availability is not guaranteed.
- Generation quality depends on input framing, occlusion, and domain gap versus training data—validate outputs before production use.
- Resolution/time numbers quoted from project materials may vary by hardware and Space configuration.
Frequently asked questions
What is TRELLIS.2?+
TRELLIS.2 is an open image-to-3D generative model from Microsoft Research collaborators, described in “Native and Compact Structured Latents for 3D Generation” (arXiv:2512.14692). Weights are published as TRELLIS.2-4B under MIT on Hugging Face.
Is TRELLIS.2 free to try?+
Yes—the official Gradio demo runs on Hugging Face Spaces (microsoft/TRELLIS.2). Local inference requires following the GitHub install instructions and accepting hardware requirements.
How does it relate to TRELLIS / SLAT?+
TRELLIS introduced structured 3D latents (SLAT) for scalable generation. TRELLIS.2 is the follow-on work emphasizing native compact structured latents and higher-fidelity image-to-3D assets.
Sources & references
Landing claims are grounded in the following primary materials. Always prefer the original paper or model card for citations.
Try TRELLIS.2 in the browser
Jump into the playground and generate a 3D model from your own image—no install required.