AI multimodale

TRELLIS.2: image-to-3D assets with materials

microsoft/TRELLIS.2

Microsoft’s 3D generation project with image-to-3D inference, PBR texturing and GLB export for GPU-equipped asset workflows.

★ 11,2KStelle
⑂ 1,4KFork
143Problemi aperti
PythonLingua
MITLicenza
Q84Punteggio editoriale

Panoramica

TRELLIS.2 uses structured 3D latent representations to generate textured shapes. The repository includes the TRELLIS.2-4B checkpoint, an image-to-3D example and a separate path for texturing existing shapes. It can turn a concept image into an initial asset for inspection in a 3D editor. Game prototypes, product visualizations and research workflows should evaluate the exported geometry and materials against the target renderer’s requirements.

Funzionalità principali

  • Generates a 3D shape from a reference image.
  • Uses a compact structured 3D representation.
  • Targets objects with complex topology.
  • Generates PBR materials and exports GLB assets.
  • Includes a texturing example for existing shapes.
  • Includes local web demos and training code.

Requisiti, installazione e avvio rapido

1. Prepare Linux and an NVIDIA GPU with at least 24GB VRAM, plus the required CUDA Toolkit.
2. Run git clone -b main https://github.com/microsoft/TRELLIS.2.git --recursive.
3. Use setup.sh as documented to install the environment and required extensions.
4. Load Trellis2ImageTo3DPipeline from microsoft/TRELLIS.2-4B.
5. Follow example.py to generate a mesh from an image and export it as GLB.
6. Run python app.py for the image-to-3D interface; use example_texturing.py or app_texturing.py for texture workflows.

Utilizzo

For a prop prototype, start with a complete subject against a simple background. Inspect both the turntable preview and the exported GLB in Blender or the destination engine. Review inferred back surfaces, small structures, material channels, polygon count and scale. Add mesh reduction and collision geometry when preparing real-time assets, retaining the input and generation settings for comparison.

Implementation notes
A single image does not reveal every surface, so inspect occluded geometry. Exported GLB files default to OPAQUE mode; assets containing alpha data may require material adjustments in the destination software.

Compatibilità del modello e casi d'uso

Upstream currently documents testing on Linux and recommends CUDA 12.4. Some extensions need compilation. The stated 24GB hardware requirement does not eliminate memory differences between resolutions and tasks.

Note su licenza e rischi

The model and code use MIT. Dependencies including nvdiffrast and nvdiffrec have separate license terms.

UniRL

Tencent-Hunyuan/UniRL

★ 854Python

diffusers

huggingface/diffusers

★ 34,5KPython

GLM-V

zai-org/GLM-V

★ 2,4KPython