TRELLIS.2-4B
Mirror of Microsoft's TRELLIS.2 4B image-to-3D model, which makes textured meshes with PBR materials from one image.
What it is for
This repo is a mirror of TRELLIS.2-4B, a flow-matching transformer with a sparse-voxel 3D VAE. It uses the O-Voxel representation to make 3D assets with arbitrary topology and PBR materials from one image. It holds the image-to-3D and texturing pipeline configs and the checkpoints with their configs. The checkpoints are the sparse-structure flow model, the image-to-shape and shape-to-texture flow models, and the shape and texture encoders and decoders. Use it with the TRELLIS.2 code for image-to-3D and for texturing an input mesh.
Provenance
- Upstream model: microsoft/TRELLIS.2-4B.
- Paper: Native and Compact Structured Latents for 3D Generation.
- Upstream code: microsoft/TRELLIS.2. Project page: microsoft.github.io/TRELLIS.2.
Use
from trellis2.pipelines import Trellis2ImageTo3DPipeline
pipeline = Trellis2ImageTo3DPipeline.from_pretrained("chibifire/TRELLIS.2-4B")
pipeline.cuda()
For install steps, rendering and GLB export, read the upstream model card and the TRELLIS.2 repo.
Licence
MIT, the upstream licence. Terms: https://opensource.org/license/mit.