TRELLIS.2-4B for mlx-serve (8-bit)

microsoft/TRELLIS.2-4B repackaged for mlx-serve on Apple Silicon. Microsoft's image-to-3D model: an image with a transparent background becomes a textured PBR GLB.

The flow-model (DiT) block projections are affine-quantized to 8 bit, group size 64; everything else is as released.
Includes the DINOv3 image encoder it conditions on (float16); its licence is in dinov3/LICENSE.md.
No background remover is included: send an image with a transparent background (alpha is the subject).

Licences

The TRELLIS.2-4B weights are MIT (LICENSE). The bundled DINOv3 encoder is Meta's, under the DINOv3 License (dinov3/LICENSE.md), which travels with any redistribution of it.

Run: mlx-serve --model-dir <folder holding this repo>, then POST /v1/3d/generations. Converted by tests/convert_trellis2_pixal3d.py in the mlx-serve repository.

Downloads last month: -; Downloads are not tracked for this model. How to track

MLX

Hardware compatibility

Quantized

Inference Providers NEW

Image-to-3D

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

Model tree for ddalcu/TRELLIS.2-4B-MLX-Serve-8bit

Base model

microsoft/TRELLIS.2-4B

Quantized

(3)

this model