TRELLIS.2 MLX 8-bit Affine for mlx-spatial

This is an unofficial quantized derivative. It is not a Microsoft, Meta, or BRIA release. Because the bundle contains RMBG-2.0, it is restricted to non-commercial use unless the user obtains separate commercial rights from BRIA.

Variants

Compatibility

Use

Bundle

Quantization

• self-attention QKV and output projections in all five TRELLIS.2 flow DiTs;
• cross-attention query, key/value, and output projections in those DiTs;
• DiT MLP input and output matrices;
• internal ConvNeXt MLP matrices in the shape and texture encoders and decoders;
• DINOv3 attention Q/K/V/output and MLP up/down projections;
• RMBG-2.0 Swin attention QKV/output and MLP FC matrices.

• all dense and sparse convolutions, including the sparse-structure decoder;
• input, output, patch, latent, subdivision, and final prediction boundaries;
• timestep and condition projections, embeddings, learned tokens, and positional tensors;
• normalization parameters, RMSNorm parameters, modulation tensors, and biases;
• RMBG stage-transition reductions;
• matrices whose input dimension is incompatible with group size 64.

Reproducing the Bundle

Verification

• Logical checkpoint inspection matched all 5,587 source tensor names, shapes, and dtypes exactly.
• The quantization inventory contains 1,546 packed matrices: 1,050 DiT matrices, 256 encoder/decoder MLP matrices, 144 DINOv3 matrices, and 96 RMBG-2.0 matrices.
• The complete mlx-spatial repository suite passed 1,175 tests; 10 tests were skipped, 50 were deselected, and 3 expected failures remained.
• A 512-resolution, 12-step run completed background handling, DINOv3 conditioning, sparse-structure sampling, shape and texture SLat sampling, both decoders, and textured GLB export with no blocker.
• That run produced 4,022 shape tokens, 1,041,541 decoded shape/texture points, and a 12,083,344-byte GLB. The rendered asset retained coherent geometry and recognizable appearance.
• The same run completed in 152.23 seconds, observed 3.516 GB peak MLX allocator use, and recorded zero swap growth.
• A separate 1024_cascade, 12-step run completed the same end-to-end path and produced a Blender-readable 12,670,448-byte GLB with 199,884 faces and embedded 1024 x 1024 PBR textures.

Limitations

• Quantization changes the sampling trajectory. Geometry, pose, topology, texture placement, material values, and unseen surfaces can differ from the source-precision model even with the same seed.
• End-to-end validation covers 512 and 1024_cascade. The standalone 1024 and 1536_cascade routes have not received an equivalent quality evaluation.
• Single-view reconstruction cannot determine unseen geometry with certainty.
• Fine detail depends on foreground extraction, cropping, occlusion, reflections, transparency, and thin structures.
• Generated meshes can contain small holes, non-manifold regions, or other topology requiring downstream cleanup.
• The packed checkpoint format requires mlx-spatial; generic safetensors readers expose the physical packed arrays rather than the logical matrices.
• This bundle supports inference, not training or fine-tuning.
• Commercial use of the bundled RMBG-2.0 weights requires separate rights from BRIA.

License and Attribution

• Microsoft TRELLIS.2 files are provided under the MIT License. A copy is included as LICENSE_TRELLIS2.
• Meta DINOv3 files and derivatives are governed by the DINOv3 License. The complete agreement is included as LICENSE_DINOV3.md, as required for redistribution.
• BRIA RMBG-2.0 is provided for non-commercial use under CC BY-NC 4.0. Its attribution and change notice are included in LICENSE_RMBG2.md. Commercial use requires separate authorization from BRIA.

Links

• App Automaton
• AppAutomaton models on Hugging Face
• appautomaton/mlx-spatial — MLX-native 3D and spatial inference for Apple Silicon.
• mlx-spatial documentation
• mlx-spatial on PyPI
• TRELLIS.2 guide
• TRELLIS.2 MLX 8-bit model
• Microsoft TRELLIS.2-4B model
• Microsoft TRELLIS.2 source
• TRELLIS.2 project page
• Meta DINOv3 ViT-L/16 model
• Meta DINOv3 source
• BRIA RMBG-2.0 model
• BRIA RMBG-2.0 source
• MLX