TRELLIS.2 — MLX pipeline weights (consolidated)
MLX-ready, consolidated pipeline weights for Pixal3D / Microsoft TRELLIS.2 single-image →
textured 3D mesh, for the Swift-MLX port
xocialize/mlx-trellis2-swift
(MLXEngine imageTo3D). One download replaces three upstream repos, with keys already in the
port's module layout so the Swift loader consumes them remap-free.
This is a pipeline artifact, not a model mirror. It includes a converted DINOv3 image conditioner (
dino.safetensors: MLX bf16, pipeline-specific normalization, port key layout) as one integral component of the TRELLIS.2 pipeline. If you are looking for DINOv3 itself, use Meta's official release —facebook/dinov3-vitl16-pretrain-lvd1689m— not this repo.
License — read before use
The DINOv3 conditioner is redistributed under the DINOv3 License (its
§1.b redistribution terms are satisfied by shipping the license and attribution with the weights —
see NOTICE). By downloading or using these weights you must comply with the DINOv3 License,
including its Acceptable Use restrictions and Trade Control terms, and derivatives must carry
"Built with DINOv3" attribution. All other components are Microsoft TRELLIS.2, MIT (LICENSE).
Consumers via MLXEngine get this surfaced automatically: the package's license gate presents the DINOv3 terms at registration, and shipping apps display the "Built with DINOv3" attribution.
Contents
| File | Component | License | dtype |
|---|---|---|---|
dino.safetensors |
converted DINOv3 ViT-L/16 image conditioner | DINOv3 License | bf16 |
struct_flow.safetensors |
sparse-structure flow DiT | MIT | bf16 |
struct_dec.safetensors |
sparse-structure decoder | MIT | fp16 |
shape_flow_512.safetensors / shape_flow_1024.safetensors |
shape SLat flow DiT (512 / 1024 tiers) | MIT | bf16 |
shape_dec.safetensors |
shape SLat decoder (FlexiDualGrid) | MIT | fp16 |
tex_flow_512.safetensors / tex_flow_1024.safetensors |
texture SLat flow DiT (512 / 1024 tiers) | MIT | bf16 |
tex_dec.safetensors |
texture SLat decoder (PBR) | MIT | fp16 |
normalization.json |
shape + texture SLat mean/std | — | — |
Total ≈ 15 GB. The _1024 flows are only needed for the res1024 / res1536 cascade tiers.
Tensor values are byte-identical to upstream; only keys/layout were adapted to the Swift module
namespace (the conditioner is not drop-in for generic DINOv3 loaders).
Usage
hf download xocialize/trellis2-mlx --local-dir trellis2-mlx
The Swift port loads every file directly (no key remap). Geometry + per-vertex texture color;
res512 / res1024 / res1536 resolution tiers.
Provenance
Derived from microsoft/TRELLIS.2-4B + microsoft/TRELLIS-image-large (MIT, see LICENSE) and
facebook/dinov3-vitl16-pretrain-lvd1689m (DINOv3 License, see NOTICE).
Port code is MIT.
Quantized