trellis2-mlx — consolidated weights for the TRELLIS.2 Apple-Silicon port

Consolidated weights for xocialize/mlx-trellis2-swift — a verified Swift-MLX re-port of Microsoft TRELLIS.2 (single image → textured 3D mesh) that runs on Apple Silicon (no CUDA). This repo bundles, as one download, the exact model files that pipeline loads.

Built with DINOv3.

⚠️ This is a pipeline artifact, not a model mirror

This repository exists only to package the weights the mlx-trellis2-swift pipeline needs into a single, engine-materializable download. The files are byte-for-byte upstream weights, renamed only by filename to the pipeline's component layout — no weights are modified.

It is not a general-purpose model distribution, and not a mirror of DINOv3. The DINOv3 image conditioner (dino.safetensors) is redistributed here solely because the TRELLIS.2 pipeline requires it, under the DINOv3 License.

Want DINOv3 on its own? Please get it from Meta's official, access-controlled repository and accept Meta's terms — do not use this repo to obtain standalone DINOv3 or to bypass Meta's gating:

https://huggingface.co/facebook/dinov3-vitl16-pretrain-lvd1689m

Contents

File	Component	Upstream
dino.safetensors	DINOv3 ViT-L/16 image conditioner	facebook/dinov3-vitl16-pretrain-lvd1689m
struct_flow.safetensors	sparse-structure DiT	microsoft/TRELLIS.2-4B
struct_dec.safetensors	sparse-structure decoder	microsoft/TRELLIS-image-large
shape_flow_512.safetensors · shape_flow_1024.safetensors	shape SLat DiT (512 / 1024 tiers)	microsoft/TRELLIS.2-4B
shape_dec.safetensors	shape FlexiDualGrid VAE decoder	microsoft/TRELLIS.2-4B
tex_flow_512.safetensors · tex_flow_1024.safetensors	texture SLat DiT (512 / 1024 tiers)	microsoft/TRELLIS.2-4B
tex_dec.safetensors	texture VAE decoder	microsoft/TRELLIS.2-4B
normalization.json	per-channel SLat mean/std	microsoft/TRELLIS.2-4B pipeline.json

Internal tensor keys are the original upstream keys (unmodified); only the file names follow the pipeline's component layout.

Usage

These weights are consumed automatically by the mlx-trellis2-swift MLXEngine imageTo3D package — first-run materialization downloads this snapshot into the engine's model store. You do not normally fetch it by hand.

Licenses & attribution

• DINOv3 weights (dino.safetensors) — DINOv3 License (Meta Platforms, Inc.), the most-restrictive component and the governing license for shipping the conditioner. Any product distributing it must display "Built with DINOv3", include the DINOv3 License, and comply with the DINOv3 Acceptable Use Policy.
• TRELLIS.2 / TRELLIS-image-large weights — MIT (Microsoft).
• Packaging metadata (this card, NOTICE, normalization.json, layout) — MIT (Xocialize), see LICENSE.

See NOTICE for the full attribution.