mickeyvanolst/trellis-text-coreml

TRELLIS Text — Core ML bundle for TD Apple ML

Converted model weights for Geo Gen's native prompt-to-mesh mode on Apple silicon and macOS 14 or newer. The bundle contains 40 Core ML packages, the CLIP tokenizer, licences, and a SHA-256 inventory (bundle.json). Download size is approximately 3.28 GB. No Python inference runtime is needed by the TouchDesigner operator.

The complete folder is required: text encoder, coarse shape flow/decoder, sparse detail flow, mesh decoder, and a separate Gaussian appearance decoder. The native helper performs sampling, sparse routing, FlexiCubes extraction, fragment removal, mesh simplification, appearance rendering, and UV baking. The model packages alone are not a standalone generation application.

Use

In an AML release supporting this bundle, choose Geo Gen → Input → Prompt, then Manage Models → TRELLIS Text → Download. Alternatively, download this repository without changing its directory layout and set Prompt Model Folder to that folder. The image-only TripoSR/SF3D downloads do not supply these models.

Start with 8 Shape Steps, 8 Detail Steps, Mesh Detail Standard, and a 1024² texture. A complete a mouse generation at seed 42 took 112.5 seconds inside TouchDesigner on an Apple M4 with 24 GB memory, with models already loaded. First-time Core ML compilation takes longer. This is a development-machine measurement, not a speed guarantee for other prompts or hardware. Increasing steps, mesh detail, and texture resolution trades time and memory for quality.

Both mesh-detail settings extract the full learned 256³ surface; Standard targets at most 60,000 triangles and a quarter of the original count, with a geometric error limit that can retain more triangles when necessary. Minimum Piece Size defaults to 0.1% of the largest connected piece's triangle count. Set it to 0 when small detached details are intentional. Generation is approximate: meshes may contain openings or other defects, and are not guaranteed printable or watertight. Colours use the appearance decoder, not the mesh head's auxiliary colour channels. Albedo is sampled per texel from rendered appearance views. The normal output is the geometric surface normal.

Provenance and compatibility

No fine-tuning or substitute model weights were used. Source revisions:

• TRELLIS text flow: microsoft/TRELLIS-text-large, 4aad9f4a110329a410974d7f41ce5333a9a1fc87.
• Shape, mesh and appearance decoders: JeffreyXiang/TRELLIS-image-large, 25e0d31ffbebe4b5a97464dd851910efc3002d96.
• CLIP text encoder/tokenizer: openai/clip-vit-large-patch14, 32bd64288804d66eefd0ccbe215aa642df71cc41.

Conversion recipes are in labs/trellis-text of TD Apple ML. Converted with coremltools 9.0, using mixed source precision and explicit float32 attention where required. The helper selects tested CPU/GPU compute paths per package; changing those paths can produce incorrect results. Native random seeds are repeatable within this implementation but do not reproduce NumPy's noise stream. Full equivalence with the upstream CUDA pipeline has not been established.

The current helper supports at most 32,768 occupied detail cells and 8,192 pooled attention tokens. It reports an error rather than truncating larger results; use another seed or the coarse preview in that case.

Licences

TRELLIS weights and adapted inference code: Microsoft Corporation, MIT (TRELLIS-LICENSE.txt). CLIP: OpenAI, MIT (CLIP-LICENSE.txt). The conversion does not change the upstream model licences. Native geometry libraries are distributed and attributed separately with the AML plugin.