The development

What happened

Hugging Face published a guide on 12 November for adapting TRELLIS.2, an image-to-3D AI model, to footwear geometry. It trains low-rank adapters in the model's structure and shape stages while keeping the original weights and texture stage frozen.

Key facts

  • The experiment used 188 footwear objects from Google Scanned Objects: 179 for training and nine held out for evaluation.

  • The adapters contain about 9.3 million trainable parameters. Texture training was excluded because the dataset lacked separate material maps and included baked lighting.

  • On the nine held-out objects, the authors report mean F-score rising from 0.276 to 0.803, using a matching threshold of 2% of each scan's bounding-box diagonal.

  • Scripts and containers support Hugging Face Jobs or local Docker. The reported structure and shape training runs took 7.2 and 3.6 hours respectively, each on one A100 80 GB GPU.

The context

Why it matters

The workflow illustrates targeted adaptation of generative AI rather than retraining an entire model. The results concern one small footwear test set, not general 3D quality or independently reproduced performance; the guide is not a new TRELLIS.2 model launch.