Hugging Face Releases Fine-Tuning Guide for TRELLIS.2

The model uses a unified latent representation to support diverse output formats. To improve accuracy on specific tasks like footwear geometry, LoRA adapters are applied to key stages—sparse structure and shape generation—while keeping other components frozen. Training leverages the GSO dataset and requires access to DINOv3 via Hugging Face account.

Summarised from the original article by an open-source model, then checked automatically: every number and English name must appear in the source. The original article is authoritative.