Hugging Face publishes TRELLIS.2 fine-tuning guide

Researchers demonstrate how to fine-tune the TRELLIS.2 3D generation model for specific object categories using Low-Rank Adaptation (LoRA) on the Google Scanned Objects dataset. By freezing the texture stage and applying LoRA adapters exclusively to the sparse-structure and shape diffusion transformers, the authors significantly improved footwear geometry accuracy on held-out evaluation objects. Open article →