Hugging Face publishes a TRELLIS.2 fine-tuning guide
Hugging Face authors Suvaditya Mukherjee and Pedro Cuenca published scripts and containers for adapting TRELLIS.2 to custom 3D assets, with workflows available through local Docker or Hugging Face Jobs. Their footwear experiment used LoRA adapters for the structure and shape stages while freezing the texture stage, using 188 Google Scanned Objects and nine held-out samples. The adapted model improved mean F-score from 0.276 to 0.803, reduced Chamfer distance from 0.174 to 0.033, and raised normal consistency from 0.530 to 0.804.
The release gives developers a concrete, reproducible workflow for task-specific TRELLIS.2 adaptation and documents measured geometry improvements from a limited footwear experiment.
#TRELLIS.2#Hugging Face#Hugging Face Jobs#PEFT#LoRA#Google Scanned Objects