From the source
Hugging Face published fine-tuning scripts and containers that let users adapt TRELLIS.2, a 3D generation model whose unified latent representation the company says avoids switching models per output format, to their own 3D assets.
- The guide answers a practical gap: a general-purpose model such as TRELLIS.2 can make errors in details that matter for a specific application, and neither TRELLIS nor TRELLIS.2 offers an easy way to adapt it to a task.
- Hugging Face tests low-rank adaptation on a subset of the Google Scanned Objects dataset with 188 footwear objects, where the pretrained model usually produced shoes that looked correct but had incorrect soles.
- LoRA trains small adapter matrices while keeping the original weights frozen, cutting trainable parameters and the memory needed for gradients and optimizer state, and the work uses Hugging Face's PEFT library to add and train them.
- The guide adds adapters only to the sparse-structure flow DiT and the shape flow DiT, leaving the material generation stage untouched, because its dataset has diffuse color maps but lacks the separate material maps that stage needs.
- The scripts can run locally with Docker or on Hugging Face Jobs, which requires an HF_TOKEN and Hugging Face account access to the gated DINOv3 model.





