Hugging Face released a technical guide on its blog to help developers fine-tune the TRELLIS.2 3D generation model. The post includes Docker containers and scripts for running the process on Hugging Face Jobs or locally. The tutorial explains how to freeze the texture flow DiT while training adapters on the sparse-structure and shape flow DiTs.
The team validated the method using a subset of 188 footwear objects from the Google Scanned Objects dataset. The full preprocessing and training run took about one hour on an A100-large job. The adapters are trained using Hugging Face’s PEFT library with specific rank settings for the different layers.
This documentation lowers the barrier for users to customise 3D assets for specific applications. The quick training time on standard hardware makes the approach practical for iterative workflows. Developers can now create specialised 3D models without building complex pipelines from scratch.
