Guide details LoRA fine-tuning for TRELLIS.2 3D generation
A new guide describes scripts and Docker containers for adapting TRELLIS.2, a 3D generation model, to specialized asset collections. The workflow can run on local hardware or through Hugging Face Jobs, provided the dataset’s meshes are uploaded to the Hugging Face Hub. The guide focuses on improving shoe geometry using 188 footwear objects selected from the 1,030-object Google Scanned Objects dataset.
The experiment uses LoRA adapters, which train small additional matrices while leaving the original model weights frozen. Adapters are added to TRELLIS.2’s sparse-structure flow DiT and shape flow DiT because these components influence coarse structure and detailed geometry. The texture flow DiT remains unchanged: the dataset contains diffuse color maps with baked lighting rather than the separate material maps needed for that stage. The configuration trains about 9.3 million adapter parameters, using rank 8 and rank 16 adapters in specified layers.
Two containers handle the process. A preprocessing container uses Blender to render 16 views per object and convert GLB or OBJ meshes into TRELLIS.2 representations. A training container then runs separate LoRA jobs for the two DiTs and can upload checkpoints to the Hub. The guide also explains configuration, monitoring with Trackio, and the model’s O-Voxel representation.
Written by Sakhanda Wire AI from reporting by Hugging Face. Facts, quotes and full context are in the original article below.