Hugging Face published TRELLIS.2 fine-tuning scripts plus preprocessing and training containers for local GPUs or Jobs. Its example uses 188 footwear objects from GSO, training roughly 9.3 million LoRA parameters in the structure and shape DiTs while freezing material generation. Dataset color maps contain baked lighting, making them unsuitable material targets that could teach incorrect surface properties or erode existing capabilities. Prerequisites include DINOv3 access. This asset-specific geometry example is not a guarantee of improved quality across all 3D tasks.
Back to Tech
Hugging Face publishes tools for shape-focused TRELLIS.2 fine-tuning
Freeze material generation when the dataset lacks valid material targets, and train only the structure and shape components needed.