The Definitive Fine-Tuning Guide for 3D Generation Models
Web page · https://huggingface.co/blog/3d-lora-finetuning
First seen by Oriel on .
AI summary
Hugging Face provides scripts and Docker containers for LoRA fine-tuning TRELLIS.2 on custom 3D assets, runnable locally or through Hugging Face Jobs.
- The example uses 188 GSO footwear objects to improve shoe geometry, particularly soles, while holding out 9 objects for evaluation.
- LoRA adapters target the sparse-structure flow DiT and shape flow DiT, while the texture flow DiT remains frozen because the dataset lacks separate material maps.
- The configuration trains approximately 9.3 million adapter parameters, using rank 8 for sparse structure and rank 16 for shape generation.
- Preprocessing renders 16 Blender views per object and saves TRELLIS.2 training latents; 188 objects took approximately one hour on an a100-large job.
- Separate preprocessing and training containers support Hugging Face Jobs or local Docker, with checkpoints, datasets, and trained adapters uploadable to the Hugging Face Hub.
#trellis.2 #lora #3d generation #hugging face jobs
Generated by gpt-5.6-sol. Check the original for details.
Aliases
| Source | External ID | URL |
|---|---|---|
| Hugging Face Blog | https://huggingface.co/blog/3d-lora-finetuning | https://huggingface.co/blog/3d-lora-finetuning Read discussion |