TRELLIS.2 Shoe LoRA Fine-Tuning Guide, F-score 0.803

·2026.11.16 12:30

Key point

Applying LoRA to the structure and shape stages of TRELLIS.2 improved the F-score to 0.803.

1 / 9

Details

  • Applied LoRA only to the structure and shape stages of TRELLIS.2, significantly improving shoe generation quality
  • F-score improved from 0.276 to 0.803, and Chamfer distance improved from 0.174 to 0.033, based on 9 evaluation shoes
  • Kept the material stage fixed and trained with structure Rank 8 and shape Rank 16
  • Training completed in approximately 11.8 hours at a cost of approximately $30 using a single A100 80GB

Training and Execution Environment

  • Extracted features using DINOv3 ViT-L/16, then sequentially executed three 1.3B parameter DiTs
  • Structure and shape DiTs were trained independently, using PEFT's inject_adapter_in_model
  • Can be run locally on an NVIDIA GPU with 24GB or more; DINOv3 access permission is required

License and Data

  • Code and model weights are under the MIT license, and data is under CC BY 4.0
  • DINOv3 is subject to Meta's separate license, so terms must be verified before service application

This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.

Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.