The Definitive Fine-Tuning Guide for 3D Generation Models
, updated First seen Updated Limited coverage, tapering
Hugging Face · ModelsGOOGL$265.29▼0.9%
Reported by Hugging Face1 source ↓
Key points
- The example fine-tunes TRELLIS.2 using 188 footwear objects from Google Scanned Objects
- LoRA adapters are added to the structure and shape models, while the texture model is frozen
- Hugging Face provides scripts for local training and Hugging Face Jobs
Details
Hugging Face reports that the example addresses a shortcoming in the pretrained model: shoes generally looked correct, but their soles were inaccurate. The guide focuses on adapting TRELLIS.2 to that footwear category using the Google Scanned Objects dataset.
The training setup uses LoRA adapters on the sparse-structure and shape flow models, which control the object's structure and geometry. Hugging Face leaves the texture model unchanged because the dataset has diffuse color maps rather than the separate material maps needed to train that component.
Written from coverage as of
Images
- Hugging Face
Stocks in this story
Change since open: 20 Nov to 21 Nov 2026
Last refresh