2026-11-12 · ← News
LoRA fixed shoe soles across 188 objects while leaving textures alone
Hugging Face has released containers and scripts for fine-tuning TRELLIS.2 on custom 3D assets. An experiment with 188 footwear models raised F-score from 0.276 to 0.803, but the broader lesson is restraint: train only the components for which the dataset provides the right signal.
The image could not be loaded.
Hugging Face has published a reproducible workflow for adapting TRELLIS.2 to a specific category of 3D objects with LoRA. The scripts and two Docker containers can run locally or through Hugging Face Jobs.
Two LoRAs change geometry while the texture stage stays frozen
TRELLIS.2 turns an input image into a textured mesh through several stages. The authors inserted LoRA adapters only into the coarse-structure model and the detailed-shape model. They froze the material and texture stage because their data contained diffuse color maps with baked lighting, not separate material maps.
That choice matters more than the recipe itself. Training the texture stage on unsuitable targets could teach the model to treat lighting as a surface property and damage capabilities the base model already handled well.
A narrow dataset raised F-score from 0.276 to 0.803
The experiment used 188 footwear objects from Google Scanned Objects. Nine were held out for evaluation and 179 were used for training. Together, the two LoRAs contained about 9.3 million trainable parameters.
Across the nine held-out objects, mean F-score rose from 0.276 to 0.803. Chamfer distance fell from 0.174 to 0.033 and normal consistency increased from 0.530 to 0.804. The adapters improved F-score on all nine test objects, with the visual examples highlighting better sole geometry.
Five examples and nine held-out objects do not establish a universal recipe
The result is persuasive for this category, not for 3D generation as a whole. The authors show visual comparisons for five of the nine held-out objects. Training loss alone does not measure geometry quality, and a footwear dataset is too narrow to support claims about furniture, robots or scenes.
The compute is also material. Preprocessing 188 objects took about 1 hour on an A100 with 80 GB. Training the coarse-structure adapter took 7.2 hours and the shape adapter 3.6 hours, each on one A100 with 80 GB.
Other categories will show whether the method travels beyond footwear
The next useful test is to repeat the same data split, metrics and evaluation discipline on a category with different geometric demands. Other 3D models may use different representations and require a different choice of blocks to freeze.
For production teams, the workflow is valuable as a decision map: name the defect, identify the pipeline component that creates it and train that component. Full fine-tuning would have been more expensive and harder to control here.
Lilith's verdict
The authors did not send the whole 3D model back to school. They took 188 shoes, pointed at the bad soles and retrained only the two components that actually draw them.