Hugging Face has published scripts and containers that let developers fine-tune TRELLIS.2, its 3D generation model, on their own 3D assets. The guide, posted November 12, says the scripts run with a few arguments once a dataset is uploaded to the Hugging Face Hub, and they can run on local hardware or on Hugging Face Jobs, its cloud container service.
To test the approach, the authors used a subset of 188 footwear objects from the Google Scanned Objects dataset. In initial tests, the pretrained model usually produced shoes that looked correct, but the soles were wrong. The team trains LoRA (low-rank adaptation) adapters, which are small trainable matrices added to a model while its original weights stay frozen, using Hugging Face's PEFT library. TRELLIS.2 runs in five stages: background removal, image encoding with DINOv3-L, coarse structure prediction, detailed geometry generation and material generation. The authors add adapters to the sparse-structure and shape-generation stages, which control overall form and geometry, and leave the material stage unchanged because their dataset lacks the separate material maps that stage needs. The excerpt reviewed does not report measured accuracy gains.
Users need a Hugging Face token and approved access to DINOv3, which is a gated model, before they can run the pipeline. For Australian businesses and professionals, this changes little today unless they produce 3D assets for product visualisation, robotics simulation or games. Cloud compute costs on Hugging Face Jobs are not priced in the article, so check current rates before running training. Check the licence terms of any scanned-object dataset and the DINOv3 access conditions before training on them. Privacy Act 1988 obligations would only arise if scans capture identifiable people, which object-only datasets are unlikely to do. It is worth a trial only for a technical team with a narrow geometry problem; otherwise no action is needed.
Key Points
Hugging Face released scripts and containers to LoRA fine-tune TRELLIS.2 on custom 3D data, runnable locally or on Hugging Face Jobs.
The test case used 188 footwear objects from Google Scanned Objects; the pretrained model produced correct-looking shoes but inaccurate soles.
LoRA adapters are added to the sparse-structure and shape-generation stages, while the material stage stays frozen.
Running the pipeline requires a Hugging Face token, approved access to the gated DINOv3 model, and custom datasets uploaded to the Hub first.