Hugging Face has published scripts, containers and a worked example for fine-tuning TRELLIS.2, its 3D-generation model, on custom assets. The practical takeaway: a small LoRA adaptation can target the parts of the model that need to change, rather than retraining everything.
Hugging Face Watch analysis
What happened
The guide walks through adapting TRELLIS.2 to generate more accurate footwear. Its example uses 188 shoes from the Google Scanned Objects dataset, with LoRA adapters applied to the model’s sparse-structure and shape components while its texture component is left untouched.
The team reports training about 9.3 million adapter parameters in total. The guide includes Docker images and commands for running preprocessing and training either on a local machine or through Hugging Face Jobs. Preprocessing the example dataset took about an hour on an A100-large job.
Why it matters
Fine-tuning a 3D generator can sound like a job for a well-funded lab. This walkthrough makes the process more concrete: prepare mesh data, render 16 views of each object, then train adapters on the parts responsible for shape. LoRA keeps the original weights frozen and reduces the number of parameters being trained.
There are practical prerequisites. Users need a Hugging Face token and access to gated DINOv3; they also need their own mesh dataset uploaded to the Hub to follow the custom-data path. The footwear example is a demonstration, not evidence that the same setup will work equally well for every category.
Our read
This is a useful piece of engineering guidance, not just a model announcement. The best detail is the decision to leave the texture stage alone because the example data lacks the separate material maps it needs. That is the sort of small, specific choice that can save a fine-tuning run from learning the wrong thing. If you work with 3D assets, the Hugging Face guide is a practical place to start, provided its access and hardware requirements fit your setup.
What to watch
- Whether the scripts and containers work smoothly with datasets beyond the footwear example.
- How much the LoRA adapters improve geometry on held-out assets.
- Whether Hugging Face adds further model-specific fine-tuning guides.
Discussion spark: For adapting a 3D model to a specialised task, would you rather use a targeted LoRA like this or retrain more of the model for greater control?
Sources and evidence
- The Definitive Fine-Tuning Guide for 3D Generation Models (12 November 2026, 00:00 UTC)
not affiliated with or endorsed by Hugging Face