Back to AI game news
AI game news

Hugging Face releases TRELLIS.2 LoRA training code to tailor 3D game props

By GameSlash

New November 12 code and containers let creators fine-tune TRELLIS.2 geometry on their own 3D meshes; a footwear test shows gains with clear hardware limits.

Hugging Face published a new TRELLIS.2 fine-tuning codebase and container workflow on November 12, 2026. Instead of accepting the same generic image-to-3D geometry each time, creators can train two compact LoRA adapters on a chosen family of objects and use them to improve the model's shape output. The workflow supports both local Docker and Hugging Face Jobs.

Official comparison of input shoes, base TRELLIS.2 meshes, alignment-adjusted base meshes, LoRA-fine-tuned meshes and reference scans for five held-out examples
Image: Hugging Face technical article — selected before/after geometry comparisons from the author's footwear experiment; not independent results or game screenshots.

What the released workflow actually does

The team provides preprocessing and training containers plus editable Python scripts. First, the pipeline imports 3D meshes and renders 16 reference views per item in Blender. It computes latents, then trains separate LoRA adapters for the sparse-structure flow and the shape flow. The texture and material generation stage stays frozen, because the experiment's dataset has only diffuse color textures without the separate physically based maps required for material training.

Hugging Face diagram showing the TRELLIS.2 image-to-3D pipeline from background removal and image encoding through structure, geometry, and texture to mesh export
Image: Hugging Face technical article — official pipeline diagram explaining which stages create structure, geometry and texture.

From a reference mesh collection to reusable adapters

Creators put their own 3D assets into a Hugging Face dataset, run the preprocessing image, then start separate sparse-structure and shape training jobs, either via HF Jobs or on a suitable GPU with Docker. The demonstration used 188 scanned footwear objects, of which 179 were used for training and nine were held out for evaluation. Access to the gated DINOv3 encoder, a token, GPU resources and correctly licensed training meshes is required; this is not a one-click mesh generator.

Official multi-angle reference renders of one footwear object produced by Blender for TRELLIS.2 LoRA conditioning
Image: Hugging Face technical article — Blender renders used as multi-view conditioning images during preprocessing.

Results: improvements on shoes, not every game asset

For nine held-out footwear scans, the authors report an average F-score@0.02 increase from 0.276 to 0.803, with Chamfer distance decreasing from 0.174 to 0.033. Two training stages used a single A100 80 GB each, taking approximately 7.2 and 3.6 hours. These are authors' results on a specific small benchmark; they do not establish performance on characters, weapons, props, or arbitrary fantasy styles.

For AI-assisted game teams, the technique may be useful when an art pipeline repeatedly needs one type of prop with consistent geometry. Test on held-out models and separately verify topology, UVs, materials, scale, polygon budget and engine import before calling a result game-ready.

Code, scripts and container instructions

Official technical announcement and result figures, November 12, 2026

View source