Open-weight 3D generation and fine-tuning

Hugging Face publishes a practical TRELLIS.2 LoRA workflow for 3D models

An open training recipe adds local and hosted fine-tuning paths for TRELLIS.2 geometry, while gated DINOv3 access and substantial GPU needs still matter.

Open Model Weights published13 Nov 2026
Primary sourceHugging Face — The Definitive Fine-Tuning Guide for 3D Generation Models
Source published2026-11-12
FOLLOW OUR COVERAGE

Follow Open Model Weights on Google See more of our open-weight AI reporting in Google Search.

Add as a Preferred Source on Google
Original glossy editorial visualization of a translucent silver 3D sneaker emerging from triangulated mesh layers with subtle mint lights, illustrating TRELLIS.2 LoRA.
Original high-gloss conceptual 3D mesh artwork illustrating geometry adaptation; not an actual TRELLIS.2 output, measurement or reproduced training result. AI-generated editorial illustration; not a documentary photograph.

Hugging Face has published a concrete fine-tuning workflow for TRELLIS.2, the image-to-3D model used to generate meshes and materials from visual references. The November 12 guide includes training and preprocessing containers, documented command-line entry points and an implementation repository. The new contribution is not a new base model release. It is a way to adapt already published model components to specialised geometry, with both self-hosted Docker and Hugging Face Jobs routes described in the original technical material. Primary source ↗

The central engineering choice is selective adaptation. TRELLIS.2 first builds a coarse structure and then refines geometry before creating textures and materials. The authors use low-rank adapters on the sparse-structure and shape-flow diffusion transformers, while leaving the texture-flow model unchanged. Their example uses a subset of Google Scanned Objects with 188 footwear assets. Sixteen rendered views provide conditioning images for training; nine of the objects are held out for evaluation. This is a focused research recipe, not proof that every 3D asset class will benefit equally.

The guide explains preprocessing into model latents, adapter insertion with PEFT and separate runs for the structure and shape components. The published example includes local GPU Docker commands and cloud jobs using A100-class hardware. That detail is useful for teams planning repeatable adaptation, because the expensive encoding stage can be reused across runs. However, readers should not confuse a published script with an effortless laptop workflow: model dependencies, memory, data conversion and licensing remain operational constraints.

One access caveat is explicit. The workflow requires access to DINOv3, which Hugging Face describes as gated, even when the training containers themselves are available. Generated mesh quality and the footwear results should be read as the authors' experiment, not an independent Open Model Weights benchmark. The registry implication is to track the underlying weights, adapter artifacts and external dependencies separately, and to mark the workflow as an open training resource rather than falsely presenting it as a newly released foundation model.

OMW REGISTRY WATCH

What this changes for the evidence layer

A reproducible fine-tuning workflow is not a new foundation-model checkpoint. Inspect TRELLIS.2 source checkpoints, adapters, gated DINOv3 dependencies, code and dataset licenses separately. The guide's footwear evaluation is an author-reported demonstration, not a general-purpose geometry-quality ranking.

PRIMARY SOURCE

Hugging Face — The Definitive Fine-Tuning Guide for 3D Generation Models

This brief is based on the cited primary source. Performance, benchmark and comparative claims remain attributed unless Open Model Weights publishes an independent measurement.

Open source article ↗