Skip to content
Main Site News Console

The Ultimate Guide to Fine-Tuning 3D Generative Models

· Hugging Face Translated
教程模型卡

Published on November 12, 2026

Image 8: Suvaditya Mukherjee's avatar Suvaditya Mukherjee suvadityamuk

Image 9: merve's avatar merve merve

Image 10: Pedro Cuenca's avatar Pedro Cuenca pcuenq

Image 11: Hero GIF

TLDR: In this blog post, we provide scripts and containers to fine-tune the popular 3D model TRELLIS.2 both locally and on Hugging Face Jobs using your own 3D assets. If you’re interested in fine-tuning, you can use our script with just a few arguments (assuming you’ve uploaded your dataset to the Hugging Face Hub), or tinker around with this repository. If you’re curious to learn more, we also break down how the script works. Let’s dive in!

Introduction

3D generation has wide-ranging applications in robotics, simulation, media, and entertainment. Across these fields, it is difficult to make generated objects adhere to strict rules and to standardize representations for models and data. Many file formats exist, such as GLB and PLY. There are also multiple representations, such as Neural Radiance Fields (NeRFs) and 3D Gaussian Splatting (3DGS). Each comes with its own trade-offs.

TRELLIS introduced a versatile internal representation called a unified latent representation. Its decoders can convert this representation into 3D models in the desired output format. As a result, you don’t need to use a separate model for each representation format.

Later, the same team released TRELLIS.2, improving how the model represents 3D structures. However, neither model offered an easy way to adapt to specific tasks.

What we want to improve

A general-purpose model like TRELLIS.2 can get details wrong that are critical for a specific application. For example, a