RELEASE · MODELS · #3483
Hugging Face releases fine-tuning scripts and containers for TRELLIS.2 3D model
Hugging Face published a blog and repository that provide scripts and containerized workflows to fine-tune TRELLIS.2 on custom 3D assets, runnable locally or on Hugging Face Jobs. The guide explains which pipeline components to freeze, how to add LoRA adapters via PEFT, and demonstrates a GSO footwear experiment to improve geometry.
KEY POINTS
- Hugging Face published a blog and repository that provide scripts and containerized workflows to fine-tune TRELLIS.2 on custom 3D assets, runnable locally or on Hugging Face Jobs.
- The guide explains which pipeline components to freeze, how to add LoRA adapters via PEFT, and demonstrates a GSO footwear experiment to improve geometry.
- Provides practical, reproducible tooling and guidance to adapt a strong 3D generative model (TRELLIS.2) to domain-specific data using parameter-efficient LoRA adapters and cloud/local workflows, which can improve geometry/material fidelity for applied use cases.
WHY IT MATTERS
Provides practical, reproducible tooling and guidance to adapt a strong 3D generative model (TRELLIS.2) to domain-specific data using parameter-efficient LoRA adapters and cloud/local workflows, which can improve geometry/material fidelity for applied use cases.