Fine-tuning Diffusion Models at Scale with NeMo Automodel

This article presents the collaboration between NVIDIA and Hugging Face, introducing NVIDIA NeMo Automodel as a solution for distributed and large-scale fine-tuning of diffusion models. NeMo Automodel is an open-source PyTorch native DTensor library, designed to seamlessly integrate with the Hugging Face Diffusers ecosystem.

The solution allows users to fine-tune image and video diffusion models (such as FLUX.1-dev, Wan 2.1, HunyuanVideo) without checkpoint conversion or code rewriting, offering a consistent interface for inference and adaptation. It handles technical challenges such as memory-efficient sharding, latent caching, and multi-resolution bucketing, and scales transparently from a single GPU to hundreds. The integration supports full fine-tuning and parameter-efficient fine-tuning (LoRA), with demonstrated performance on H100 GPUs.