Nvidia and Hugging Face let developers fine-tune video and image models at scale without rewriting code

10:00 AM · July 18, 2026

Nvidia and Hugging Face released a collaboration bringing production grade, distributed diffusion model training to any Diffusers format model on the Hugging Face Hub, according to a joint blog post. The integration, built around Nvidia's NeMo Automodel library, lets developers fine-tune pretrained image and video generation models such as Qwen-Image, FLUX, Wan2.1, Wan2.2, and Hunyuan directly, with pretrained weights from the Hub working out of the box and fine-tuned checkpoints loading straight back into a standard inference pipeline or back to the Hub for sharing, without any manual checkpoint conversion step. Developers can switch between different parallelism strategies, including fully sharded data parallel, tensor parallel, expert parallel, context parallel, and pipeline parallel, as a configuration choice rather than a code rewrite, and the system supports both full fine-tuning and more efficient LoRA style parameter efficient fine-tuning depending on whether a project needs maximum quality or maximum efficiency. The companies say that when a new diffusion model architecture lands in the Diffusers library, enabling it in NeMo Automodel now requires only a small, contained code addition, a data preprocessing handler and a model adapter, rather than a full custom training script written from scratch. The release lowers a real practical barrier for smaller teams that want to customize open image and video models on their own data without building distributed training infrastructure themselves.