NVIDIA and Hugging Face launch NeMo Automodel for distributed diffusion fine-tuning

NVIDIA and Hugging Face have announced a collaboration bringing production-grade, distributed fine-tuning for diffusion models to the Diffusers ecosystem via the NeMo Automodel open-source library. This integration eliminates the need for checkpoint conversion—users can point pretrained_model_name_or_path to any Diffusers model ID on the Hugging Face Hub and start training immediately. NeMo Automodel supports flow-matching models and uses PyTorch DTensor-native parallelism, allowing configurations like FSDP2, tensor parallel, expert parallel, context parallel, and pipeline parallel without code rewrites. Supported models include FLUX.1-dev (12B), FLUX.2-dev (32B), Wan 2.1 T2V (1.3B/14B), Wan 2.2 T2V (27B MoE, 14B active), HunyuanVideo 1.5 (13B), and Qwen-Image (20B). Many offer LoRA recipes, enabling efficient fine-tuning on consumer GPUs (e.g., Wan 1.3B fits on a single 40GB A100).

The practical benefits are significant: pretrained weights work out of the box, fine-tuned checkpoints load directly into DiffusionPipeline for inference, and downstream tools like quantization, compilation, and custom samplers remain compatible. The integration is documented in the Diffusers training guide and is fully open-source under Apache 2.0. NVIDIA provides ready-to-use fine-tuning recipes, including pre-encoding datasets with VAE latent caching and multiresolution bucketed dataloading to accelerate throughput. This lowers barriers for enterprises and researchers to adapt state-of-the-art video and image models at scale, from a single GPU to hundreds, without losing access to the broader Hugging Face ecosystem.