NVIDIA and Hugging Face announced integration of NVIDIA’s NeMo Automodel framework with the Diffusers library, enabling distributed training of diffusion models such as FLUX.1-dev, Wan 2.1, and HunyuanVideo directly from Hugging Face Hub checkpoints without format conversion. The framework supports full fine-tuning and parameter-efficient LoRA adapters using FSDP2 parallelism, latent caching, and multiresolution bucketing. Benchmarks on eight H100 GPUs show throughput of 35.51 to 53.73 images per second for text-to-image training, with a Pythonic API planned to supplement the current YAML configuration interface.
