Fine-tune video and image models at scale with NVIDIA NeMo Automodel and 🤗 Diffusers

July 17, 2026 | Source: Hugging Face Blog

Tags: NVIDIA, NeMo Automodel, Hugging Face, Diffusers, FLUX.1, fine-tuning, distributed training, video generation

NVIDIA and Hugging Face integrate NeMo Automodel with Diffusers, enabling distributed fine-tuning of FLUX.1-dev, Wan 2.1, and HunyuanVideo at any scale — switch between FSDP2, tensor, and pipeline parallelism via YAML config rather than code rewrites. Apache 2.0 licensed, no checkpoint conversion needed.

Details

NVIDIA's NeMo Automodel library is now natively integrated with Hugging Face Diffusers, making production-scale distributed training available for any diffusion model on the Hub. The core promise: point pretrained_model_name_or_path at any Diffusers model ID and start training immediately — no checkpoint conversion, no model rewrites required. Parallelism in NeMo Automodel is declared via configuration rather than hardcoded into training scripts. Developers can switch between FSDP2, tensor parallel, expert parallel, context parallel, and pipeline parallel by editing YAML files, letting the same recipe scale from a single GPU to hundreds. The library currently supports flow-matching models: FLUX.1-dev for text-to-image, and Wan 2.1 and HunyuanVideo for text-to-video. Training efficiency features include latent caching, multiresolution bucketing, and memory-efficient sharding — critical for video model fine-tuning where GPU memory is the primary bottleneck. Checkpoints round-trip cleanly back into standard Diffusers pipelines, preserving full compatibility with the broader Hugging Face ecosystem. The integration is fully open-source under Apache 2.0 and documented in the official Diffusers training guide. NVIDIA indicates Pythonic recipe APIs are coming next to further reduce the setup overhead for new fine-tuning jobs.