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

Source: HuggingFace Date: 2026-07-17 URL: https://huggingface.co/blog/nvidia/scale-diffusers-finetuning-nemo-automodel

Summary

NVIDIA’s NeMo Automodel now plugs directly into Hugging Face Diffusers, letting distributed fine-tuning (FSDP2, tensor/pipeline parallelism, multi-node via SLURM) run against Diffusers-format checkpoints with no conversion step. It covers FLUX.1 (12B–32B), Wan 2.1/2.2 video (1.3B–27B), HunyuanVideo 1.5 (13B), and Qwen-Image (20B), reporting 35–53 images/sec and 1–2 video clips/sec on 8×H100.

Implications

• Local models thread: lowers the infrastructure bar for fine-tuning large open video/image diffusion models — LoRA + full-parameter paths on standard Hub checkpoints means smaller teams can adapt frontier-scale diffusion models without bespoke training code.
• Model capability clocks: video generation (Wan, HunyuanVideo) is maturing fast enough that vendors are now racing to make it trainable, not just usable — a sign the video-model layer is following the same open-weights-plus-tooling pattern that text models went through.