Fine-tune video and image models at scale with NVIDIA NeMo Automodel and ๐ค Diffusers
Summary
NVIDIA and Hugging Face announce a NeMo Automodel integration that lets you fine-tune Diffusers-format diffusion models โ including FLUX, Wan 2.1/2.2, HunyuanVideo, and Qwen-Image โ at scale without checkpoint conversion or model rewrites. Training parallelism is driven by YAML config, not code changes.
NVIDIA and Hugging Face jointly announced that NeMo Automodel, NVIDIAโs open-source PyTorch DTensor-native training library, now integrates directly with Diffusers-format models on the Hugging Face Hub. The collaboration targets the pain point of scaling diffusion model fine-tuning beyond single-GPU setups without having to rewrite training code or convert checkpoints.
Whatโs actually new
The core selling point is zero checkpoint conversion: you point pretrained_model_name_or_path at a Hub model ID, train, and the resulting checkpoint loads back into a standard DiffusionPipeline for inference or re-upload. NeMo Automodel ships ready-made fine-tuning recipes for seven model configurations spanning text-to-image and text-to-video โ Wan 2.1 (1.3B and 14B), Wan 2.2 (A14B MoE), FLUX.1-dev, FLUX.2-dev, HunyuanVideo 1.5, and Qwen-Image. Both full fine-tuning and LoRA are supported across most of these. The library currently targets flow-matching models only, using latent-space training with pre-encoded VAE outputs and multiresolution bucketed dataloading. Adding support for a new Diffusers model requires only a data preprocessing handler and a model adapter rather than a complete custom training script.
What it means for your config
This is where NeMo Automodelโs design is worth paying attention to from a config perspective. Parallelism strategy โ FSDP2, tensor parallel, context parallel, pipeline parallel, expert parallel โ is declared in YAML configuration files rather than baked into training scripts. The blog walks through a concrete example using the checked-in flux_t2i_flow.yaml with run-specific values passed as CLI overrides (cache directory, resolution, learning rate schedule, checkpoint intervals). That means switching from single-node LoRA to multi-node full fine-tuning is a config change, not a code change.
If youโre already running Diffusers training scripts (the train_dreambooth_lora.py-style examples), this is a different orchestration layer. NeMo Automodel doesnโt replace those scripts for simple single-GPU jobs โ itโs aimed at the point where you need FSDP sharding, multi-node SLURM jobs, or models that donโt fit on one device. The YAML configs live in examples/diffusion/finetune/ in the Automodel repo, and the post notes Kubernetes orchestration support is coming but isnโt available yet.
One practical detail: installation via the NeMo Automodel Docker container (nvcr.io/nvidia/nemo-automodel:26.06) is the recommended path since it bundles PyTorch, TransformerEngine, and CUDA dependencies. A pip install option exists but expect the usual CUDA build headaches for TransformerEngine if you go that route.
Recommended next step
If youโre fine-tuning diffusion models larger than what fits comfortably on a single GPU, clone the Automodel repo and look at the YAML recipes in examples/diffusion/finetune/ for your target model. The pre-encoding step for latent caching is mandatory and model-specific, so start there before attempting a training run. If youโre happy with single-GPU LoRA via the existing Diffusers training scripts, thereโs no urgent reason to migrate โ but bookmark this for when you scale up. The post also mentions upcoming โPythonic recipe APIs,โ which may change the config surface; worth watching.
Read the full announcement on Hugging Face Blog โ Fine-tune video and image models at scale with NVIDIA NeMo Automodel and ๐ค Diffusers
More Hugging Face Updates
The OlmoEarth Platform: Geospatial inference at planetary scale
Allen Institute for AI details the infrastructure behind OlmoEarth Platform, which runs geospatial foundation model inference at continent scale by splitting work across CPU and GPU stages with heavy parallelism.
NVIDIA Cosmos-H-Dreams: Bringing Real-Time Generative Simulation to Surgical Robotics
NVIDIA releases Cosmos-H-Dreams, a distilled world model that runs action-conditioned surgical simulation at ~160 fps on a single GPU, along with FlashDreams inference engine and a recipe for adapting to custom embodiments.
Bringing Nunchaku 4-bit Diffusion Inference to Diffusers
Hugging Face integrates Nunchaku Lite into Diffusers, enabling native loading of SVDQuant W4A4 checkpoints via from_pretrained() with no local CUDA compilation. This cuts diffusion model VRAM usage roughly in half while also delivering inference speedups.