Browse the models and styles behind stories like this one — free account, instant gallery.
NVIDIA and Hugging Face have jointly released a NeMo Automodel integration with the Diffusers library that lets anyone fine-tune Flux image models and Wan video models across multiple GPUs — without writing a line of distributed training code.
nemo-automodel package.Historically, scaling a Diffusers fine-tune beyond a single GPU meant wrestling with PyTorch's DistributedDataParallel, Accelerate configs, or DeepSpeed YAML files — none of which are friendly to someone whose main skill is prompting and curating training images. NeMo Automodel wraps all of that into a single AutoModel.from_pretrained()-style call. You point it at a Flux or Wan checkpoint, pass your dataset, set a GPU count, and the library handles tensor parallelism, gradient checkpointing, and bf16/fp8 mixed precision automatically.
The practical upshot: a fine-tune that previously required a beefy single A100 can now be spread across several smaller cards — or finished faster on the same hardware. For creators building custom LoRA-style adaptations of Flux.1 Dev on a subject, style, or product, that means shorter iteration cycles between training runs and prompt tests.
Flux.1 Dev has become one of the most popular base models for custom image generation — you can browse community fine-tunes in the Charmloop model catalog to see how wide the style range already is. The NeMo integration adds a credible path to training your own Flux adapter without cloud-vendor lock-in: the checkpoint format stays standard Diffusers, so the result works anywhere that runs Diffusers inference.
Wan 2.1 support is the more novel angle. Video fine-tuning has lagged image fine-tuning by roughly a year in terms of accessible tooling, and multi-GPU video training has been even more restricted. Being able to fine-tune a video diffusion model on custom footage — a specific motion style, a character, a product — using the same API as an image model is a meaningful step toward video generation workflows that actually respond to creator-specific training data.
The output of a NeMo Automodel run is a standard Diffusers checkpoint directory. That means no conversion step, no format wrangling — load it with DiffusionPipeline.from_pretrained() and generate. For teams already using Diffusers in a production pipeline, the integration is effectively zero-friction on the inference side.
A range of LoRA and full fine-tune outputs demonstrating style and subject consistency achievable with NeMo Automodel.
Image: Hugging Face Blog
Creators who want to experiment with fine-tuning before committing to a training run can start with the Charmloop image generator to benchmark what a base Flux model produces on their target subject — then use those prompt results to curate a tighter training set. The guides section also covers dataset preparation fundamentals that apply directly to this kind of custom model work.
The nemo-automodel package is available on PyPI now. NVIDIA and Hugging Face have indicated that additional model support — likely including other video diffusion architectures — will follow as the integration matures.