Hugging Face and NVIDIA have released an open-source integration between the Diffusers ecosystem and NVIDIA NeMo Automodel, giving developers a distributed training path for image and video diffusion models stored on the Hugging Face Hub. The integration is intended to scale from a single GPU to multi-node training without converting checkpoints into a separate format.
The companies say the work is available under the Apache 2.0 licence and documented in the Diffusers training guide. It brings NeMo Automodel's distributed training features to models that use Diffusers classes and pipelines, while keeping the resulting checkpoints compatible with the wider Hugging Face ecosystem.
Hub-native model loading
NeMo Automodel can load a supported model by pointing its pretrained model setting at a Diffusers model identifier on the Hub. Training uses Diffusers model classes, and generation uses the matching Diffusers pipeline. Fine-tuned checkpoints can be loaded for inference or shared back through the Hub without a separate conversion stage.
That compatibility matters because conversion steps can introduce operational friction and break downstream tooling. Hugging Face and NVIDIA say existing quantisation, compilation, adapter and sampler tools can continue to work with the resulting checkpoints.
Recipes for current image and video models
The initial integration includes fine-tuning recipes for Wan 2.1 and Wan 2.2 text-to-video models, FLUX.1-dev and FLUX.2-dev text-to-image models, HunyuanVideo 1.5 and Qwen-Image. Some recipes support low-rank adaptation, allowing developers to tune a smaller set of parameters when a full model update would require too much memory or compute.
Model size still sets practical hardware requirements. The smaller Wan 2.1 1.3-billion-parameter model can fit on a single 40 GB NVIDIA A100 according to the announcement, while models with tens of billions of parameters need substantially more capacity. NeMo Automodel currently supports flow-matching diffusion models rather than every architecture available through Diffusers.
Parallelism as configuration
The integration exposes distributed strategies including FSDP2, tensor parallelism, expert parallelism, context parallelism and pipeline parallelism. Developers select these approaches through configuration rather than rewriting the model for each cluster layout.
It also supports pre-encoding training data into cached variational-autoencoder latents and text embeddings. That avoids repeating the same encoding work during every training step. Multiresolution bucketing can group samples by shape, while sharding and checkpointing help larger jobs span multiple GPUs and nodes.
The published workflow demonstrates full fine-tuning of FLUX.1-dev using a small public image dataset, checked-in YAML configuration and command-line overrides. NVIDIA recommends its NeMo Automodel container because it includes CUDA-compiled dependencies, though installation from Python packages or source is also available.
What the release changes
For Hugging Face users, the main change is a more direct route from a model on the Hub to NVIDIA's distributed training stack. For NVIDIA, the integration broadens the set of community diffusion checkpoints that can use NeMo Automodel without waiting for a bespoke training implementation.
The companies say support for a newly added Diffusers model requires a preprocessing handler and model adapter rather than a full training script. That should reduce integration effort, but it is not automatic support for every new checkpoint. Developers must still confirm that the architecture uses the supported objective and that an appropriate recipe exists.
Limits and evaluation
Distributed training efficiency depends on model architecture, interconnect, precision, dataset and selected parallelism. The availability of configuration options does not guarantee that a job will scale efficiently without tuning. Teams should benchmark throughput, memory use, checkpoint reliability and output quality on their target hardware.
The release is most relevant to developers who need to adapt open image or video models beyond what a single machine can comfortably handle. Its open licence and checkpoint compatibility lower the barrier to experimentation, while production users still need to manage training data rights, model licences, safety evaluation and the substantial compute cost of full fine-tuning.