NVIDIA and Hugging Face team up to make training giant image and video AI models far easier

Key points

• NVIDIA and Hugging Face jointly released NeMo Automodel, a free, open-source training tool, in June 2025.
• The tool supports fine-tuning of FLUX.1-dev (12 billion parameters) and HunyuanVideo (13 billion parameters), two of the most capable open image and video generators available.
• Developers can start training directly from models hosted on the Hugging Face Hub without converting files to a special format first.
• The software is free under the Apache 2.0 licence, meaning anyone can use it commercially or build on it.

Teaching an AI to generate pictures in a specific style is expensive, technically demanding work. Until now, doing it at scale typically meant rewriting software from scratch every time a new model came along.

NVIDIA and Hugging Face want to change that. The two companies have published NeMo Automodel, an open-source library that connects NVIDIA's industrial-scale training machinery to the Hugging Face Diffusers library. Diffusers is the most widely used toolkit for running diffusion models, the class of AI that generates images and video by gradually refining random noise into a finished picture.

In plain terms: a developer who wants to teach FLUX.1-dev, one of the best publicly available text-to-image generators, to draw in the style of vintage tarot cards can now do that without touching any specialised code. Point the tool at a dataset, pick a configuration file, and run it.

What does this actually change for people building AI tools?

Before this release, moving a model from the Hugging Face library into a training system often required converting the model's saved weights, the numerical values the AI has learned, into a different file format, then converting back after training. NeMo Automodel skips that entirely. A fine-tuned model loads straight back into standard tools the moment training finishes.

Scale is handled automatically too. A small team with one powerful machine and a large research group with hundreds of GPUs, the specialised chips that do the heavy number-crunching AI needs, can both use the same configuration files. Switching between different ways of splitting a model across many chips is a settings change, not a coding project.

For developers with limited hardware, the tool supports LoRA, a technique that fine-tunes only a small fraction of a model's values rather than all of them, cutting memory use sharply while keeping quality high.

The announcement walks through a worked example: fine-tuning FLUX.1-dev on 78 public-domain tarot card images. Training runs for 200 steps and produces a model that generates tarot-style images on demand. It's a deliberately modest dataset, which is the point: the same workflow scales up without rewriting anything.

Supported models at launch include FLUX.1-dev for images and both Wan 2.1 and HunyuanVideo for video generation. Our earlier look at Apple's single-layer approach to image-generation training on 15 July showed a different angle on the same pressure: researchers across the industry are racing to make fine-tuning cheaper and faster.

For most ordinary users, none of this changes anything directly today. The people it affects are developers who build the image tools and creative apps that ordinary people eventually use. Better, cheaper training pipelines tend to mean more specialised AI features arriving faster and at lower cost. That's the real bet here: commodity infrastructure that lowers the floor without capping the ceiling.

The software is available now on GitHub and is documented in the official Diffusers training guide.