Generative artificial intelligence models are not new to many, but the University of Surrey may have something very unique on its hands.

The institution has released a new generative AI that runs locally on devices. However, the University of Surrey's NitroDiffusion model is able to generate images in real time as the user enters prompts, thanks to a number of wizardry.

If the user enters “a picture of a meerkat floating in space, wearing sunglasses,” the image is generated as specified.

The code is available on GitHub, where the team explains: “NitroFusion is a fundamentally different approach from single-step diffusion, which achieves high-quality generation through a dynamic adversarial framework.”

The team also explains that “NitroFusion is a new approach to image generation, which is based on a single-step diffusion approach.

“While one-step methods offer dramatic speed advantages, they generally suffer from quality degradation compared to multi-step methods.

The approach they took relied on the concepts of art critics who provide comprehensive feedback on composition, color, and technique. They also maintain a large pool of specialized identification heads that guide the generative process.

In short, the model can be thought of as being chock full of mini-AI art critics who guide the generation process when words are entered via prompts.

According to the team, these discriminators, once sampled, return increased knowledge to the discriminator pool.