This weird camera turns photos straight into AI art

AI art is cropping up everywhere. And if the explosion of AI image generators weren't enough, an engineer has just designed a bizarre compact 'AI camera' that turns shots of real scenes straight into novelty AI-generated images.

DreamGenerator runs Stable Diffusion (see our pick of the best AI art generators) to transform images as soon as they're taken. Described as "a toy that sparks new ideas and adds an unexpected twist to everyday photography", it provides 30 themes to choose from. But why?

Currently only a prototype, DreamGenerator is a strange proposition. The prompt 'themes', which include 'retro', 'sci-fi', 'fairytale' and 'pirate', feel like an extension of the concept of smartphone filters, although they transform images in more radical ways. In theory, the generated images reference the composition and pose of the original photo, but some examples appear to show generations that look completely different from the source.

The bigg question is why put this in a camera body rather than in an app for the best cameraphones? Or even why not just make the AI transformation in post-production on another device?

Well, the camera's creator Kyle Goodrich, a 3D artist and augmented-reality product designer at Snapchat, says the aim is to make AI art simpler. He says that many AI image generators offer complicated user experiences. The simplicity of a point-and-shoot camera would make AI art generation more intuitive.

"AI image generators are a game changer, but coming up with the perfect prompt is tough," Goodrich writes on Twitter. "Plus, navigating some of these platforms feels like solving a puzzle. I wanted to simplify this process and create a tool that combines intuitive design with the power of AI."

He added: "I love my DSLR, but lugging around a bulky camera can feel like a burden. Smart phones do a decent job, but there’s something undeniable about the simplicity of a point-and-shoot... and I find that I take a lot more photos."

DreamGenerator looks and works much more like a traditional camera than a previous example of an AI camera that we saw – the Paragraphic camera, which has no lens and Google Street View data rather than images to generate AI art. Goodrich's device has a basic camera module that feeds images to a Raspberry Pi computer running an instance of Stable Diffusion and ControlNet. A simple interface allows a user to take a photo and then toggle through the prompt themes. The original image appears in the viewfinder but is quickly transformed according to the theme chosen. Different buttons let you generate variations of an image. 

But while the UX is simpler and more intuitive, the results would seem to be much more limited, even if a final product introduced more than the 30 themes used on the prototype. It's not clear if the original image is also saved, and Goodrich hasn't clarified where the computing takes place – given the size of the device we presume it offloads to the cloud rather than processing the images on the device itself (if you're looking to use AI art in post-production, see our guide to the best AI art tutorials).