A research team led by Stanley Chan, Elmore Professor from the Elmore Family School of Electrical and Computer Engineering at Purdue University has introduced a groundbreaking method that enables existing Generative AI models to understand camera settings while maintaining scene consistency. The work, which allows for precise control over real-world camera effects such as bokeh, focal length, shutter speed, and color temperature, marks a significant step toward AI-assisted professional photography.

The research overcomes a key limitation in current generative models: While generative AI methods today excel at creating images and videos, they lack an understanding of camera physics. For example, if a user wants to generate images with different fields of view—such as those captured by a 24mm lens versus a 70mm lens — the computer fails to understand the difference. To professional photographers, this is a significant limitation because the AI-generated photos today do not respect camera physics.

To address this, the team introduces Generative Photography, a new framework designed to incorporate camera intrinsic settings into content generation. The core innovations of this work include a new concept known as dimensionality lifting, which transforms the traditional spatial-only problem to a spatial-camera joint problem, and a new technique known as contrastive camera learning, which disentangles camera physics from the scene content.

The work has received widespread discussion on social media across academia and industry.

Yu Yuan, a second-year PhD student at Purdue, is the lead student on the project.

"Generative photography not only enhances creative control but also pushes AI to understand the physical world more deeply,” said Yuan. “We will continue exploring ways to create a physically consistent generative world."

With this breakthrough, Purdue University’s research team is setting the stage for a new era in AI-powered photography, where generative models are no longer black boxes but instead intelligently control camera physics for scene-consistent, high-quality visual outputs.

This research will be presented at the IEEE Conference on Computer Vision and Pattern Recognition 2025, one of the top conferences in computer vision. 

Source: Generative Photography Scene-Consistent Camera Control for Realistic Text-to-Image Synthesis