New AI Rescues the Blown-Out Sky and Dark Shadows in Your Photos

Have you ever taken a photo of a sunset and gotten either a washed-out white sky or a pitch-black beach? That's because phone cameras capture a narrow slice of brightness — a limited dynamic range. Real scenes have a much wider range, so something gets sacrificed. High dynamic range (HDR) images keep both the bright and dark parts visible, closer to what your eyes actually see. Until now, getting HDR usually meant snapping several shots at different exposures and stitching them together.

A team of researchers has published a method called ExpandDiff that does the job from just one picture. It uses a diffusion model — the same family of AI that generates images from text prompts — but here it's pointed at a repair job rather than a creative one. The AI looks at a normal photo and predicts the detail that was clipped away, filling in bright highlights and dark shadows in a single pass rather than two separate steps. A key trick called Dynamic Clipping Synthesis teaches it to handle many different degrees of damage, so it isn't thrown off by an unusually harsh or gentle exposure.

On the standard test they used, the researchers report their method is 3.43 decibels better than the strongest competing approach, and 7.34 decibels better in the hardest cases where both ends of the brightness range are damaged. In plain terms, decibels are a quality score — a jump that large is not a small nudge, it's the difference between a usable photo and a murky one. The code and extra material have been released publicly, and the paper has been submitted to a major speech and image processing conference.

The catch: this is a research paper, not a product you can download today. It needs to be trained on paired data — ordinary photos and their proper HDR versions — and running a diffusion model takes real computing power, so it won't land in your phone's camera app overnight. Still, the direction is clear. The next time your shot of a bright window and a dark room comes out looking wrong, software like this may quietly fix it for you.