AI Now Rebuilds Pictures From Just 10% of the Data

Collecting data costs money. Every extra pixel in a camera, or extra measurement in an MRI machine, means more power, more storage, and more time. But cutting corners carelessly throws away details doctors or analysts later need. Two researchers, An Vuong and Thinh Nguyen, wrote a paper on arXiv tackling that trade-off: let the AI choose what to measure, moment by moment, instead of measuring everything and sorting it out later.

The trick is a diffusion model — the same type of AI that powers image generators like DALL·E. These models learn what real images look like by cleaning up random noise. Here, that knowledge is used backwards: given a few measured pixels, the model guesses what the rest of the picture probably looks like. That guess tells it which new pixels would be most informative to grab next. Think of a doctor examining a blurry scan and asking the technician for one more specific angle, rather than redoing the whole thing.

The results were solid but not magical. On MNIST, a standard set of 70,000 handwritten digits used to test AI, the method made 8 times fewer errors than random selection while measuring only 10% of the pixels. On CIFAR-10, a set of small photos of everyday objects, it improved image quality by 0.9 to 3.4 decibels. Most promising practically: on fastMRI, a public collection of real knee and brain scans, a single fixed measurement pattern designed this way beat well-known techniques.

The catch is speed and complexity. Sequential acquisition means the AI picks one measurement, looks, then picks the next — great for accuracy, but slower than taking everything at once. The one-shot version, which decides everything up front, only beat random guessing when the researchers added careful math to keep it honest. So don't expect your phone camera to shrink overnight. But expect medical imaging and sensor design to borrow this idea soon.