New AI Codec Shrinks Images 8% With No Quality Loss

Every photo you snap, screenshot you save, or image you load on a website has been squeezed down in size by a compression system. The challenge is always the same trade-off: make the file smaller, and it usually looks worse. A team of researchers has published a new method, called LC3EM, that nudges that trade-off in your favor — getting smaller files while keeping the picture looking identical to your eyes.

Here is the plain-English version of how it works. Older systems like COOL-CHIC use a tiny AI model that studies one specific image and learns to guess what patterns come next, kind of like predicting the next word in a sentence. But it only looks at nearby pixels, so it misses big repeating patterns across the whole image. The new method adds a second predictor based on simple math that extends patterns across long distances — perfect for things like flat backgrounds, straight lines, and grid layouts. A built-in switch automatically picks whichever predictor does better for each patch of the image, and it barely costs any extra data to store.

The results: when plugged into the existing COOL-CHIC 4.0 system, files shrank by 3.4% and 7.7% on two standard test sets, and by about 3% when paired with the newer COOL-CHIC 5.0 — all at the same picture quality. The gains were strongest on highly regular, computer-generated images, such as interface mockups, diagrams, and synthetic graphics.

The catch is that these are single-digit improvements, and the benefit is smaller for ordinary photographs, which are messier and less predictable. The code is not yet publicly available, and even great compression research can take years to show up in the formats your browser and phone actually use. Still, the direction is clear: smaller images mean faster pages, cheaper storage, and less data burned on your phone plan.