PicoDiet tries several encodings per image and keeps the smallest one that still hits your quality score

Most image squeezers hand you a slider and a prayer. A JPEG quality setting is one number applied to every photo, so the foggy beach shot gets crushed exactly as hard as the screenshot full of sharp text, and one of them looks fine while the other looks like it was faxed.

PicoDiet skips the guessing. It tries several settings per image, scores each result, and keeps the smallest file that clears the bar you set. The default bar is 80, and you can move it anywhere from 10 to 95.

The score comes from SSIMULACRA2, which the page calls “an open metric built to predict what people see.” The page also translates the numbers: “90 and up is usually indistinguishable, 80 looks the same at normal viewing distance.” The metric’s own repository says it was developed by Jon Sneyers at Cloudinary between July and October 2022, and updated in April 2023. It labels 80 as “very high quality” and 90 as “visually lossless.”

Why not just use the older yardstick? Sneyers, in a Cloudinary blog post about the metric’s predecessor, wrote that “A simple metric like PSNR does not really take human perception into account.” He also pointed out that a score “is still based on averaging the score over the entire image,” so SSIMULACRA adds “local artifact detection, which looks for local “lumps” of errors.” In plain terms: one ugly patch of sky shouldn’t hide behind nine hundred fine ones.

The tool builds that idea into its viewer. Open any result and it shows you the original next to the new file, zoomed to the spot that changed the most, with a draggable divider and a small corner map of where the differences sit. The score picks the file. Your eyes get the last word on the worst corner.

It takes PNG, JPEG, WebP, GIF (animated too), AVIF, HEIC, BMP and TIFF, and writes JPEG, PNG, WebP or AVIF. Animations can come out as GIF, WebP or animated PNG. Each image can be up to 25 megapixels and 30 MB. The page says “There is no sign-up and no watermark,” which in 2026 counts as a feature list.

Everything is processed on its server, not in your browser. The page says uploads are encrypted in transit and deleted within 3 minutes after you finish, that images are “Never shared, sold or used for training,” and that EXIF and GPS metadata is stripped from what it produces. Those are the maker’s claims, and nothing I found checks them from outside. The page credits the developer as @tevkert on X, and that’s the only name attached.

A score of 80 also isn’t a promise. The metric predicts what people notice, and a prediction can miss. I found no published figure for how much smaller PicoDiet’s files come out than a plain slider’s, either, so the saving on your own folder is still a thing you have to measure yourself.

Expect more of this. A public perceptual metric means anyone can check a compressor’s guess, and that tends to spread. My money says the next round of image tools will advertise a target score the way cameras advertise megapixels, and the first one that shows its worst-looking corner by default will earn some trust.

Diet starts Monday.