Tech Specs

Shadan

How it works

It learns what you hide. When you hide an image, Shadan immediately hides exact copies and reposts of it. It then shows you similar images from your recent browsing; each one you confirm or reject teaches it what that topic looks like, even in new poses, outfits or scenes.
It’s careful when unsure. Confident matches are hidden; uncertain ones are blurred so you can check with one click, and your answer to “Hide images like this?” makes it more accurate.
It remembers numbers, not pictures. To spot look-alikes, it keeps a compact numeric fingerprint of each image you scroll past on the sites you allow, plus its web address for the review screen, for up to 30 days. Fingerprints can’t be turned back into images.
It all happens on your Mac. Recognition runs on your Mac with Apple’s Core ML; your images, browsing and choices never go to any server, including ours.

Copies and reposts: about 97% caught, even with new captions, crops or screenshots; about 1 in 800 other images wrongly hidden.
Characters and topics: about 88% caught in new poses, outfits or scenes after 5 confirmed examples, with 1 in 100 other images blurred for a check.
Recognition: about 7 ms per image on Apple M5 (2.8 ms DINOv2, 4.2 ms SigLIP 2), using Core ML. Images seen before are answered instantly; new images are usually limited by downloading, not recognition.
Models: DINOv2-small by Meta (43 MB); SigLIP 2 ViT-B/16 by Google (185 MB); plus a small per-topic classifier trained on your answers.
Storage: about 2.3 KB of fingerprints per image seen, kept up to 30 days (46 MB at most).

Measured on small test sets of popular memes and one AI mascot; results on your own browsing will vary.

macOS 15 or later, Safari, and a Mac with Apple silicon.