How to check part of an image for AI

The instinct to zoom into the one spot that looks off is the right one, and it's also how you get a better answer from a detector. Checking a cropped region often beats scanning the whole picture. Here's why that works, and how to do it well.

Why the whole image dilutes the signal

When a detector scores a full image, it's averaging evidence across everything in the frame: the subject, the background, borders, watermarks, interface chrome if it's a screenshot, and any untouched areas. If only one part of the image is suspicious, say a face that was swapped into an otherwise real photo, its signal gets watered down by all the genuine pixels around it. The average comes back ambiguous, and you learn nothing. Cropping removes the noise and lets the detector concentrate on the pixels you actually care about.

Resolution is the other half of the story

Detectors read fine, low-level texture, and that texture only survives at adequate resolution. When you feed in a large image, it typically gets scaled down before analysis, and scaling down destroys exactly the fine-grained fingerprint the detector depends on. A face that occupies 5% of a big image might be just a few dozen pixels wide after downscaling, far too coarse to judge. Crop to that face first and it fills the frame at full resolution, giving the detector real detail to work with instead of a blurry thumbnail.

What to crop to

Choose the region most likely to carry a verdict. For a suspected fake person, crop tightly to the face, skin texture, eye catchlights, and the hairline are dense with tells. For a photo you suspect was AI-edited, crop to the specific object that looks wrong: the added or removed thing, the hand, the piece of text. For a scene you think is fully generated, a patch of complex background, a crowd, a shelf of products, a stretch of architecture, often gives a cleaner read than the polished subject the model spent its effort on. The rule of thumb: crop to where a human eye would go to be suspicious.

How to do it with the crop tool

Our detector has the crop tool built in so you don't need a separate editor. Upload the image, open the crop tool, and drag a box around the region you want to check. The detector then scores that selection at full resolution and returns a likelihood for that specific area. Because it takes seconds, you can check several regions in turn, the face, then a hand, then the background, and see whether the suspicion holds up across the parts of the image that matter.

Reading regional results

Cropping also lets you localize a problem, which is useful for a specific and increasingly common case: a real photo with an AI-edited region. If the face scores high but the rest of the scene scores low, that pattern points toward a genuine photo with a synthetic element added, a very different conclusion from "the whole image is fake." Checking multiple crops turns a single yes/no into a small map of where the synthetic signal actually lives, which is often the more truthful answer.

A few practical tips

Start from the best-quality source you have, since cropping can't recover detail that a screenshot already threw away. Keep the crop tight but not so tight that it's a handful of pixels, you want the region to fill the frame with real texture. And when results across crops disagree, don't force a single verdict; a mixed picture is itself informative, and often the honest answer is that part of the image is synthetic and part isn't.