AI image detector
Tells whether a photo came from a camera or a generator, all in your browser.
Drop a photo or an image
or click to choose (.jpg, .jpeg, .png, .webp, .avif, .gif, .bmp), or paste from the clipboard (Ctrl+V). Up to 10 MB per file.
The model judges photographs. Drawings, screenshots, collages and heavily compressed images come out unreliable.
This is not proof. No detector can tell where an image came from with certainty. The model learned from photos and generator output of its time; a newer generator, a collage, a drawing or heavy compression confuse it both ways. Treat the number as a prompt to check the source, not as a verdict about the author.
The whole check runs in your browser using the Nonescape mini model (Aedilic, Apache 2.0 licence). Images never leave your device.
- Runs right in your browser
- Free, no sign-up
More about AI image detector
How the AI image detector works
A neural network judges the image. It was trained on pairs: real photographs and the output of generators such as Stable Diffusion, Midjourney and DALL·E. It does not look at what the picture shows but at how it is put together: the faint traces in noise, texture and gradients that generators leave behind and a camera does not. The result is a probability from 0 to 100 % that the image came from a generator.
We use the Nonescape mini model (Aedilic, Apache 2.0 licence), a lightweight detector built on the EfficientNetV2 network. It runs entirely in your browser: it downloads on first use, then stays stored, and further checks are instant. Images never leave your computer.
Reading the result
Up to 30 % the marks point to a photograph, above 70 % to a generator, and in between the result is inconclusive. The middle is wide on purpose: the model can be confident where it should not be, and “inconclusive” is a more honest answer than a guess.
The result is not proof. The model learned from photos and generator output of its time. A newer generator may pass as a photo, and a heavily edited, compressed or downscaled photograph may come out as synthetic. Treat the number as a prompt to check the source: where the image first appeared, whether a larger version exists, whether the author has the original.
Frequently asked questions
Is the image uploaded to a server?
No. The check runs in your browser and the image is never sent anywhere. The only download is the model, and only the first time.
Which images work?
Photographs at ordinary resolution, ideally as they came out of the camera or the generator, without resampling and without cropping to a small detail. Drawings, screenshots, collages, images with text and heavily compressed social-media thumbnails come out unreliable. The model takes the centre of the image, so for a collage it judges only what is in the middle.
Does it recognise Midjourney, DALL·E or Flux?
Unedited output of earlier versions usually yes. Every new generator changes its traces and a detector is always a step behind. The safest approach combines several clues: the detector, the file metadata, a reverse image search and common sense.
Can I accuse someone based on this?
No. No detector, paid ones included, has the reliability such a decision would need. A false accusation does more harm than an unrecognised generated image. If you need to know how an image was made, ask the author for the original and the process.