Raccognize - have AI companies stolen my images?

We built a little detective for images, with which we can see if an image is in the training data of a diffusion model (Like DALL-E, stable diffusion or midjourney). We do this by taking your image that you wanted to be checked, describe it (think auto alt text) and then feed it a number of times into the diffsuion model (in our code for test purposes SDXL), then we pick the best visual match. How do we determine the best visual match? Well theres been a lot of research papers on this topic that dont combine each others statstics, so we thougt we should fix that, and used all of them we thougt are relevant (9 in total). With the best image choen, we then trained an ML model on the weights for those statstics, so how significant each statistic is, and give you one final confience score that shows you how confident that AI is in the image being in the dataset the AI used to train. With only 200 images used in our training set for this model, we already got an accuracy of 77%, with much more to gain.

All of this fancy technology gets sent to a (very pretty) flutter frontend, that shows this data in nice chunks, and explains what the additional values mean. It also shows you the "most visually similar" image the AI generated. The data gets sent between the backend and frontend using fastAPI.

If you want to try it out, the APK in the release section of our github will work until Saturday Nov23 early mroning europe time, because then our credits for hosting the api and backend will run out.