If it could, that would suggest Meta is quite incompetent. If there is an AI detector that can tell the difference between your AI generated images and real images, you should be using that to train the AI model.
Oh it’s able to, it just doesn’t.
It was today that I realized I could so I did blacklist meta with pihole on my personal vLANs.
Lol this garbage again. Even an untrained network will have an input that is valid for any realistic combination of pixel values with the correct seed and string input. How the fuck was this supposed to work? I will bet that they are putting pictures back through their network and seeing if it correlates with certain common phrases. A single pixel change might be enough to throw this thing off. Changing from PNG to jpeg would be enough to ruin this.
Oh. So someone programmed that in? “Look how good I am!”. SMH.
Sounds like a feature
