Critique AI-generated images on a scale of 1 to 100 points.

[HIRO0132]

This is just a what if.
The good parts of the SD-generated image are masked in red and scored from 51 to 100.
Likewise, the bad parts of the SD-generated image are masked in blue and scored from 1 to 49.
The unmasked parts are given a score of 50.
Finally, the intensity of the masked colors tells the SD which parts are unnatural or natural.

Can a human critique the image in this way so that it can be used to generate the next image for SD?
Can we tell AI what should not be generated in unnatural hand parts or body shapes and their positions?

I am using DeepL to translate.

[grexzen]

I used a data set of embeddings of hundreds of thousands of gen images scored by people to train a scorer that tells you if the pic is good or not, grade 0-10. But your approach sounds more fundamental, hope you can take it there.

Also this belongs in discussions.

[HIRO0132]

This proposal is only intended to be a temporary critique as a stronger command to SDs per environment.
Humans give SDs CARROT AND STICK