Distortion/Disfigurement detector for output

[BenjaminIrwin]

I was thinking it would be really useful to have a simple distortion detector, which returns True (with a confidence score) if an image of a person is distorted (i.e. extra limbs, mangling, disfigurement etc.).

This could be very useful for A1111, Stable Diffusion and image gen as a whole. Two basic use cases I can think of: building massive datasets automatically and controlling quality of images at inference time.

Has anyone encountered something like this before? Or does anyone have any tips to build it or general thoughts?

[catboxanon]

How do you determine if there's distortion though?

That being said, what you want is RLHF, which Stability is currently doing, and was done on a smaller scale here: https://github.com/AUTOMATIC1111/stable-diffusion-webui/discussions/9084