I'm using the Automatic1111 Stable Diffusion XL. Whatever settings I select, if I use it for a period of a couple of days (say, 30-50 images generated--I'm just playing with it right now), the images drift from photorealistic to absurdly abstract. There's also a weird thing where you prompt for one person, but get a set of near-identical twins. This one is the result of the prompt, "Nervous-looking man in swim trunks on a diving board, knees slightly bent and hands together."
The abstraction seems unaffected by any WebUI settings. The only way I've found to get back to photorealistic imagery is to delete configuration files (config.json) ... but manually changing the settings doesn't work. Is this a bug? Is there some magic setting I'm missing?
That's Automatic1111 Stable Diffusion XL on Debian GNU/Linux. It doesn't recognize my Radeon GPU, so currently using CPU only. Flags are
--skip-torch-cuda-test --no-half --opt-sdp-attention --opt-split-attention
Help much appreciated.