When I used to finetune ESRGAN on its own seemingly flawless looking output, the "feedback" would create artifacts in the subsequent output. Sometimes I could

When I used to finetune ESRGAN on its own seemingly flawless looking output, the "feedback" would create artifacts in the subsequent output. Sometimes I could see why if I went back and pixel peeped the training data, but sometimes I couldn't. I wonder if something similar can be applied to LLMs? Maybe running its own output through the net would "amplify" it in a detectable way.