santa hat
deerdeer nosedeer glow
Sign In

some models/Loras seem to "pass" the good quality point into over generated bad images

So this has been bugging me on some models/Loras. (Some people with lower end graphics cards may not see this). During the Generation, I see the image built in steps from a blur to a full image in like 6 steps for 512x512. Sometimes it reaches a "best quality point" at the end and creates a great picture. Sometimes that "best quality point" seems to hit around step 4/6. then the fianl 2 steps start disorting the image and making thicker edges and oversaturated colors to the point the image is ruined. I see examples with great looking images and sometimes are unable to reproduce the same or even similar quality. I know some Loras are trained on specific models and will usually create less quality images with different models. I know an OK amount about the generation process but not enough to really know what's going on to be able to reach that "best quality point" and stop it there. Interrupting or skippping on the good spot just causes the image to distort really bad. I hope my explanation is clear enough as I don't really know what to look up or where to go to find help with this.

2 Answers

You can try the anti-burn extension, which was an extension made to stop a few steps before completion:
https://github.com/klimaleksus/stable-diffusion-webui-anti-burn

try set CFG scale to 7 or even lower, and u need 30 step at least to see normal thing

Your answer