Hires fix usually does the trick

@zArvy18 Hires fix is utilized just like the PNG info suggests. The rest of the picture looks crisp as it should though. Am I missing something here? @FIKOsofter

@Nachttot3r Make sure you're using a non-Latent upscaler for hires fix. BSRGAN or ESRGAN for instance. A 2x upscale with somewhere between a 0.4-0.6 denoise setting should work. You might also need to set the number of steps to ~30 for both the hires fix and the original image. Finally, prompts matter a lot -- I use some variation of XpucT's negative prompt to help guide Stable Diffusion into a corner of its latent space with high resolution pictures.

Remember that the model is trained on both "good" and "bad" pictures. You can encourage it to use good pictures with prompts, textual inversions (basically a vector representing a specific prompt that guides the generation), a LORA (which locks certain transformations to force a kind of output), etc. Multiple LORAs will fight each other, a CFG scale set too high will "deep fry" the image, etc., etc. Lots can go wrong. Find a picture on the site, save the png, edit it in notepad to see the header (with all the prompt info), and follow that prompt exactly. Then work towards your desired result to figure out where things are going wrong for you. Good luck!

@Nachttot3r this can happen because you are using some additional face restoration like CodeFormer or GFPGAN when processing. especially on some higher quality models like this one.

using vae-ft-mse-840000-ema-podado.safetensores

Hi! I don't see the point of this.