>They were mostly fine as thumbnails but blown up to ~500x500 the the eyes and mouth are rough. That's the result of automated training. If you want good resu

>They were mostly fine as thumbnails but blown up to ~500x500 the the eyes and mouth are rough. That's the result of automated training. If you want good results you need to properly label your dataset (15 or more images of yourself) and then use a proper upscaling solution for it. Dreambooth-based images usually need some manual fine-tuning (and sometimes inpainting) to look presentable at higher resolutions. You can use latent upscaling for adding details but it'd require some manual supervision to weed out the bad generations. Using upscalers like Lanczos is usually foolproof.