Image Scale
Most models are trained on images either 512x512 or 768x768 so they are often best at making images at that size. Smaller sizes often look very pixelated.
While larger images can have "artifacts" appear. These artifacts are often doubling the amount of people, or alternatively, giving double heads/torsos.
Modern models seem to have a much lower chance of these artifacts appearing.
If you intend to make images larger than 1024x1024 it is recommended to turn on "Highres Fix" in settings.
Please note that for inpainting to work properly the "location" you work on is still under these constraints. So if you generated an image at 768x768 with a character in the distance, and it´s face just being 30x30 pixels, then inpainting is not going to work. Its best to upscale the image and then inpaint.
Last updated
