Upscaling and post-processing

Let's continue exploring AI generated images with stable diffusion. With high resolution 4k monitors it can get somewhat underwhelming with a 512x512 pixel images. With AI up-scaling and post-processing we can add depth and details and remove some of those annoying artifacts.

Face improvements with GFPGAN🔗

Humans are very good at detecting small deviations in symmetry, especially in faces. If something is just a little off we perceive the face as that of Quasimodo. There is a GFPGAN model that can be used to improve faces. The effect can be seen below. It has a tendency to smooth quite a lot when cranked up to 100%. This does mess up the art-style somewhat. Something to use in moderation.


Upscaling Real-ESRGAN🔗

This is a model that performs 2x/4x resolution up-scaling and noise-reduction and sharpening. This model does not produce new details but enhance what is there. It's quite quick to run and produces overall good results. It can be quite smoothing and not work that well on a highly detailed image. But with highly stylized content the result is amazing.


Upscaling go-big🔗

This up-scaling scheme is essentially running Real-ESRGAN to increase resolution and dividing the image in sections re-running stable diffusion on each section generating new details. This gives the possibility to feed it a prompt when doing this step. This does redraw the image quite a bit as seen below. A zoomed in view is seen comparing the original with the up-scaled to the right. It's essentially reinterpreting every little detail but with higher resolution. It's also very slow, it took a half an hour just to run on a single image.