I don’t know how this model works, and maybe this idea can’t be used for it. So my ideas are just based on my speculation about how things work 
But if this tool works a bit like all of these AI “diffusion” techniques, wouldn’t it be possible to lead the recovery with a description of what’s on the image?
By default, it would do its thing automatically, and if the results are not good enough, we could try to lead the recovery process with key words and choose their weight with a slider to be more precise.
I tested this model on old, low-resolution, and very compressed images I had. Sometimes it works very well, but mostly for the face. And the rest still has the compression artifacts, but more defined.
After the recovery process, wouldn’t it be possible to use some AI segmentation to further recover the image in a faster way?
- It cleans up the very low resolution and blurry/compressed image to its final size.
- It uses segmentation to get a more precise idea of what’s on the image, like this:
then it would work out all the edges and remove the persistent compression artifacts that were just sharpened at the first stage.
The general idea would be to drive the recovery process based on a strong segmentation model. And maybe this can be used for the other model too? If this technique is not already used by Topaz of course.
