"banish the plastic AI look" is a bold claim here.

I would gladly download this, we need a LORA that fits this purpose. Problem is, the example images for Z Image look more plastic than the default model! So much so it's tempting to download the LORA anyway and use it at negative magnitude...

Chances are, there are dozens of terrible retouch examples in the dataset. Stuff like Chinese skin smoothing filters full of "soap", old frequency separation nonsense from ancient Photoshop workflows, things like that.

If you want a good high end retouch dataset, you'll have to be picky. I am not sure you can scrape Vogue magazine or something, but maybe that's the general direction you should try? Or photographer platforms and awards in fashion or portrait category... But definitely not Instagram or what have you!

Alternatively, you might get away with artificial dataset, using something like Retouch4Me plugin at a moderate strength. It's not ideal, but it's definitely much better than frequency separation technique that produced what appears to be 90% of the dataset here.