Two new Engadget articles discuss research projects that train neural networks to in one case generate 'fake celebrity' photographs & in the other to try to give photos taken with smartphones 'DSLR quality.' Obviously, the results are far from perfect but imagine how as this & related technology becomes more mature it could be applied to graphics software in the not so distant future.
Inpainting & auto-trace algorithms could "see" more than the pixels in a raster image & adapt accordingly; color corrections & RAW development could create 'looks' otherwise impossible for existing adjustments to achieve; blending elements from different photos could be done 'automagically' to correct shadow & perspective differences; & entirely new types of filters could be developed that could do things like provide time of day adjustments.
The mind boggles.