Long time no see!
Don't worry, I've been busy ;)
I've done some fruitful experiments with Machine Learning - Super Resolution.
Here're the results!
The images go in the following order:
- a lowres 800x800 input image
- a bicubic upsampling to 1600x1600
- a 'preserve details' upsampling to 1600x1600 (Photoshop)
- model's prediction from input 800x800 image
- ground truth hires image