High resolution AI pictures
A recent update of NovelAI brought new functions for large images: among the standard sizes, the small size is not listed anymore, and there is a new large+ size and a wallpaper size. The latter comes with a whopping 1920x1088 pixels and is prohibitively expensive: 20 iterations cost 30 units (which is around ten times as much as a normal picture).
I've tried out the new sizes. They have, however, a problem... as you can see in these pictures for the prompt:
Azuki is a girl, {{older}}, looking to the side, very long hair, pink hair, blue eyes, smile, hair ribbon, {{pink skirt}}, white blouse,blue collar,{{sneakers}},looking to the side,small breasts,{{older}}, {{cat ears}},eye focus, legs crossed,school uniform, cherry blossoms, flowers,bushes, light rays, sunlight,floating hair, pose, butterfly ,sneakers,blue sky
Instead of just one Azuki, we suddenly have three of them! (That they have six legs in total is good, the distribution of the legs to the Azuki, however, could be improved...)
It seems that the huge wallpaper canvas lets the AI draw more than one copy of the motif. Is this better in large+?
Yes, it's better: we are down to two Azuki now. (The legs are down to three and a half.) These problems disappear when we go to normal size:
And if we increase the iterations (which is still cheap in that size) we end up with a pretty decent picture (just that the butterfly got lost):
If I were to use this picture I would now further improve it, in particular the hands and the color of the skirt, but these are fairly minor details. Overall it looks very nice already! – However, it's only 768x512 pixels.
I tried finally another approach: I uploaded this picture as a starting point, used the same prompt and run the Ai with a strength of 0.75. Okay, the result is highres (1536x1024), has only one Azuki on it, but the poor girl had an accident with her leg and her breast and some other minor injuries. At least the butterfly is resurrected...
Finally, I've applied good old waifu2x on the 768x512 picture. Here's the result, and Azuki and I think that is still the preferred method to produce high res pictures...
AI still isn't up to the task to generate high res pictures from the scratch!