Starting with this txt2img I wanted to make a Sunny Winter Wishday picture, so after some liberal usage of img2img I got something a little less burnt out. In order to add the beanie (because of course no AI has knowledge of something so recent) I had to photoshop a rough approximation (and to a lesser extent the scarf) on to then feed into the inpainting tool. Of course i’m unhappy that I was unable to capture to pale lime green colour or to maintain the string from which the star hangs, still I did get one ugly green toque. After that minor things like the star latch on the satchel were added, then the image was upscaled to its current resolution from 512x768. At that point I broke the image up into 1024/1024 chunks in photoshop and regenerated aspects that were artifacted, overlapping or otherwise felt improper.

For a long time I was suspicious of the line of the belly into the hind leg and whether or not it should be visible, which resulted in a painful hour of outpainting and regenerating until I had that anatomy correct, placing it over the original images frame and… finding out yes the original was apparently fine. funny world.