yadaさんのイラストまとめ

sd

sd

sd

working on an actual anime-ish #sdxl finetune, work in progress with around 350k images for training / 1.2m images seen:

sd

sdxl

First tries on SDXL.
from the paper it's using penultimate layers so it's probably better to add clip skip2 on comfy? Didn't check the exact impl here. The last sd1 i2i is to remove the refiner "artstation-ish" bias. seems to work pretty well for hand fixes
#AIイラスト #AIArt

#sdxl

#AIイラスト #AIart

基底图:
- text -> 448p (DDIM 50it cfg 5.5 model a)
- 448p -> 704p (eulera 20it cfg6 model b)
然后用第三方blend node可以两张图叠加在一起, 再i2i后升分辨率, 总共4pass
说实话sd能有个blender一样的界面还是很舒服啊, 加上完善的math node基本就是真 · procudural了(
https://t.co/I4z30iD4M6