V 21 Realistic loras a bit overtuned
V 21
V 20
V 19
Same workflow same prompt and that tested with a flat anime color lora at 0.3 strength on top of it when I asked for an anime illustration.
Edit: I figured out what happened here. I had the words "highly detailed" in my Prompt to help improve the quality of the image. In V 19 and V 20 Worked fine. But in V 21 it likely triggered one of the inbuilt realisms Loras.
Still had to increase the Anime lora strength to get it closer to a proper anime style like in V 19 which started affecting prompt adherence.
I did a bit more testing and noticed that V 19 followed the pose prompt a bit better, example: I asked for both fists up.
V 21
V 19
I'll try to tweak my loras a bit but you could consider adding at least one or 2 anime loras, even if at low strength for V 22.
The Lora I use is Anime Flat Style by Agino.
https://civitai.com/models/1952560?modelVersionId=2559263
I'm not sure there is a great balance between anime and realistic Qwen Image Edit. Raw Qwen Image Edit prefers "waxy" and plastic skin, but great anime. Adding the realism LORAs pushes it toward more realistic generations while still allowing anime (but does make it harder to have completely flat 2D anime).
Also, just use v20 or v19 if you prefer. There is no reason to use v21 if you think v20 fits your style and prompt adherence better. I'm not sure it is going to be possible to have both perfect anime and perfect realistic styles at the same time without any additional LORAs.
I've updated the notes on the version history to mention v21 does tend more realistic styles, something that I think was lacking from v20 and v19. However, that means if you want anime, v20 and v19 is better.