When using 1-boy pants as a prompt, women's leggings will dominate in image generation.
So... Are you gonna give us some prompt or anything, OP?
Sorry, I didn't explain. The image I uploaded includes the workflow in ConfyUI.
@ starshadowmagician, @ mgong520, @ naoko \(naonocoto\), @ rinotuna, @ suiso \(owp\), @ sheya tin, @ kalak 39,
1boy, solo, expressionless, (blush:0.5), tachi-e, standing, grey background, simple background, jacket, short hair, shirt, :D, full body, pants,
and
monochrome, darkness,
Bad prompt. Spamming artists doesn't work well, this model can't mix artists well. Using only 1-2 artists and leggings are much less frequent. Or you can also just put leggings in the negative or you can just describe the image a bit more, like specifying blue pants.
Leggings are, ultimately, pants.
I certainly don't think tight pants should be the default
By default, male characters do not get leggings. Just try it out yourself and see... The default is not prompting with 7 artists.
By default, when not specifically mentioning leggings, male characters do not get leggings.
Red - leggings
Yellow - unsure
Green - pants
Top 30: No negative, no artists, with your prompt.
Middle 30: Your negative, no artists, with your prompt.
Bottom 30: Your negative, 1 artist (kalak 39), with your prompt.
Instead of pointless redundant custom nodes, I used torch compile and spectrum https://github.com/ruwwww/comfyui-spectrum-sdxl . Still with anima preview 2.
66 green, 14 yellow, 10 red. Even if we assume every time I was unsure if they are leggings that they are, that's still ~3:1 regular pants:leggings. Very much default not leggings.
Your last 3 images have no workflow. Why? I am not checking every single image. Why do you keep using 1600x640? The resolution used for SDXL models, which implicitly anima was trained on, was 1536x640.
At best, it's possible the model performs worse at an extremely tall aspect ratio, however I am not digging through danbooru for tall images with those specific artists/other parts of the prompt to tell if that is normal or not.
Thanks for your testing.
I used 1060 × 640 in NovelAI, so I used it in Anima. I didn't realize the problem.But I will try 1536 × 640.
And yes, I think you're right—the default is not leggings.
As you said about the leggings issue, the model doesn’t mix artists well.
I will give up on mixing artists in Anima.
I will give up on mixing artists in Anima.
You should also avoid weighting things above 1.0. Weighting below 1.0 will work roughly as expected, but above 1.0 will not work well. The poor weighting and artist mixing are due to this model not having CLIP.
The weighting might be fixed in the future, there might be other ways to do it than how it was done for CLIP. Artist mixing is less likely to be fixable.
Style loras are more mixable.