Maya Yotsuba from Mahouka Koukou no Rettousei (魔法科高校の劣等生)
This model was trained on anime screencaps and two game CGs.
Base Model: Pony Diffusion V6 XL (https://civarchive.com/models/257749?modelVersionId=290640)
Prompts:-
Trigger Words:
maya_olis, 1girl, black hair, drill sidelocks, drill hair, purple eyes, mole under eyeOutfit:
cleavage, red dress, long dressAccessory:
earringsExtra / Optional:
solo
anime coloring (can slightly alter the style, but may not have a significant effect)
(you may refer to the sample images)Negative Prompts (optional):
blurry background
chromatic aberration (can reduce the chance of her having hair highlights like those in the last season of the anime)Known Issue:
Random clothes with long sleeves could have "sleeves past wrists" that look like those in her outfit. If any of you mind, I can train another model after adding the tag sleeves past wrists to my dataset. Doing this should solve the problem, I believe.
Weight: 1 (feel free to lower it, especially if you don't care about her outfit)
Afterword:
I have been experimenting with Euler Ancestral CFG++ sampler and the Detail Daemon extension in my sample images. With this sampler, I can safety set a high Detail Amount to increase image details.
Hopefully some of you will like this model.
Description
FAQ
Comments (10)
Please make SD1.5Hi,
Someone already did it:
https://civitai.com/models/342686/maya-yotsuba-mahouka-koukou-no-rettousei
your sample picture are amazing!!!
where can I get Euler Ancestral CFG++?
and how can you use CFG with decimal? (the smallest value that I can use it 1)
when I set it as 1, the image look more like anime style(I like it) but it will become blur, but in your sample image, it looks also anime style "and" not blur. how to do this?
Hi, thanks for your comments.
First, Euler Ancestral CFG++ has been added to WebUI reForge recently (https://github.com/Panchovix/stable-diffusion-webui-reForge). I'm not sure if AUTOMATIC1111's WebUI has it.
Second, it was suggested to use 0-1 CFG with CFG++ samplers. To use a lower CFG in WebUI, you have to edit the ui-config.json file located in your WebUI folder. The settings you need to adjust are:
"txt2img/CFG Scale/minimum"
"txt2img/CFG Scale/step"
"img2img/CFG Scale/minimum"
"img2img/CFG Scale/step"
Remember to restart the WebUI and refresh the WebUI web page if you have them on.
Third, don't trust some of settings in my sample images, such as Steps. This is because I did img2img upscaling twice and then manually inpainted the eyes (iris only actually). The original prompts were gone after inpainting. Therefore, I decided to regenerate the image with a denoising value of 0 to preserve them, and I used a random sampling step for this. One more thing, people recommended using 50 sampling steps with CFG++. This takes longer and may not suit you.
@JasonCMK
thanks for your reply!!!
1. I'm using origin Forge, it didn't have Euler Ancestral CFG++
I can find it in current reForge. Let me try to use it!!!
2. it works, thanks!
@badkevin
No problem.
Though I haven't noticed a clear quality increase over Euler A after spending more time experimenting with the CFG++ version.
@JasonCMK
thanks for your help :)
I can generate her like anime style so much
but one more quest.
1st step, I use t2i with 512x768 and hires 2 (= 1024x1536).
the picture looks good (human logic always good)
2nd, I try to use i2i with 1024x1535 noise 0.5~0.7 resize 1.5 (= 1536x2304)
the picture 80% good but sometimes the human logic is wrong (ex:2 head, 3 hands)
not sure it is CFG problem or because I try to generate big size picture :(
(both cases use CFG 0.6 and Euler Ancestral CFG++)
@badkevin
I usually use denoising strength <=0.4 for the second i2i.
@JasonCMK
anyway, I tried to use the same prompt and checkpoint and other parameter.
except sampler is Euler Ancestral CFG++ and Euler a.
only Euler Ancestral CFG++ can reproduce anime style perfectly
@badkevin
Maybe you can try adding "3d, realistic" to negative prompts if you want a more accurate anime style. Besides, I have seen people suggesting removing "score_8_up" from positive prompts. I am not a pro with the score tags, though. Also, add "anime coloring" to the positive prompts if you haven't already, since I tagged the anime screenshots with it.
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.