This is a style model that does tworships style art, the artist tends to draw mature faces, large breasts, and school uniforms. Tends to work best at 0.6. This model was trained on NAI anime-full-pruned and generated in AnythingV5. Made with the Kohya tool for AI Training through google colab. Any suggestions to improve or recommended things for the data set would be appreciated.
You can find the artist here.
https://twitter.com/tworships?lang=en
Description
Trained on more than 300 images and a separated 75 headshots.
credits:
-EDG for several outfit lora's
Some things that I plan to work on when I can:
-Update to the cutesy style
-Lunatic from black rock shooter Dawn Fall (if no one beats me to it)
-Update to the Trapped in phone
FAQ
Comments (5)
I'm a noob at this and am having issues generating images like this. I'm using automatic1111, have the lora installed, am using the same prompts with the same seeds, but all the images are super generic. I know the lora is doing something, because when it's removed the image changes. Was there anything different that was needed to get images to generate like this?
Well welcome to stable diffusion I'll try to tell you any tricks to it and hopefully one solves your problem.
First of all check your model and make sure your not using a realism, most any anime model should be able to use this style. The cover images used AnythingV5 which I think the version I used to use isn't around but you can use anylora https://civitai.com/models/23900/anylora-checkpoint, I can confirm that this works in cuteyuki special chapter since my recent generations have been using that one https://civitai.com/models/28169?modelVersionId=195410, but again any anime model should be good to go.
Next make sure your hi-res is on in the main window I recommend setting it to 'Latent (nearest-exact)' and make sure your step count is at least 12-20 but you can leave the hi-res steps at 0.
Next thing to check is to use a vae images without a vae look dim and washed out cause there lacking color. I tend to just use the orange mix vae, but thats personal preference https://huggingface.co/WarriorMama777/OrangeMixs/tree/main/VAEs.
If its none of those you can check your sampling method in the main window I recommend 'DPM++ 2M Karras'.
Finally and most importantly make sure your using the lora correctly, this lora has no keywords so your good there. but make sure the file is in the lora tab and when clicked you see somthing like this in your prompt window '<lora:tworship-10:0.6>' I recommend 0.6-0.8 for best usage.
Also I doubt this is your problem but this is a stable diffusion lora not XL or pony.
Hopefully something in there helps.
@rulles Your first suggestion did the trick. I was still using SD1.5 as the base. I've changed the base and now it's making much better images. Thanks!
For the resolution, do you set them to 1024x1536 to start, or do you start at 512x512 and modify them from there? I get some strange results at 1024x1536.
@itman3333 my images end at 1024x1536 after hi-res, the image is set to generate at 512x768 and when it goes through hi-res it doubles.