With Pony as a base, the goals were to get real/true anime like characters by lessening the 2.5D/3D in order to get less shading/shadows and get smoother coloring of the skin. The second goal was to get prompt simplicity similar to that of SD 1.5, While this model still allows most pony prompts including quality and source ones. They are not needed and instead the use of the recommended settings is better. This model began as a trained model, but I soon found it better to train loras and merge them in. It is not perfect yet, but this has better results than anything I could find for these use cases.
Recommended settings:
Steps: 20
CFG: 7
Sampler: Euler A
Resolution: 1024 x 1024
with HiRes Fix by 1.5 up to 1536 x 1536, Latent upscaler and 0.7 Denoising
Prompts:
"lora-trigger-words, image/character descriptions, anime screencap, anime coloring, anime screencap"
Negative Prompts:
"3d, 3d, shadows, shades, 3d, 3d, 3d, realistic, shiny, shiny_skin, tan skin, shine"
If the character you are looking for needs a darker skin, remove "tan skin" from the negatives.
Check sample images for example prompts.
follow me for more content:
https://linktr.ee/AIOverlord
Description
FAQ
Comments (5)
Can this be made available to run onsite?
I have no control over that, sorry.
you should upload more sample images bro, like in different poses and concepts. since there are no on-site generation, you will definitely get more downloads, this seems really good.
Thanks for your interest. Honestly I did not want to put too much effort into this version as I deem it incomplete. It gives pretty decent results but lora usage varies a lot based on their training data(tags used). I don't have a good gpu so I can only do so much with free resources, so progress is slow and I havent been focused on generation as much. However, given that im not the only one interested, I will look into it.
@AlOverlord I understand bro. I'm actually loving this model. My gpu is quite horrible, i can only manage to train lora with it. But i merged this checkpoint in with a few to test out the results, I'll be uploading it in Civit Ai as a merger. Will include a link to the orginal as well. Cheers Mate!!
Details
Files
Available On (2 platforms)
Same model published on other platforms. May have additional downloads or version variants.