This is a v-prediction model, please make sure you're using a compatible UI.
I wanted to make a merge that gave a solid 2D anime screenshot base style without needing to throw in a ton of extra tags. See galleries for my recommended prompt structure, all of my images are a single pass with no upscalers or refiners.
All credit goes to the original checkpoint authors and whoever they credit for their own models.
Recommended Settings
V4
Steps: 20-30
CFG: 3-5
Sampler: Euler Ancestral
V3
Steps: 20-30
CFG: 2-3
Sampler: Euler Ancestral
Description
Sorry for another update so soon, this should be the last one for a while!
Updated the merge recipe for better hands without hand negatives + a less AI-shaded aesthetic. New settings in model description.
FAQ
Comments (9)
Thank you very much for all your updates and for always aiming for improvements!
I really like this checkpoint. Thank you once again for sharing it.
Wish you a great day!
The best anime checkpoint by far
can you share a list of characters or tags that is used to train this model?
it should be everything with more than a couple hundred tags on danbooru, so lots! I don’t have a list though
@aimaliasart can you share the tag files? like the text files only?
@krigeta here you go! The tags beyond danbooru2024 might be diluted, but I've tried some characters with under 100 images there and gotten decent results: danbooru2024, e621, e621_newest, yande_full
@aimalias yo thanks for this!
It always bugged me how most checkpoints opt to go for as much shading as possible, even anime-themed checkpoints, and other checkpoints or loras tend to still have some excess shading that just doesn't fit the anime theme at all.
When I saw this, I felt like it might be the same case, but no; it took little effort to achieve the intended result, and the results come out looking like TV broadcast episodes with minor imperfections. The shading is accurate to modern anime, and the colors are vibrant even at low CFG. This quickly became one of my favorite checkpoints after only half an hour of finding that sweet spot. I'm sad it took me so long to find this checkpoint!
For some reason this model appears to be good at nearly everything. Not just anime screenshots (though its really good at that). This thing can do 3d, and somehow just works with any lora I throw at it. Literal magic.