Krea 2 Anime is a Krea 2 Turbo's fine-tuned model.
It is a merge between a mix of loadstone's Kroma v0.3, Krea 2 base turbo and rank-256 lora trained with over 20k diverse anime images. The refusals should also be tuned quite down.
Recommended Settings
Resolution: buckets of 1024 - 2048
Steps: 8-12
Cfg: 1.0
Sampler: Euler A
Scheduler: Beta
Prompting Guides:
(Type of art) (Character) (Location)
(Clothing)
(Actions)
(Style)
Example:
A polished digital cowboy shot image from thigh up of a black haired cat girl in a cafe.
She wears a maid dress with detached collars.
She is holding a broom sweeping the floor.
This is a highly detailed digital anime illustration rendered with soft gradients.
v1.4:
Mainly just aesthetics improvements and better prompt adherence at shorter length.
NSFW prompts is possible without further tuning, but this model is generally not built for it so your mileage may vary.
Due to hardware and financial limitation, the dataset is still limited so certain style and concepts can bleed depending on the character appearance or poses.
Future work:
Getting a more stable default style while allowing for more fine grained art style changes.
Credits:
https://huggingface.co/lodestones/Kroma
https://huggingface.co/krea/Krea-2-Turbo
License:
https://www.krea.ai/krea-2-licensing
Description
Tuned the style to better fit my vision. I skipped posted v1.1 - v1.3 because they had too much weird quirks.
FAQ
Comments (5)
1.4 seems to be less capable than 1.0 and is also more likely to censor your prompt regardless.
Thanks for the feedback. Mind sharing more details regarding the capability drop?
feel the same. The style is better, but it's less likely to follow the prompt.
@FarewellYhl Thanks. It's likely due to the style merge, I know what to work on now.
Very good at complex composition, poses, and positions. The weak parts are the artist style, and also the model doesn't know many characters.
I will rate it as 8/10, since the model can do fine pose tuning that other models like Anima, SDXL, or NovelAi cannot do.




