Download
1 variant available
The Anima Model is licensed by CircleStone Labs LLC. Copyright CircleStone Labs LLC. IN NO EVENT SHALL CIRCLESTONE LABS LLC BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH USE OF THIS MODEL.
Built on NVIDIA Cosmos
Activator is @geffstyle
Weight of lora - 0.7-0.8, at 1.0 it usually becomes sushi yuusha for the most part.
Overview:
A style mix made from mostly my generated images - mostly based on this collection. Main styles in mix are Sushi Yuusha Toro/Karasu Raven with some addition of others, and even some 3d-like images. Idea - to try to make a tool that can make some things look like scribbles, while keeping others more or less clean (mostly a combination of "scribble-like-background + normal looking character")
Usage:
Put it into LoRA folder in ComfyUI / A1111 (I heard one of versions supports Anima), I guess, and use xD
Weights - experiment. In images above I used different weights. At 1.0 it looks much closer to SYT, try first at 0.6-0.7. I added more comparisons in preview images
Technical info:
Batch size 4, 800 steps, 90 images in dataset + 90 generated with anima from descriptions of these images as a regularization dataset. I used Anima TrainFlow by @ThetaCursed - thanks to them for that tool and for the Style Explorer too xD
All the settings were default (Prodigy, lr = 1 and so on). For tagging I used combination of WD14 + JoyCaption, the first one gives base for image, while the second allows to recreate the positioning and things like that. DM me if you need some more info
Thoughts:
0) "Waow lora trains and works 0_0"
1) Not ideal, leans too heavily to scribbly style
2) Adding images generated from the same exact prompt - but without the style - is a great way to make sure the activator actually does something
3) Sometimes WD14 + JoyCaption allows Anima to recreate image (save the style) almost identically