This model is based on Anima General Furry Multi-Style/Concept LoRa (160kE6Anima). Thanks @garaggge for this awesome LoRA.
160kE6Anima "was trained on around 160k images from e621". It has extended knowledge and 3k+ style tags. Check its model page above for more info (list of styles, etc.).
About RDBT model and dataset: https://civarchive.com/models/2356447.
For best diversity, both 160kE6Anima and RDBT does not provide overfitted default style. You MUST add style tags in prompt, or stack style LoRA. Otherwise you will get undefined averaged style.
Versions:
v2 base calibrated:
"Calibrated" from RDBT v2. The final ~10% steps was trained on a ckpt with the 160kE6Anima LoRA merged. This fixed inner conflicts when you stacking two LoRAs trained with large datasets. Improved the model quality to the same level the RDBT model should be on original anima base.
Settings: CFG 4, 24+ steps, 1024px.
You don't need the high-res LoRA (mentioned in 160kE6Anima page because it was trained at 512/768px). The RDBT model was calibrated/trained at 1024px.
Description
FAQ
Comments (4)
Edit: (score_2:20) worked. Regular anima only needs about up to 3 prompt weight, this checkpoint just needed a lot more.
Original: I dont know how to just get 2D digital art.
I seem to be getting mostly either pencil-art/traditional or 3D.
"2D digital art" just seems to increase chances of 3D.
You need very specific style prompt, like @artist tags, or style LoRA. "2D digital art" is too ambitious. This model doesn't provide overfitted default style.
@reakaakasky nah it worked with putting a lot of weight on score_x tag
@artist tags works well too though.
This model has potential to become the SeaArt Furry XL 1.0 of Anima Based Models.
I like that finally there's a Furry Anima based that isn't a Fine-Tune Merge but a Trained Fine-Tune.