Ivy Anima (aesthetic lora)

Ivy Anima is a LoRA finetune of Anima that aims to improve the general quality and aesthetic of the model. More specifically, the focus is on improved base styles/aesthetic, poses, lightning and backgrounds.

This LoRA is trained using over 750 manually selected images with a focus specifically on anime-like illustrations with one or more characters. The dataset leans towards SFW illustrations and mainly female characters.

Usage

When it comes to settings, you should generally follow the same guidelines as for the base model. er_sde with a CFG of 4-5 and 30-50 steps is a good default. As with the base model, euler_a and euler_a++ lean a bit more towards '2.5D' images.

For prompts, natural language may work best but tags (or a combination) will also work. No trigger word is needed. I prefix my positive prompts with the following:

newest, masterpiece, absurdres

The following negative prompt is a good baseline:

score_1, score_2, score_3, worst quality, low quality, jpeg artifacts, old, early, artist name

Known Issues

Generations may contain unprompted chromatic aberration. You can optionally put this in your negative prompt to prevent this. This can also help prevent the model from breaking in some cases.