Experimental mix. Uses zero terminal SNR and v-prediction (https://arxiv.org/abs/2305.08891)

It can make 2D characters on photo-like backgrounds, 2D art, or semi-realistic art. Prompting specific character will make 2D face, prompting "1girl" without name will make semi-realistic face usually. Model knows a lot of vtubers.

You need to download .yaml config (in Files tab on the right) and put it into the same folder as model.

Recommended settings:

Recommended CFG Rescale: 0.5-0.7 (use https://github.com/Seshelle/CFG_Rescale_webui extension for A1111, or sampler_rescalecfg for comfyui). Depending on other settings, 0-0.5 may work too. 0 is better if using a lot of LoRAs.

Sampler: DPM++ 2M Karras or Restart, CFG scale: 7-9, Clip skip: 2, ADetailer for faces, VAE: WD or ft-mse-840000

For nice landscapes try tags like "(photo background), scenery, shadow, reflection" and horizontal resolutions like 768x576