Description

Since there is less fanart of Yurisaki Mika, I trained this LoRA based on Anything V4.5. Also because of it, overfitting is seriously occured while using this model. Including but not limit to the following would happend:

• Gesture will be the same if you have low weight of gesture prompt or it even doesn't exist (You may use ControlNet to fix that)

• Pose may be strange

Limitations

And limited to the limited train dataset, the following characteristics may be not accurate:

• Halo

• Sailor collar

• Number of flower ornaments

• Number of wings

Hints

Trigger word is `yurisaki_mika`, and the recommended weight is 0.6~0.8. Here are some prompts that help you generate more accuratly:

• `mika_serafuku`: specify the costume, Heaven 3 Mika

• `mika_halo`: sometimes will generate a good Mika's halo

• `low_twintails`: seldom is needed

• `multiple_wings`

• `beret`

Surely you can add some negative prompt to avoid prompt polluting, or use some other prompts to specify Mika's costume. Generlization has been tested.

A sample prompt using Counterfeit V2.5 with VAE:

((masterpiece,best quality)), 1girl, (close_up), head_sculpture, solo, (low twintails:1.1), light_smile, <lora:yurisaki_mika_20:0.8>, yurisaki_mika, mika_serafuku, mika_halo, beret, multiple_wings, blue_sky, clouds

Negative prompt: EasyNegative, extra fingers,fewer fingers

Steps: 28, Sampler: DPM++ 2M Karras, CFG scale: 6, Seed: 167769193, Size: 512x512, Model hash: a074b8864e, Denoising strength: 0.6, Clip skip: 2, ENSD: 31337, Hires upscale: 1.75, Hires steps: 20, Hires upscaler: R-ESRGAN 4x+ Anime6B

This model was tested on Anything V4.5, Counterfeit V2.5, AOM V3, PastelMix, and all perform well as except.