[Learning Model]
[Generation confirmation model]
JANKU v4 NSFW Trained + NoobAI EPS + RouWei Illustrious XL (v4.0)
Unholy Desire Mix - Crimson Seduction (NoobAI)
私のおすすめの生成方法は表現が豊かなJANKUを使用し、LoRAのウエイトを0.6ぐらいまで落として元の画像を生成。それをよりLoRAの情報を反映させて再現性が良いWAIを使用してi2iでアップスケールする方法です
My recommended generation method is to use the expressive JANKU with the LoRA weight reduced to around 0.6 to generate the base image. Then, upscale it with i2i using WAI, which better reflects the LoRA information and offers higher reproducibility.
[Generate positive prompts]
highres,very detailed,amazing quality,masterpiece,very aesthetic,anime screencap,shiny skin,
[Generate negative prompts]
3D,deformation,sketch,wip,aax,(worst quality),(low quality),(normal quality),lowres,normal quality,monochrome,grayscale,watermark,dull quality,3d model,flat colour,no shading,
[Creation parameters]
Clip skip : 1
Sampling method : Euler a
Schedule type : Karras
Sampling steps : 30
Width : 896
Height : 1152
（1024,1024）
CFG Scale : 5
[Upscale parameters]
Clip skip : 1
Sampling method : Euler a
Schedule type : Karras
Sampling steps : 30
CFG Scale : 5
Denoising strength : 0.3~0.6
[ControlNet v1.1.455]
Pixle Perfect : on
Preprocessor : tile_colorfix+sharp
Model : control_v11fle_sd15_tile
Control Weight : 0.6
Script : SD upscale
Upscaler : R-ESRGAN 4x+Anime6B
LoRA LIST
https://docs.google.com/spreadsheets/d/1NRgxEhNyCmYIJfT8429uSv4BCE_IwSKr8stTsDUk2-M/edit?usp=sharing