Chinese Animation: Thousand Illusions | Xian Ni - Li Muwan (Newly Established Model)

Recommendation weight: 1.0

Recommendation sample:DPM++ 3M SDE Karras iteration around 40

Trigger word for Limuwan: Limuwan

Image restoration for Limuwan: a girl with long black hair wearing white tulle and  blue long skirt, clothing colors can be customized

Anime original scene: outside the heavenly palace

Parameter recommendations

①For SDXL output, please do not enable VAE, output should be larger than 768*1024

②Weights: text to image 0.9, image to image redraw can be 1.0, better transformation 0.9

③Sampling: DPM++ 2M SDE Karras (for finer output), iterations 25-45. Euler a sampling can also be used to improve the accuracy of text-to-image but output may not be as detailed, especially for backgrounds

④CFG: 6.5. In special cases, try 6.0 or 7.0, CLIP: 1

⑤HD restoration: algorithm R-ESRGAN 4x+ Anime6B, around 20 iterations, HD redraw around 0.4, increase redraw if blurry, decrease if distorted

⑥AD face retouching: algorithm face_yolov8s.pt, redraw level around 0.35, adjustment techniques similar to HD restoration, not recommended for close-ups

⑦If AD face retouching still cannot restore the face shape, recommend using redraw with image-to-image generation or upscaling scripts, redraw level 0.3-0.4, use Euler a sampling

Other suggestions

Model recommendation for output: For 3D, recommend using DreamWorks Industrial V5, V4, 2D output not yet tested

Aesthetic transformation: can use positive descriptors (eye makeup: 0.7), NSFW, cleavage, shiny skin, and other aesthetic vocabulary

Common positive words: highly detailed, ultra-high resolutions, 32K UHD, best quality, masterpiece

Common negative words: bad hand, blurry