OIA Illustrator

This is a large illustration-style model developed by training and integrating with the Flyingsnow 2.0Hill base model. During the training, 1500 high-definition materials were used, comprising 500 each of gouache, watercolor, and flat illustrations. The final tuning was slightly biased towards gouache. As for the style, let's tentatively call it: 'what's style'. Anyway, hope you like it. The actual effect of the generated images can be seen in the sample.Recommended Sampling Method: Use DPM++2MKarras, Eulera, or Eulera. All of them can quickly produce high-quality images; Sampling Steps: Suggested 20-30, unnecessary to waste time with too many steps; CFG Scale: Recommended 6 to 9; Clip Skip: Suggested 2 to 5, though higher values like 6 to 10 can sometimes offer surprises; VAE: Built-in; Hiresfix: After obtaining a satisfying result, enable Hiresfix for enhanced clarity, saving a lot of time. Upscaler: Recommended using 4x-AnimeSharp, R-ESRGAN4x+, R-ESRGAN4x+Anime6B; Sampling Steps 20-30 suffice; Denoising Strength: Recommended 0.3-0.4 (too high values may deviate from the original image) Output Size: 768x512, 512x768, 512x512, or 512xN, Nx512 (where N represents any number). That is, retaining a width or height of 512 yields better results. Since the original materials of Stabledifussion1.5's base model were 512x512, keeping one dimension as 512 enhances the outcome (true for all 1.5 models).这一个以Flyingsnow 2.0Hill为底模训练+融合而成的一个插画风格的大模型。模型在训练的时候，采用了1500张高清素材，水粉、水彩、平面插画各500张，最终调教的时候，稍偏一点点水粉，至于像什么风格，暂且叫：what's style吧。Anyway，希望你们喜欢，具体生图效果看样张。 Sampling Method（采样方法）建议使用DPM++2MKarras、Eulera或者Eulera。都可以速度跑出高质量的图片； Sampling Steps（采样步长值）建议20-30，不需要太多的步长值浪费跑图的时间； CFG Scale（提示词相性）建议6到9； Clip Skip建议2到5即可，太高的值有时候也给你一些惊喜你可以试试6到10； VAE：已内置； Hiresfix（高清修复），抽出自己满意的效果后，再打开Hiresfix（高清修复）。这样可以省去很多时间。 Upscaler（修复方式）建议使用4x-AnimeSharp、R-ESRGAN4x+、R-ESRGAN4x+Anime6B； 采样步长值20-30即可；DenoisingStrength（重绘强图）建议0.3-0.4（太高的值会偏离原图） 出图尺寸：768x512、512x768、512x512、或者512xN、Nx512（N代表任意数）。即，保留、宽高有一个值为512，会得到更好的效果。因为Stabledifussion1.5的底模训练的原始素材就是512x512的，所以保留一个值为512会得到更好的效果（所有1.5大模型全部如此）