SDXL-Anime | 天空之境

注意事项（必看）：

1.原始分辨率请设置为1024x1024以上，由于画布较大，prompt要尽可能的多一些，否则会崩坏，Hiresfix倍数可以调低一些，Steps: 25, Sampler: DPM++ SDE Karras, CFG scale: 7,Clip:2

2.请务必在出图后对图片使用SD upscale进行分块放大，此步骤可以显著提升图片质量（具体参数见下方评论区）

3.SDXL系二次元模型拟合相对困难，很难做到所谓的“1girl出好图”，所以对使用者的prompt书写能力会有所要求。prompt学习指路→元素同典：确实不完全科学的魔导书  https://docs.qq.com/doc/DWFdSTHJtQWRzYk9k -元素法典制作组

4.请务必使用SDXL的修复版VAE: https://huggingface.co/bdsqlsz/stable-diffusion-xl-base-1.0_fixvae_fp16 -青龙圣者
5.训练底模推荐： https://civitai.com/models/125721/finalyuananimeandcg -【Final】制作组

模型介绍：SDXL-Anime | 天空之境 是个人探索SDXL二次元风格训练流程的大胆尝试。重新制定了训练流程，使用了1000+张高质量图片，经过6次迭代总计60w步的训练，3.0版本已经可以在SDXL上还原1.x系模型的画风特化效果，并且拥有更强的笔触还原，更泛化的脸部，更严格的prompt要求以及更广的风格词兼容

特别鸣谢：风吟的算力支持，T4的镜像支持，秋麒麟的底模支持

Precautions (must see):

1. The original resolution should be set to 1024x1024 or more, due to the larger canvas, the prompt should be as much as possible, otherwise it will collapse, Hiresfix multiplier can be adjusted down, Steps: 25, Sampler: DPM++ SDE Karras, CFG scale: 7, Clip: 2

2. Please be sure to use SD upscale to enlarge the picture after the image, this step can significantly improve the quality of the picture (see the comment area below for specific parameters)

3. SDXL is relatively difficult to fit the quadratic model, and it is difficult to achieve the so-called "1girl out of a good picture", so the user's prompt writing ability will be required. Prompt Learning Guide→ Elemental Classics: A Magic Guide Book that Is Really Not Fully Scientific https://docs.qq.com/doc/DWFdSTHJtQWRzYk9k - Elemental Codex Making Team

4. Be sure to use SDXL's restored VAE: https://huggingface.co/bdsqlsz/stable-diffusion-xl-base-1.0_fixvae_fp16 - Blue Dragon Sage

5. Recommended training base: https://civitai.com/models/125721/finalyuananimeandcg -[Final] production team

Model description: SDXL-Anime | Sky Realm is a bold attempt by individuals to explore the SDXL two-dimensional style training process. Reformulated the training process, using 1000+ high-quality images, after 6 iterations of training with a total of 60w steps, version 3.0 can already restore the style specialization effect of the 1.x series model on SDXL, and has stronger stroke reproduction, more generalized faces, stricter prompt requirements and wider style word compatibility

Special thanks: Fengyin's computing power support, T4 image support, and Qiu Qilin's base mold support