模型介绍（Chinese Part）

I 引言

在本介绍中，您将了解：

1. 模型介绍（见 II 部分）；

2. 使用指南（见 III 部分）；

3. 训练参数（见 IV 部分）；

4. 触发词列表（见附录 A 部分）

II 模型介绍

动漫插画设计XL，或称 AIDXL 是一款专用于生成二次元插图的模型。它内置了 200 种以上（随着更新越来越多）的插画风格，依靠特定触发词（见附录 A 部分）触发。

优点：构图灵活多变，没有摆拍感，主体突出，没有过多繁杂的细节，认识很多动漫人物（依靠角色日文名拼音触发，例如，“ayanami rei”对应角色“绫波丽”，“kamado nezuko”对应角色“祢豆子”）。

III 使用指南（将与时俱进）

推荐使用 ComfyUI 生成图像。工作流参数如图 1 所示。

图1：ComfUI 工作流

1 提示词书写

如您所见，SDXL 的正负面提示词都分为 text_g 和 text_l 两个部分。假设您用于生成图像的描述提示词为 {main_prompt}（不包含质量描述语），质量提示词为 {quality_prompt}，负面提示次为 {neg_prompt}。则就经验来说，正面提示词的 text_g 处填写 “{main_prompt}”，text_l 处填写 “{quality_prompt}, {main_prompt}”；负面提示次的 text_g 和 text_l 都填写 “{neg_prompt}”。

另外，为了强调生成图像的风格，风格触发词的位置应尽量提前，例如， by xxx, 1girl, solo。

1.1 示例

质量提示词示例： a masterpiece with best quality, detailed 4k illustration

负面提示次示例： (low quality:1.4), bad anatomy, worst quality, ugly, disgusting, fused hands and fingers, bad hands, deformed hands and fingers, blurry, lowres, malformed, morbid, overexposure, too colorful and over-saturated, chromatic aberration, jagged

1.2 工作流

我在 Suggested Resources 内添加了 ComfyUI 的工作流 json 文件。

2 生成参数

建议图像总分辨率（总分辨率=高度x宽度）大于 1024×1024，否则生成的图像可能质量不高。不进行“Clip Skip”操作。采用 “dpmpp_2m” 采样器（sampler），搭配 “karras” 调度器（scheduler），该组合在 webui 里称为 DPM++ 2M Karras。在 8 CFG 上采样 50 步。所有涉及图像宽度（width）和高度（height）的参数均应保持一致，即所有宽度一致，所有高度一致。

IV 训练参数

以 SDXL1.0 为底模，使用大约 2w 张自己标注的图像在 5e-6 学习率，总长为 1 的余弦调度器上训练了约 100 期得到模型 A。之后在 2e-7 学习率，其余参数相同的条件下，训练得到模型 B。将模型 A 与 B 混合后得到 AIDXLv0.1 模型。

V 对比基于 SD1.5 的 AID

2023/08/08：AIDXL 使用与 AIDv2.10 完全相同的训练集进行训练，但表现优于 AIDv2.10。AIDXL 更聪明，能做到很多以 SD1.5 为底模型无法做到的事。它还能很好地区分不同概念，学习图像细节，处理对 SD1.5 来说难于登天的构图，几近完美地学习旧版 AID 无法完全掌握的风格。总的来说，它绝对拥有比 SD1.5 更高的上限，我会继续更新 AIDXL。

Model Introduction (英文部分)

I Introduction

In this introduction, you’ll learn about:

Model presentation (see Section II);

Instructions for use (see Section III);

Training parameters (see Section IV);

List of Trigger Words (see Appendix Part A)

II AIDXL

Anime Illustration Design XL, or AIDXL, is a model dedicated to generating anime illustrations. It has more than 200 (with more and more updates) built-in illustration styles, which are triggered by specific trigger words (see Appendix A).

Advantages: flexible composition, no sense of posing, prominent subject, not too many complicated details, familiar with many anime characters (triggered by the character’s name, for example, “ayanami rei” corresponds to the character “Ayanami Rei”, “kamado nezuko ” corresponds to the character “Nezuko”, “lucy \(cyberpunk edgerunners\) corresponds to the character “Lucy”).

III User Guide

(Keep pace with the times)

It is recommended to use ComfyUI to generate images rather than WebUI for higher generation speed, flexibility and even image quality. The workflow parameters are shown in Figure 1.

Figure 1: ComfUI workflow

1 Guidance of Prompt Writing

As you can see, SDXL Clip has both positive and negative prompts divided into two parts, text_g and text_l. Assume that the description prompt you use to generate images is {main_prompt} (without the quality prompt), the quality prompt is {quality_prompt}, and the negative prompt is {neg_prompt}. The rule of thumb is, for positive prompts, filling in “{main_prompt}” in text_g, and “{quality_prompt}, {main_prompt}” in text_l. For negative prompts, filling in “{neg_prompt}” in both text_g and text_l.

In addition, in order to emphasize the style of the generated image, the position of the style trigger word should be as early as possible, or even be the first tag. For example, by yoneyama mai, 1girl, solo.

1.1 Examples

Quality prompts: a masterpiece with best quality, detailed 4k illustration

Negative prompts: (low quality:1.4), bad anatomy, worst quality, ugly, disgusting, fused hands and fingers, bad hands, deformed hands and fingers, blurry, lowres, malformed, morbid, overexposure, too colorful and over- saturated, chromatic aberration, jagged

1.2 Workflow

I added ComfyUI’s workflow json file inside the suggested resources tab.

2 Suggested Generation Parameters

It is recommended that the total image resolution (total resolution = height * width) be greater than 1024×1024, otherwise the resulting image may not be of high quality.

“Clip Skip” operation is not performed.

Use “dpmpp_2m” sampler, with “karras” scheduler, this combination is called “DPM++ 2M Karras” in WebUI.

Sample 50 steps over 8 CFG.

All parameters related to image width and height should be consistent, that is, all widths are the same, and all heights are the same.

IV Training Parameters

Using SDXL1.0 as the base model, using about 22k labeled images to train about 100 epochs on a cosine scheduler with a learning rate of 5e-6 and a total length of 1 to obtain model A. Then, using a learning rate of 2e-7 and the same other parameters to obtain model B. The AIDXLv0.1 model is obtained by merging model A and B.

V AIDXL vs SD1.5 based AID

2023/08/08. AIDXL is trained on the same training set as AIDv2.10, but outperforms AIDv2.10. AIDXL is smarter and can do many things that SD1.5-based models cannot. It also does a really good job of distinguishing between concepts, learning image detail, handling compositions that are difficult or even impossible for SD1.5 and AID. Overall, it is absolute potential. I’ll keep updating AIDXL.

Appendix / 附录

A. Trigger Words List / 触发词列表

v0.1: by 35s00, by 3meiji, by 5eyo, by 7nu, by 7thknights, by adenim, by agm, by ajimita, by akizero, by ame929, by anmi, by anteiru, by arutera, by ask, by atelier irrlicht, by bunbun, by caaaaarrot, by camu, by canking, by ccroquette, by chi4, by chicken utk, by chon, by cola, by cutesexyrobutts, by darumakarei, by dino, by dora, by dsmile9, by ei maestrl, by ekita kuro, by ekita xuan, by eku uekura, by fadingz, by fajyobore, by foomidori, by freng, by fuzichoco, by gesoking, by gomzi, by hachisan, by hakuhiru oeoe, by hamukukka, by haru, by hata, by hidulme, by hikinito0902, by hinaki, by hitoimim, by hitomio16, by hizumi, by homutan, by hotatenshi, by houk1se1, by hyatsu, by icecenya, by ichigo ame, by inoriac, by iromishiro, by iwzry, by jnthed, by joezunzun, by junsui0906, by karohroka, by kaya7hara, by kazari tayu, by killow, by kin, by kinta, by kishiyo, by kitada mo, by kkuni, by konya karasue, by kooork55, by kot rou020, by krenz, by kurige horse, by kuroume, by lalalalack, by lemoneco, by lm7, by lovelymelm, by lpmya, by mar takagi, by matcha, by matsukenmanga, by melowh, by menou, by midori xu, by mika pikazo, by misumigumi, by miv4t, by mochizukikei, by mogumo, by momoco, by momoku, by morikuraen, by mqkyrie, by muina, by munashichi, by muryou tada, by myaru, by myc0t0xin, by myung yi, by nack, by naji yanagida, by nanmo, by nardack, by narue, by nekojira, by netural, by nezukonezu32, by nico tine, by nikuzume, by nine, by nineo, by ninev, by niwa uxx, by nixeu, by noco, by noodle4cool, by nounoknown, by noyu, by oda non, by omutatsu, by onineko, by palow, by panp, by pikuson, by poharo, by poire, by potg, by pro-p, by qooo003, by rai hito, by rattan, by reiko, by rella, by rhtkd, by rin7914, by roitz, by ryuseilan, by saberiii, by sais, by sakiika, by samip, by sanosomeha, by say hana, by scottie0073, by senryoko, by serie niai, by seuhyo99, by shal-e, by shimanun, by shirabii, by shiraishi kanoya, by shiren, by shirentutu, by sho, by sia, by siki, by silver, by solipsist, by some1else45, by sonomura00, by sooon, by star furu, by starshadowmagic, by starzin07, by sui 0z0, by sul, by sushi0831, by suzukasuraimu, by taiki, by takumi bis, by teffish, by tidsean, by tira27, by tsukiho tsukioka, by tsvbvra, by ttosom, by tukumi bis, by uiiv, by ukiatsuya, by umaiyo puyoman, by void, by wait ar, by walzrj, by wanke, by whoisshe, by wlop, by xilmo, by yejji, by yogisya, by yohan, by yomu, by yoneyama mai, by yosk6000, by yumenouchi, by yun216, by yunikon147, by yunsang, by ziyun, by zumoti4