这是一个基于KXL eps rev3改进的图像生成模型，使用了约780万张图像在1张4090上进行训练，极大改善了人物的肢体表现，现在更容易生成准确的人体结构，例如：更好的手指，更好的脚以及脚趾，以及更容易生成准确多人的图像，对于双人拥抱之类的动作均有改善。并且在艺术风格上更加多元，可以像Novel ai V3那样使用artist tag来生成你的喜欢的艺术风格，而且模型也可以生成更多的动漫角色。
This is an image generation model improved upon KXL eps rev3. It was trained on approximately 7.8 million images using a single NVIDIA 4090 GPU, greatly enhancing the representation of human limbs and body structure. Now it's easier to generate accurate human anatomy, such as better fingers, feet, and toes. It also facilitates the creation of precise multi-person images, with improvements in depicting actions like two people embracing. Moreover, the model offers a wider variety of artistic styles. Similar to Novel AI V3, you can use artist tags to generate your preferred art styles. The model is also capable of producing a broader range of anime characters.
描述文字
描述文字
模型发布声明 | Model Release Statement
本模型仅在以下网站正式发布：
This model is officially released only on the following websites:
1<1girl/1boy/1other/...>,2<character>,<series>,<artists>,3<general tags>,4masterpiece, best quality, absurdres, highres, sensitive, newest,
Negative prompt (short):
lowres, low quality, worst quality, normal quality, text, signature, jpeg artifacts, bad anatomy, old, early, mini skirt, chibi, multiple girls, multiple boys, multiple tails, multiple views, copyright name, watermark, artist name, signature
Negative prompt (long):
lowres, bad anatomy, blurry, (worst quality:1.8), low quality, hands bad, face bad, (normal quality:1.3), bad hands, mutated hands and fingers, extra legs, extra arms, duplicate, cropped, jpeg, artifacts, blurry, multiple view, Reference sheet, long body, multiple breasts, mutated, bad anatomy, disfigured, bad proportions, bad feet, ugly, text font ui, missing limb, monochrome, bad anatomy, blurry, (worst quality:1.8), low quality, hands bad, face bad, (normal quality:1.3), bad hands, mutated hands and fingers, extra legs, extra arms, duplicate, cropped, jpeg, artifacts, blurry, multiple view, long body, multiple breasts, mutated, disfigured, bad proportions, duplicate, bad feet, ugly, missing limb,
负面提示词必须包含: lowres (由于使用了大量低分辨率图片进行训练)。worst quality 和 low quality 可根据个人需求选择是否添加。也可以使用例图里的负面提示词(long)，这串提示词是内部成员测试时使用的的，它非常的"屎山"，我们也无法确认其效果如何，但"it just work."。
Negative prompts must include: lowres (due to the use of a large number of low-resolution images for training). "worst quality" and "low quality" can be added based on personal preference. You can also use the negative prompts (long) shown in the example images. This string of prompts was used by internal members during testing. It is quite "shitty", and we cannot confirm its effectiveness, but "it just works."
例如要生成一张东风谷早苗的图像，可以使用以下prompt（换行是非必要的）：
For example, to generate an image of Sanae Kochiya, you can use the following prompt (line breaks are optional):
python
11girl,2kochiya sanae, touhou
3ask (askzy),4solo, green hair, green eyes,5masterpiece, best quality, absurdres, newest, safe, highres
When using Stable Diffusion models to generate character images, due to the characteristics of the training data and methods, it is best to provide the character's name along with their feature tags. If only the character's name is provided without feature tags, details such as hair color and eye color generated by the model may be inaccurate, especially when the character's features are more complex. However, for ordinary users, it is not easy to supplement and complete all the necessary feature tags.
为了解决这一问题,我们强烈推荐使用DTG(Danbooru Tag Generator),它是一个可以在AUTOMATIC1111的WebUI上使用的插件。DTG本质上是一个语言模型(LLM),可以根据输入的角色名自动扩写提示(prompt),补充角色的特征标签以及更多的细节描述。使用DTG可以更好地发挥像SanaeXL这样的高质量模型的性能。
To solve this problem, we strongly recommend using DTG (Danbooru Tag Generator), which is a plugin that can be used on AUTOMATIC1111's WebUI. DTG is essentially a language model (LLM) that can automatically expand prompts based on the input character name, supplementing the character's feature tags and more detailed descriptions. Using DTG can better leverage the performance of high-quality models like SanaeXL.
"1girl/1boy + character name + series name + action/scene perspective + quality words"After providing basic prompt information in this format, you can leave the rest of the work to DTG to complete automatically. DTG can generate additional relevant tags based on the given prompts, helping AI models generate more accurate and specific images.
We have reason to speculate that NovelAI3 internally may also use a similar mechanism. After the user provides the initial prompt, NovelAI3 may connect to a tag generator to automatically supplement and expand the prompt information, achieving better guidance effects.
This "prompt + DTG" combination approach allows users to control the content and style of AI-generated images in a more concise and efficient manner, without the need to manually add a large number of complex tags, greatly improving the convenience and accuracy of using Stable Diffusion models for character image generation.
Because the number of training images is very large, both models have a large number of effective tags, especially artists and characters with over 1,000 images on Danbooru, which can achieve good results. It should be noted that the training set is up to February 2024, so the model's support for recent characters is not ideal.
Below are links to recommended wildcards for artists and characters:
Based on our CCIP evaluation of the model, we have provided a list of excellently performing roles for you to choose from.
Please note that this list only includes outstanding performers from within our evaluation scope (3711 roles) and does not cover all roles.
For some less common roles, you may need to test their effectiveness yourself.
We adhere to strict evaluation standards and will not claim that the model supports a role simply because related data for that role was added to the training dataset without proper evaluation.
If we were to adopt such a loose standard, our model could theoretically claim to support all roles (>12000), but this would obviously be meaningless.
Using the built-in high-resolution fix (Highres.fix) feature of webui at 1024×1024 resolution can better leverage the performance of sanaeXL. The recommended fix ratio is between 1.2 and 1.5 times.
There are no special requirements for samplers. If the output effect is not satisfactory, you can appropriately increase the number of sampling steps and the value of CFG.
Taking Euler A as an example, the number of sampling steps can be between 20-30. A higher number of steps will yield better results. The recommended CFG is between 7-12.
CCIP is a model that can effectively evaluate character fitting. Although characters were not trained separately, thanks to effective training, the model did not suffer from catastrophic forgetting like starryXL. In fact, it has become the best-performing model in CCIP evaluation at the time of writing. 68.77% of characters obtained a CCIP score above 0.8, with a median of 0.94. In this evaluation set composed of 3,711 characters, the vast majority of characters can be directly generated through prompt combinations.
