Considerations: This model can easily generate characters and figures that are protected by copyright in Japan and other countries. Please strictly comply with relevant national laws and regulations, and do not use it for commercial or malicious purposes, as this could lead to legal risks, including infringement. Users are advised not to misuse the model.
1. Model Overview
IllustriousXL is an image generation model fine-tuned by the South Korean company OnomaAI based on the SDXL model kohaku-xl-beta5. Released on September 30, 2024, its V0.1 version quickly gained popularity among users and became a hot new model. It has a high potential, suitable for creating various art styles, characters, and compositions, and derivative models are gradually emerging. Currently, many users are exploring different creative possibilities using this model, from illustrations to character design. The application scope of IllustriousXL is expanding rapidly. The diversity and flexibility of the model enable it to perform excellently across different styles of artistic creation, especially for users aiming to generate high-quality images. As time goes on, more and more derivative models are being developed to meet users' needs for different types and styles of images.
In Japan, IllustriousXL is commonly known as "イラストリアス," which reminds many players of a certain ship game character. This familiar name has helped boost the model's recognition, gaining more attention among anime enthusiasts and illustrators.
2. Model Features
IllustriousXL was trained using Danbooru resources, and its prompt syntax is similar to SD1.5 or Animagine series, making it easy for users to transition from those models to IllustriousXL. The model has undergone extensive training on characters, compositions, and other aspects, making it capable of generating diverse styles and high-quality images. Thanks to the rich training data, IllustriousXL can generate works with high levels of detail and rich composition that meet user expectations, whether it be complex backgrounds or multi-character scenes.
2.1 Recommended Settings for Image Generation
Sampling Method: Euler a
Sampling Steps: 20-28
CFG: 5-7.5
Recommended Size: Similar to other SDXL models, the recommended size is 1024x1024, with additional supported sizes including 896x1152, 832x1216, 768x1344, etc. Using these recommended sizes helps generate more balanced and high-quality images, avoiding common distortion issues.
Future Version Updates: From version V1.0 onwards, the training image size is expected to increase to 1536x1536, which may bring changes to the output settings. This will offer higher resolution and richer details in future versions, enhancing the visual performance of images.
3. Label Order During Training
During training, IllustriousXL labeled images in the following order:
Subject Details (e.g., 1boy, 1girl, no human)
Character Name and Work Name
Rating (e.g., general, sensitive, questionable, explicit)
General Elements
Artist Name
Quality
Era
This label order helps the model better understand and process image content. Although the prompt order does not need to strictly follow the training label order, adjusting the order can sometimes improve results when the generated image is unsatisfactory. A well-organized prompt order can significantly impact the overall quality of the generated image, especially for complex scenes or multi-character compositions.
4. Rating System
IllustriousXL's rating system follows Danbooru's standards, classified into the following four categories:
general (General)
sensitive (Sensitive)
questionable (Questionable)
explicit (Explicit Adult Content)
If you want to generate images with more adult content, you can directly add the "explicit" tag. The rating system allows users to better control the scale and style of generated content, ensuring that the images align with the expected content type.
5. Quality Tags and Recommendations
The model classifies image quality during training, and it is usually recommended to add the top three quality tags in prompts to improve image quality, while including the following negative tags to eliminate undesirable effects:
Recommended Quality Tags: masterpiece, best quality, good quality
Negative Quality Tags: bad quality, worst quality
In addition, the model supports other quality tags like "very aesthetic" and "absurdres," which can be used as needed. By appropriately using these quality tags, users can finely control the details, composition, and overall style of the image, ensuring that the final generated work meets expectations.
6. Usage Tips and Considerations
Negative Prompts: It is recommended to use the following prompts to avoid undesirable image generation:
lowres, (bad), text, error, fewer, extra, missing, worst quality, jpeg artifacts, low quality,
watermark, unfinished, displeasing, oldest, early, chromatic aberration, signature, extra digits,
artistic error, username, scan, abstract,These negative prompts help remove common flaws in images, such as watermarks, low resolution, and incomplete details.
Space vs. Underscore in Tags: IllustriousXL handles spaces and underscores differently. Many users have reported that replacing underscores with spaces often improves the success rate when using copyright characters. This issue is particularly pronounced when drawing copyright characters with less extensive training. Proper use of spaces and underscores in fine character designs and outfit tags can significantly impact the output quality of the image.
Tips for Drawing Copyright Characters:
Follow the tag structure used during model training.
Use related tags for clothing and hairstyles to enhance the representation of specific characters.
Look up accurate character tags on Danbooru. Spelling mistakes in character names or work names may lead to drawing failures.
Characters with weaker training can be enhanced by strengthening tag weights to improve accuracy.
Adding "male focus" can enhance the accuracy of male character outputs.
Experiment with different combinations of tags to adjust details and ensure that each part of the character and scene is accurately represented.
7. Unique Advantages of IllustriousXL
IllustriousXL has significant advantages in character generation, namely, it can draw numerous trained characters using tags alone without LoRA. Compared to mainstream SDXL models like Animagine or Pony, IllustriousXL supports a wider variety of character generation, especially due to its training based on Danbooru resources, making it much easier and more efficient to generate yuri or multi-character images. Additionally, the fewer tags required for generation help conserve tokens and improve the influence of other detail tags. When generating multi-character scenes or complex backgrounds using the IllustriousXL model, users can experience remarkable convenience and efficiency.
7.1 Model Limitations
Although IllustriousXL has high flexibility and versatility, its wide training range means that stability in art style can be relatively weak. Therefore, when using this model, it is recommended to use detailed prompts and negative prompts to enhance the stability of the artwork. Additionally, it may be helpful to use some derivative models to compensate for this instability. For complex scenes, achieving ideal results may require multiple prompt adjustments, necessitating a certain level of patience and fine-tuning skills from users.
