Illustrious XL is the Illustration focused Stable Diffusion XL model which is continued from Kohaku XL Beta 5. The model focuses on utilizing large-scale annotated dataset, Danbooru2023. We release the v0.1 and v0.1-GUIDED model here, under fair public ai license, however discourages the usage of model over monetization purpose / any closed source purposes. For full technical details, please refer to our technical report.

📄 Documentation: https://arxiv.org/html/2409.19946v1

Model Information:

Name: Illustrious-XL-v0.1
Model Type: Stable Diffusion XL Model
Dataset: Fine-tuned on Danbooru2023 Dataset

Description

Illustrious-XL is a powerful generative model series, fine-tuned on the comprehensive Danbooru2023 dataset and its variants. It includes a wide variety of character designs, styles, and artistic knowledge derived from the dataset, making it suitable for creative and artistic AI generation tasks.
Illustrious-XL-v0.1 is untuned BASE model, which works as possible base for all future model variants. LoRAs / Adapters can be trained on this model, ensuring future usecases. The model is research-only purpose, as not tuned for aesthetics / preferences.
Illustrious-XL-v0.1-GUIDED is minimally safety controlled model, which works as better option for usual usecases.

We plan to release several aesthetic-finetuned model variants in near future.

Technical Details:

Full technical details and specifications will be released soon.
Expected Release Date: September 30, 2024, 00:00 -> 06:00 UTC. (Arxiv database is down)

Terms and Conditions:

We recommend to use official repositories, to prevent malicious attacks.
Users must agree with LICENSE to use the model. As mentioned in LICENSE, we do NOT take any actions about generated results or possible variants.
As mentioned in LICENSE, users must NOT use the generated result for any prohibited purposes, including but not limited to:
- Harmful or malicious activities: This includes harassment, threats, spreading misinformation, or any use intended to harm individuals or groups.
- Illegal activities: Using generated content to violate any applicable laws or regulations.
- Unethical, offensive content generation: Generating offensive, defamatory, or controversial content that violates ethical guidelines.

By using this model, users agree to comply with the conditions outlined in the LICENSE and acknowledge responsibility for how they utilize the generated content.

Safety Control Recommendation:

Generative models can occasionally produce unintended or harmful outputs.
To minimize this risk, it is strongly recommended to use the GUIDED model variant, which incorporates additional safety mechanisms for responsible content generation.
By choosing this variant, users can significantly reduce the likelihood of generating harmful or unintended content.
We plan to update GUIDED model variants and its methodologies, with extensive research.

Training/Merging Policy: You may fine-tune, merge, or train LoRA based on this model. However, to foster an open-source community, you are required to:

Openly share details of any derived models, including references to the original model licensed under the fair-ai-public-license.
Provide information on datasets and "merge recipes" used for fine-tuning or training.
Adhere to the fair-ai-public-license, ensuring that any derivative works are also open source.

Uploading / Generation Policy: We do not restrict any upload or spread of the generation results, as we do not own any rights regard to generated materials. This includes 'personally trained models / finetuned models / trained lora-related results'. However, we kindly ask you to open the generation details, to foster the open source communities and researches.

Monetization Prohibition:

You are prohibited from monetizing any close-sourced fine-tuned / merged model, which disallows the public from accessing the model's source code / weights and its usages.
As per the license, you must openly publish any derivative models and variants. This model is intended for open-source use, and all derivatives must follow the same principles.

Usage: We do not recommend overusing critical composition tags such as 'close-up', 'upside-down', or 'cowboy shot', as they can be conflicting and lead to confusion, affecting model results. Recommended sampling method: Euler a, Sampling Steps: 20–28, CFG: 5–7.5 (may vary based on use case). We suggest using suitable composition tags like "upper body," "cowboy shot," "portrait," or "full body" depending on your use case. The model supports quality tags such as: "worst quality," "bad quality," "average quality," "good quality," "best quality," and "masterpiece (quality)." Note: The model does not have any default style. This is intended behavior for the base model.