Animagine XL V3.1

Animagine XL is an open-source, anime-themed text-to-image model generating high-quality anime-style images. It features a broader range of characters from popular anime series, an optimized dataset, and new aesthetic tags for improved image creation. Built on Stable Diffusion XL, it aims to be a valuable resource for anime fans, artists, and content creators.

Model Details:

Developed by: Cagliostro Research Lab, in collaboration with SeaArt.ai

Model type: Diffusion-based text-to-image generative model

Description: Generates high-quality anime images from textual prompts. Includes enhanced hand anatomy, improved concept understanding, and advanced prompt interpretation.

License: Fair AI Public License 1.0-SD

Usage Guidelines:

Tag Ordering: Use a structured prompt template for optimal results (details not provided here).

Special Tags: Utilize special tags for quality, rating, creation date, and aesthetics to improve results. Simplified rating and quality tags are now used. The year modifier helps target specific anime art styles (modern or vintage).

Aesthetic Tags: A Vision Transformer model (shadowlilac/aesthetic-shadow-v2) was used to pre-classify images for aesthetic value, ensuring visual appeal.

Recommended Settings: For high-aesthetic images, use specific negative prompts (not listed here). For higher quality, prepend prompts with specific phrases (not listed here). A lower CFG scale (5-7), fewer sampling steps (below 30), and the Euler Ancestral sampler are recommended.

Multi-Aspect Resolution: Supports various image dimensions (specific dimensions not listed here).

Acknowledgements: The model's development benefited from contributions from SeaArt.ai, Shadow Lilac, Derrian Distro, Kohya SS, Cagliostrolab Collaborators, early testers, and NovelAI.

Limitations:

Anime-Focused: Specifically designed for anime-style images; not suitable for realistic photos.

Prompt Complexity: Requires detailed and specific prompts for high-quality results; short or simple prompts may yield suboptimal results. Optimized for Danbooru-style tags.

Anatomy and Hand Rendering: While improved, suboptimal results may still occur.

Dataset Size: Trained on approximately 2.1 million images (combination of datasets); while substantial, it may be considered limited for an "ultimate" anime model.

NSFW Content: May generate NSFW results even without explicit prompting.

License: The Fair AI Public License 1.0-SD requires sharing modifications and providing source code accessibility for network-accessible versions. Distribution must be under this or a similar license. Non-compliance must be addressed within 30 days.

Contact: Join the Cagliostro Lab Discord server: https://discord.gg/cqh9tZgbGc Donations are welcome.