Animagine XL

Animagine XL is an open-source anime-themed text-to-image model family built upon Stable Diffusion XL. The latest version, Animagine XL v3.1, was released on April 3, 2024, bringing significant improvements to anime image generation capabilities.

Model Evolution and Training

The Animagine XL family has evolved through several versions, each bringing substantial improvements. Version 3.0 marked a shift from purely aesthetic improvements to enhanced concept learning, utilizing approximately 1.2 million training images. Version 3.1 further expanded the training dataset to include an additional 870,000 images, bringing the total to around 2.1 million images.

The training process for version 3.0 involved approximately 500 GPU hours on 2x A100 80GB GPUs over 21 days, implementing a three-stage approach:

1. Base stage for feature alignment and UNet refinement
2. Curated stage for aesthetic tuning
3. Final refinement stage

The development process involved collaboration with SeaArt.ai and utilized the aesthetic classification model shadowlilac/aesthetic-shadow-v2. The training leveraged modified Kohya SS scripts and incorporated a custom learning rate scheduler from Derrian Distro.

Features and Capabilities

Key improvements in the latest versions include:

• Enhanced hand anatomy rendering
• Better concept understanding and prompt interpretation
• Wider range of recognizable anime characters
• Improved image quality and consistency

The model supports various aspect ratios, including:

• Square (1:1)
• Portrait (7:9, 13:19, 4:7)
• Landscape (9:7, 19:13, 7:4)
• Panoramic (12:5, 5:12)

Usage and Optimization

For optimal results, the following parameters are recommended:

• CFG scale: 5-7
• Sampling steps: Less than 30
• Preferred sampler: Euler Ancestral

The model utilizes special tags for controlling various aspects:

• Image quality tags
• Rating controls (safe, sensitive, nsfw, explicit)
• Creation year modifiers (newest, recent, mid, early, oldest)
• Aesthetic modifiers (very aesthetic to very displeasing)

Prompt structure follows a specific template:

1girl/1boy, character name, from what series, additional details in any order

Licensing and Availability

The model is released under the Fair AI Public License 1.0-SD, which requires users to share modifications and source code for network-accessible versions. This license is compatible with Stable Diffusion's base license while maintaining openness and protecting contributors.

References

• Civitai Model Page - Official model release page
• Cagliostro Research Lab - Model developers' organization
• Animagine XL 3.0 - Previous version documentation
• Fair AI Public License - Model license details
• Cagliostro Lab Blog - Official release announcement
• SeaArt.ai - Development collaboration partner