AstraXL

AstraXL is a specialized anime-focused checkpoint designed to bridge the gap between high-fidelity background rendering and stylistic 2D character art.

While many models sacrifice background depth for character detail (or vice versa), AstraXL is tuned to ensure that environments feel "lived in" and deep, while maintaining a crisp, flat 2D aesthetic for subjects. It has been extensively tuned to respond to Natural Language prompts, allowing for complex scene generation without relying heavily on traditional "tag salads."

Key Features

• Deep Backgrounds: The model excels at architectural and natural scenery, providing a sense of distance and atmosphere often lost in standard anime mixes.
• Natural Language Support: You don't need to depend solely on Danbooru tags. Sentences like "A girl walking through a rural village" work just as well, if not better, than tag lists.
• Vivid Color & Line Work: Tuned for "main line" clarity (reducing fuzziness around contours) and ensuring colors pop without oversaturation.
• Hybrid Workflow: Works with both Euler a for softer, painterly results and DPM++ 2M for cleaner, sharper digital illustration styles.

Recommended Settings

To get the results intended during training, please use the following configurations:

• Steps: 25 to 40
• CFG Scale: 3.5 to 6.0
• Sampler: Euler a (for soft/painterly) or DPM++ 2M (for crisp/digital)
• Scheduler: Simple or Karras

Prompting Guide

AstraXL is designed to be flexible. You can use standard tagging or natural sentences.

Natural Language Approach (Recommended)

A beautiful girl standing in a library, warm lighting, books everywhere, detailed background.

Standard Tagging Approach

1girl, library, books, depth of field, warm lighting, masterpiece, best quality

Quality Tags (Optional) While the model is tuned to not strictly require them, adding these can push the aesthetic: masterpiece, best quality, amazing quality, very aesthetic, absurdres, highres, newest

Negative Prompt negativeXL_D, NEGATIVE_HANDS, average quality, bad quality, worst quality

Limitations

• Stylistic Constraints: This model is heavily optimized for 2D anime aesthetics. It may struggle with photorealistic generation or styles that deviate significantly from modern digital anime art.
• Text Rendering: Like many diffusion models in this class, it does not natively render intelligible text within images.

License

This project is licensed under the Apache 2.0 License.