|
|
--- |
|
|
license: apache-2.0 |
|
|
tags: |
|
|
- stable-diffusion |
|
|
- stable-diffusion-diffusers |
|
|
- text-to-image |
|
|
- anime |
|
|
- waifu |
|
|
- diffusers |
|
|
- sdxl |
|
|
- sdxl-base |
|
|
- illustrous |
|
|
language: |
|
|
- en |
|
|
base_model: |
|
|
- stabilityai/stable-diffusion-xl-base-1.0 |
|
|
pipeline_tag: text-to-image |
|
|
--- |
|
|
|
|
|
# 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. |
