Animagine Xl 3.1

🚀 Animagine XL 3.1

Animagine XL 3.1 is an update in the Animagine XL V3 series. It's an open - source, anime - themed text - to - image model, generating high - quality anime - style images, which is valuable for anime fans, artists, and content creators.

🚀 Quick Start

Animagine XL 3.1 is an update in the Animagine XL V3 series, enhancing the previous version, Animagine XL 3.0. This open - source, anime - themed text - to - image model has been improved for generating anime - style images with higher quality. It includes a broader range of characters from well - known anime series, an optimized dataset, and new aesthetic tags for better image creation. Built on Stable Diffusion XL, Animagine XL 3.1 aims to be a valuable resource for anime fans, artists, and content creators by producing accurate and detailed representations of anime characters.

✨ Features

• Generates high - quality anime images from textual prompts.
• Boasts enhanced hand anatomy, improved concept understanding, and advanced prompt interpretation.
• Utilizes special tags for better image quality, rating, creation date, and aesthetic control.
• Supports multi - aspect resolution image generation.

📦 Installation

🧨 Diffusers Installation

First install the required libraries:

pip install diffusers transformers accelerate safetensors --upgrade

Then run image generation with the following example code:

import torch
from diffusers import DiffusionPipeline

pipe = DiffusionPipeline.from_pretrained(
    "cagliostrolab/animagine-xl-3.1", 
    torch_dtype=torch.float16, 
    use_safetensors=True, 
)
pipe.to('cuda')

prompt = "1girl, souryuu asuka langley, neon genesis evangelion, solo, upper body, v, smile, looking at viewer, outdoors, night"
negative_prompt = "nsfw, lowres, (bad), text, error, fewer, extra, missing, worst quality, jpeg artifacts, low quality, watermark, unfinished, displeasing, oldest, early, chromatic aberration, signature, extra digits, artistic error, username, scan, [abstract]"

image = pipe(
    prompt, 
    negative_prompt=negative_prompt,
    width=832,
    height=1216, 
    guidance_scale=7,
    num_inference_steps=28
).images[0]

image.save("./output/asuka_test.png")

💻 Usage Examples

Basic Usage

The basic usage is shown in the installation code above. You can generate images by providing a prompt and a negative prompt.

Advanced Usage

To guide the model towards generating high - aesthetic images, use negative prompts like:

nsfw, lowres, (bad), text, error, fewer, extra, missing, worst quality, jpeg artifacts, low quality, watermark, unfinished, displeasing, oldest, early, chromatic aberration, signature, extra digits, artistic error, username, scan, [abstract]

For higher quality outcomes, prepend prompts with:

masterpiece, best quality, very aesthetic, absurdres

It’s recommended to use a lower classifier - free guidance (CFG Scale) of around 5 - 7, sampling steps below 30, and to use Euler Ancestral (Euler a) as a sampler.

📚 Documentation

Model Details

Gradio & Colab Integration

Try the demo powered by Gradio in Huggingface Spaces:

Or open the demo in Google Colab:

Usage Guidelines

Tag Ordering

For optimal results, it's recommended to follow the structured prompt template because we train the model like this:

1girl/1boy, character name, from what series, everything else in any order.

Special Tags

Quality Modifiers

Quality tags now consider both scores and post ratings to ensure a balanced quality distribution. We've refined labels for greater clarity, such as changing 'high quality' to 'great quality'.

Rating Modifiers

We've also streamlined our rating tags for simplicity and clarity, aiming to establish global rules that can be applied across different models. For example, the tag 'rating: general' is now simply 'general', and 'rating: sensitive' has been condensed to 'sensitive'.

Year Modifier

We've also redefined the year range to steer results towards specific modern or vintage anime art styles more accurately. This update simplifies the range, focusing on relevance to current and past eras.

Aesthetic Tags

We've enhanced our tagging system with aesthetic tags to refine content categorization based on visual appeal. These tags are derived from evaluations made by a specialized ViT (Vision Transformer) image classification model, specifically trained on anime data. For this purpose, we utilized the model [shadowlilac/aesthetic - shadow - v2](https://huggingface.co/shadowlilac/aesthetic - shadow - v2), which assesses the aesthetic value of content before it undergoes training. This ensures that each piece of content is not only relevant and accurate but also visually appealing.

Recommended settings

This model supports generating images at the following dimensions:

Training and Hyperparameters

Animagine XL 3.1 was trained on 2x A100 80GB GPUs for approximately 15 days, totaling over 350 GPU hours. The training process consisted of three stages:

• Pretraining: Utilized a data - rich collection of 870k ordered and tagged images to increase Animagine XL 3.0's model knowledge.
• Finetuning - First Stage: Employed labeled and curated aesthetic datasets to refine the broken U - Net after pretraining.
• Finetuning - Second Stage: Utilized labeled and curated aesthetic datasets to refine the model's art style and improve hand and anatomy rendering.

Hyperparameters

Model Comparison (Pretraining only)

Training Config

Source code and training config are available here: https://github.com/cagliostrolab/sd - scripts/tree/main/notebook

Acknowledgements

The development and release of Animagine XL 3.1 would not have been possible without the invaluable contributions and support from the following individuals and organizations:

• SeaArt.ai: Our collaboration partner and sponsor.
• Shadow Lilac: For providing the aesthetic classification model, [aesthetic - shadow - v2](https://huggingface.co/shadowlilac/aesthetic - shadow - v2).
• [Derrian Distro](https://github.com/derrian - distro): For their custom learning rate scheduler, adapted from [LoRA Easy Training Scripts](https://github.com/derrian - distro/LoRA_Easy_Training_Scripts/blob/main/custom_scheduler/LoraEasyCustomOptimizer/CustomOptimizers.py).
• [Kohya SS](https://github.com/kohya - ss): For their comprehensive training scripts.
• Cagliostrolab Collaborators: For their dedication to model training, project management, and data curation.
• Early Testers: For their valuable feedback and quality assurance efforts.
• NovelAI: For their innovative approach to aesthetic tagging, which served as an inspiration for our implementation.
• KBlueLeaf: For providing inspiration in balancing quality tags distribution and managing tags based on Hakubooru Metainfo

Thank you all for your support and expertise in pushing the boundaries of anime - style image generation.

Collaborators

• Linaqruf
• ItsMeBell
• Asahina2K
• DamarJati
• Zwicky18
• Scipius2121
• Raelina
• Kayfahaarukku
• Kriz

Limitations

⚠️ Important Note

While Animagine XL 3.1 represents a significant advancement in anime - style image generation, it has the following limitations:

1. Anime - Focused: This model is specifically designed for generating anime - style images and is not suitable for creating realistic photos.
2. Prompt Complexity: This model may not be suitable for users who expect high - quality results from short or simple prompts. The training focus was on concept understanding rather than aesthetic refinement, which may require more detailed and specific prompts to achieve the desired output.
3. Prompt Format: Animagine XL 3.1 is optimized for Danbooru - style tags rather than natural language prompts. For best results, users are encouraged to format their prompts using the appropriate tags and syntax.
4. Anatomy and Hand Rendering: Despite the improvements made in anatomy and hand rendering, there may still be instances where the model produces suboptimal results in these areas.
5. Dataset Size: The dataset used for training Animagine XL 3.1 consists of approximately 870,000 images. When combined with the previous iteration's dataset (1.2 million), the total training data amounts to around 2.1 million images. While substantial, this dataset size may still be considered limited in scope for an "ultimate" anime model.
6. NSFW Content: Animagine XL 3.1 has been designed to generate more balanced NSFW content. However, it is important to note that the model may still produce NSFW results, even if not explicitly prompted.

License

💡 Usage Tip

Based on Animagine XL 3.0, Animagine XL 3.1 falls under Fair AI Public License 1.0 - SD license, which is compatible with Stable Diffusion models’ license. Key points:

1. Modification Sharing: If you modify Animagine XL 3.1, you must share both your changes and the original license.
2. Source Code Accessibility: If your modified version is network - accessible, provide a way (like a download link) for others to get the source code. This applies to derived models too.
3. Distribution Terms: Any distribution must be under this license or another with similar rules.
4. Compliance: Non - compliance must be fixed within 30 days to avoid license termination, emphasizing transparency and adherence to open - source values.