cagliostrolab/animagine-xl-4.0

Animagine XL 4.0

Overview

Animagine XL 4.0, also stylized as Anim4gine, is the ultimate anime-themed finetuned SDXL model and the latest installment of Animagine XL series. Despite being a continuation, the model was retrained from Stable Diffusion XL 1.0 with a massive dataset of 8.4M diverse anime-style images from various sources with the knowledge cut-off of January 7th 2025 and finetuned for approximately 2650 GPU hours. Similar to the previous version, this model was trained using tag ordering method for the identity and style training.

Model Details

• Developed by: Cagliostro Research Lab
• Model type: Diffusion-based text-to-image generative model
• License: CreativeML Open RAIL++-M
• Model Description: This is a model that can be used to generate and modify specifically anime-themed images based on text prompt
• Fine-tuned from: Stable Diffusion XL 1.0

Downstream Use

1. Use this model in our Hugging Face Spaces
2. Use it in ComfyUI or Stable Diffusion Webui
3. Use it with 🧨 diffusers

🧨 Diffusers Installation

1. Install Required Libraries

pip install diffusers transformers accelerate safetensors --upgrade

2. Example Code

The example below uses lpw_stable_diffusion_xl pipeline which enables better handling of long, weighted and detailed prompts. The model is already uploaded in FP16 format, so there's no need to specify variant="fp16" in the from_pretrained call.

import torch
from diffusers import StableDiffusionXLPipeline

pipe = StableDiffusionXLPipeline.from_pretrained(
    "cagliostrolab/animagine-xl-4.0",
    torch_dtype=torch.float16,
    use_safetensors=True,
    custom_pipeline="lpw_stable_diffusion_xl",
    add_watermarker=False
)
pipe.to('cuda')

prompt = "1girl, arima kana, oshi no ko, hoshimachi suisei, hoshimachi suisei \(1st costume\), cosplay, looking at viewer, smile, outdoors, night, v, masterpiece, high score, great score, absurdres"
negative_prompt = "lowres, bad anatomy, bad hands, text, error, missing finger, extra digits, fewer digits, cropped, worst quality, low quality, low score, bad score, average score, signature, watermark, username, blurry"

image = pipe(
    prompt,
    negative_prompt=negative_prompt,
    width=832,
    height=1216,
    guidance_scale=6,
    num_inference_steps=25
).images[0]

image.save("./arima_kana.png")

Usage Guidelines

1. Prompt Structure

The model was trained with tag-based captions and the tag-ordering method. Use this structured template:

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

2. Quality Enhancement Tags

Add these tags at the start or end of your prompt:

masterpiece, high score, great score, absurdres

3. Recommended Negative Prompt

lowres, bad anatomy, bad hands, text, error, missing finger, extra digits, fewer digits, cropped, worst quality, low quality, low score, bad score, average score, signature, watermark, username, blurry

4. Optimal Settings

• CFG Scale: 5-7 (6 Recommended)
• Sampling Steps: 25-28 (25 Recommended)
• Preferred Sampler: Euler Ancestral (Euler a)

5. Recommended Resolutions

Orientation	Dimensions	Aspect Ratio
Square	1024 x 1024	1:1
Landscape	1152 x 896	9:7
	1216 x 832	3:2
	1344 x 768	7:4
	1536 x 640	12:5
Portrait	896 x 1152	7:9
	832 x 1216	2:3
	768 x 1344	4:7
	640 x 1536	5:12

6. Final Prompt Structure Example

masterpiece, high score, great score, absurdres, 1girl, firefly \(honkai: star rail\), honkai \(series\), honkai: star rail, casual, solo, looking at viewer, outdoors, smile, reaching towards viewer, night

Special Tags

The model supports various special tags that can be used to control different aspects of the image generation process. These tags are carefully weighted and tested to provide consistent results across different prompts.

Quality Tags

Quality tags are fundamental controls that directly influence the overall image quality and detail level. Available quality tags:

• masterpiece
• best quality
• low quality
• worst quality

Sample image using "masterpiece, best quality" quality tags with negative prompt left empty.	Sample image using "low quality, worst quality" quality tags with negative prompt left empty.

Score Tags

Score tags provide a more nuanced control over image quality compared to basic quality tags. They have a stronger impact on steering output quality in this model. Available score tags:

• high score
• great score
• good score
• average score
• bad score
• low score

Sample image using "high score, great score" score tags with negative prompt left empty.	Sample image using "bad score, low score" score tags with negative prompt left empty.

Temporal Tags

Temporal tags allow you to influence the artistic style based on specific time periods or years. This can be useful for generating images with era-specific artistic characteristics. Supported year tags:

• year 2005
• year {n}
• year 2025

Sample image of Hatsune Miku with "year 2007" temporal tag.	Sample image of Hatsune Miku with "year 2023" temporal tag.

Rating Tags

Rating tags help control the content safety level of generated images. These tags should be used responsibly and in accordance with applicable laws and platform policies. Supported ratings:

• safe
• sensitive
• nsfw
• explicit

Training Information

The model was trained using state-of-the-art hardware and optimized hyperparameters to ensure the highest quality output. Below are the detailed technical specifications and parameters used during the training process:

Parameter	Value
Hardware	7 x H100 80GB SXM5
Num Images	8,401,464
UNet Learning Rate	2.5e-6
Text Encoder Learning Rate	1.25e-6
Scheduler	Constant With Warmup
Warmup Steps	5%
Batch Size	32
Gradient Accumulation Steps	2
Training Resolution	1024x1024
Optimizer	Adafactor
Input Perturbation Noise	0.1
Debiased Estimation Loss	Enabled
Mixed Precision	fp16

Acknowledgement

This long-term project would not have been possible without the groundbreaking work, innovative contributions, and comprehensive documentation provided by Stability AI, Novel AI, and Waifu Diffusion Team. We are especially grateful for the kickstarter grant from Main that enabled us to progress beyond V2. For this iteration, we would like to express our sincere gratitude to everyone in the community for their continuous support, particularly:

1. Moescape AI: Our invaluable collaboration partner in model distribution and testing
2. Lesser Rabbit: For providing essential computing and research grants
3. Kohya SS: For developing the comprehensive open-source training framework
4. discus0434: For creating the industry-leading open-source Aesthetic Predictor 2.5
5. Early testers: For their dedication in providing critical feedback and thorough quality assurance

Contributors

We extend our heartfelt appreciation to our dedicated team members who have contributed significantly to this project, including but not limited to:

Model

• KayfaHaarukku
• Raelina
• Linaqruf

Gradio

• Damar Jati

Relations, finance, and quality assurance

• Scipius
• Asahina
• Bell
• BoboiAzumi

Data

• Pomegranata
• Kr1SsSzz
• Fiqi
• William Adams Soeherman

Fundraising Are Now Open Again!

We’re excited to reopen Fundraising to fund new training, research, and model development. Your support helps us push the boundaries of what’s possible with AI.

You can help us with:

• Donate: Contribute via ETH or USDT to the address below.

• Share: Spread the word about our models and share your creations!

• Feedback: Let us know how we can improve.

Donation Address:

ETH/USDT/USDC(e): 0xd8A1dA94BA7E6feCe8CfEacc1327f498fCcBFC0C

Why do we use Cryptocurrency?

When we initially opened fundraising through Ko-fi and using PayPal as withdrawal methods, our PayPal account was flagged and eventually banned, despite our efforts to explain the purpose of our project. Unfortunately, this forced us to refund all donations and left us without a reliable way to receive support. To avoid such issues and ensure transparency, we have now switched to cryptocurrency as the way to raise the fund.

Want to Donate in Non-Crypto Currency?

Although we had a bad experience with Paypal, and you’d like to support us but prefer not to use cryptocurrency, feel free to contact us via [Discord Server](https://discord.gg/cqh9tZgbGc) for alternative donation methods.

Join Our Discord Server

Feel free to join our discord server

Limitations

• Prompt Format: Limited to tag-based text prompts; natural language input may not be effective
• Anatomy: May struggle with complex anatomical details, particularly hand poses and finger counting
• Text Generation: Text rendering in images is currently not supported and not recommended
• New Characters: Recent characters may have lower accuracy due to limited training data availability
• Multiple Characters: Scenes with multiple characters may require careful prompt engineering
• Resolution: Higher resolutions (e.g., 1536x1536) may show degradation as training used original SDXL resolution
• Style Consistency: May require specific style tags as training focused more on identity preservation than style consistency

License

This model adopts the original CreativeML Open RAIL++-M License from Stability AI without any modifications or additional restrictions. The license terms remain exactly as specified in the original SDXL license, which includes:

• ✅ Permitted: Commercial use, modifications, distributions, private use
• ❌ Prohibited: Illegal activities, harmful content generation, discrimination, exploitation
• ⚠️ Requirements: Include license copy, state changes, preserve notices
• 📝 Warranty: Provided "AS IS" without warranties

Please refer to the original SDXL license for the complete and authoritative terms and conditions.