High-quality anime-style illustrations made easy! The latest image generation AI model 'Animagine XL 4.0' has arrived

On January 28, 2025, Cagliostro Research Lab released 'Animagine XL 4.0,' the latest image generation AI model specialized for anime images. This model is a powerful tool that allows you to easily generate high-quality anime illustrations.

Cagliostro Research Lab has released 'Animagine XL 4.0,' the latest image generation AI model specialized for anime images. This model is a powerful tool capable of generating high-quality anime illustrations that surpass previous models. In this article, the AICU media editorial department will explain in an easy-to-understand manner the features of Animagine XL 4.0, how to use it, and how to actually generate images.

The 'Shared ComfyUI' released by AICU now also includes Animagine XL 4.0!
A free evaluation period is currently underway! We look forward to your use.

*Regarding image generation using characters, trademarks, or proper nouns from anime, games, etc., 'similarity and reliance' may become an issue under Japanese copyright law. The AICU editorial department introduces this model with the position that the user is responsible for the use of openly released models, and it is necessary to be very careful regarding its use and generation.

Although the original site introduces it with anime and game characters, the following will be presented by 'Full Affirmation Boyfriend-kun [LuC4]'.

[PR] The LuC4 Illustration Contest is now being held!

What kind of model is Animagine XL 4.0?

Animagine XL 4.0 (also called Anim4gine) is the latest model based on Stable Diffusion XL (SDXL) specialized for generating anime illustrations. This model, which further evolves the previous version, is based on Stable Diffusion XL 1.0 and has been retrained on 8.4 million anime-style images.

Main features of Animagine XL 4.0

Specialized for anime: It has been trained on a vast dataset of anime images and supports a diverse range of anime styles.

High quality: Based on SDXL, it can generate beautiful, highly detailed illustrations. Control via tags: Because it is trained with tag-based captions and a tag ordering method, you can finely control the illustrations generated using tags.

Approximately 2,650 GPU hours of training: By training for 2,650 hours on high-end GPUs, it is possible to generate higher-quality anime illustrations.

Open source: It is released on Hugging Face and can be used by anyone for free.

How do I use it?

Animagine XL 4.0 can be used in the following ways.

Hugging Face Space: You can easily try it out in your web browser. https://huggingface.co/spaces/cagliostrolab/animagine-xl-4.0

Diffusers: This is a Python library intended for developers, so we will skip it this time!

ComfyUI or WebUI (AUTOMATIC1111 or Forge)

This article explains how to use it with ComfyUI.

For Google Colab, which is explained in Episode 1 of ComfyUIMaster, click here.

'Downloading various models'

By writing the wget command, you can download and use the Animagine XL 4.0 model directly on Google Colab.

If you can select 'Animagine-xl-4.0.safetensors' in LoadCheckpoint, you have succeeded.

Don't forget to set the latent image resolution to 1024x1024.

This is because this model is based on SDXL and its 'Training Resolution' is 1024x1024.

Now, I actually tried generating images, and the atmosphere is completely different even with the same prompt. While it is manga or anime-style, it is, how should I put it, 'so bad it's good'.

It seems to have slightly different characteristics from the previous Animagine series.
Let's continue to explore prompts.

Recommended settings and how to write prompts

I will explain how to write prompts and settings to make the most of Animagine XL 4.0.

Basic structure of prompts

Animagine XL 4.0 is trained with tag-based captions and tag ordering. Using the following template allows you to control the illustrations you want to generate more precisely.

1girl/1boy/1other, character name, work name, other tags in any order

For example, you can consider a prompt like the following.

Tags to improve image quality

Adding the following tags to the beginning or end of your prompt will improve image quality.

masterpiece, high score, great score, absurdres

Negative Prompt

Negative prompts that specify unwanted elements are also effective. The following are recommended negative prompts.

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

Recommended Settings

• CFG Scale: 5-7 (6 is recommended)

• Sampling Steps: 25-28 (25 is recommended)

• Sampler: Euler Ancestral (Euler a)

Recommended Resolutions

The following resolutions are recommended.

• Square: 1024 x 1024 (aspect ratio 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)

Prompt Example

Here is a prompt example based on the above.

Don't forget the negative prompt

More precise control! Special tags

Animagine XL 4.0 supports special tags that allow for even more precise control.

Quality tags

You can directly control the quality of the image.

• masterpiece

• best quality

• low quality

• worst quality

If you don't specify a negative prompt, using 'low quality, worst quality' with the same prompt allows you to generate images that look intentionally rough.

This is quite an interesting technique.

Score tags

You can adjust the quality more precisely than with quality tags.

• high score

• great score

• good score

• average score

• bad score

• low score

By specifying "bad score, low score," you can controllably generate completely different results compared to not using a negative prompt with the same prompt.

Temporal Tags

The era tags, which were popular in AICU's previous Animagine articles, have become even more detailed. From 2005 onwards, you can express specific years or eras on a yearly basis.

• year 2005

• year {n} (any year)

• year 2025

I would like to conduct a separate experiment on this!

Rating Tags

Controls the safety level of the generated images.

• safe

• sensitive

• nsfw

• explicit

These are useful for controlling the content safety level of generated images. These tags should be used responsibly in accordance with applicable laws and platform policies. Of course, the technique of using them in negative prompts is also effective.

Training Information

Animagine XL 4.0 was trained in the following environment.

• Hardware: 7 x H100 80GB SXM5 (high-performance GPU)

• Number of training images: 8,401,464 images

• Training Parameters:

  • 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

Acknowledgments and Fundraising

In developing Animagine XL 4.0, Cagliostro Research Lab expresses its gratitude for the technology and contributions of Stability AI, Novel AI, and the Waifu Diffusion Team.
They also extend their thanks to Moesacpe AI, Lesser Rabbit, Kohya SS, discus0434, and all early testers.

Cagliostro Research Lab has resumed fundraising for future development. This time, they are also accepting donations in cryptocurrencies such as ETH and USDT. They previously collected donations via Ko-fi, but it seems they experienced issues where PayPal suspended their account when they attempted to withdraw funds.

Limitations

Animagine XL 4.0 has the following limitations.

• Prompt format: It is optimized for tag-based prompts, so inputting in natural language may not be effective.

• Human anatomy: It may struggle with complex anatomical details such as hand poses or the number of fingers.

• Text generation: Rendering text within images is currently not supported.

• New characters: Recently introduced characters may have lower accuracy due to a lack of training data.

• Multiple characters: You may need to refine your prompts when featuring multiple characters.

• High resolution: At high resolutions such as 1536x1536, quality may decrease due to the resolution constraints used during training.

• Style consistency: Since it is trained with a focus on character individuality rather than maintaining a specific style, you may need to use style tags to maintain a particular look.

License

Animagine XL 4.0 adopts the CreativeML Open RAIL++-M license.
Commercial use, modification, distribution, and private use are permitted, but illegal acts, the generation of harmful content, discrimination, and exploitation are prohibited.
Additionally, you must include a copy of the license, describe any changes made, and retain the notices.
Please check here for the detailed license terms.
https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0/blob/main/LICENSE.md

Summary

Animagine XL 4.0 is a powerful tool capable of generating high-quality anime illustrations. Please try using this model to create your own original illustrations!
*Regarding image generation using characters, trademarks, or proper nouns from anime, games, etc., 'similarity and reliance' may become an issue under Japanese copyright law. The AICU editorial department introduces this model with the stance that the user is responsible for the use of publicly released models, and one should be sufficiently careful regarding its use and generation.

Animagine XL 4.0 has also been added to the Shared ComfyUI we released the other day!
We look forward to everyone using it.

LuC4 Illustration Contest is now being held

Regarding the 'Full-Power Affirmation Boyfriend-kun [LuC4]' used in the sample this time, we are currently holding an illustration contest.

[How to participate] Post a still image or video on X that includes the specified hashtag " #ValentinesGiftFromBoys #StValentine #AIART #AICU" and, depending on the category, "#SetsunaCreation" or " #FullSupportBoyfriend #LuC4", and submit the required information via this entry form by the deadline.
[Entry Period] January 22, 2025 – February 16, 2025, 22:00

Let's all create lots of LuC4 character illustrations and share them with LuC4 fans! The application form is https://j.aicu.ai/C25 .

About AICU

With the vision of "creating creators," AICU delivers the latest generative AI and creative AI technologies to everyone.

📝 On our note, we share the latest information and tutorials on generative AI. Please follow us to check for the latest updates!
https://note.com/aicu

📩 In our email newsletter "Weekly AICU News," we deliver AI-related news, trends, and event information every week. Subscribe here!
https://corp.aicu.ai/ja

📚️ We are currently distributing "Monthly AICU Magazine" on the Kindle Store! You can read special features on AI technology and in-depth commentary articles.
https://j.aicu.ai/kindle

We are also sharing information on X! Follow us 😉
https://x.com/AICUai

#GenerativeAI #StableDiffusion #ImageGenerationAI #AIImageGeneration #Prompt #ComfyUI #SDXL #StableDiffusionWebUI #AIIllustration #Generation #Diffusers #AnimagineXL #CagliostroResearchLab #AnimeIllustration