Animagine XL 4.0

Animagine XL 4.0

Generates anime-style images from tag-based text prompts, a Stable Diffusion XL fine-tune retrained on a large anime dataset. For artists and developers using diffusers, ComfyUI or Stable Diffusion WebUI.

Made by cagliostrolab on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: Animagine XL 4.0 is an anime-themed text-to-image model from Cagliostro Research Lab. The team started over from Stable Diffusion XL 1.0 and trained on 8.4M anime-style pictures, which took roughly 2650 GPU hours. It reads tag-style prompts that name the character, the series and the look you want.

Who it suits: Illustrators, hobbyists and developers who want anime-style pictures from text. An Opt release adds a refined dataset, which the card says improves stability, anatomy, noise, saturation and color accuracy. Read the CreativeML Open RAIL++-M license terms before using the output.

Title: Who uses it, and how
Industry | Who uses it | How they use it | Example
Illustration | Artists | Draft character art from tag prompts | Write a tag prompt for a smiling character outdoors at night
Game and story design | Indie creators | Make concept art for characters and scenes | Generate several poses of one original character
Software | Developers | Run the model through diffusers in a script | Generate an 832 by 1216 portrait from a prompt
Hobby art | Anime fans | Experiment with tags and quality settings | Compare results at CFG 4 and CFG 7

Skip it if you want photos or non-anime styles, because the model is built for anime-themed images.

Where to use it:

• Python with the diffusers library, on a CUDA GPU
• ComfyUI
• Stable Diffusion WebUI
• The card's Hugging Face Space for a quick try

Use case | Who | What they get back
Character art | Illustrators | An anime-style image from your tag prompt
Batch generation | Developers | Many images from a script with set size and steps
Prompt testing | Hobbyists | Images that show how tags and quality words change the result

How to set it up: You need Python, pip and a CUDA GPU for the card's example. Hugging Face lists about 5.1 GB for the FP16 weights.

• Python and pip
• A CUDA GPU
• The diffusers, transformers, accelerate and safetensors libraries

1. Install the libraries.
2. Load the pipeline, which downloads the weights.
3. Write a prompt and generate.

pip install diffusers transformers accelerate safetensors --upgrade

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')

How to use it: Write a tag-based prompt in the card's order, add the quality tags at the end, and pass the recommended negative prompt. The card suggests CFG 5 and 28 sampling steps.

1girl/1boy/1other, character name, from which series, rating, everything else in any order and end with quality enhancement

masterpiece, high score, great score, absurdres

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=5,
    num_inference_steps=28
).images[0]

Tool | What it does
lpw_stable_diffusion_xl | A pipeline that handles long, weighted prompts better
CFG scale | The card recommends 5, within a range of 4 to 7
Sampling steps | The card recommends 28, within a range of 25 to 28

Example: You write the prompt "1girl, looking at viewer, smile, outdoors, night, masterpiece, high score, great score, absurdres" and pass the recommended negative prompt. The pipeline returns one 832 by 1216 anime-style image.

For newcomers and experts: If you are new, copy the loading code, write a short tag prompt and generate one image. Then change one setting at a time.

For people in the field, the model keeps the SDXL architecture and uses a tag-ordering method for identity and style during training. The example uses the long-prompt-weighting pipeline, and the weights already ship in FP16.

Write a tag prompt
Add the quality tags
Add the negative prompt
Run the pipeline on a GPU
Save the image

Good to know:

• License: CreativeML Open RAIL++-M, so read its use terms
• Cost to run: none named, you use your own GPU
• Data it reaches: it downloads the weights from Hugging Face, and the card names nothing else it touches
• Not for: photographic images or styles other than anime
• The prompt style is tag-based, so plain sentences work less well than tags

On Hugging Face, 13 November 2026

• Task: text to image
• Library: diffusers
• Parameters: 2.6B
• Weights: about 5.1 GB in F16; about 1.3 GB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 1,026,807
• Likes: 582
• License: openrail++
• Languages: en
• Based on: stabilityai/stable-diffusion-xl-base-1.0
• Quantized versions: 3
• Fine-tunes: 70
• Tags: diffusers, safetensors, text-to-image, stable-diffusion, stable-diffusion-xl

Links

• Model: https://huggingface.co/cagliostrolab/animagine-xl-4.0
• Files: https://huggingface.co/cagliostrolab/animagine-xl-4.0/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download cagliostrolab/animagine-xl-4.0

Then follow the usage in its model card. Nothing in this kit runs by itself.