README.md · animagine-xl-4.0

1

---

2

language:

3

- en

4

tags:

5

- text-to-image

6

- stable-diffusion

7

- safetensors

8

- stable-diffusion-xl

9

widget:

10

- text: >-

11

1girl, green hair, sweater, looking at viewer, upper body, beanie, outdoors,

12

night, turtleneck, masterpiece, high score, great score, absurdres

13

parameter:

14

negative_prompt: >-

15

lowres, bad anatomy, bad hands, text, error, missing fingers, extra digit,

16

fewer digits, cropped, worst quality, low quality, low score, bad score,

17

average score, signature, watermark, username, blurry

18

example_title: 1girl

19

- text: >-

20

1boy, male focus, green hair, sweater, looking at viewer, upper body,

21

beanie, outdoors, night, turtleneck, masterpiece, high score, great score,

22

absurdres

23

parameter:

24

negative_prompt: >-

25

lowres, bad anatomy, bad hands, text, error, missing fingers, extra digit,

26

fewer digits, cropped, worst quality, low quality, low score, bad score,

27

average score, signature, watermark, username, blurry

28

example_title: 1boy

29

license: openrail++

30

base_model:

31

- stabilityai/stable-diffusion-xl-base-1.0

32

---

33

34

# Animagine XL 4.0

35

36

![image/png](https://cdn-uploads.huggingface.co/production/uploads/6365c8dbf31ef76df4042821/_tsxjwf3VPu94xh9wJSbo.png)

37

38

## Overview

39

40

**Animagine XL 4.0**, also stylized as **Anim4gine**, is the ultimate anime-themed finetuned SDXL model and the latest installment of [Animagine XL series](https://huggingface.co/collections/Linaqruf/animagine-xl-669888c0add5adaf09754aca). Despite being a continuation, the model was retrained from [Stable Diffusion XL 1.0](https://huggingface.co/stabilityai/stable-diffusion-xl-base-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.

41

With the release of **Animagine XL 4.0 Opt (Optimized)**, the model has been further refined with an additional dataset, improving **stability**, **anatomy accuracy**, **noise reduction**, **color saturation**, and **overall color accuracy**. These enhancements make **Animagine XL 4.0 Opt** more consistent and visually appealing while maintaining the signature quality of the series.

42

43

44

## Changelog

45

- 2025-02-13 – Added Animagine XL 4.0 Opt

46

- Better stability for more consistent outputs

47

- Enhanced anatomy with more accurate proportions

48

- Reduced noise and artifacts in generations

49

- Fixed low saturation issues, resulting in richer colors

50

- Improved color accuracy for more visually appealing results

51

- 2025-01-24 – Initial release

52

53

54

## Model Details

55

56

- **Developed by**: [Cagliostro Research Lab](https://github.com/cagliostrolab)

57

- **Model type**: Diffusion-based text-to-image generative model

58

- **License**: [CreativeML Open RAIL++-M](https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0/blob/main/LICENSE.md)

59

- **Model Description**: This is a model that can be used to generate and modify specifically anime-themed images based on text prompt

60

- **Fine-tuned from**: [Stable Diffusion XL 1.0](https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0)

61

62

## Downstream Use

63

64

1. Use this model in our [`Hugging Face Spaces`](https://huggingface.co/spaces/cagliostrolab/animagine-xl-4.0)

65

2. Use it in [`ComfyUI`](https://github.com/comfyanonymous/ComfyUI) or [`Stable Diffusion Webui`](https://github.com/AUTOMATIC1111/stable-diffusion-webui)

66

3. Use it with 🧨 `diffusers`

67

68

## 🧨 Diffusers Installation

69

70

### 1. Install Required Libraries

71

72

```bash

73

pip install diffusers transformers accelerate safetensors --upgrade

74

```

75

76

### 2. Example Code

77

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.

78

79

```python

80

import torch

81

from diffusers import StableDiffusionXLPipeline

82

83

pipe = StableDiffusionXLPipeline.from_pretrained(

84

"cagliostrolab/animagine-xl-4.0",

85

torch_dtype=torch.float16,

86

use_safetensors=True,

87

custom_pipeline="lpw_stable_diffusion_xl",

88

add_watermarker=False

89

)

90

pipe.to('cuda')

91

92

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"

93

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"

94

95

image = pipe(

96

prompt,

97

negative_prompt=negative_prompt,

98

width=832,

99

height=1216,

100

guidance_scale=5,

101

num_inference_steps=28

102

).images[0]

103

104

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

105

```

106

107

## Usage Guidelines

108

109

The summary can be seen in the image for the prompt guideline.

110

111

![image/png](https://cdn-uploads.huggingface.co/production/uploads/64c088660a4d02f37a965f6c/YPe3MCnQAHM7nCZ1vQ7vI.png)

112

113

### 1. Prompt Structure

114

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

115

116

```

117

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

118

```

119

120

### 2. Quality Enhancement Tags

121

Add these tags at the end of your prompt:

122

123

```

124

masterpiece, high score, great score, absurdres

125

```

126

127

### 3. Recommended Negative Prompt

128

```

129

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

130

```

131

132

### 4. Optimal Settings

133

- **CFG Scale**: 4-7 (5 Recommended)

134

- **Sampling Steps**: 25-28 (28 Recommended)

135

- **Preferred Sampler**: Euler Ancestral (Euler a)

136

137

### 5. Recommended Resolutions

138

139

| Orientation | Dimensions | Aspect Ratio |

140

|------------|------------|--------------|

141

| Square | 1024 x 1024| 1:1 |

142

| Landscape | 1152 x 896 | 9:7 |

143

| | 1216 x 832 | 3:2 |

144

| | 1344 x 768 | 7:4 |

145

| | 1536 x 640 | 12:5 |

146

| Portrait | 896 x 1152 | 7:9 |

147

| | 832 x 1216 | 2:3 |

148

| | 768 x 1344 | 4:7 |

149

| | 640 x 1536 | 5:12 |

150

151

### 6. Final Prompt Structure Example

152

```

153

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

154

```

155

156

## Special Tags

157

158

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.

159

160

### Quality Tags

161

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

162

- `masterpiece`

163

- `best quality`

164

- `low quality`

165

- `worst quality`

166

167

| <img src="https://cdn-uploads.huggingface.co/production/uploads/6365c8dbf31ef76df4042821/bDdKraYxjiReKknlYJepR.png" width="100%" style="max-height: 400px; object-fit: contain;"> | <img src="https://cdn-uploads.huggingface.co/production/uploads/6365c8dbf31ef76df4042821/mAgMMKL2tBj8oBuWHTYUz.png" width="100%" style="max-height: 400px; object-fit: contain;"> |

168

|---|---|

169

| 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. |

170

171

### Score Tags

172

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:

173

- `high score`

174

- `great score`

175

- `good score`

176

- `average score`

177

- `bad score`

178

- `low score`

179

180

| <img src="https://cdn-uploads.huggingface.co/production/uploads/6365c8dbf31ef76df4042821/PXK6D1yhD8SND-VHFQOXD.png" width="100%" style="max-height: 400px; object-fit: contain;"> | <img src="https://cdn-uploads.huggingface.co/production/uploads/6365c8dbf31ef76df4042821/0uUw7DQ9IMiSNE_MZ9Uyf.png" width="100%" style="max-height: 400px; object-fit: contain;"> |

181

|---|---|

182

| 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. |

183

184

### Temporal Tags

185

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:

186

- `year 2005`

187

- `year {n}`

188

- `year 2025`

189

190

| <img src="https://cdn-uploads.huggingface.co/production/uploads/6365c8dbf31ef76df4042821/weRv0BmfkZrBhcW5NxXAI.png" width="100%" style="max-height: 400px; object-fit: contain;"> | <img src="https://cdn-uploads.huggingface.co/production/uploads/6365c8dbf31ef76df4042821/WwFoeLrbN2MkXuGHh91Ky.png" width="100%" style="max-height: 400px; object-fit: contain;"> |

191

|---|---|

192

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

193

194

### Rating Tags

195

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:

196

- `safe`

197

- `sensitive`

198

- `nsfw`

199

- `explicit`

200

201

## Training Information

202

203

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:

204

205

| Parameter | Value |

206

|-----------|--------|

207

| Hardware | 7 x H100 80GB SXM5 |

208

| Num Images | 8,401,464 |

209

| UNet Learning Rate | 2.5e-6 |

210

| Text Encoder Learning Rate | 1.25e-6 |

211

| Scheduler | Constant With Warmup |

212

| Warmup Steps | 5% |

213

| Batch Size | 32 |

214

| Gradient Accumulation Steps | 2 |

215

| Training Resolution | 1024x1024 |

216

| Optimizer | Adafactor |

217

| Input Perturbation Noise | 0.1 |

218

| Debiased Estimation Loss | Enabled |

219

| Mixed Precision | fp16 |

220

221

## Acknowledgement

222

223

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:

224

225

1. [**Moescape AI**](https://moescape.ai/): Our invaluable collaboration partner in model distribution and testing

226

2. **Lesser Rabbit**: For providing essential computing and research grants

227

3. [**Kohya SS**](https://github.com/kohya-ss): For developing the comprehensive open-source training framework

228

4. [**discus0434**](https://github.com/discus0434): For creating the industry-leading open-source Aesthetic Predictor 2.5

229

5. **Early testers**: For their dedication in providing critical feedback and thorough quality assurance

230

231

## Contributors

232

233

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

234

235

### Model

236

- [**KayfaHaarukku**](https://huggingface.co/kayfahaarukku)

237

- [**Raelina**](https://huggingface.co/Raelina)

238

- [**Linaqruf**](https://huggingface.co/Linaqruf)

239

240

### Gradio

241

- [**Damar Jati**](https://huggingface.co/DamarJati)

242

243

### Relations, finance, and quality assurance

244

- [**Scipius**](https://huggingface.co/Scipius2121)

245

- [**Asahina**](https://huggingface.co/Asahina2K)

246

- [**Bell**](https://huggingface.co/ItsMeBell)

247

- [**BoboiAzumi**](https://huggingface.co/Boboiazumi)

248

249

### Data

250

- [**Pomegranata**](https://huggingface.co/paripi)

251

- [**Kr1SsSzz**](https://huggingface.co/Kr1SsSzz)

252

- [**Fiqi**](https://huggingface.co/saikanov)

253

- [**William Adams Soeherman**](https://huggingface.co/williamsoeherman)

254

255

## Fundraising Has New Methods!

256

257

We're excited to introduce new fundraising methods through GitHub Sponsors to support training, research, and model development. Your support helps us push the boundaries of what's possible with AI.

258

259

**You can help us with:**

260

261

* **Donate**: Contribute via ETH, USDT, or USDC to the address below, or sponsor us on GitHub.

262

263

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

264

265

* **Feedback**: Let us know how we can improve.

266

267

**Donation Address**:

268

269

ETH/USDT/USDC(e): ```0xd8A1dA94BA7E6feCe8CfEacc1327f498fCcBFC0C```

270

271

**Github Sponsor**: [https://github.com/sponsors/cagliostrolab/](https://github.com/sponsors/cagliostrolab/)

272

273

274

<details>

275

<summary>Why do we use Cryptocurrency?</summary>

276

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.

277

</details>

278

279

<details>

280

<summary>Want to Donate in Non-Crypto Currency?</summary>

281

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.

282

</details>

283

284

## Join Our Discord Server

285

Feel free to join our discord server

286

<div style="text-align: center;">

287

<a href="https://discord.gg/cqh9tZgbGc">

288

<img src="https://discord.com/api/guilds/1115542847395987519/widget.png?style=banner2" alt="Discord Banner 2"/>

289

</a>

290

</div>

291

292

## Limitations

293

294

- **Prompt Format**: Limited to tag-based text prompts; natural language input may not be effective

295

- **Anatomy**: May struggle with complex anatomical details, particularly hand poses and finger counting

296

- **Text Generation**: Text rendering in images is currently not supported and not recommended

297

- **New Characters**: Recent characters may have lower accuracy due to limited training data availability

298

- **Multiple Characters**: Scenes with multiple characters may require careful prompt engineering

299

- **Resolution**: Higher resolutions (e.g., 1536x1536) may show degradation as training used original SDXL resolution

300

- **Style Consistency**: May require specific style tags as training focused more on identity preservation than style consistency

301

302

## License

303

304

This model adopts the original [CreativeML Open RAIL++-M License](https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0/blob/main/LICENSE.md) from Stability AI without any modifications or additional restrictions. The license terms remain exactly as specified in the original SDXL license, which includes:

305

306

- ✅ **Permitted**: Commercial use, modifications, distributions, private use

307

- ❌ **Prohibited**: Illegal activities, harmful content generation, discrimination, exploitation

308

- ⚠️ **Requirements**: Include license copy, state changes, preserve notices

309

- 📝 **Warranty**: Provided "AS IS" without warranties

310

311

Please refer to the [original SDXL license](https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0/blob/main/LICENSE.md) for the complete and authoritative terms and conditions.