Syn-Vis-v0

# Syn-Vis-v0: A Dataset of Synthetic Faces

Syn-Vis-v0 (Synthetic Visage Version 0) is a dataset of 480 synthetic faces generated with Qwen-Image and Qwen-Image-Edit-2509.


- **Diversity**:
- The dataset is balanced across ethnicities - approximately 60 images per broad category (Asian, Black, Hispanic, White, Indian, Middle Eastern) and 120 ethnically ambiguous images.
- Wide range of skin-tones, facial features, hairstyles, hair colors, nose shapes, eye shapes, and eye colors.
- **Quality**:
- Rendered at 2048x2048 resolution using Qwen-Image-Edit-2509 (BF16) and 50 steps.
- Checked for artifacts, defects, and watermarks.
- **Style**: semi-realistic, 3d-rendered CGI, with hints of photography and painterly accents.
- **Captions**: Natural language descriptions consolidated from multiple caption sources using GPT-OSS-120B.
- **Metadata**: Each image is accompanied by ethnicity/race analysis scores (0-100) across six categories (Asian, Indian, Black, White, Middle Eastern, Latino Hispanic) generated using DeepFace.
- **Analysis Cards**: Each image has a corresponding analysis card showing similarity to other faces in the dataset.
- **Size**: 1.6GB for the 480 images, 0.7GB of misc files (analysis cards, banners, ...).



## Dataset Structure

```txt
Syn-Vis-v0/
├── images/
│ └── base/ # Main dataset images
├── metadata.csv # Root-level metadata file for Hugging Face preview
├── dataset_info.json # Schema definition for image and metadata fields
├── misc/ # Analysis cards, banners, etc.
└── README.md
```

## Metadata Format

The `metadata.csv` contains the following columns:

- `file_name`: Image filename (e.g., "001-0042.png")
- `caption`: Consolidated natural language description
- `race_asian`: Asian demographic score (0-100)
- `race_indian`: Indian demographic score (0-100)
- `race_black`: Black demographic score (0-100)
- `race_white`: White demographic score (0-100)
- `race_middle_eastern`: Middle Eastern demographic score (0-100)
- `race_latino_hispanic`: Latino Hispanic demographic score (0-100)
- `dominant_race`: Primary predicted demographic category

## Caption Quality

Captions are consolidated from multiple JoyCaption-Beta-One outputs using GPT-OSS-120B and feature:

- **Natural language**: Start with "The woman..." or "A woman..."
- **Specific descriptions**: Physical features described precisely rather than using broad demographic categories
- **Structured order**: Face → hair/physical features → clothing → background → lighting
- **Style-neutral**: Remove technical photography terms and medium references
- **Flowing narrative**: Natural sentences without section headers

Example caption:
> "The woman has a smooth medium‑brown complexion that catches a gentle, even glow. Her eyes are large, dark brown and framed by thick, dark lashes, giving them a calm, slightly serious look as she gazes directly forward..."

## Use Cases

- Raw training data for small models
- Base images for image-to-image generation tasks
- Base images for style transfer
- Whatever you want!

## Statistics

- **Ethnicities**: (by Dominant Race Counts; some faces show similar scores across multiple categories)
- White: 94 images
- Latino Hispanic: 93 images
- Asian: 90 images
- Indian: 70 images
- Black: 68 images
- Middle Eastern: 65 images
- **Skin tones**: Full spectrum from very light to very dark
- **Facial features**: Wide variety of eye shapes, nose shapes, lip shapes
- **Hair styles**: Various textures, colors, and arrangements
- **Backgrounds**: Dark and light, plain and scenic
- **Ages**: Almost exclusively 30 ± 5 years according to analysis with DeepFace.

> Anecdotally, Asian, White, and Black, were predicted with a single high score (85+) much more frequently than Latino Hispanic, Indian and Middle Eastern.

## Ethical Considerations and Other Notes

- The ethnicity/race labels are generated by automated analysis and should not be considered ground truth for real-world applications involving human subjects. Their primary purpose is to ensure coverage of wide range facial features.
- Only female-presenting individuals are included. I decided against including male-presenting individuals because **beards** - I didn't know how well the classifiers would handle them (obscured features), so I decided to avoid that complexity.
- All faces were explicitly declared female-presenting (in the prompt and the caption), however, DeepFace occasionally suggested some images may be male-presenting.
- The dataset has a strong beauty bias and the faces are unusually symmetrical.

## Creation Process

1. **Initial Image Generation**: Generated an initial set of 5,500 images at 768x768 using Qwen-Image (FP8). Facial features were randomly selected from lists and then written into natural prompts by Qwen3:30b-a3b. The style prompt was "Photo taken with telephoto lens (130mm), low ISO, high shutter speed".
2. **Initial Analysis & Captioning**: Each of the 5,500 images was captioned three times using JoyCaption-Beta-One. These initial captions were then consolidated using Qwen3:30b-a3b. Concurrently, demographic analysis was run using DeepFace.
3. **Selection**: A balanced subset of 480 images was selected based on the aggregated demographic scores and visual inspection.
4. **Enhancement**: Minor errors like faint watermarks and artifacts were manually corrected using GIMP.
5. **Upscaling & Refinement**: The selected images were upscaled to 2048x2048 using Qwen-Image-Edit-2509 (BF16) with 50 steps at a CFG of 4. The prompt guided the model to transform the style to a high-quality 3d-rendered CGI portrait while maintaining the original likeness and composition.
6. **Final Captioning**: To ensure captions accurately reflected the final, upscaled images and accounted for any minor perspective shifts, the 480 images were fully re-captioned. Each image was captioned three times with JoyCaption-Beta-One, and these were consolidated into a final, high-quality description using GPT-OSS-120B.
7. **Final Analysis**: Each final image was analyzed using DeepFace to generate the demographic scores and similarity analysis cards present in the dataset.

### Models and Tools Used

- **Qwen-Image**: Image Generation
- **Qwen-Image-Edit-2509**: Image Refinement/Upscaling
- **JoyCaption-Beta-One**: Captioning
- **Qwen3:30b-a3b**: Prompt Writing & Initial Caption Consolidation
- **GPT-OSS-120B**: Final Caption Consolidation
- **Tools**: vLLM, DeepFace, Python, R, GIMP, ComfyUI

## Projects That Use Syn-Vis-v0

- Coming soon!
- Your project here?

## Known Issues

- 001-0309: Appears to be wearing a mask, likely introduced during the image-to-image upscaling step.

## License

- **Images**: CC0 (Public Domain) - Individual synthetic images are released to the public domain
- **Dataset compilation, metadata, and documentation**: CC-BY-SA-4.0 - The curation work, analysis, and documentation

You may use these images and this dataset for any purpose, including commercial use. If you use this dataset, I will appreciate attribution.

## Citation

```bibtex
@dataset{syn-vis-v0-2025,
title={Syn-Vis-v0: A Synthetic Face Dataset},
author={Wyss, Reto},
year={2025},
url={https://huggingface.co/datasets/retowyss/Syn-Vis-v0},
note={Images: CC0 (Public Domain); Dataset compilation and documentation: CC-BY-SA-4.0}
}
```

# Syn-Vis-v0：合成人脸数据集（Syn-Vis-v0）

Syn-Vis-v0（Synthetic Visage Version 0，即合成人脸版本0）是一个包含480张合成人脸的数据集，由Qwen-Image和Qwen-Image-Edit-2509生成。


- **多样性**：
该数据集在种族分布上保持均衡——六大宽泛类别（亚洲人、黑人、西班牙裔拉丁人、白人、印度人、中东人）各约60张图像，另有120张种族模糊的图像。
涵盖丰富的肤色、面部特征、发型、发色、鼻型、眼型及虹膜颜色。
- **质量**：
采用Qwen-Image-Edit-2509（BF16精度）以2048×2048分辨率渲染，共50步推理。
已检查过伪影、缺陷与水印问题。
- **风格**：半写实的3D渲染计算机生成图像（Computer Generated Imagery, CGI），兼具摄影质感与绘画艺术笔触。
- **描述文本**：通过GPT-OSS-120B整合多来源的自然语言描述生成。
- **元数据**：每张图像均附带由DeepFace生成的六大类别（亚洲人、印度人、黑人、白人、中东人、拉丁裔西班牙人）的种族/人种分析得分（0-100区间）。
- **分析卡片**：每张图像均配有对应的分析卡片，展示其与数据集中其他人脸的相似度。
- **规模**：480张图像占用1.6GB存储空间，附属文件（分析卡片、横幅等）占用0.7GB。



## 数据集结构

txt
Syn-Vis-v0/
├── images/
│ └── base/ # 主数据集图像目录
├── metadata.csv # 用于Hugging Face预览的根级元数据文件
├── dataset_info.json # 图像与元数据字段的架构定义文件
├── misc/ # 分析卡片、横幅等附属文件目录
└── README.md

## 元数据格式

`metadata.csv`包含以下字段：
- `file_name`：图像文件名（例如："001-0042.png"）
- `caption`：整合后的自然语言描述文本
- `race_asian`：亚洲人种人口统计得分（0-100）
- `race_indian`：印度人种人口统计得分（0-100）
- `race_black`：黑人种人口统计得分（0-100）
- `race_white`：白人种人口统计得分（0-100）
- `race_middle_eastern`：中东人种人口统计得分（0-100）
- `race_latino_hispanic`：拉丁裔西班牙籍人种人口统计得分（0-100）
- `dominant_race`：主要预测人口统计类别

## 描述文本质量

描述文本由GPT-OSS-120B整合多份JoyCaption-Beta-One生成的结果而来，具备以下特点：
- **自然语言风格**：以"The woman..."或"A woman..."开头
- **精准描述**：对物理特征进行细致刻画，而非使用宽泛的人口统计类别
- **结构化顺序**：按面部→毛发/物理特征→衣着→背景→光线的逻辑排布
- **风格中立**：移除专业摄影术语与媒介相关表述
- **流畅叙事**：自然语句，无分段标题

示例描述文本：
> "这位女性拥有光滑的中等棕褐色肤色，在柔和均匀的光线映衬下泛着光泽。她的眼睛大而深邃，呈棕褐色，眼周环绕着浓密的深色睫毛，直视镜头时神情沉静且略带严肃..."

## 应用场景

- 小型模型的原始训练数据
- 图像到图像生成任务的基底图像
- 风格迁移任务的基底图像
- 任意你所需的用途！

## 统计数据

- **种族分布（按主导种族计数；部分面孔在多个类别中得分相近）**：
白人：94张
拉丁裔西班牙人：93张
亚洲人：90张
印度人：70张
黑人：68张
中东人：65张
- **肤色**：覆盖从极浅到极深的全光谱范围
- **面部特征**：包含丰富多样的眼型、鼻型与唇型
- **发型**：涵盖多种纹理、发色与造型
- **背景**：涵盖明暗各异、简洁与场景丰富的多种背景
- **年龄**：经DeepFace分析，几乎全部为30±5岁的成年人。

> 据观察，亚洲人、白人与黑人的单一高得分（85+）预测频率远高于拉丁裔西班牙人、印度人与中东人。

## 伦理考量与其他说明

- 种族/人种标签由自动化分析生成，不应作为涉及人类主体的真实世界应用中的真值（ground truth）。其核心用途是确保覆盖广泛的面部特征范围。
- 本数据集仅包含女性呈现特征的个体。未纳入男性呈现特征的个体，原因是胡须可能遮挡面部特征，且暂不确定分类器对这类遮挡的处理效果，因此选择规避该复杂度。
- 所有生成人脸均在提示词与描述文本中明确标注为女性呈现特征，但DeepFace偶尔会将部分图像判定为男性呈现特征。
- 该数据集存在较强的审美偏向，人脸对称性异于常规。

## 生成流程

1. **初始图像生成**：使用Qwen-Image（FP8精度）生成5500张768×768分辨率的初始图像。面部特征从预设列表中随机选取，再由Qwen3:30b-a3b转化为自然语言提示词。风格提示词为"使用130mm长焦镜头拍摄，低ISO，高快门速度"。
2. **初始分析与描述**：对5500张图像各使用JoyCaption-Beta-One生成3次描述，再由Qwen3:30b-a3b整合为最终初始描述。同时使用DeepFace进行人口统计分析。
3. **筛选**：基于聚合后的人口统计得分与人工视觉检查，筛选出均衡分布的480张图像子集。
4. **修正**：使用GIMP手动修正轻微瑕疵，如淡水印与伪影。
5. **放大与优化**：使用Qwen-Image-Edit-2509（BF16精度）将选中图像放大至2048×2048分辨率，共50步推理，CFG（Classifier-Free Guidance）值为4。提示词引导模型将风格转化为高质量3D渲染计算机生成肖像，同时保留原始人像特征与构图。
6. **最终描述生成**：为确保描述文本准确反映放大后的最终图像，并适配可能存在的轻微视角变化，对480张图像重新进行描述：各使用JoyCaption-Beta-One生成3次描述，再由GPT-OSS-120B整合为高质量的最终描述。
7. **最终分析**：使用DeepFace对最终图像进行分析，生成人口统计得分与数据集中的相似度分析卡片。

### 所用模型与工具

- **Qwen-Image**：图像生成
- **Qwen-Image-Edit-2509**：图像优化与放大
- **JoyCaption-Beta-One**：图像描述生成
- **Qwen3:30b-a3b**：提示词编写与初始描述整合
- **GPT-OSS-120B**：最终描述文本整合
- **工具**：vLLM、DeepFace、Python、R、GIMP、ComfyUI

## 已使用Syn-Vis-v0的项目

- 即将上线！
- 你的项目？

## 已知问题

- 001-0309：图像中人物似乎佩戴口罩，大概率是图像到图像放大步骤引入的。

## 许可协议

- **图像**：CC0（公共领域）—— 单张合成图像已发布至公共领域
- **数据集编译、元数据与文档**：CC-BY-SA-4.0—— 数据集的整理工作、分析内容与文档

您可将本数据集与图像用于任何用途，包括商业用途。若您使用本数据集，我们将感谢您的署名。

## 引用格式

bibtex
@dataset{syn-vis-v0-2025,
title={Syn-Vis-v0: 合成人脸数据集},
author={Wyss, Reto},
year={2025},
url={https://huggingface.co/datasets/retowyss/Syn-Vis-v0},
note={图像：CC0（公共领域）；数据集编译与文档：CC-BY-SA-4.0}
}