kaikaiyao/ffhq-image-attribution

---
configs:
- config_name: stylegan2-ffhq-256
data_files:
- viewer-stylegan2-ffhq-256.parquet
- config_name: stylegan3-ffhq-256
data_files:
- viewer-stylegan3-ffhq-256.parquet
- config_name: r3gan-ffhq-256
data_files:
- viewer-r3gan-ffhq-256.parquet
- config_name: cips-ffhq-256
data_files:
- viewer-cips-ffhq-256.parquet
- config_name: ganformer-ffhq-256
data_files:
- viewer-ganformer-ffhq-256.parquet
- config_name: styleswin-ffhq-256
data_files:
- viewer-styleswin-ffhq-256.parquet
- config_name: vqvae-ffhq-256
data_files:
- viewer-vqvae-ffhq-256.parquet
- config_name: nvae-ffhq-256
data_files:
- viewer-nvae-ffhq-256.parquet
- config_name: vdvae-ffhq-256
data_files:
- viewer-vdvae-ffhq-256.parquet
- config_name: adm-ffhq-256
data_files:
- viewer-adm-ffhq-256.parquet
- config_name: ldm-ffhq-256
data_files:
- viewer-ldm-ffhq-256.parquet
- config_name: ncsnpp-ffhq-256
data_files:
- viewer-ncsnpp-ffhq-256.parquet
license: other
language:
- en
task_categories:
- image-classification
tags:
- ffhq
- model-attribution
- generated-images
- face-generation
- gan
- vae
- diffusion
- benchmark
---

# FFHQ Image Attribution

[](https://huggingface.co/datasets/kaikaiyao/ffhq-image-attribution)
[](https://huggingface.co/datasets/kaikaiyao/ffhq-image-attribution)
[](https://huggingface.co/datasets/kaikaiyao/ffhq-image-attribution)
[](https://huggingface.co/datasets/kaikaiyao/ffhq-image-attribution)

A public benchmark for **FFHQ model attribution**, built from twelve face generators spanning GAN, VAE, and diffusion families.


> Version `v2` includes `10,000` images from each of `12` FFHQ-trained generators for a total of `120,000` images.

## Why this dataset

- Same image domain across multiple FFHQ generators makes source attribution cleaner and easier to study.
- Public metadata links each image to its source model, family, release, seed, and file integrity hash.
- The dataset viewer uses lightweight embedded thumbnails so you can browse each model subset quickly.
- The release is deliberately simple: only model identity varies, without prompt or text metadata.

## At a glance

| Images | Models | Families | Version |
|---:|---:|---:|---|
| **120,000** | **12** | **3** | **v2** |

## Coverage

All images are `256x256` RGB face images from FFHQ-trained generators. Each subset corresponds to one model family member.

| Model | Family | Checkpoint |
|---|---|---|
| `stylegan2-ffhq-256` | `gan` | `https://nvlabs-fi-cdn.nvidia.com/stylegan2-ada-pytorch/pretrained/paper-fig7c-training-set-sweeps/ffhq140k-paper256-noaug.pkl` |
| `stylegan3-ffhq-256` | `gan` | `https://api.ngc.nvidia.com/v2/models/nvidia/research/stylegan3/versions/1/files/stylegan3-t-ffhqu-256x256.pkl` |
| `r3gan-ffhq-256` | `gan` | `brownvc/R3GAN-FFHQ-256x256` |
| `cips-ffhq-256` | `gan` | `https://drive.google.com/file/d/1JRd4ZpMDmlkbNlxnVvZx77Eyfac53KSq/view?usp=sharing` |
| `ganformer-ffhq-256` | `gan` | `https://drive.google.com/uc?id=1-b0vwevUQs6LI_EybdO8XJ5uYSx63vEa` |
| `styleswin-ffhq-256` | `gan` | `https://drive.google.com/file/d/1OjYZ1zEWGNdiv0RFKv7KhXRmYko72LjO/view` |
| `vqvae-ffhq-256` | `vae` | `kohido/ffhq256_vqvae_mhvq` |
| `nvae-ffhq-256` | `vae` | `https://drive.google.com/uc?id=1lQzywY5O71Z5NqAUJUcWy2Q1K2hPFO6j` |
| `vdvae-ffhq-256` | `vae` | `https://openaipublic.blob.core.windows.net/very-deep-vaes-assets/vdvae-assets/ffhq256-iter-1700000-model-ema.th` |
| `adm-ffhq-256` | `diffusion` | `xutongda/adm_ffhq_256x256` |
| `ldm-ffhq-256` | `diffusion` | `kaayaanil/ldm-ffhq-256` |
| `ncsnpp-ffhq-256` | `diffusion` | `https://drive.google.com/uc?id=1-mtdSwuefIZA0n85QWScQo2WRvJNWwUy` |

## Gallery


## How to use

Load a viewer subset with `datasets`:

```python
from datasets import load_dataset

ds = load_dataset("kaikaiyao/ffhq-image-attribution", "stylegan2-ffhq-256", split="train")
print(ds[0]["source_id"], ds[0]["seed"])
```

Or read the full metadata table with `pandas`:

```python
import pandas as pd

df = pd.read_parquet(
"https://huggingface.co/datasets/kaikaiyao/ffhq-image-attribution/resolve/main/metadata/all.parquet"
)
print(df[["source_id", "seed", "image_path"]].head())
```

## Metadata

`metadata/all.parquet` is the main table for the release.

- Identity: `source_id`, `family`, `seed`
- Image and integrity: `image_path`, `image_size`, `sha256`
- Release: `release`

Each subset also has:

- `metadata/by_model/<source_id>.parquet`
- `viewer-<source_id>.parquet`

## Uses and limitations

- Intended for research on model attribution, image provenance, model fingerprinting, and generated-face forensics.
- All images are `256x256` and come from FFHQ-trained generators.
- `vqvae-ffhq-256` is reconstruction-based, unlike the other models that directly sample generated images.
- This public release includes `10,000` images per model from the current FFHQ bank snapshot.
- Use of this dataset remains subject to the licenses and terms of the upstream model checkpoints and FFHQ-related resources.

## Citation

If you use this dataset, please cite:

```bibtex
@dataset{yao2026ffhq_image_attribution,
author = {{Kai Yao}},
title = {{FFHQ Image Attribution}},
year = {2026},
publisher = {{Hugging Face}},
url = {https://huggingface.co/datasets/kaikaiyao/ffhq-image-attribution},
note = {Version: v2}
}
```

配置项：
- 配置名称：stylegan2-ffhq-256
数据文件：
- viewer-stylegan2-ffhq-256.parquet
- 配置名称：stylegan3-ffhq-256
数据文件：
- viewer-stylegan3-ffhq-256.parquet
- 配置名称：r3gan-ffhq-256
数据文件：
- viewer-r3gan-ffhq-256.parquet
- 配置名称：cips-ffhq-256
数据文件：
- viewer-cips-ffhq-256.parquet
- 配置名称：ganformer-ffhq-256
数据文件：
- viewer-ganformer-ffhq-256.parquet
- 配置名称：styleswin-ffhq-256
数据文件：
- viewer-styleswin-ffhq-256.parquet
- 配置名称：vqvae-ffhq-256
数据文件：
- viewer-vqvae-ffhq-256.parquet
- 配置名称：nvae-ffhq-256
数据文件：
- viewer-nvae-ffhq-256.parquet
- 配置名称：vdvae-ffhq-256
数据文件：
- viewer-vdvae-ffhq-256.parquet
- 配置名称：adm-ffhq-256
数据文件：
- viewer-adm-ffhq-256.parquet
- 配置名称：ldm-ffhq-256
数据文件：
- viewer-ldm-ffhq-256.parquet
- 配置名称：ncsnpp-ffhq-256
数据文件：
- viewer-ncsnpp-ffhq-256.parquet
许可协议：其他
语言：英语
任务类别：图像分类
标签：
- ffhq
- 模型归因
- 生成图像
- 人脸生成
- 生成对抗网络（GAN）
- 变分自编码器（VAE）
- 扩散模型
- 基准测试

# FFHQ 图像归因数据集

[](https://huggingface.co/datasets/kaikaiyao/ffhq-image-attribution)
[](https://huggingface.co/datasets/kaikaiyao/ffhq-image-attribution)
[](https://huggingface.co/datasets/kaikaiyao/ffhq-image-attribution)
[](https://huggingface.co/datasets/kaikaiyao/ffhq-image-attribution)

本数据集是面向**FFHQ（Flickr-Faces-HQ）模型归因**的公开基准测试集，由12款基于FFHQ训练的人脸生成模型构建而成，涵盖生成对抗网络（GAN）、变分自编码器（VAE）以及扩散模型三大模型族。


> 版本`v2`包含12款FFHQ训练生成器各自产出的10000张图像，总计120000张图像。

## 数据集设计初衷

- 所有图像均来自同一FFHQ图像域，使得模型归因研究更加清晰便捷。
- 公开的元数据可将每张图像与其源模型、模型族、发布版本、随机种子以及文件完整性哈希值相关联。
- 数据集查看器内置轻量级缩略图，支持快速浏览各模型子集。
- 本次发布设计简洁：仅模型类别存在差异，不包含提示词或文本元数据。

## 概览

| 图像数量 | 模型数量 | 模型族数量 | 版本 |
|---:|---:|---:|---|
| **120,000** | **12** | **3** | **v2** |

## 数据集覆盖范围

所有图像均为256×256分辨率的RGB人脸图像，均来自FFHQ训练的生成模型。每个子集对应一个模型族成员。

| 模型名称 | 模型族 | 模型权重检查点 |
|---|---|---|
| `stylegan2-ffhq-256` | 生成对抗网络（GAN） | `https://nvlabs-fi-cdn.nvidia.com/stylegan2-ada-pytorch/pretrained/paper-fig7c-training-set-sweeps/ffhq140k-paper256-noaug.pkl` |
| `stylegan3-ffhq-256` | 生成对抗网络（GAN） | `https://api.ngc.nvidia.com/v2/models/nvidia/research/stylegan3/versions/1/files/stylegan3-t-ffhqu-256x256.pkl` |
| `r3gan-ffhq-256` | 生成对抗网络（GAN） | `brownvc/R3GAN-FFHQ-256x256` |
| `cips-ffhq-256` | 生成对抗网络（GAN） | `https://drive.google.com/file/d/1JRd4ZpMDmlkbNlxnVvZx77Eyfac53KSq/view?usp=sharing` |
| `ganformer-ffhq-256` | 生成对抗网络（GAN） | `https://drive.google.com/uc?id=1-b0vwevUQs6LI_EybdO8XJ5uYSx63vEa` |
| `styleswin-ffhq-256` | 生成对抗网络（GAN） | `https://drive.google.com/file/d/1OjYZ1zEWGNdiv0RFKv7KhXRmYko72LjO/view` |
| `vqvae-ffhq-256` | 变分自编码器（VAE） | `kohido/ffhq256_vqvae_mhvq` |
| `nvae-ffhq-256` | 变分自编码器（VAE） | `https://drive.google.com/uc?id=1lQzywY5O71Z5NqAUJUcWy2Q1K2hPFO6j` |
| `vdvae-ffhq-256` | 变分自编码器（VAE） | `https://openaipublic.blob.core.windows.net/very-deep-vaes-assets/vdvae-assets/ffhq256-iter-1700000-model-ema.th` |
| `adm-ffhq-256` | 扩散模型 | `xutongda/adm_ffhq_256x256` |
| `ldm-ffhq-256` | 扩散模型 | `kaayaanil/ldm-ffhq-256` |
| `ncsnpp-ffhq-256` | 扩散模型 | `https://drive.google.com/uc?id=1-mtdSwuefIZA0n85QWScQo2WRvJNWwUy` |

## 图库


## 使用方法

使用`datasets`库加载指定的查看器子集：

python
from datasets import load_dataset

ds = load_dataset("kaikaiyao/ffhq-image-attribution", "stylegan2-ffhq-256", split="train")
print(ds[0]["source_id"], ds[0]["seed"])

或者使用`pandas`读取完整元数据表：

python
import pandas as pd

df = pd.read_parquet(
"https://huggingface.co/datasets/kaikaiyao/ffhq-image-attribution/resolve/main/metadata/all.parquet"
)
print(df[["source_id", "seed", "image_path"]].head())

## 元数据说明

`metadata/all.parquet` 是本次发布的主元数据表，包含以下字段：
- 身份信息：`source_id`（源模型ID）、`family`（模型族）、`seed`（随机种子）
- 图像与完整性信息：`image_path`（图像路径）、`image_size`（图像尺寸）、`sha256`（SHA256完整性哈希）
- 发布信息：`release`（发布版本）

每个模型子集还包含以下元数据文件：
- `metadata/by_model/<source_id>.parquet`
- `viewer-<source_id>.parquet`

## 使用场景与局限性

- 本数据集旨在用于模型归因、图像溯源、模型指纹识别以及生成人脸取证等相关研究。
- 所有图像均为256×256分辨率，均来自FFHQ训练的生成模型。
- `vqvae-ffhq-256` 属于基于重构的生成模型，与其他直接采样生成图像的模型有所不同。
- 本次公开发布包含来自当前FFHQ数据集快照的每个模型对应的10000张图像。
- 使用本数据集需遵守上游模型权重检查点以及FFHQ相关资源的许可协议与使用条款。

## 引用方式

若您使用本数据集，请引用以下文献：

bibtex
@dataset{yao2026ffhq_image_attribution,
author = {{Kai Yao}},
title = {{FFHQ Image Attribution}},
year = {2026},
publisher = {{Hugging Face}},
url = {https://huggingface.co/datasets/kaikaiyao/ffhq-image-attribution},
note = {Version: v2}
}