Datasets:

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

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:

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:

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:

@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}
}