FFHQ Image Attribution: 24,000 Generated Faces from 12 Models

A public benchmark for model attribution built from twelve face generators trained on the FFHQ dataset. Version v1 includes 2,000 images from each of the 12 generators, spanning GAN, VAE, and diffusion families, for a total of 24,000 images. The dataset was created by kaikaiyao and last updated on Hugging Face in April 2026.

Use Cases

• Train classifiers for source model attribution based on the provided metadata linking each image to its generator.
• Benchmark attribution methods across different generative architectures (GAN, VAE, diffusion) using the same image domain.
• Study the unique artifacts or fingerprints of specific FFHQ-trained models mentioned in the description.
• Analyze the difficulty of attribution across different model families as suggested by the dataset's purpose.

Strengths

• Contains 24,000 images, with a structured 2,000 images per generator.
• Built from 12 distinct generators, providing diversity across GAN, VAE, and diffusion model families.
• Public metadata links each image to its source model, family, release, and seed for traceability.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is known, but specific image features, file formats, and license details are not provided in the input.
• Data may reflect the bias inherent to the FFHQ source domain and the specific 12 generators selected.

Provenance

Source

Hugging Face dataset by author kaikaiyao.

Collection Method

Images generated from twelve FFHQ-trained models, with metadata linking each image to its source.

Time Range

Version v1; specific generation dates not provided.

Freshness

Last updated 2026-04-02 18:52:19; freshness should be verified.

Geography

Spatial coverage is not applicable for synthetic face images.