Request access to VendorBench-100
This repository is publicly accessible, but you have to accept the conditions to access its files and content.
VendorBench-100 is released for non-commercial research and benchmarking only. The corpus contains AI-generated and manipulated images together with authentic photographs used as controls. All third-party generator outputs and any images of external origin remain the property of their respective owners; inclusion here transfers no rights. Access is reviewed and granted per request. Please complete every field below and agree to all terms.
Log in or Sign Up to review the conditions and access this dataset content.
VendorBench-100
A compact, deliberately adversarial corpus for benchmarking deepfake / AI-generated-image detectors. It accompanies the paper "VendorBench-100: A Unified Cross-Paradigm Benchmark for Deepfake Image Detection" and the evaluation code at https://github.com/sharayu-20/vendorbench-100.
Access is gated. Complete the request form above; access is granted after review.
What this is
VendorBench-100 is built for difficulty, not scale: a hand-picked set of hard, modern fake images plus authentic negative controls, used to compare three detection paradigms (commercial APIs, zero-shot vision LLMs, open-source detectors) on common ground.
- 8 edge-case failure families: live / virtual face-swap smear · near-duplicate face-swap selfies · letterboxed text-to-video stills (Sora/Veo) · AI-avatar compositing seams (HeyGen) · fully synthetic text-to-image · on-device AI photo-edits of real scenes · opaque / unknown provenance · compressed research-dataset frames (DF40, FaceForensics++).
- Provenance registry documenting the source group and verification status of each image.
- Anti-leakage: models are served neutral numeric filenames; descriptive
fake_<GROUP>_NNNlabels are post-hoc join keys only.
Structure (once published)
Source/
├── fake/ label-named fake images (fake_<GROUP>_NNN.*)
├── real/ authentic negative controls (real_NNN.*)
├── neutral/ numeric-named copies served to models (anti-leakage)
└── Provenance/ per-image provenance registry
Intended use & restrictions
Research and benchmarking of detection systems only. Not for training generative/deepfake models, not for redistribution, not for any use that harms the individuals depicted. See the gated terms above.
Citation
@misc{deshmukh2026vendorbench100,
author = {Deshmukh, Sharayu N. and Deshmukh, Nilesh K.},
title = {VendorBench-100: A Unified Cross-Paradigm Benchmark for Deepfake Image Detection},
year = {2026}
}
- Downloads last month
- 6