OmniFake-SET

Paper & Project Page: Omni-Fake: Benchmarking Unified Multimodal Social Media Deepfake Detection

Multimodal forgery detection: image, audio, video and audio-visual talking-head. Labels are REAL, FULL_SYNTHETIC and TAMPERED; tampered images and videos have a box ([x1, y1, x2, y2], 0-999), tampered audio has an interval ([start, end], seconds). OOD test set: OmniFake-OOD. Model: OmniFake.

Config	train	test
image	343,138	77,448
audio	106,057	13,798
video	76,994	18,255
audiovisual	5,678	636

Config	Fields	Media
image	id, label, box, image	384x384
audio	id, label, interval, duration, audio	16 kHz mono FLAC
video	id, label, box, video	448x448, 24 fps, H.264
audiovisual	id, label, video	H.264 with audio, up to 12 s

SFT

Examples with explanations, in sft/{modality}.jsonl (ms-swift format). Media are in sft/media/{modality}.tar, extract them inside sft/media/.

TAMPERED localization ground truth

All released TAMPERED samples have localization ground truth:

• Image: binary 384x384 masks in the image_masks config (id, box, mask), 120,749 rows.
• Video: binary 448x448 first-frame masks in the video_masks config (id, box, mask), 10,000 rows.
• Audio: temporal ground truth is the interval field in the main audio config; no mask is needed.
• Audio-visual: this release has no TAMPERED samples.

The sidecar id is identical to the corresponding main-config sample ID. White pixels are manipulated; black pixels are unmodified. Each sidecar box is derived directly from its mask. Video masks use the exact square geometry of the released 448x448 video first frame.