DATASHEET.md

MOZAIC: Datasheet

1. Why it was made

MOZAIC is for training and testing audio-visual deepfake detectors in a controlled way. The same real clip appears in all four classes, every clip goes through the same final encoding, and known shortcuts were measured and removed where possible, so a detector has to find the fake, not a side effect of how the data was made. Made by: Vector Institute with AIXPERT. Funding:

2. What is in it

• Clips: 137,745 talking-face videos with sound.
• People: 1,000 from the VoxCeleb2 dev set, half men and half women: public figures in interview and talk videos from YouTube.
• Pairs: a source clip (person A speaking) and a target person (B, same gender). One pair gives four classes: RVRA, RVFA, FVRA, FVFA. 7,590 pairs at 77 frames, 7,424 at 153 frames.
• Arms: main, and vocoded for RVRA and FVRA (the same clip with its voice passed through BigVGAN; the label does not change).
• Lengths: 77 frames (3.08 s) and 153 frames (6.12 s); the two lengths use different source clips. Animate is 77 frames only: at 6 s its output blurred, which would let a detector find it by blur alone.
• Splits: by person, with equal numbers of men and women. Both people of a pair are in the same split, and nobody appears in two splits.
• Per clip: label, generators, length, arm, split, whose face, whose voice, gender.
• Per pair: the source and target clips used, and which generators kept it.
• Personal data: VoxCeleb2 ids and gender (from VoxCeleb2's metadata). No names.
• Real clips are stored once per length and shared by all generators.
• Questions: four tasks, in qas/ and qas_small/: T1 which part of a clip is fake, T2 how the video was made, T3 whether a clip is real or fake and the same person as a second clip, T4 a written explanation (MOZAIC-Small only). T1 to T3 answers come from the clip's labels. T4 answers were written by language models from those labels and checked by code and a second model. Counts are in README.md.
• MOZAIC-Small: 16,995 clips from 1,858 whole pairs, drawn per split to mirror the full mix (seed 7), with every person the source of at least one pair. It shares the full set's videos.

3. How it was made

• Source: VoxCeleb2 dev set, from a public mirror, checked against the official VoxCeleb2 checksums. Each clip was cut from its start to 77 or 153 frames, 224x224, 25 fps, 16 kHz mono.
• Fake faces: FaceFusion (inswapper_128_fp16) swaps the inner face. Wan2.2-S2V-14B (480p, 40 steps, cfg 4.5) makes the whole video from one picture of the target plus the audio, at 16 fps, retimed to 25 fps. Wan2.2-Animate-14B, replace mode, 480p.
• Fake voices: Seed-VC (5 steps, cfg 0.3) turns the source's speech into the target's voice, using one reference recording of the target (another of the target's VoxCeleb2 clips, never from the session the face picture came from); it is listed per pair in pairs.jsonl.
• Vocoded voices: real audio passed through BigVGAN v2 (22 kHz), the same vocoder and settings Seed-VC uses, without the conversion.
• Final step, the same for every clip of every class: brightness, colour and loudness brought to fixed targets; H.264 (crf 18) video and PCM audio; no silence trimmed. For the release, encoder version text was removed from every file by rewriting the container (no re-encoding; every file was checked to decode to exactly the same picture, sound and timing as before). Then every clip got the same 1:1 aspect-ratio tag and the same start of sound (50 ms of silence, then a 10 ms fade-in); every file was checked to decode to exactly its source picture, and to its source sound with that start.
• Reproducibility: the pair draw used seed 42; every Seed-VC, S2V and Animate job used its own fixed seed, recorded in our build records; FaceFusion has no random sampling.

4. Quality gate

Every pair was checked separately for each generator; pairs that failed were removed.

Check	What it tests	Tool
Face	The fake looks like the target, not the source	ArcFace (insightface buffalo_l), 8 frames
Picture	Flicker, melting and sharpness, compared with the real clip	our own measures
Voice	The fake voice sounds like the target's reference recording; the vocoded voice still sounds like the source	ECAPA speaker model (SpeechBrain)
Lip sync	Lips match the sound	SyncNet, on a sample of about 10% of pairs

• Limits: taken from real clips only (what 95%, or 99% for flicker, of real recordings do), except three judgement calls: face at least 0.30, melt at most 2.0, sharpness at least 0.5. Frozen on 5 Sep 2026, last changed on 14 Sep 2026; the first batch was used to set them.
• Two tests are switched off for one generator each: "too much of the source face left" for FaceFusion (it keeps the source's hair and head by design) and "melting" for S2V (it invents its own head movement).
• Repeatability: scoring the same clips twice on different machines gave identical numbers.
• Voice scores were measured again on the released sound, after the final start-of-sound step, with the gate's own model and code; re-scoring the sound the gate checked reproduced its numbers within 0.0003. The 15 pairs that then fell just below the voice limit were removed, so every released pair passes every gate rule on the released files.
• Human checks: people did not watch every clip; visual checks were done on samples.

Pairs kept per generator:

Generator	Length	Pairs kept
facefusion	77f	7,020 of 8,000 (88%)
facefusion	153f	6,835 of 8,000 (85%)
s2v	77f	6,360 of 8,000 (80%)
s2v	153f	6,279 of 8,000 (78%)
animate	77f	4,407 of 8,000 (55%)

5. Shortcuts we checked

Measured on this release (October 2026) with simple linear classifiers that see only cheap clues: file statistics, picture statistics, frozen ImageNet image features, sound statistics, and the silence at the start of the sound. They were always tested on people they had not seen. AUC 0.50 means no better than guessing and 1.00 a perfect shortcut; 95% ranges in brackets; all clips together unless a generator is named.

Fixed:

Shortcut	Before	Now	Fix
How the audio is packed in the file	0.79 (pilot)	0.50	PCM audio with fixed framing
A header tag (aspect ratio) set on real clips only	1.00	gone: every clip has the same file and stream fields	the same 1:1 tag in every clip
Converted voices fade in over their first 50 ms	0.74 (0.73-0.75)	0.57 (0.56-0.57)	50 ms of silence, then a 10 ms fade-in, at the start of every clip
Brightness and colour	0.77 (pilot)	0.49-0.50 for each	per-clip brightness and colour targets
Vocoder sound	0.83 (pilot)	0.50 on the vocoded view, 0.61 as shipped	the vocoded real arm

Not fixable:

Clue	All clips	FaceFusion	Animate	S2V
How the video is packed in the file	0.86 (0.85-0.86)	0.73-0.75	0.86	1.00
Picture statistics (sharpness, noise, colour)	0.71 (0.71-0.72)	0.68	0.73	0.79-0.80
Frozen ImageNet features (ResNet-18, 8 frames)	0.87 (0.86-0.87)	0.82-0.83	0.93	0.94-0.96
Converted voice, sound statistics	0.90 (0.89-0.90)	same voices for every generator

Why they stay: generated video compresses differently from camera video (no encoder setting removed this; only uncompressed video would, about 1.2 TB); the generators draw faces with less fine detail than a camera records; and the voice converter leaves its own trace in the sound.

Advice: train on decoded frames, not on file bytes or sizes. To test audio without the vocoder trace, use the vocoded view (RVRA vocoded against RVFA, FVRA vocoded and FVFA). Report results per generator and per length.

6. Known limitations

1. The way the video is packed in the file still differs between real and generated clips (0.86 AUC from file statistics alone; S2V 1.00). Train on decoded frames.
2. Animate in replace mode redraws the whole frame on VoxCeleb2's tight face crops, so its fakes are closer to a full re-animation (background and clothes can change) than a local face swap.
3. FaceFusion keeps the source's hair and head shape.
4. The voice model used for scoring was trained on VoxCeleb, so it has probably heard these people; its scores may read high.
5. Lip sync was checked on about 10% of pairs only.
6. Three gate limits are judgement calls, and the first batch both set the limits and is released.
7. Each generator kept different pairs, so per-generator sets differ; use pairs.jsonl.
8. Same-gender pairs only; gender as recorded by VoxCeleb2 (two values); no age or ethnicity labels.
9. YouTube interview footage of mixed quality; some clips have on-screen text, black borders, or are greyscale. A few clips with very dark, greyscale or peak-limited sources could not reach the brightness, colour or loudness target exactly.
10. Short (3 and 6 seconds), low-resolution (224x224) clips.
11. Generated faces have less fine detail than camera video; generic image features find them (0.87 AUC; FaceFusion 0.82-0.83, Animate 0.93, S2V 0.94-0.96).
12. Converted voices can be told from real ones by simple sound statistics (0.90 AUC).
13. The first 60 ms of every clip's sound was changed (50 ms of silence, then a 10 ms fade-in) to remove a start-of-sound shortcut; a weak trace remains (0.57 AUC).
14. VoxCeleb2 lists one person under two ids (id04048 and id05211, both in training), so there are 999 different people. Two YouTube videos appear under people in different splits (id04177 and id04887; id01036 and id00577); a face check and the gender labels show these are different people.

7. Uses

• For: training and testing deepfake detectors (audio, video or both); studying shortcuts; telling which generator made a fake; studying the effect of clip length.
• Not for: recognising or identifying people; making deepfakes of the people shown; commercial use; anything that could harm the people in it.

8. Ethics and consent

• The people in VoxCeleb2 are public figures in public videos; they did not consent to this use. Like other VoxCeleb2-based deepfake datasets (FakeAVCeleb, LAV-DF, AV-Deepfake1M), MOZAIC relies on research-only use and gated access; it adds a removal process.
• Every fake shows a real person, so misuse is possible; this is limited by the non-commercial licence, access only after accepting the terms, and short low-resolution clips.
• Removal: contact TO FILL. We remove that person's clips, and every fake made with them, in the next version.

9. Distribution

Hugging Face. Private now; gated with manual approval when public. Licence: TO FILL, plus terms of use; credit VoxCeleb2. Version 1.0, released TO FILL.

10. Maintenance

Maintained by Vector Institute with AIXPERT. Errors are fixed in new versions (1.1 and later) with a change log; removed clips are listed by id.