Deep-Fake/mozaic-preview

    

MOZAIC

MOZAIC is an audio-visual deepfake dataset built from VoxCeleb2, with question files for testing models. Every clip is a talking face with sound, in one of four classes: real or fake face, crossed with real or fake voice. The fakes come from three video generators and one voice converter.

• Pairs at two lengths: 7,590 at 77 frames (3.08 s) and 7,424 at 153 frames (6.12 s)
• Split by person: nobody in the test split appears in training
• Four tasks with question files: T1 to T3 on the full set and on MOZAIC-Small, T4 on MOZAIC-Small
• MOZAIC-Small: a smaller set of whole pairs that mirrors the full one

DATASHEET.md says how it was made, what was checked and what its limits are. summary.json has every count.

Made by Vector Institute with AIXPERT.

Statistics

	Clips	Hours	People	Pairs	Questions	Tasks
MOZAIC	137,745	169.9	1,000 ids	15,014	289,587	T1 to T3
MOZAIC-Small	16,995	21.0	the same 1,000 ids	1,858	47,087	T1 to T4

Split	Men	Women	Pairs	Clips
train	350	350	10,541	96,957
val	75	75	2,232	20,553
test	75	75	2,241	20,235

Gender is as recorded by VoxCeleb2. Across all clips, 51.4% show men and 48.6% show women. MOZAIC has no age or race labels. The 1,000 ids are 999 people, because VoxCeleb2 lists one person under two ids.

The four classes

Label	Face	Voice	Clips
RVRA	real	real	30,028
RVFA	real	fake	15,014
FVRA	fake	real	61,802
FVFA	fake	fake	30,901

Generators

Name	Makes	What it does	Clips
facefusion	fake face	Swaps the inner face; hair and head stay the original person's	41,565
s2v	fake face	Wan2.2-S2V: makes the whole video from one picture of the target plus the audio	37,917
animate	fake face	Wan2.2-Animate (replace mode); 77 frames only	13,221
seed_vc	fake voice	Seed-VC voice conversion	45,915

Source and target: the source is the person in the original clip; the target is the person the fake turns them into. (FaceFusion's own documentation uses these two words the other way round.)

Arms

• main: the clip as made.
• vocoded: the same clip with its real voice passed through BigVGAN, the vocoder Seed-VC uses. The label does not change: a vocoded RVRA clip is still real. This stops "vocoder sound = fake" from being a shortcut.

Tasks

Task	Question	Answers	Input	Full: train / val / test	Small: train / val / test
T1 spot the fake	Which part of this audio-visual clip, if any, is AI-generated or manipulated?	the video, the audio, both, neither	frames and sound	64,638 / 13,702 / 13,490	6,664 / 1,332 / 3,334
T2 manipulation type	Which part of this video, if any, was generated or altered by AI?	real, face swap, full-body swap, fully generated	frames	32,319 / 6,851 / 6,745	3,332 / 666 / 1,667
T3 identity check	Is the face (or voice) in clip 1 real or fake, and is it the same person as in clip 2?	real, same person; real, different person; fake of this person; fake of a different person	frames of two clips, or sound of two clips	106,802 / 22,586 / 22,454	11,012 / 2,200 / 5,550
T4 explanation	Which parts of this audio-visual clip, if any, were generated or altered by AI, where, and what in the clip shows it?	a written answer, with its facts in gt_target	frames and sound	none	6,664 / 1,332 / 3,334

Which clips each task asks about:

• T1 and T4: the vocoded view: RVRA vocoded, RVFA main, FVRA vocoded and FVFA main.
• T2: RVRA main and FVRA main. It is only about the video.
• T3: clip 1 is RVRA main or FVRA main for a face question, RVRA vocoded or RVFA main for a voice question. Clip 2 is a real recording (RVRA main) of the pair's source or target, of the same length, from a YouTube video the pair never used. Each clip 1 is asked twice, once for each person. Four val pairs (source id09056, target id05938) are left out of T3 because the target has no such recording; T1 and T2 still use them.

T1 to T3 answers come from the clip's labels. T4 answers were written by language models from the clip's facts, then checked by code and by a second model; each row's meta names the models, and those models should not be tested on T4. meta.human_checked is true on 400 test rows, the set three people reviewed.

Do not show a model the file names or question ids: they contain the label.

MOZAIC-Small

Whole pairs, each Small split drawn from the same full split, in the same mix as the full set, with every person the source of at least one pair. Its files sit next to the full ones with _small added, and every row is the full set's row, unchanged.

Small has no clips of its own: it uses the same videos/. T3 on Small also uses 2,913 real clips outside metadata_small.jsonl as clip 2. They are listed in metadata.jsonl.

Folders

mozaic/
├── README.md, DATASHEET.md, requirements.txt, assets/
├── metadata.jsonl           one row per clip with everything about it
├── metadata_table.parquet   the same rows, typed, for fast loading
├── pairs.jsonl              one row per pair: its people, clips and the generators that kept it
├── splits.json              person -> train, val or test
├── summary.json             version, settings, generators, all counts
├── the same four files for Small, with _small in their names (no splits file: Small uses the same one)
├── SHA256SUMS               the sha256 of every clip, manifest and question file
├── qas/                     T1, T2 and T3: one folder per task, with train, val and test files
├── qas_small/               the same for Small, and T4
└── videos/
    ├── train.tar            26.2 GB
    ├── val.tar              5.6 GB
    └── test.tar             5.5 GB

Download and unpack

The whole dataset is 38 GB. Unpacking needs the same space again, until you delete the tars.

pip install huggingface_hub
hf download vector-institute/Mozaic --repo-type dataset --local-dir mozaic
cd mozaic
for t in videos/*.tar; do tar -xf "$t"; done
sha256sum -c SHA256SUMS

Every clip then sits at the path in file_name: videos/<label>/<generator>/<length>/<arm>/*.mov, where <generator> is real for RVRA and seed_vc for RVFA. A tar holds each clip unchanged, so its bytes are the ones SHA256SUMS lists.

To get only the test split's clips (5.5 GB):

hf download vector-institute/Mozaic --repo-type dataset --local-dir mozaic --exclude "videos/train.tar" --exclude "videos/val.tar"
cd mozaic
tar -xf videos/test.tar
sha256sum -c --ignore-missing SHA256SUMS

To get one clip, take its split and file_name from metadata.jsonl. It is in videos/<split>.tar:

tar -xf videos/test.tar videos/FVFA/animate/77f/main/FVFA__animate__77f__main__id00024_3ysHAigPe68_00005__id00330.mov

File names

<label>__<generator>__<length>__<arm>__<source_clip>__<target_id>.mov

Example: FVFA__s2v__153f__main__id00058_6PiUqI-q-1w_00025__id03180.mov

A clip id such as id00058_6PiUqI-q-1w_00025 is the VoxCeleb2 clip id00058/6PiUqI-q-1w/00025: the person is the first 7 characters, the segment the last 5, and the YouTube video id the part in between. Video ids can contain _ and -, so do not split names on underscores; read the fields from metadata.jsonl.

Video and audio

.mov, H.264, 224x224, 25 fps; audio is uncompressed PCM, 16 kHz, mono. Every clip was brought to the same brightness, colour and loudness targets (see summary.json), and encoder version text was removed from every file. Every clip carries the same 1:1 aspect-ratio tag, and every clip's sound starts with 50 ms of silence and a 10 ms fade-in (converted voices used to fade in on their own). This largely removes these differences between real and fake; DATASHEET.md (section 5) lists what was measured and what remains.

Question files

One row per question, one file per task and split. Clip paths are relative to the dataset folder.

Field	Meaning
id	the question's id
task	the task's name
split	train, val or test
clips	the clip paths: one, or two for T3
question	the question text
options	the answers to choose from, shuffled per question; for T4 the regions
part	T3 only: face or voice
gt_answer	the right answer; for T4 the written answer
gt_label	the clip's label: RVRA, RVFA, FVRA or FVFA
gt_target	the answer as data; for T4 the fake parts, their regions and each part's facts
gt_target_name	T2 and T3 only: the answer as an id
meta	the clip's generators, frames, arm, people and pair; T3 also says whose recording clip 2 is

Manifests

• metadata.jsonl, one row per video (137,745 rows), everything about it in one place.

Field	Type	Meaning
file_name	string	path of the video inside the dataset folder
label	string	RVRA, RVFA, FVRA or FVFA
video_generator	string	real, facefusion, s2v or animate
audio_generator	string	none, seed_vc or bigvgan
length	string	77f (3.08 s) or 153f (6.12 s)
arm	string	main or vocoded
split	string	train, val or test
face_id	string	whose face you see (VoxCeleb2 person id)
voice_id	string	whose voice you hear (VoxCeleb2 person id)
gender	string	m or f
face_real	bool	the face is real
voice_real	bool	the voice is real
pair_id	string	the pair the video belongs to
source_id	string	the source person: who is in the original clip
target_id	string	the target person: who the fake turns them into
source_clip	string	the VoxCeleb2 clip the video was made from
target_face_clip	string	the target's clip the face picture came from
target_voice_clip	string	the target's recording Seed-VC used as its voice reference
kept_by	list of strings	the generators that kept the pair
sha256	string	checksum of the video file

• metadata_table.parquet: the same rows and fields as metadata.jsonl, with the types above stored in the file, so it loads much faster.
• pairs.jsonl, one row per pair: pair_id, length, split, gender, source_id, target_id, source_clip, target_clips (face: the target clip the face picture came from; voice: the target recording Seed-VC used as its voice reference) and kept_by (the generators that kept the pair).
• splits.json: person -> split.
• summary.json: version, settings, generators and every count.
• SHA256SUMS: the sha256 of every clip, manifest and question file. It leaves out README.md, DATASHEET.md, summary.json and summary_small.json, which are text and may still be edited.

Loading

requirements.txt lists the packages these examples use. The fastest way to read the clip table is metadata_table.parquet:

import pandas as pd

clips = pd.read_parquet("mozaic/metadata_table.parquet")
test_fakes = clips[(clips["split"] == "test") & ~clips["face_real"]]

A task's questions:

from datasets import load_dataset

t1_test = load_dataset("vector-institute/Mozaic", "t1_spot_the_fake", split="test")
row = t1_test[0]
print(row["question"], row["options"], row["gt_answer"], row["clips"])

One generator only. Real clips are stored once and shared by all generators, but each generator kept different pairs. To build a clean set for one generator, keep the pairs it kept (kept_by):

facefusion_77f = clips[(clips["length"] == "77f")
                       & clips["kept_by"].apply(lambda kept: "facefusion" in kept)
                       & clips["video_generator"].isin(["real", "facefusion"])]

Good to know

• Splits are by person: no face or voice in the test split appears in training.
• Train on decoded frames, not on file bytes or sizes: how the video is packed still differs between real and generated clips (DATASHEET.md, section 5).
• To test audio without the vocoder trace, use the vocoded view: RVRA vocoded against RVFA, FVRA vocoded and FVFA.
• Known limitations are listed in DATASHEET.md.

Access and licence

This repository is private for now. When it is public, access will be gated with manual approval.

Licence: TO FILL. MOZAIC is built on VoxCeleb2, so please also cite VoxCeleb2.

If you appear in a clip and want it removed, contact: TO FILL. We remove that person's clips, and every fake made with them, in the next version.

Citation

MOZAIC: TO FILL

VoxCeleb2: Chung, Nagrani, Zisserman. VoxCeleb2: Deep Speaker Recognition. Interspeech 2018.

Changes

• 1.0: first version.