shainaraza/faceswap-corpus-v5

Faceswap Corpus v5

A four-class audio-visual deepfake corpus derived from VoxCeleb2, built so that neither stream alone carries the label, across three forgery families. Every driver clip appears in all four combinations of real/swapped video and real/converted audio, and all four go through one identical terminal chain — same crop, same loudness solve, same gamma and chroma correction, same encoder settings.

from datasets import load_dataset

ff = load_dataset("shainaraza/faceswap-corpus-v5", "ff", split="train")
ff[0]["video"]          # PCM master, 77 frames, 3.08 s
ff[0]["video_preview"]  # same pixels, AAC audio — plays in a browser

Configs

config	generator	clips	pairs
ff	FaceFusion (inswapper128 fp16)	459	74
s2v	Wan-S2V 14B	496	80
animate	Wan-Animate 14B, replace mode	496	80

1,451 clips total. Audio conversion is Seed-VC in all three.

Classes

RVRA  real video    + real audio        label: real
RVFA  real video    + converted audio   label: fake
FVRA  swapped video + real audio        label: fake
FVFA  swapped video + converted audio   label: fake

Every clip is exactly 77 frames at 25 fps = 3.08 s, verified on all 1,451 files. Uniform length is enforced, not incidental: if duration varied by class, duration alone would leak the label.

Pairing: two same-gender rings of five

Ten identities in two rings — one male, one female. Each borrows a face from the four others in its own ring. Rings are same-gender by construction: a cross-gender swap would make the classes separable by pitch alone.

Read this before you train on it

The s2v config is perfectly separable from container statistics. A linear probe on packet sizes alone distinguishes real from generated at 1.000 separability — no pixels decoded:

config	container: real vs generated	container: label real vs fake
ff	0.724	0.657
animate	0.743	—
s2v	1.000	0.861

Real camera footage carries sensor noise that x264 spends bits on; a diffusion render does not. FaceFusion keeps the real background, so its leak is partial. Wan-S2V synthesises every pixel, so the gap is total. A detector trained on s2v as shipped can reach perfect accuracy by counting bits.

Re-encoding at a constant bitrate does NOT fix it — we measured. Forcing every clip to the same byte budget brought the mean packet size to within 0.8% (2950 B real against 2927 B generated) and left the separation intact (AUC 0.693 on 20 v 20 clips, against 0.593 before). The signal is not total size but the shape of the allocation: even on a fixed budget the encoder spends bits differently on a smooth synthetic frame than on a grainy real one. Anyone planning to normalise this away should know that the obvious fix has been tried.

Until something is shown to work, treat s2v as test-only, or train on it with the container features from the forensics config regressed out. A uniform synthetic noise floor — raising every clip's bit cost equally — is the untested candidate.

Real clips are shared across configs. The RVRA arm comes from the same crops through the same chain in all three builds — the generator never touches it. Pool the configs naively and the real class is triple-counted while the fake classes genuinely triple.

Ten identities, one split. Every probe here is linear and small-sample. A low number bounds nothing about what a CNN could find.

Refusals fell entirely on the female ring. FaceFusion's content analyser refused 7 of 248 render attempts — 0 of 124 male, 7 of 124 female. Six pairs lost their swapped-video arms and were dropped whole, which is why ff ships 459 and 74 pairs rather than 496 and 80. All attrition is on one ring; three identities fall from 16 pairs to 11, 12 and 13. Treat the rings as unequally sampled, and do not read the gap as a property of the faces. s2v and animate had no refusals.

Measured quality

config	self-drive PSNR / SSIM	FID	identity swapped	target median
ff	32.39 dB / 0.947	17.4	74/74	0.855
s2v	16.49 dB / 0.540	42.1	80/80	0.718
animate	18.16 dB / 0.604	61.8	79/80	0.538

Self-drive PSNR is not comparable across families: FaceFusion swaps a face and keeps the real background, while S2V and Animate regenerate every pixel. FID and identity strength are the fair comparisons. One animate pair failed to reach the identity floor.

All three builds pass the full battery: structure, shortcut probes, difficulty, visual quality, identity, speaker similarity and lip-sync — zero structural violations, nothing partial.

The forensics config

Per-clip measurements for all 1,451 clips, so the shortcuts are measurable rather than merely warned about. Join to any config on sample_id.

f = load_dataset("shainaraza/faceswap-corpus-v5", "forensics", split="train")

columns	what they are
n_packets, pkt_mean, pkt_std, pkt_p90, pkt_max, bytes_total, keyframe_ratio, file_bytes	video-stream packet statistics — the features the container leak lives in
identity_target, identity_leak, identity_swapped	per-pair face similarity to donor and residual to driver
sync_offset_frames, sync_confidence	SyncNet lip-sync, on the gate's sampled clips
sv_target_cosine, sv_source_cosine, sv_timbre_moved	speaker verification, keyed on the conversion

Mean packet size alone separates real from generated video by:

config	AUC on pkt_mean	real mean	generated mean
ff	0.578	1183 B	1111 B
animate	0.540	1189 B	1161 B
s2v	0.713	1189 B	982 B

Use it to stratify, to regress the container signal out, or to check a detector is not simply reading it.

Coverage is uneven and that is not a defect. identity_* is per pair and so covers every clip; sync_* covers 366 of 1,451 and sv_* covers 492, because the lip-sync and speaker-verification gates score a sampled subset rather than the whole corpus. Treat missing values as "not scored", never as zero.

Files

data/<config>/train-*.parquet   video (PCM master) + video_preview (AAC) + 16 metadata columns
data/forensics/train-*.parquet  1,451 rows of per-clip measurements, no media
preview/<config>/<CLASS>/*.mp4  browser-playable copies, one per clip
preview/pairs_index.json        234 pairs, for the review Space

Masters are MOV, H.264 video with PCM 16 kHz mono audio — uncompressed on purpose, so audio artefacts are the generator's and not a codec's. No browser decodes PCM in MOV, which is why video_preview exists; its video stream is copied without re-encoding, so its pixels are the master's.

Terms

Source video is VoxCeleb2, which carries its own licence and terms of use. This repository redistributes derived clips: cropped, re-encoded, and in three of the four classes altered by face swap and/or voice conversion. Downstream users are responsible for complying with the original VoxCeleb2 terms in addition to anything stated here. If you are the rights holder of a depicted identity and want a clip removed, open a discussion on this repository.

Provenance

Builds shaina_v5_ff, shaina_v5_s2v, shaina_v5_animate, seed 20260908, rendered 8–9 September 2026 on the Vector Institute Bon Echo cluster.