shainaraza/faceswap-corpus-v5-review

Faceswap Corpus v5 — human review

Plays every pair of the shainaraza/faceswap-corpus-v5 build four classes at a time, off a single clock, so a reviewer can judge the swap and the voice conversion against the real clip they were made from.

• —Play all starts the four tiles together; the RVRA tile is the clock and the others are nudged back if they drift more than ~2 frames.
• —Listen moves the audio to one tile. Four tracks at once is unlistenable, so only one is unmuted.
• —Good / Flag records a judgement per pair, with an optional note. Keyboard: g good, f flag, ← → to move, space to play.
• —Export CSV writes pair_id, ring, driver, face, verdict, note.

Judgements live in the reviewer's own browser (localStorage) and are never sent anywhere. Export before clearing site data, and expect an empty sheet in a private window.

Layout

                audio: real     audio: converted
video: real        RVRA              RVFA
video: swapped     FVRA              FVFA

Rows differ only in the face; columns differ only in the voice. The picture is identical across a column, frame for frame.

Previews, not masters

The tiles stream AAC copies from preview/ in the dataset repo, because no browser decodes the corpus's uncompressed PCM-in-MOV audio. The video stream is copied without re-encoding, so the pixels are the master's. Judge content here; measure on clips/.