File size: 2,003 Bytes
c4721c6
 
 
 
 
 
 
b108d2b
 
 
 
 
 
a85d8c1
 
 
 
c4721c6
 
 
 
 
 
 
 
b108d2b
c4721c6
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
# Paired synthetic clips and clean original windows

`metadata.jsonl` contains one row per synthetic clip. `video` is the synthetic
MP4; `original.video` is its exact clean driving window, not the full raw source
video and not the donor identity portrait. Both paths are repository-relative.
For A's face on B's video, the original is the selected window of B with B's audio.

Schema 4 uses the existing `clips/<pair>/` directory. For each sample,
`<sample>-<version>.mp4` is the synthetic clip,
`<sample>-<version>.original.mp4` is its clean original window, and
`<sample>-<version>.mp4.json` is the metadata sidecar, all in the same directory.
There are no new per-sample folders. Existing synthetic paths are reused when
their checksums match. The global index points to the current version. Older
schema 1/2/3 rows may remain during migration. With replacement enabled, obsolete
versions of each validated sample (including its previous `samples/` layout) are
removed in the same commit as its replacement. Samples awaiting validation are
left untouched.

Originals are uploaded byte-for-byte from the frozen semantic clips, preserving
their existing resolution, frames and audio. Source timestamps describe the raw
video interval; playback of the uploaded window starts at zero. These real clips
are marked `ai_generated: false`; synthetic clips remain `ai_generated: true`.

`originals.jsonl` groups original content by SHA-256, listing all per-sample
`videos`, associated `sample_keys`, and source-window references. Each sample
has its own `.original.mp4` even when multiple swaps reuse identical content.
Unique-original counts therefore need not match the number of original files.

Legacy runs, full raw videos, reference portraits and model files are excluded.
Source-media rights remain the uploader's responsibility. Automated clean-window
screening is not a guarantee that every frame is free of overlays. Account for
shared identities and source windows when creating train/test splits.