Why two lines are harder than one
A single-line caption occupies roughly eight percent of a 1080p frame. A bilingual pair commonly takes fifteen to eighteen percent, and that larger area has to be rebuilt from surroundings further away from the centre of the hole.
The lines also interact. If you mark only one of them, the model samples the other as if it were real background, and the repaired strip inherits ghosting from the line you left in place.
- Mark the full band, including the gap between the lines
- Do not run two passes — the second pass repairs against a frame the first pass already altered
- A bilingual band leaves less clean background to sample from, so textured footage degrades faster
What we have tested, and what we have not
Our published tests used two-line captions at 46 px with a 3 px stroke — the same geometry as a bilingual pair, so the two-line layout is covered directly by the numbers below.
What we have not tested is mixed scripts. Chinese glyphs carry more ink per character than Latin ones, and a CJK line with a heavy outline can be twice as dense as the Latin line above it. We have no measurement for that case, so treat it as unverified rather than supported.
When it goes wrong, and what it looks like
The failure is usually a horizontal seam where the two lines met, or a strip that is uniformly soft compared with the rest of the frame. Both mean the band was under-marked.
Re-run with a box that includes the outline and any drop shadow. A three-pixel stroke is part of the text — leaving it behind leaves a ghost of the letters in the shape of their edges.