Taking a finished video apart into layers
I take a vertical video with the banner and word-by-word captions already baked into the pixels and pull it back apart into layers. Once it is apart, the video can be rebuilt: another language, another voice, and your logo where someone else's used to be.
Clean footage
The banner and captions are painted out frame by frame, leaving the original picture without them. Everything else is then laid back on top of this.
The banner as its own layer
With transparency, with a separate image of every state it had, and a schedule of the moments each state changed.
Captions as text, not pixels
Recognised word by word: the text, the timing of each word, size, colour, outline thickness, position and the curve it pops in on.
Translated as one piece
Words are gathered into phrases by pauses and cuts, and the monologue is translated whole rather than line by line, which is what keeps the sense and the tone connected.
Dubbed in a chosen voice
Speech is synthesised per phrase and fitted to the original timing. The captions are then rebuilt against the new voice rather than the old one.
Your brand where theirs was
The logo takes the same zone of the frame and switches at the same moments the original banner flipped.
If you want it in one line: this is not translating a video, it is dismantling one. Ordinary localisation works on top of a finished file — it lays a track over the audio and nails new captions over the ones already burned in. Here the video is first taken back apart, and after that you can change any piece of it, because the pieces are editable objects again rather than pixels.
Why this is not “translating a video”
The captions show the difference. In an ordinary file they are drawn into the frame: they are pixels, and no word in them can be changed. All you can do is cover them with a new block — a patch over a patch.
After the teardown the captions are text again. Each word sits on the timeline with its own entry time, its own size, colour and outline thickness, all taken from the original. That is why the new language looks the way the original looked, rather than like a sticker over it.
The banner is the same story. It comes out with transparency, together with every state it had and the moments those states changed. Your logo takes its place and switches on the same seconds, so the rhythm of the video survives.
What happens to the sound
The speaker’s voice is not muted wholesale; it is muted against the caption timing. Under the speech the track drops away, and in the pauses it stays. That matters more than it sounds: the pauses hold laughter, the room’s reaction and the background, and a video without them goes sterile. If the original track is not wanted at all it can be removed entirely — but that is a decision, not the default.
The new speech is synthesised per phrase and fitted to the original timing. The captions are then rebuilt against that new dub: the words land where they are actually spoken rather than where they used to be.
Where it breaks
Honestly, because a demo never shows this part.
Text recognition depends on the font and the outline. White words with a dark stroke all the way round read reliably; a thin unoutlined face over a busy background reads worse, and some words will need fixing by hand.
Painting out leaves traces on a complicated background. If there was a small, detailed scene behind the banner, the restored footage will be softer there than the original.
A translation is almost always longer than its source. When a phrase does not fit its slot it runs into the pause before the next one. On a tight edit that shows, and those places get checked by eye.
The teardown is not free in time: twelve seconds took about a minute of machine time in my own run, and it grows with the length of the video.
What I need from you
The source file at decent quality, your logo as a file, and a decision about the voice. If a translation already exists it goes in ready-made, and no machine translation is used at all.
The rights to the source are a separate conversation. Taking someone else’s video apart is technically possible; publishing the result is only possible when you hold the right to do it.