How fast are Kaption’s subtitles?

Genshin types dialogue out letter by letter. A line takes about one to two seconds to finish, and for most of that time what is on screen is half a sentence. You cannot translate half a sentence reliably. "Yes, y" is the start of a hundred different lines, and picking one of them is just guessing.

So Kaption waits for the text to stop changing before it commits. That wait is most of the delay you notice, and it comes from the game, not from us. Star Rail and Zenless put the whole line up in one frame, which is why subtitles feel instant there and a bit more patient in Genshin.

We record real gameplay and replay it through the same code that ships. Eight clips across the three games, about 4,700 frames. Replaying a recording instead of playing live means we can measure the exact same footage again after every change, so the numbers actually compare.

The number we care about is how long after a line settles before its translation is on screen. Across six Genshin clips the typical line landed at about 170 ms, and 9 out of 10 landed inside 374 ms. Star Rail came out at effectively zero: the subtitle was there on the first frame the line existed.

These come from one computer, so treat them as a reference point rather than a promise about yours. A slower GPU reads each frame slower and everything stretches. We simulated machines up to five times slower to check, and it gets worse gradually instead of falling off a cliff.

Reading the screen sixty times a second would be a waste and would not help anyway. Dialogue does not change sixty times a second, and every read costs real GPU work.

Kaption watches the dialogue area and reads only when something has actually changed, then waits a moment for it to settle before trusting what it sees. Between lines, when nothing is moving, it does close to nothing. That is what buys it the headroom to be quick when it counts.

Two rounds of optimisation in August 2026 moved these numbers a lot. The first made reading itself about 3.5x faster by fixing how the text-recognition model gets fed on the GPU. Same text out, fraction of the time.

The second used the harness above to retune how long it waits before reading. On Genshin that worked out as roughly 3x better worst case and about a quarter more lines translated, because lines that used to get missed entirely now get read in time.

Roughly in order of how much they matter. Exclusive fullscreen stops the overlay drawing properly, so run the game borderless or windowed. A capture region drawn much bigger than the dialogue box means more pixels to chew through on every read. And an older GPU makes each read take longer, though Kaption handles that by reading less often rather than falling behind.