Six seconds in, five seconds out, eleven seconds billed.
That was the first surprise. I had assumed I was paying for the shot I received. Video-to-video charges the footage I sent as well, which makes the source clip a budget line rather than a free input.
The second surprise was better: the motion survived. My handheld walk through a car park came back as the same walk, in a different decade.
The short version. H3 video-to-video runs at 768P and 2K, takes a 2–15 s mp4 or mov clip, and bills input plus output seconds — a 6-second source into a 5-second output is 11 seconds.
What footage does video-to-video accept?
768P or 2K only. An mp4 or mov clip, two to fifteen seconds, fifty megabytes at most, with an aspect ratio between 0.4 and 2.5. Output is four to fifteen seconds. Anything outside those bounds is rejected before the job exists, so a bad file costs a click rather than credits.
The arithmetic, at two cents a second for 768P and ten for 2K:
- 4 s in, 4 s out — 8 seconds billed, 16 cents at 768P.
- 6 s in, 5 s out — 11 seconds, 22 cents.
- 10 s in, 10 s out — 20 seconds, 40 cents, or two dollars at 2K.
Ten seconds of restyle at 2K costs the same as a minute of fresh 768P footage. I restyle cheap, approve the look, then regenerate at the higher tier if the shot needs the resolution.
What survived, what did not
Survived: a slow walk, a head turn, a hand opening a door, the camera drifting through a corridor. Any continuous move held its timing, which is what makes the restyle useful — you keep the performance and change the world.
Turned to soup: a whip pan, and a hard cut inside the source clip. The reconstruction needs one continuous move; give it two and it merges them into something with no geography. My fix is to cut the source to a single move first, even if that means a four-second clip.
Also worth knowing: the output carries its own audio. The source clip's sound does not come through, so plan the sound for the output, not the input.
Can you chain one shot into the next?
Chaining. Take the last frame of a finished shot as the first frame of the next request and the composition carries over. At 2K a single first frame is accepted; at 480P and 768P the first-frame work uses the two-frame form.
I chained ten of these into a sequence that stays in one room, and it worked better than any single long prompt. It also drifted: by the eighth link the face had softened into a cousin. Keeping a reference sheet alongside the chain fixed that, at the cost of one more thing to manage.
When not to bother
If you have no footage, everything here is more expensive than text-to-video at the cheap tier. If what you actually want is a different camera move, generated footage beats reconstructed footage — the model invents a better move than it can recover from yours.
And if the clip is under two seconds, it is rejected outright, which rules out using a single flash as a source.
Video-to-video tiers, clip rules and the per-second quote are on the H3 page.
I work on Callirra, so the clip rules come from the validator that rejects them. If your restyle kept the source audio, that contradicts what I see, and it is worth posting.