MiniMax H3 Video to Video AI

MiniMax H3 video to video AI reads your clip as a condition and generates a fresh one that obeys it. The action, the framing and the timing carry over because you asked for them, not because the pixels were preserved. Give it up to three sources, 2 to 15 seconds each and 15 seconds combined, and you get back 4 to 15 seconds at 768P or 2K with sound in the same MP4.

Source clips

3

Source clips

Across all sources

15 s

Across all sources

Out, at 24 fps

4–15 s

Out, at 24 fps

Resolution

768P / 2K

Resolution

• No edit endpoint
• No timeline
• No file of yours is changed

• Restyle — the look changes, the action stays

  [video editing]

  Describe only the new look and leave the rest alone; anything you mention is fair game to change. Motion and timing are the parts that survive best, surfaces and lighting the parts that move.

  Survives

  Action, timing, composition

  Moves

  Surfaces, lighting, grade

• Re-voice — new words, same face

  [video editing + audio reuse]

  Write the line short enough to fit the seconds you asked for, or the lips will not land.

  Survives

  Identity, scene

  Moves

  Lip sync, the line

• Continue — pick up where the clip stopped

  [video continuation]

  Your clip is the run-up; describe what happens after the last frame.

  Survives

  Subject, scene, style

  Moves

  What happens next

• Move the camera — borrow the move, drop the subject

  [reference generation]

  Your clip supplies only the movement. Describe the new subject and say nothing about the move itself — name it twice and you can no longer tell which instruction the camera followed.

  Survives

  Camera path and rhythm

  Moves

  Subject and scene, entirely

  Officially this one is reference generation. It is taught on reference to video.

The bracketed names are the official `summary` prefixes, not our shorthand. Restyle, re-voice and continue also open with `The target video is an edited version of <Video 1>.`

Your source is billed by its own length, on top of the output. Ten seconds in and ten seconds out at 2K is charged twice over. Trim the source to the seconds that matter and the same run costs a third less.

Frame size runs 256 to 5,760 pixels on each side and the aspect has to sit between 0.4 and 2.5, which almost anything shot on a phone already satisfies. Two more: reference video replaces first and last frame control, so you cannot use both, and output is always 24 fps. Full rates are on pricing.

This is why a voice you never uploaded turns up in the result, and it carries gender and accent with it. A laugh recorded from a woman still reads as a woman's laugh on a man's face.

• Keep it whole

  The original track becomes the finished track.

  fully_copy

• Keep it under new dialogue

  The music survives, mixed beneath lines you write.

  partially_copy

• Motion only

  The audio is dropped and the clip conditions on movement alone.

  weak_reference

These are three of the eight retention markers H3 reads; the other five matter more when you are building a scene from scratch.

• Video editing (with audio)

  First place(rank range 1–2)

  #1

  9001500

  Elo

  1,127

  95% CI

  ±6

  Votes

  11,474

  Rank range

  1–2