Craft
One motion per MiniMax H3 take
A spicy MiniMax H3 take is 5–15 seconds. That is not a limitation you work around with a denser prompt. It is the unit of the job. The takes that hold are the ones that ask for one motion: one camera move, or one body action — not both, and not a sequence pretending to be one verb.
This is the craft note people skip when they paste a paragraph that would take a minute of coverage. The model does not compress a scene into 8s. It samples 8s of whatever you named, and if you named three actions, none of them finish.
What “one motion” actually is
Write the take as a director of photography would write a slate:
- Camera only — “slow push-in, she does not move.”
- Body only — “locked camera, she turns her head to camera.”
- Not both — “orbit while she undresses and he walks in” is three takes pretending to be one.
Intimate work is where this fails most often. A sex scene is a chain of blocks, not a single verb. Arrive in the position, or move once inside it. Changing position is the next take. The dictionary for those blocks is spicy positions; the camera language is camera moves.
How long is a take, for real
Duration is a schema enum, not a wish. The H3 family on this site:
| H3 tier | per second | resolution | duration |
|---|---|---|---|
| MiniMax H3-Developer Text-to-Video | $0.030 | 480P, 768P, 1440p-sr, 4k-sr | 4–15s |
| MiniMax H3-Developer Image-to-Video | $0.030 | 480P, 768P, 1440p-sr, 4k-sr | 4–15s |
| MiniMax H3-Developer Reference-to-Video | $0.030 | 480P, 768P, 1440p-sr, 4k-sr | 4–15s |
| MiniMax H3 Max Turbo Text-to-Video | $0.036 | 480P, 768P | 5–15s |
| MiniMax H3 Max Turbo Image-to-Video | $0.036 | 480P, 768P | 5–15s |
| MiniMax H3 Text-to-Video | $0.057 | 480P, 768P, 2K | 4–15s |
| MiniMax H3 Image-to-Video | $0.057 | 480P, 768P, 2K | 4–15s |
| MiniMax H3 Reference-to-Video | $0.057 | 480P, 768P, 2K | 4–15s |
| MiniMax H3 Fast Text-to-Video | $0.066 | 480P | 5–15s |
| MiniMax H3 Fast Image-to-Video | $0.066 | 480P | 5–15s |
| MiniMax H3 Fast Reference-to-Video | $0.066 | 480P | 5–15s |
| MiniMax H3 Max Text-to-Video | $0.072 | 480P, 768P, 1440p-sr, 4k-sr | 5–15s |
| MiniMax H3 Max Image-to-Video | $0.072 | 480P, 768P, 1440p-sr, 4k-sr | 5–15s |
Pick a length that matches the motion. A 5s crash-zoom is a sentence. A 15s orbit is a paragraph. Asking for a 5s “they move from the door to the bed and undress” is how you get smear. Iterate on the cheap tier — five seconds is about 15 cents at $0.030 per second — until the one motion is right, then render.
A brief that fits 8 seconds
Use the seven locks from how to prompt adult MiniMax H3 video, then cut the action until only one remains:
Two adults, tungsten hotel, 35mm, 8s, 9:16. She is already sitting on the bed edge. He takes one step in and stops. Do not change her face, the mole on the left collarbone, the rain on the window. Native audio: rain and cloth, no score.
That is a take. The kiss, the garment, the next blocking each get their own request. Stack them on the eight-cell board.
When you are tempted to stack
If the sentence has “then”, it is two takes. If it has “while” and both sides are large motions, it is two takes. “While” only survives when one side is tiny: breathing, rain, a hand tightening on a sheet. Put the tiny thing in audio or in “must not change”, and keep the slate to one verb. Run it on spicy text-to-video or, if the still already locks the bodies, image-to-video.