MiniMax H3 Prompt Guide: Structure, Camera Moves and Common Mistakes

Why one-line prompts that work elsewhere fail here

This guide is based on MiniMax's official prompt-writing guides for the open-weight H3 model, the ComfyUI docs, community reports and our own tests. Rules we checked ourselves say so.

MiniMax's own web app and hosted API run a prompt rewriter in front of the model: your one sentence is expanded into a long, structured English prompt before H3 sees it. The open-weight model has no such rewriter; MiniMax says it matters a lot for quality, and it isn't open source. So a viral "one sentence, amazing clip" example often comes out vague on open weights. The fix is to write what the rewriter would have written: concrete, ordered and complete.

The structure: four beats in time order

1. Anchor the opening frame: who, where, wearing what, in what light. A medium shot frames the young woman standing by the river at dusk, in a long beige coat.
2. The action starts: She slowly turns her head toward the camera…
3. It develops: …as the wind lifts her hair and she pulls the coat tighter.
4. The result or reaction: She smiles softly and looks back at the water.

Then the camera: The camera holds a static shot. Every sentence should map to something visible or audible. Words like beautiful, cinematic feel or amazing give the model nothing to draw.

Length: 60–150 English words per shot. Too short and the model improvises; too long and instructions near the end get ignored, so repeat key requirements in each shot instead of adding global rules at the end.

Shot size: always state it (close-up, medium close-up, medium shot, medium-wide shot, wide shot) and match it to the action. Don't squeeze a face and feet into the same shot.

Camera moves: type + amplitude + speed

Add amplitude (with small amplitude / with large amplitude) and speed (at slow speed / at fast speed); medium and normal can be left out. Write it as a sentence in the flow of the action, The camera pushes in with small amplitude at slow speed toward the letter in her hands., not as tags piled at the end.

What about bracket commands like [Pan left]? They come from earlier Hailuo models, and MiniMax's hosted API docs still mention them. The official open-weight H3 guides write camera moves as full sentences instead, and that is the form we use.

Only write what you want

H3 has no negative prompt. A list of "no …" and "don't …" tends to pull the forbidden things into the picture. For a still camera, write The camera holds a static shot, not no camera movement. To avoid subtitles, describe the scene fully (see below); writing "no subtitles" doesn't work.

One shot or several

• H3 tends to cut on its own. For a single take, write in one continuous shot.
• Cut only when the subject, place, time or viewpoint changes; for small changes, move the camera instead.
• Write later shots as [Shot 2] At 00:04.000, the camera cuts to …, with strictly increasing single time points, not ranges. In our tests cuts land on the written frame; still put key actions in order.
• First + last frame clips can cut too, even at 4 s: write [Shot 2] with a time point, or the model bridges the two images with a dissolve or push-in. Start the shot after the cut in a pose different from the last frame and move into it, or it freezes on the last frame.
• Rough guide: 4–6 s: 1–2 shots; 7–10 s: 2–3; 11–15 s: 3–5.

People

• Small faces in wide shots blur: attach a sharp close-up reference image at the angle the person faces in that shot (front when facing the camera, side in profile), and distant faces stay clear. Without one, keep faces to close-ups and medium close-ups, and use backs or three-quarter backs in wide shots.
• One person? Say so: the same single man throughout, or turning and walking past the camera can produce a duplicate.
• For the same face across shots, use reference photos and don't describe looks. See character consistency.
• On-screen text goes in double quotes after a word saying it is written: a red neon sign reading "OPEN".

Dialogue and sound

• Lines stay word for word in their own language (Chinese lines stay Chinese) and should end with a full stop, or the speaker keeps mumbling.
• Keep lines short enough for the duration: for Chinese, at most about 7 characters per second.
• Outside the lines, avoid words like talk, chat or scold: they summon extra voices.
• Describe ambient sound concretely: Soft wind over the river, distant birds, fabric rustling. Concrete sounds give fewer stray noises than leaving it empty.
• In two-person dialogue, a scene of only one or two sentences often gets the lines burned in as subtitles. Describing the full scene (shot size, set dressing, light, positions, clothes) plus a silent closing action after the last line fixed it in our tests.

For singing and talking to a recorded voice track, see lip sync.

Styles

Start the description with a style word if you want something other than live action: Cinematic, Live-action, 2D-animated, 3D CG, claymation, watercolor, vintage film. With a first frame, the style must match the image.

Common problems and fixes

On XXT: you write the scene, the platform adds the format

H3's full prompt format has an alignment sentence for the first and last frame, [Shot 1] markers, speaker tags, reference-image notes and separate sound and music fields. On XXT you write only the scene description; the platform translates Chinese into English, wraps it in the official format and adds the rest. When you need exact control (several speakers, voiceover, music), add --raw and write the complete prompt yourself.

h3 submit --mode i2v --first-frame river.png --aspect 9:16 --seconds 6 \
  --prompt "A medium shot frames the young woman standing by the river at dusk in a long beige coat. She slowly turns her head toward the camera and smiles softly as the evening breeze lifts her hair. The camera holds a static shot." \
  --sound "Soft wind over the river, water lapping against the bank, distant birds." \
  --wait --out ./out

Videos are 4–60 seconds, in 9:16, 16:9 or 1:1, at 768P or 1080P (1080P costs double). At $0.009 per second at night and $0.0135 during the day, iterating is cheap: fix the seed with --seed and change one thing at a time.

Your agent doesn't need this page: the full rules it follows are in our prompting reference. See generating videos with Claude Code or Codex.

Videos, songs and uploaded files are kept for 8 hours after the job ends, then deleted automatically. Please download them in time.

FAQ