MiniMax H3 (also called Hailuo 3) follows plain, specific sentences better than lists of keywords. It also generates the sound together with the picture, so a good prompt describes what you should hear, not only what you should see.
The five parts of a good prompt
Write one or two sentences that cover, in this order:
- Subject — who or what is on screen: "a golden retriever in a tiny yellow raincoat".
- Action — what happens, with a verb: "splashes through puddles on a city street".
- Camera — how it's filmed: "slow motion", "the camera circles her", "drone shot gliding over".
- Light and mood — "at dusk", "warm morning light", "neon reflections".
- Sound — "cheerful ukulele music", "birdsong and a gentle breeze", "sizzling kitchen sounds".
Example: A street dancer spins on a rooftop at dusk, the camera circles her, city lights flicker on below, a punchy hip-hop beat.
Describe one moment, or a short sequence
H3 can render several shots in one clip (up to 15 seconds), but each extra shot needs room. For a 5-second video, describe one moment. For 10–15 seconds, you can add a second beat with "then": …mist rolling between the hills, then a slow tilt up to the glowing sky.
Starting from your own image
In Image mode, your picture becomes the first frame, so don't describe what it already shows. Describe what changes: The camera slowly pushes in while she turns and smiles; wind in her hair. Add a last frame if you want the video to end on a specific shot. The video takes the shape of your first image.
Using references
In Reference mode, images, clips and audio guide the character, style, motion or voice without becoming frames. Say which file does what: Use the character from image 1 in the style of the reference video. Audio needs at least one image or clip with it.
Common mistakes
- Keyword soup. "4k, cinematic, masterpiece, epic" adds little; one concrete detail adds more.
- Too many subjects. Two or three people or objects is the comfortable maximum for a short clip.
- Forgetting sound. Without a sound cue you get a generic soundtrack; name the music or ambience you want.
- Asking for real people or brands. Prompts and images of public figures or copyrighted characters are rejected by the content check, and those attempts cost nothing.
Try it
Start with H3 Max at 480P to test ideas cheaply, then render the one you like on H3 at 2K. Open the generator.