How to Generate Video with MiniMax Hailuo 02 from Text
MiniMax Hailuo 02 is a text-to-video and image-to-video model that outputs 6 or 10 second clips at 768p or 1080p. It is best known for handling physical motion — cloth, splashes, falling objects — better than earlier models. You need an account on the MiniMax platform and credits; nothing runs locally.
This guide covers the prompt structure that works, the settings that actually change output, and the credit maths before you burn a month of budget on test renders.
What can Hailuo 02 actually generate?
Hailuo 02 accepts a text prompt, or a still image plus a prompt, and returns a silent clip. Supported output: 512p, 768p and 1080p, at 6 or 10 seconds. Frame rate is 24 fps. Sound is not generated — you add audio in an editor.
Text rendering inside the frame works for short Latin strings at 1080p. Chinese characters render more reliably than English, which is a real limitation if your ad copy is in English.
On the MiniMax web console the 768p 6-second render costs about 1 unit of credit and 1080p 6-second roughly 2 units; 10-second 1080p is the heaviest option. Credit packs are the usual purchase path, so check the current rate card before you promise a client a re-render budget.
How do you write a Hailuo 02 prompt that holds together?
The reliable formula is subject → action → environment → camera → light → style, in that order, one clause each. Long poetic prompts do worse than five blunt clauses.
Example that rendered cleanly on the first attempt after two revisions:
A street vendor in his sixties flips a steel wok over a gas flame; market passage in Hanoi at dusk; slow push-in from waist height; warm tungsten light with blue ambient fill; documentary film look, 35mm.
Why it works: one subject, one action, a defined camera move, and a light source. Adding a second character or a second action is where limbs start melting.
Prompts that fail usually contain three verbs in one sentence — “runs, jumps and turns to look back” — or request a camera move that changes mid-clip.
Which settings matter, and which are noise?
Aspect ratio matters most. Vertical 9:16 for social, 16:9 for anything shown on a screen. Decide before you write the prompt; the crop is not a free change afterwards.
Image-to-video is the most reliable upgrade path. Generate one strong still in an image model, then animate. Motion is smaller but far less likely to corrupt faces.
Duration is a budget lever, not a creative one. Ten seconds costs roughly 1.7× a six-second clip for 66% more runtime. If your edit needs eight seconds, render ten and trim.
Hop on MiniMax Hailuo 02 and run the same prompt twice before you commit — variance between two renders of an identical prompt is noticeable, especially with hands and water.
What does a realistic production loop look like?
- Write three prompt variants, each with a different camera move.
- Render at 768p, 6 seconds, to check composition. Cheap and fast.
- Keep the best composition; re-render it at 1080p with the same seed if the platform exposes one.
- For anything human, switch to image-to-video with a clean still.
- Trim and grade in your editor; Hailuo clips grade well because the output is fairly flat.
Budget rule of thumb from my own tests: about one usable clip per three renders. Plan credits for 3× what your timeline shows.
Where Hailuo 02 struggles
Hands are still the weak point — not every time, but often enough that a shot where a character holds an object needs three tries. Fast lateral camera moves smear background detail. Complex text overlays should be done in the editor, not prompted.
It also will not generate clips longer than 10 seconds, and it has no audio. If you need dialogue, sync a separate voice track — the timing will not match by itself.
Who should use it, and who should not
Use it for mood shots, ads, product b-roll and social content where motion quality sells the idea. Skip it for anything needing precise character consistency across shots, since identity drift between renders is still real. If your workflow needs one character through eight shots, a model with reference-image locking is a better fit.
Quick Summary
- Hailuo 02 outputs 6 or 10 second clips at up to 1080p, 24 fps, without audio.
- Prompt order: subject, action, environment, camera, light, style — one clause each.
- Image-to-video is more stable than pure text for anything involving faces.
- Budget roughly three renders per usable clip; 1080p 10-second renders cost about twice a 768p 6-second one.
- Hands, fast lateral moves and embedded English text are the known failure modes.
