Hailuo AI: 9 Things to Know Before You Generate

Everyone has seen a Hailuo clip. Fewer people can tell you which Hailuo made it.

That's the quiet problem with the name. "Hailuo AI" is a consumer app, a family of video models, a run of version numbers (01, 02, 2.3), a Fast tier, a Live model for illustrations and, as of this summer, a successor called H3. Type the name into a search bar and you'll land on all of them at once. Then you generate, the clip looks fine but not like the one that made you search, and you assume the model is overrated. Usually it isn't. You just picked the wrong Hailuo, or asked it the wrong way.

So here's the short version of what we'd tell a colleague before their first run of Hailuo AI: nine things, in the order they'll matter to you.

1. Hailuo AI Is an App and a Model Family

Hailuo is MiniMax's video brand. On one side there's the Hailuo AI app and website, where you click Create Video and get a clip. On the other there's the model line underneath it, which MiniMax also ships through its own API platform. When people say "Hailuo" they might mean either, and it's worth separating the two, because the app wraps the models in templates and tools while the API hands you the models directly.

The lineage is short and fast. MiniMax released Hailuo 02 on June 18, 2025, built on an architecture it calls Noise-aware Compute Redistribution, with native 1080p output, three times the parameters of Hailuo 01 and four times the training data. Hailuo 2.3 followed on October 28, 2025, together with a Hailuo 2.3 Fast variant, rolling out across the web app, the mobile app and the API at once.

The app is a front door. The models are the engine, and results improve the moment you know which one you're driving.

2. Hailuo 2.3 Is the Workhorse, H3 Is the Next Generation

On July 31, 2026, MiniMax published MiniMax H3 and framed it explicitly as the next step in the same line: "From Hailuo 01 and Hailuo 02 to H3." If you open the Hailuo AI site today, H3 is the video model it leads with. So in practice, the newest Hailuo isn't called Hailuo at all.

That doesn't retire 2.3. The two models solve different problems. Hailuo 2.3 is a focused motion model: one prompt or one image in, 6 or 10 seconds of video out, tuned for bodies, faces and camera moves. H3 is a much bigger tool. It generates 4 to 15 seconds at 768P or 2K with native stereo sound in the same pass, accepts prompts up to 7,000 characters, and in its reference mode takes up to nine images, three video clips and three audio tracks to hold a character steady across shots. We covered all of that in our MiniMax H3 deep dive.

Our take: reach for H3 when the job is a sequence, a recurring character or a clip that needs sound. Reach for 2.3 when the job is one strong, physically convincing shot and you want to iterate on it quickly.

3. Pick the Right Hailuo Before You Write a Word

Most disappointing Hailuo runs are a routing problem, not a prompting one. Here's the lineup on each::labs, straight from the request schemas.

Hailuo 2.3 Standard comes as text-to-video and image-to-video, renders at 768p, and lets you choose 6 or 10 seconds. It's the right default for exploring an idea.

Hailuo 2.3 Pro is the 1080p tier, again in text-to-video and image-to-video flavors. The Pro endpoints don't expose a duration setting, and MiniMax's own docs list 1080p on Hailuo 2.3 at 6 seconds, so plan Pro shots as short, finished beats.

Hailuo 2.3 Fast is image-to-video only, in Standard (768p, 6 or 10 seconds) and Pro (1080p) versions. MiniMax built it to generate faster for batch work. If you're animating a whole product catalog or testing twenty openings, Hailuo 2.3 Fast is the one that keeps up.

Hailuo 02 is still available too, with 6 or 10 second clips and, in its image-to-video version, a choice of 768P or 512P. Use it to match footage you already made with it.

Hailuo Live is the odd one out, and we'll come back to it in item 8.

The pattern that works: explore on Standard, batch on Fast, finish on Pro, and step up to H3 when the shot needs sound or continuity.

4. Motion Is the Point, So Ask for Real Motion

Every video model claims realism. Hailuo's specific claim is about how things move. MiniMax pitched Hailuo 02 on "extreme physics mastery" and said that for intricate scenarios like gymnastics, artists considered it the only model able to deliver. With 2.3, the stated upgrades are more fluid and precise complex body movements, more natural facial performances and micro-expressions, better response to motion commands, and lighting, shadow and color that hold up even under dynamic camera movement.

Read that list as a prompting instruction. A model trained to be good at bodies, faces and physical cause and effect is wasted on a slow drift across a static landscape. Give it a backflip, a sprinter leaving the blocks, a dancer's spin with the fabric lagging behind, a face that goes from polite to furious in two seconds. That's where you'll see what it's for.

Style is the other half. MiniMax says 2.3 handles anime, illustration, ink wash painting and game CG better than before, and beta creators making e-commerce ads reported a significant jump in their success rate. Stylized motion isn't a side feature here.

5. Write the Prompt Like a Shot, Not a Wish List

Hailuo 2.3 accepts prompts up to 2,000 characters. That's plenty, and the temptation is to fill it with adjectives. Don't. A Hailuo prompt works best when it reads like a line from a shot list: who, doing what, where, how the camera behaves, what the light is doing. In that order.

A gymnast in a red leotard runs down the mat and launches into a round-off back handspring, landing with arms raised. Side-on medium shot, the camera tracks with her run. Bright arena lighting, crowd softly out of focus behind her. Realistic, high frame clarity.

Notice what's missing: "stunning", "cinematic masterpiece". Those words describe nothing renderable. The verbs carry the clip. "Runs, launches, lands" gives Hailuo three physical events to solve, and that's exactly what it's built to solve.

For performance, write the emotion as a change rather than a state:

Close-up of an elderly fisherman on a dock at dawn. He reads a letter, his jaw tightens, then his eyes fill and he looks out at the water. Soft golden light, gentle breeze moving his hair. Static camera.

A change gives the model a micro-expression arc to animate. A state gives it a frozen face. For more patterns that transfer across models, our collection of AI video prompt examples that actually move is a good companion.

6. Use Hailuo's Bracket Camera Commands

This is the Hailuo trick most people never hear about. MiniMax's API docs define fifteen camera commands for Hailuo 2.3 and Hailuo 02, written in square brackets right inside the prompt: [Truck left], [Truck right], [Pan left], [Pan right], [Push in], [Pull out], [Pedestal up], [Pedestal down], [Tilt up], [Tilt down], [Zoom in], [Zoom out], [Shake], [Tracking shot] and [Static shot].

You can combine moves inside one bracket (MiniMax recommends no more than three) or chain them in sequence. Natural language still works, but the docs are explicit that the bracket commands give more accurate results.

A vintage red convertible drives along a coastal cliff road at sunset [Tracking shot], then the car rounds a bend and the ocean opens up below [Pull out, Pedestal up]. Warm backlight, long shadows, sea spray catching the light.

Two rules of thumb. Give each camera move a reason in the scene (the bend in the road is why the camera pulls out). And don't stack three moves on a shot that's also full of subject motion: one moving thing at a time reads as intentional, two reads as chaos.

7. Know When to Turn the Prompt Optimizer Off

Every Hailuo endpoint on each::labs has a prompt_optimizer switch, and it's on by default. With it on, MiniMax rewrites your prompt before generating, which is great when you type five words and want a polished clip back.

It's less great when your wording is the spec. If you've written exact bracket commands, a precise sequence of actions or a specific framing a client approved, the optimizer can smooth it into something more generic. MiniMax's own guidance says it plainly: set it to false for more precise control and stricter adherence to your instructions.

Our rule: leave it on while you're exploring, switch it off the moment you start directing. If a carefully written prompt keeps coming back "close but not quite", this switch is the first thing to check.

8. The "Hailuo AI Photo" Trick: Animate the Still, Don't Describe It

A lot of people searching for "Hailuo AI photo" want one thing: to make a photo move. That's image-to-video, and it's arguably where Hailuo is strongest, because the image locks the look and the model can spend everything on motion.

The mistake is describing the photo back to the model. It can already see the photo. Your prompt should describe only what changes:

She turns her head toward the camera and smiles slowly, a strand of hair lifting in the wind. Background leaves sway. [Push in]

Mind the input, too. MiniMax's docs ask for JPG, PNG or WebP under 20MB, with a short edge above 300 pixels and an aspect ratio between 2:5 and 5:2. Crop to the frame you actually want before you upload, because the photo is the composition.

No photo yet? Make one. MiniMax Text to Image on each::labs returns up to four images at a time in seven aspect ratios from 1:1 to 9:16, so you can design the first frame and hand it straight to Hailuo 2.3.

And if your still is a drawing rather than a photo, try Hailuo Live. MiniMax designed it to bring 2D illustrations to life, optimized for stability and subtle expression, which is exactly what a character portrait or a mascot needs: a blink, a breath, a head turn, not a stunt.

9. Run Hailuo AI by API and Give It a Voice

Every model above runs through the same each::labs API. Switching from Hailuo 2.3 Standard to Fast to H3 is a change of model name and a few input fields (image_url for Hailuo 2.3 image-to-video, first_frame_url and an optional last_frame_url for H3), not a new integration. The call returns a prediction ID that you poll or receive by webhook, so build for asynchronous results from day one. When you're ready to chain steps, each::labs flows let you wire image, video and audio models into one pipeline.

That pipeline is where sound comes in. The Hailuo 2.3 endpoints take no audio settings, so when a shot needs a voice or a score, you add it. Stay in the MiniMax family and the pieces fit: MiniMax Speech 2.8 HD turns a script into a voiceover with an emotion setting (happy, sad, angry, surprised and more) and pronunciation boosts for dozens of languages, and MiniMax Music 3.0 writes a full song from a style prompt and lyrics with tags like [Verse] and [Chorus], or an instrumental if you'd rather. Generate the picture, the line and the bed separately, then cut them together.

Or skip the assembly and use H3, which generates stereo audio with the video. Again, it's routing. Knowing that is most of what there is to know about Hailuo AI.

Frequently Asked Questions

Is Hailuo AI the same thing as MiniMax?

Think of MiniMax as the company and Hailuo as its video brand. MiniMax builds the models (Hailuo 02, Hailuo 2.3, Hailuo 2.3 Fast, and now H3) and also makes speech, music and language models. The Hailuo AI app is MiniMax's consumer product for video, and the same models are available by API on each::labs.

Which Hailuo AI model should I use in 2026?

Start from the job. One strong shot of a body or face in motion: Hailuo 2.3 Standard, then Pro for the 1080p final. Lots of image-to-video runs: Hailuo 2.3 Fast. A multi-shot sequence, a recurring character or a clip that needs sound: MiniMax H3. Animating a drawing: Hailuo Live.

Can Hailuo AI turn a photo into a video?

Yes, and it's one of its best uses. Pass the photo to a Hailuo 2.3 image-to-video endpoint, describe only the motion you want, and add a bracket camera command if you want a specific move. The photo is your composition, so crop it to the frame you need before you upload.

How long are Hailuo AI videos?

It depends on the model. Hailuo 2.3 Standard and Hailuo 2.3 Fast Standard give you 6 or 10 seconds at 768p, and MiniMax lists 1080p at 6 seconds. MiniMax H3 runs from 4 to 15 seconds at 768P or 2K. For anything longer, generate shots and edit them together.