Hailuo 02 Image-to-Video
MiniMax's second-generation Hailuo model turns a first frame, and an optional last frame, into a clip of 6 seconds at up to 1080P or 10 seconds at 768P.
Overview
MiniMax announced Hailuo 02 in June 2025 as the successor to Hailuo 01. The company says it tripled the parameter count and trained on four times as much data, using a new architecture it calls Noise-aware Compute Redistribution (NCR). The launch post stresses instruction following and physics, singling out gymnastics as a hard case the model can handle. MiniMax also reports that an early version placed second on the Artificial Analysis Video Arena.
On this page the first image opens the clip and an optional second image closes it. The prompt covers what happens in between, including camera moves.
Choosing a variant: Hailuo 02 is the only one of the four that can end on a frame you supply, and the only one with resolution and duration settings. The two I2V-01 pages stay at 6 seconds and 720P: Hailuo I2V-01-Live is tuned for illustrated characters with small, steady motion, and Hailuo I2V-01-Director is the earlier generation's dedicated camera-command model. Hailuo S2V-01 starts from a face photo and generates the whole scene around that person.
What you can build
- Transitions between two known states, such as a product box closed and then open, or a street by day and by night, by setting both frames.
- Sports and action clips where bodies have to move in a physically believable way.
- 10-second social clips and ad cutdowns made from a single key visual.
- 1080P hero videos for landing pages, made from product or campaign photos.
Inputs
- Image First (required). The opening frame; the video takes its aspect ratio. Upload a JPG, JPEG, PNG or WebP smaller than 20 MB, with more than 300 px on the short side and a shape from 2:5 (tall) to 5:2 (wide).
- Image Last (optional). The closing frame, with the same file limits. When its dimensions differ from the first image, MiniMax crops it to match.
- Prompt (optional, 2,000-character limit). Action, mood and camera. The default describes a low tracking shot of a runner's legs crossing a finish line.
- Prompt Optimizer (advanced, on by default). Lets MiniMax rework the prompt to improve the result. Turned off, the model follows your wording more strictly.
- Resolution (advanced): 768P (default) or 1080P.
- Duration (advanced): 6 seconds (default) or 10 seconds.
Outputs
One video at the chosen resolution and length, framed like the first image. The four samples on this page are MP4 clips just under 6 seconds, three of them square and one wide.
Recommended settings
- When exact wording matters, switch the optimizer off and write a more detailed prompt for the model to follow.
- Give the last frame the same dimensions as the first, so nothing important is lost to the crop.
- Camera moves use the same 15 bracket commands as Director. Up to 3 inside one bracket run together, and separate brackets play in order.
Limitations
- 1080P is available only at 6 seconds, and 10-second clips only at 768P.
- Each slot takes one image. There are no separate character or style reference inputs.
Safety & compliance
Prompts caught by MiniMax's sensitive-content check are refused. If either frame shows a real person, have that person's consent first.
Example prompts
Great starting points for image-to-video-hailuo-02.
API quick start
Run image-to-video-hailuo-02 with a single API call.
{
"prompt": "The camera tracks low across the finish l…",
"inputImage": "https://your-cdn.com/input.png",
"duration": 6,
"inputImage2": "https://your-cdn.com/input.png"
}curl -X POST "https://api.wiro.ai/v1/Run/minimax/image-to-video-hailuo-02" \
-H "Content-Type: application/json" \
-H "x-api-key: YOUR_WIRO_API_KEY" \
--data-binary @- <<'JSON'
{
"prompt": "The camera tracks low across the finish l…",
"inputImage": "https://your-cdn.com/input.png",
"duration": 6,
"inputImage2": "https://your-cdn.com/input.png"
}
JSONOther models you might like
Explore similar video generation models.