# Wan 3.0

> Wan 3.0 by Alibaba on FlyMy.AI: video with sound from a prompt, from frames or from reference images, clips and audio; 2-30 seconds, 480P to 1080P, priced per generated second.

Wan 3.0 is Alibaba's video model. It writes a video with sound from a text
prompt, animates a first frame (and can end on a last frame), or keeps the
characters, places, style and voice of reference images, clips and audio.
Clips run 2 to 30 seconds, or the model picks the length.

| Mode | Endpoint id | Model page |
| --- | --- | --- |
| Text to video | `flymyai/wan-3_0-t2v` | https://app.flymy.ai/models/flymyai/wan-3_0-t2v |
| Image to video | `flymyai/wan-3_0-i2v` | https://app.flymy.ai/models/flymyai/wan-3_0-i2v |
| Reference to video | `flymyai/wan-3_0-ref2v` | https://app.flymy.ai/models/flymyai/wan-3_0-ref2v |
| Edit a video | `flymyai/wan-3_0-video-edit` | https://app.flymy.ai/models/flymyai/wan-3_0-video-edit |
| Extend a video | `flymyai/wan-3_0-video-extend` | https://app.flymy.ai/models/flymyai/wan-3_0-video-extend |

Edit changes a clip you give it from a plain instruction (objects, style,
lighting or dialogue), and extend continues it forward or backward, up to 30
seconds in all. Both take a `video` (MP4 or MOV, 1-15 s, up to 100 MB) and a
`prompt`, and bill the seconds of the clip you give plus the clip you get, at
the rates in [Pricing](#pricing).

## Generate a video

Making a video takes a minute or more, so use the async route: submit, then
poll the result. Inputs are form fields.

```bash
curl -X POST https://api.flymy.ai/api/v1/flymyai/wan-3_0-t2v/predict/async/ \
  -H "X-API-KEY: fly-***" \
  -F 'prompt=A red fox runs through fresh snow at dawn, low tracking shot' \
  -F 'resolution=720P' -F 'duration=10' -F 'ratio=16:9'
# {"prediction_id": "02ecd473..."}

curl "https://api.flymy.ai/api/v1/flymyai/wan-3_0-t2v/predict/async/result/?request_id=02ecd473..." \
  -H "X-API-KEY: fly-***"
```

The result route answers `425` while the video is being made. When it is
ready, it returns:

```json
{
  "inference_responses": [
    {
      "response": {
        "video": [{"url": "https://storage.googleapis.com/...", "content_type": "video/mp4"}],
        "calculated_price": "1.000000"
      },
      "infer_details": {"status": 200, "inference_time": 103.1, "request_id": "02ecd473-..."}
    }
  ]
}
```

`url` is a signed link valid for 24 hours; download the file to keep it.
`calculated_price` is what the run cost. The synchronous `/predict` route works
too, but it returns the video inline as base64, so prefer async for anything
but short clips.

## Image to video

Give the first frame, and optionally the last one, as a file upload or a public
image URL (JPEG, PNG, BMP or WEBP, up to 20 MB):

```bash
curl -X POST https://api.flymy.ai/api/v1/flymyai/wan-3_0-i2v/predict/async/ \
  -H "X-API-KEY: fly-***" \
  -F 'prompt=The cat slowly turns its head toward the camera and blinks' \
  -F 'first_frame_image=@cat.jpg' -F 'resolution=720P' -F 'duration=5'
```

`ratio` defaults to `adaptive`, which keeps the shape of your image.

## Reference to video

Pass references as public URLs, one URL or a JSON array per field, and say in
the prompt who and what they are and how to use them:

| Field | Limits |
| --- | --- |
| `reference_images` | Up to 10 images (JPEG, PNG, BMP, WEBP) |
| `reference_videos` | Up to 5 clips (mp4, mov), 1-15 s each, 15 s in total |
| `reference_audios` | Up to 5 clips (wav, mp3), 1-15 s each, 15 s in total |

```bash
curl -X POST https://api.flymy.ai/api/v1/flymyai/wan-3_0-ref2v/predict/async/ \
  -H "X-API-KEY: fly-***" \
  -F 'prompt=The woman from the photo walks through a night market and waves' \
  -F 'reference_images=["https://example.com/woman.jpg"]' \
  -F 'resolution=720P' -F 'duration=8'
```

At least one reference is required; reference to video takes no first or last
frame.

## Parameters

| Parameter | Values |
| --- | --- |
| `prompt` | What happens in the video, up to 20,000 characters |
| `resolution` | `480P`, `720P` (default) or `1080P` |
| `ratio` | `adaptive` (when omitted), `21:9`, `16:9`, `4:3`, `1:1`, `3:4`, `9:16` |
| `duration` | 2 to 30 seconds (5 when omitted); `-1` lets the model choose the length |
| `audio` | `true` (default) adds sound, `false` makes a silent video |
| `prompt_extend` | `true` (default) lets the model expand the prompt |
| `watermark` | `false` (default) or `true` |
| `seed` | 0 to 2147483647, for repeatable takes |

Invalid values are refused before the video is started, so they cost nothing.

## From an AI assistant

An assistant connected to FlyMy.AI finds these models with
`recommend_model`, quotes them with `get_pricing` and makes the video with
`run_model` (the same input as the REST fields). The result is a link to the
video file.

## Pricing

| Resolution | Price per generated second |
| --- | --- |
| 480P | $0.05 |
| 720P | $0.10 |
| 1080P | $0.20 |

You pay for the seconds Wan 3.0 generates; a run that fails costs nothing. A
run with the defaults (720P, 5 seconds) costs $0.50, a 10-second 720P clip
$1.00. Edit and extend also bill the seconds of the clip you give: a 5-second
clip edited at 720P is 10 billed seconds, about $1.00. Before a run starts, a request whose top price
is more than your balance is refused with `403 Insufficient funds`; with
`duration: -1` the top price counts 30 seconds. Every charge is in your
[usage](https://docs.flymy.ai/api/usage.md).
