Wan Text To Video
Wan2.2 S2V 14B turns one image and an audio track into a video of that character speaking or singing in sync, with camera moves and body motion aimed at film-style shots.
Overview
Wan2.2 S2V 14B is the speech-to-video model in Wan-AI's Wan2.2 release. It takes a still image of a character and an audio track and returns a video of that character speaking or singing to the sound. The Wan-S2V paper behind it is aimed at film and TV work, where a shot needs body movement and camera work on top of lip sync.
The prompt and the audio do different jobs. The authors use text for the broad dynamics of a shot, such as camera movement and where the character goes, and leave finer detail like expressions and hand gestures to the audio. They report that it outperforms Hunyuan-Avatar and OmniHuman in their comparisons.
What you can build
- Presenter or explainer clips from one portrait photo and a recorded voice-over
- Music video shots where a performer sings to a supplied track
- Single-character dialogue shots for short films and animatics, with the camera move written in the prompt
- Talking animals, like the default sample of a cat in sunglasses on a surfboard
Inputs
- Image: the character to animate, one file in JPG, JPEG, PNG, GIF, WebP or HEIC.
- Audio (required): one speech or song file in MP3, WAV, M4A, WEBM or OPUS. It drives the lip sync, and the video runs as long as the audio does.
- Prompt (optional): the scene, the action or the camera move.
- Resolution: 480P, 580P or 720P, plus Default, which is 480P. The value only caps the pixel area at about 832x480, 1024x704 or 1280x720. The video takes its shape from the image, so the 16:9 and 9:16 options give the same result.
- Seed: a number from 0 to 9999999, 42 by default.
Outputs
An MP4 video with the input audio as its soundtrack. In the default sample the cat holds its pose and moves its mouth along a five-second clip. The second animates a photo of a man walking beside railway tracks with a red and yellow locomotive already behind him. The camera and background stay put while he performs, hands at his chest at first and then arms opened wide.
Recommended settings
- Write the camera move and the body action into the prompt. The audio already covers the mouth and face.
- Crop the image to the frame shape you want before uploading, since the video keeps the image's aspect ratio.
- 580P matches the 1024x704 size used in the official example commands.
Limitations
- The authors call nuanced interaction between several people an open problem, along with precise camera control from audio alone. Dance generation is listed as future work.
- There is no duration setting. Trim the audio to the part you need before uploading.
- The open-source release can also follow a pose video and generate the speech itself from text with CosyVoice. Neither option is available on this page.
Safety & compliance
The weights are released under the Apache 2.0 license. The model card holds users accountable for their use and rules out content that breaks the law, harms people or groups, spreads misinformation, shares personal information to cause harm or targets vulnerable groups. If the image shows a real person or the audio is a real person's voice, get that person's consent first.
Example prompts
Great starting points for Wan2.2-A14B-Speech-to-Video.
API quick start
Run Wan2.2-A14B-Speech-to-Video with a single API call.
{
"prompt": "Summer beach vacation style, a white cat …",
"inputAudio": "https://your-cdn.com/input.mp3",
"inputImage": "https://your-cdn.com/input.png",
"resolution": "832*480"
}curl -X POST "https://api.wiro.ai/v1/Run/wan-ai/wan2-2-a14b-speech-to-video" \
-H "Content-Type: application/json" \
-H "x-api-key: YOUR_WIRO_API_KEY" \
--data-binary @- <<'JSON'
{
"prompt": "Summer beach vacation style, a white cat …",
"inputAudio": "https://your-cdn.com/input.mp3",
"inputImage": "https://your-cdn.com/input.png",
"resolution": "832*480"
}
JSONOther models you might like
Explore similar video generation models.