Wan 3.0 Video Generation by Alibaba

Overview

Wan 3.0 is Alibaba’s Wan video generation model, hosted as an API product. It turns a written description into an MP4 video clip with coherent motion, camera movement, and optional audio. You can also anchor the shot with a first frame, or with both first and last frames, to keep composition stable and guide the motion between endpoints.

It’s useful when you need a full short sequence from one prompt, not a single still. It’s also useful for product and character shots where the first frame must match a reference.

What you can build

• 5 to 30 second ad-style product clips with consistent framing
• Short narrative shots with time-coded beats and sound cues described in text
• Image-anchored animations where the opening frame must match a supplied still
• First-and-last frame transitions for controlled motion between two keyframes
• Vertical social clips (9:16) and ultrawide compositions (21:9)

Inputs

• A required video description written in plain text. Keep it specific about subject, action, camera, and lighting. Very long prompts may be truncated.
• An optional first frame image to lock the opening composition. Upload a JPG, JPEG, PNG (no transparency), BMP, or WEBP image. Keep each side between 240 and 8000 pixels. Keep the aspect ratio within 8:1. Keep the file under 20 MB.
• An optional last frame image to lock the ending composition. Use the same file types and size limits as the first frame.
• A target duration for the generated clip, chosen in seconds.
• A quality tier for output resolution, chosen from 480p, 720p, or 1080p.
• An output aspect ratio setting. You can let the model choose, or force a ratio like 21:9, 16:9, 4:3, 1:1, 3:4, or 9:16.
• An optional audio toggle that controls whether the exported video includes an audio track.
• An optional prompt expansion toggle that rewrites short prompts to add detail.
• An optional numeric seed to repeat results more reliably. Use an integer between 0 and 2,147,483,647.

Outputs

• A downloadable MP4 video.
• The response includes a file URL plus video metadata such as width, height, frame rate, duration, file size, and frame count.
• The response also returns the seed used for generation.
• If prompt expansion runs, the response may include the rewritten prompt that the model actually used.

Limitations

• Maximum output length is 30 seconds per generation.
• Output is produced at a fixed 30 fps.
• Output resolution is limited to 480p, 720p, or 1080p.
• First and last frame guidance depends on the quality of the input stills. Low-resolution images, heavy compression, or extreme aspect ratios can reduce stability.
• Transparent PNGs are not supported for frame inputs.
• On-screen text and small UI details can be inconsistent across frames. Review any clip that contains numbers or fine typography.
• Audio quality can vary by prompt. Treat generated speech, music, and effects as draft audio unless you review it.
• Messy or low-quality inputs raise failure risk. Scanned PDFs, photo screenshots of slides, or heavily stylized text layouts can reduce document understanding.

Safety & compliance

• This model can run with content moderation enabled for both inputs and outputs.
• Don’t upload images you don’t have rights to use. Don’t request copyrighted characters or logos for commercial work without permission.
• Don’t include private personal data in prompts or frame images.
• Keep a human review step for sensitive topics, realistic people, and any clip intended for public release.