WAN 3 vs WAN 3 Prime: Inputs, Audio and Duration

WAN 3 and WAN 3 Prime both offer routes for text, image and reference-guided video, but their field names and reference limits are not identical across hosted operations. For a production workflow, compare the actual input contract and quote rather than treating “Prime” as a drop-in switch.

This guide compares the current Raywake operations on 19 October 2026. It makes no measured claim that one variant preserves identity or motion better than the other.

Image-to-video: the important mapping

The reviewed WAN 3 image-to-video operation accepts start_image_url, while Raywake's WAN 3 Prime image-to-video operation uses image_url. Both expose an optional final image, resolution choices and an audio control, but the control names differ.

Sources: WAN 3 contract, Prime operation documentation. The public Raywake model pages are WAN 3 and WAN 3 Prime.

Reference-to-video is a separate operation

Choose reference-to-video when the brief needs separate images for identity, clips for motion or audio as context. An image-to-video endpoint may support a first and last frame without exposing those additional reference inputs.

The reviewed WAN 3 reference operation accepts up to ten images, five video references and five audio references. Its documentation limits combined reference-video duration and combined reference-audio duration to fifteen seconds each. It also states that input video duration is billed in addition to output duration.

The reviewed Prime reference operation exposes ten images and up to five video and audio files, with different documented duration limits. Its audio and video fields are named audio_urls and video_urls, rather than WAN 3's reference_audio_urls and reference_video_urls.

Sources: WAN 3 reference contract, Prime reference contract. Keep these limits attached to the named operation; do not combine the most generous values into an imaginary common contract.

A practical migration checklist

Begin with a copy of the brief, not a blind copy of its request body. Map every input to the new operation and identify any behavior that changes. In particular, check starting-image naming, audio control, duration type and reference-media limits.

Then quote the new request. A quote for one model is not approval to run another. Save the selected model and exact input with the result so a later production run can be traced to the same configuration.

For a small product animation, use a short fixed duration while evaluating. Smart duration is useful when the model can choose the timing, but it makes a controlled cost and output-length comparison less straightforward.

How to assign reference roles

Give each media item a purpose. An image may establish the product, a video may establish the camera move and audio may establish a rhythm or sound reference. State those roles explicitly and remove inputs that compete with them.

For example, ask for the bottle from the image to rotate with the camera movement in the video reference, while keeping its label unchanged. Do not assume that a motion reference also supplies the product identity or exact final composition.

Keep the reference clip as short as the useful motion segment permits. This reduces irrelevant visual information and, where reference duration is billed, avoids paying to process footage that contributes nothing to the brief.

Compare accepted clips, not names

Use the same subject, motion instruction, output duration and resolution where both operations allow them. Review the entire clip for label drift, added parts, unstable shadows and unexpected cuts. Evaluate audio separately if it matters.

Record the settled charge and whether the clip was accepted. A provider's public rate is not the Raywake customer price. Use the current Raywake quote workflow and the cost guide for the intended settings.

Common questions

Can Prime read a document or public webpage?

The reviewed Prime reference operation exposes those inputs and applies additional thinking behavior. That does not mean the image-to-video operation accepts the same fields.

Does enabling sound reproduce a reference recording exactly?

Do not assume so. Generated sound and exact soundtrack reuse are different production requirements.

Should I choose the longest supported duration first?

Use the shortest duration that can establish whether the shot works. Longer output magnifies the cost of a brief that still needs correction.

Reviewed on 19 October 2026. The cover is an editorial illustration generated with Nano Banana 2; it is not a test result from the models discussed.