Wan 3.0 reference videos: 5 clips, 15 s total, 16 fps floor
Wan 3.0 on Sume takes up to 5 reference videos, 15 seconds combined, each at 16 fps or higher. Check clip count, length and frame rate before you submit.

Wan 3.0 (wan-3.0) on Sume accepts up to 5 reference videos, with no more than 15 seconds of footage across all of them, and each clip must run at 16 frames per second or higher. Images are capped at 10 and audio files at 5 (15 seconds combined). These limits come from the Video Router catalog, so a request that goes over them fails before it reaches the provider.
The part that surprises people is the frame-rate floor. A screen recording or a time-lapse exported at a low rate can pass every other check and still be a poor reference. The 15 seconds is a sum, not a per-clip cap: three 6-second clips are over the limit even though each one looks small.
The limits in one table
Sume lists the caps as constraints on the wan-3.0 catalog row. Read the live row with GET /v1/video-router/models/wan-3.0 before you hard-code anything, because the limits can change.
| Reference type | Cap on wan-3.0 | Extra rule |
|---|---|---|
Images (reference_image_urls) | 10 | none listed |
Videos (reference_video_urls) | 5 | 15 s total, 16 fps or higher |
Audio (reference_audio_urls) | 5 | 15 s total |
| Output length | 2 to 30 s | 480p, 720p or 1080p |
A pre-flight check
Measure the clips you plan to send. video_inspect returns a probe for a clip hosted on Sume, so you can read duration and frame rate first, then add up the seconds yourself. If the total is over 15, trim the longest clip with video_trim rather than dropping a whole reference.
Two habits save failed jobs. First, count seconds across all video references, not per clip. Second, treat the 16 fps floor as a filter on your source library: phone slow-motion exports and some screen captures are the usual offenders.
Where Wan sits next to the other ids
Not every model takes the same references. In the catalog, kling-3 takes none, gemini-omni-flash-1.1 takes images and videos but no audio, and the Seedance 2.x ids and both MiniMax H3 ids take all three types. Each id has its own numeric caps, so a reference set that works on one id can be rejected on another.
Choose Wan when you need a long clip (up to 30 s) and a generous reference budget in the same request. Choose another id when your references are mostly short audio stems or when you need a different aspect ratio list. Wan also bills by resolution, so pinning 720p keeps a reference-heavy clip affordable.
Whatever you choose, test with the smallest reference set that carries the look. One clean 6-second clip usually tells the model more than five noisy ones, and it leaves room under the 15-second budget for audio.
Sources
Related posts
More in Models
- AI video models with sound: which toggle it, which are always on
Seedance 2.5, Wan 3.0, Kling 3 and MiniMax H3 all return audio on Sume. Kling prices audio separately; H3 and Omni cannot turn it off. What each does.
- Which new video model for which job: Seedance, Wan, Kling, H3, Omni
A decision table for the five new video models on Sume, by clip length, resolution, sound and price per second. Veo 3.1 and LTX-2.5 are not on Sume.
- Which Sume image models reject input_references: five text-only rows
Five of 19 Sume image models are text-to-image only and return 400 unsupported_parameter if you send input_references. The other 14 take 5, 10 or 16 references.
- Which AI video models take an end frame or reference clips on Sume
Sume's catalog flags show who accepts text-to-video, an end frame, reference images, clips and audio, and who edits video. Grok is image-to-video only.
Written by Sume