Wan 3.0 reference limits on Sume: 10 images, 5 videos, 5 audios

Wan 3.0 on Sume accepts up to 10 reference images, 5 reference videos totalling at most 15 seconds at 16 fps or more, and 5 reference audios totalling at most 15 seconds. If you are porting a Sora call that sent a single input image, you are far inside the limits; the question is only what to trim when you send a library of references.

The limits

Seconds, not just counts

Totals matter more than counts. Five reference videos of 3 seconds fit exactly; five videos of 4 seconds do not. Trim the clips before upload instead of waiting for a 400.

A simple pre-flight check in your own code saves failed submissions: sum the durations of the reference videos, sum the audio durations, and compare against 15 seconds before calling the API.

Mapping a Sora input

OpenAI's guide, read 2026-10-07, says the Sora 2 models and Videos API were shut down on September 24, 2026. An input image maps to a first frame on Sume. The field names are frame_images for first and last frames and input_references for the rest.

What to send

Each reference is a URL the service can fetch, or an asset you have uploaded; the media inputs docs describe both. Keep references to what the clip needs. More images do not make Wan hold a character better; consistent, clearly lit images do.

• Send one reference per character plus one for the setting.
• Use short video references for motion, under 15 seconds in total.
• Use an audio reference only when voice or music timing matters.

What a clip costs

Wan bills by the second at the resolution, $0.0625 at 480p, $0.125 at 720p and $0.25 at 1080p. The extra-image fee is a minimax-h3 rule, not a Wan one; Wan's listed price is per second by resolution. For a 10 second 720p clip that is $1.25.

Reference videos need at least 16 frames per second, which excludes some time-lapse exports. Re-encode those at 24 fps before upload.

Next step

The Sora versus Wan post compares the two on length and price, and the video docs list every field.