Wan 3.0 reference videos on Sume: 5 clips, 15 s total, 16 fps floor

On Sume, wan-3.0 accepts up to 5 reference videos, 10 reference images and 5 reference audios. The catalog adds that the reference videos total at most 15 seconds and run at 16 fps or faster. Every other model that takes reference videos stops at 3, and a sixth Wan clip fails with "model wan-3.0 accepts at most 5 reference_video_urls."

What Wan 3.0 accepts on the Video Router

Wan 3.0 is a catalog model with text-to-video, image-to-video, an end frame and all three reference types. Its duration range is 2 to 30 seconds, its resolutions are 480p, 720p and 1080p, and its aspect ratios include 16:9, 4:3, 1:1, 3:4 and 9:16 but not 21:9. The reference lists are flat fields on POST /v1/video-router/generate; on the OpenRouter-shaped route the same items are input_references entries with video_url types.

Count is enforced, seconds are on you

The schema counts files, but it cannot see inside them at validation time. The 15-second total and the 16 fps floor are recorded as catalog constraints for the model, so inspect each clip before you upload it. video_inspect on a stored clip reports its probe data, and it is the natural place to read duration and frame rate when the clips already live on media.sume.com.

Pre-flight check for a reference set

The snippet validates a local description of your clips against the Wan numbers, so you find a bad set before paying for a job. It uses plain dictionaries and sends nothing.

def check_wan_refs(videos: list[dict]) -> list[str]:
    problems = []
    if len(videos) > 5:
        problems.append("more than 5 reference videos")
    if sum(v["seconds"] for v in videos) > 15:
        problems.append("reference videos exceed 15 s combined")
    for v in videos:
        if v["fps"] < 16:
            problems.append(f"{v['name']} is below 16 fps")
    return problems

print(check_wan_refs([
    {"name": "walk.mp4", "seconds": 6, "fps": 24},
    {"name": "turn.mp4", "seconds": 8, "fps": 12},
    {"name": "wave.mp4", "seconds": 4, "fps": 30},
]))

Limits and alternatives

If you need a motion reference that preserves the source clip's length and framing, the Genjutsu motion-transfer model is the specialist: it takes one source video plus 1 to 8 reference images. If you need to edit an existing video with a prompt, that is gemini-omni-flash-1.1 in edit mode. Wan's reference videos are guidance for a new clip, not a source to be edited.

For a 20 second result that carries references, Wan 3.0 and Seedance 2.5 are the two general text-to-video catalog models with a 30 second ceiling that also take video references. Seedance 2.5 caps references at 3 videos.