Reference videos in AI video on Sume: Wan, H3, H3 Max and Omni limits

Four Video Router models take reference media, and their limits are not interchangeable: wan-3.0 accepts up to 10 images, 5 videos and 5 audio files; minimax-h3 and minimax-h3-max accept 9 images, 3 videos and 3 audio files; and gemini-omni-flash-1.1 accepts 10 images and 3 videos but no audio. kling-3 accepts no references at all. The numbers come from the constraint lines in Sume's catalog, which the Video Router docs tell you to read per model.

If you build one request that you replay across models, these differences are where it breaks first.

What are the exact limits?

Video and audio references carry length rules as well as counts. The table copies them from the catalog.

Which limit do people trip over?

On the H3 family, audio cannot be your only reference, and the cap of 12 applies to images, videos and audios together, so 9 images plus 3 videos already fills it. On Omni each reference video can be at most 3 seconds, which is much shorter than the 15-second videos Wan takes in total. On Wan the 5-video limit is a count while the 15-second figure is a combined length, so five 4-second clips is already over it.

On H3, reference-to-video also has a price detail: the first 5 reference images are free and each additional image carries an $0.08 list add-on, while on H3 Max the catalog says reference-to-video is billed as output seconds at list times 1.25 and fal's reference-token overage is not reserved. Check the line for the specific id before you budget a 9-image job.

How do I check a request before I send it?

The numbers above fit in a small function that counts what you plan to send. It runs offline and prints the problems, so you find a 6th Wan video before the API does.

LIMITS = {
    "wan-3.0": (10, 5, 5),
    "minimax-h3": (9, 3, 3),
    "minimax-h3-max": (9, 3, 3),
    "gemini-omni-flash-1.1": (10, 3, 0),
    "kling-3": (0, 0, 0),
}

def problems(model, images, videos, audios):
    mi, mv, ma = LIMITS[model]
    out = []
    for n, cap, kind in ((images, mi, "image"), (videos, mv, "video"), (audios, ma, "audio")):
        if n > cap:
            out.append("%s: %d %s refs, limit %d" % (model, n, kind, cap))
    if model.startswith("minimax-h3") and images + videos + audios > 12:
        out.append("%s: more than 12 refs in total" % model)
    if model.startswith("minimax-h3") and audios and not (images or videos):
        out.append("%s: audio cannot be the only reference" % model)
    return out

print(problems("wan-3.0", 4, 6, 0))
print(problems("minimax-h3-max", 0, 0, 2))

What about the edit and motion-transfer rows?

Two catalog rows work from a source video rather than from references, and they are separate from the table above. gemini-omni-flash-1.1 supports video_to_video: send a prompt and one video_url, and the model edits the clip. The docs say video_url is the edit source, not a reference, and cannot be combined with image_url, end_image_url or reference_*_urls. In edit mode resolution is optional with a default of 720p, aspect_ratio is rejected, and a duration is only a hint for the reserve.

higgsfield-genjutsu (Motion Transfer) takes one source video and 1 to 8 reference images, at 480p or 720p, for 4 to 30 seconds, and the docs say it is listed only when its provider is configured. h3-max-recast swaps the people in a source video for 1 to 4 reference photos, one photo per person. Neither is a general reference-to-video model, so they sit outside the limits table.

How are the limits enforced?

Counts and lengths are limits Sume documents, and some are enforced by the provider. The catalog notes the Omni rule that each reference video must be at most three seconds as enforced by fal and documented by Sume, so a request that breaks it can fail at the provider rather than at admission. Treat the table as the contract and test your longest and heaviest combination once before a batch.

Where should I send what?

If you hold a library of up to 10 product shots and a few short clips, Wan 3.0 has the widest limits and the 30-second ceiling. If your references are a few photos and a voice track, H3 Max takes them natively with stereo sound in the output. If the clip you hold is a 3-second move you want continued, Omni is the model with the 3-second rule. Anything with no references in the plan can go to the cheaper Kling 3 row.