Recast a video with a shot over 15 seconds: split it first on Sume

Sume's H3 Max Recast row documents a source of 5 to 30 seconds with no shot over 15 seconds, so a 40 second single take, or any clip with a 20 second unbroken shot, needs to be cut into pieces before you recast it. Trim the take into near-equal ranges of at most 15 seconds, recast each piece with the same photos in the same order, then join the results in a Timeline. Expect a visible seam where a cut lands inside a continuous shot, and test it before you run the whole file.

The limits are from Sume's OpenAPI reference and video trim docs, read on 2026-10-03.

What exactly is the limit on shots?

The OpenAPI description for h3-max-recast says it requires video_url together with reference_image_urls, one photo per new person (1 to 4), and a 5 to 30 second source with no shot over 15 seconds. The output keeps the source length. The docs state the limit but do not describe what the model does with a longer shot, so treat it as a hard edge and do not test it with an expensive clip.

A shot here means an uninterrupted stretch between edits. A 30 second clip made of four 7 second shots is inside the limit. A 30 second clip with one 20 second shot is not, even though the total length is fine.

Why do cuts matter to the swap?

Reported by The Neuron, fal describes the same MiniMax H3 Max Recast as preserving motion, gestures, camera movement, cuts, timing and soundtrack. That is a vendor-side summary, not a Sume spec. Whatever the model does inside a piece, a boundary you add yourself is a cut the original did not have, so choose boundaries where something already changes: a pause in action, a head turn, a camera move.

How do I split a long take?

Pick the number of pieces as the total divided by 15, rounded up, then make them equal so that none falls under 5 seconds. The small script prints the ranges for a 40 second take as three pieces of about 13.3 seconds. Feed each range to video trim with precision: "exact", which is the frame-accurate re-encode, so the second piece starts where the first ended.

import math

def split(total, cap=15, floor=5):
    """Cut a long take into near-equal ranges of floor..cap seconds."""
    n = math.ceil(total / cap)
    size = total / n
    if size < floor:
        raise ValueError("source too short to split")
    return [(round(i * size, 2), round(min((i + 1) * size, total), 2))
            for i in range(n)]

for start, end in split(40):
    print(f"video-trim start={start} end={end}")

How do I keep the same people across the pieces?

Send the same reference_image_urls in the same order on every piece. The docs tie one photo to one new person, so a changed order can change who is mapped to whom. Run the pieces as separate jobs with separate Idempotency-Keys, then check each result before joining.

Join the pieces with the Timeline 1.0 render: one video[] slot per piece, start values that increase from 0, and POST /v1/timeline-1.0/plan first, which is unbilled and creates no job. A short fade transition of a few hundred milliseconds is one way to soften a mid-shot seam. The render is $0.10 per output minute, so a 40 second result is one billed minute. For what the soundtrack does, see does H3 Max Recast keep the original audio.

• Each piece must be between 5 and 15 seconds, and a split should not leave a stub under 5.
• Reuse the same photos, in the same order, on every piece.
• Trim with exact precision so pieces meet without a gap.
• Plan the Timeline first, then render.
• Compare one seam on a single pair of pieces before running the full clip.

What does the extra work cost?

Splitting does not add billable seconds to the recast itself, since the pieces add up to the source length, and each trim is $0.02. Reported by The Neuron, fal prices Recast at $0.30 per second at 768p and $0.45 at 1080p; Sume's own rate is in GET /v1/catalog, and the stored post video-to-video AI cost per clip works through Sume's numbers. For sources over 30 seconds, see recast a video longer than 30 seconds, and to compare person-swap rows see Recast vs Genjutsu.