Guides

    How to direct cuts inside one Seedance pass

    Seedance 2.5 will hold 30 seconds at 720p and 47 credits if the brief is one shot with tempo shifts, not a hidden movie. Timestamp the beats in the prompt; do not ask it to recast at 0:12.

    Versely Team6 min read

    Seedance 2.5 will hold 30 seconds at 720p and 47 credits if the brief is one shot with tempo shifts, not a hidden movie. Timestamp the beats in the prompt; do not ask it to recast at 0:12.

    Seedance 2.5 is ByteDance's long single-shot row: native audio, 4s through 30s, 720p, 47 credits. It reasons about the whole shot at once so motion, lighting, and identity stay coherent from first frame to last. A "cut" here is a beat inside that take — camera speed, a pause, a look — not a new person, room, or wardrobe at 0:12. Storyboarded multi-angle sequences are a different family. This page marks tempo on the long take you bought.

    One shot is the product

    Thirty seconds in one pass is expensive attention, not a free movie. The model holds one subject, one light, one space. A hidden movie — kitchen wide, smash to a different actor in a car, smash to a studio pack — is a recast. Recasting is how second 14 stops matching second 3.

    Durations run every second from 4s to 30s. Resolution tops out at 720p; this slug is not a 4K master. Native audio is on, so a line, room tone, and foley you name can land in the same file. Forty-seven credits is the listed generate. A 30-second miss is a 30-second miss. Start shorter until identity holds, then stretch.

    This is the opposite of single-pass multi-shot directing. Kling-style storyboards and Flux-style multi-angle prompts treat a generation as a sequence of shots. Those models cut. Seedance 2.5 holds. Borrowing their shot-list syntax without borrowing their job is how a 30-second take becomes three cousins in similar coats.

    The AI video generator is the door. Pick Seedance 2.5. Write one continuous space. Then mark time.

    Timestamp the beats, do not recast

    A beat is a change the camera or subject can do without becoming a different scene: locked-off then a slow push then a hold; walk, pause, turn; a practical that flickers in the same setup; a line at 0:08 and a door at 0:18 under the same room tone.

    Forbidden, even labeled "SHOT 2": a new face, a new location, a wardrobe change, a macro of an object that was not in frame, "now we are in a different scene."

    The failure at 0:12 is almost always a recast you did not admit. "Cut to a close-up of her sister" is a second movie. "Camera dollies from a wide to a medium of the same woman, same navy coat, same lamp" is a beat. Write the second sentence. If a true cut to a new setup is the brief, generate a second clip and edit. One hold plus a hard cut in the timeline is cheaper to trust than one generate that tried to be an edit.

    Timestamp in clock time. "Later she turns" lands whenever the model feels like a turn. "0:18–0:22: she stops and turns toward the lamp, same coat, same rain" is a direction. Repeat the locked details at every beat. Redundancy on the page keeps the second beat from re-deciding the first.

    A prompt that actually directs

    Write the hold first, then the clock:

    One continuous shot. Do not cut. Do not change identity.
    Subject: woman, 30s, wet navy trench, same person the whole take.
    Place: night crosswalk, sodium lamp backlight, rain. Same lamp, same street.
    Camera: 40mm, hip height, start locked-off wide.
    Audio: rain, footsteps. One line at 0:20: "We're late." Same voice.
    0:00–0:08: she enters camera left and walks the crosswalk, slow.
    0:08–0:18: measured dolly in. Pace unchanged. Coat, lamp, rain unchanged.
    0:18–0:24: hold closer. She stops, turns to the lamp, speaks the line.
    0:24–0:30: hold profile. No new character. No new location.
    

    Four beats, zero recasts: enter, push, stop, hold. A watch insert is a second generate — or a crop through image-to-video — not a new identity at 0:12. Prove the hold at 15s first. When coat and lamp survive, lengthen the same prompt. Do not add a second person as a reward.

    What to do when you actually need a cut

    A 30-second social spot is often two or three shots pretending to be one generate. Split honestly.

    1. Write the master as one Seedance pass with timestamps, same subject, same space. That file is the spine.
    2. Generate coverage as separate clips, not recasts inside the spine. Insert, reverse, title card — each on the row that owns that job.
    3. Cut in the editor. A hard cut hides a seam. A fake cut inside one denoise pass has to survive identity, light, and audio across the boundary.
    4. Do not extend to dodge a cut. Extending another model from the last frame is a continuation tax. If Seedance already holds 30 seconds, use the 30. If the idea is 45 seconds of new coverage, you are in an edit.

    A talking 4 / 6 / 8 second close-up belongs on a talking row, not on a 30-second wide you punch in on. After every generate, pause at 0:12. Same person, coat, and lamp as 0:02? The timestamps worked. If not, you asked for a recast. Take the new identity out before you add another beat.

    FAQ

    Can Seedance 2.5 cut to a new person at 0:12?

    Not as a hold. Timestamp a tempo shift on the same subject — a stop, a turn, a camera push. A second character is a second clip. The 30-second pass is one shot.

    Why 47 credits for 30 seconds at 720p?

    That is the listed generate on Seedance 2.5: native audio, 4s–30s, 720p, 47 credits. You are paying for a single-pass hold, not a 4K master and not a hidden movie. A recast you throw away still spent the 47.

    Do I still need to stitch if I want 30 seconds?

    Not for one continuous take. Thirty seconds is a published duration. Stitch when you need coverage the hold cannot do — a new angle, a new object, a punch-in that was not in frame. Do not stitch four 8-second clips to fake a hold this row already sells.

    How is this different from Kling-style multi-shot prompting?

    Kling-style multi-shot is a sequence of camera setups. Seedance 2.5 is one shot reasoned from end to end. Directing "cuts" here means marking beats on that shot. Mixing the two syntaxes is how 0:12 becomes a different film.