Most disappointing generations are not model failures — they are under-specified prompts. Seedance 2.5 accepts up to 50,000 characters and up to 9 reference images plus 3 audio inputs, so there is plenty of room to be specific without being vague.
The five things every video prompt needs
- Subject — who or what, with one or two distinguishing details.
- Action — a single continuous motion, not a sequence of events.
- Camera — shot size and movement (handheld tracking shot, slow drone pull, locked-off wide).
- Light — time of day and quality (golden hour, overcast, practical neon).
- Look — lens and texture (35mm film grain, shallow depth of field, 16mm handheld).
Use the reference inputs deliberately
- Image-to-video: 9 reference images available; first-frame and first/last-frame modes reduce the count to 1 or 2.
- Video-to-video: 1 source video plus up to 9 stills and 3 audio inputs.
- Reference audio drives lip-sync and rhythm — supply it when the beat matters.
Content rules worth knowing before you write
- Prompts are screened before the job starts, and an additional content check can run on the output as well.
- Sexual content involving minors, non-consensual sexual content, and instructions to bypass safety systems are rejected outright.
- Prompts that try to defeat moderation are rejected even when the underlying idea would be allowed — describe the shot, not the workaround.
- The audit trail stores a hash of the prompt, not its text.
If a prompt is rejected, nothing is charged. If a generation fails, the reservation is refunded automatically — check Settings → Credits for the ledger entry.
A skeleton that survives the first render
We write prompts in this order, one clause per element, and it is the difference between a shot that lands on the first try and five attempts that all miss in different ways: "[shot size] of [subject with one distinguishing detail], [single continuous action], [camera movement], [light and time of day], [lens and texture]." Everything after the subject is what the model uses to keep the frame coherent — drop the camera clause and you get a wandering frame, drop the light clause and you get midday flatness.
- Subject first, and make it specific: "a fishmonger in a rust-coloured apron", not "a person".
- One action per shot. "She lifts the crate and sets it down" is one shot; "she lifts the crate, walks out, and drives away" is three, and a single generation will smear them together.
- Name the blocking when it matters: where the subject starts in frame and where it ends up is the cheapest way to control a handheld or tracking shot.
- State the end frame if the clip has to match the next one — Seedance 2.5 can return the last frame, which makes shot-to-shot continuity a matter of reusing it as the next prompt’s first frame.
Camera and motion vocabulary that actually lands
- Shot size: extreme close-up, close-up, medium, wide, establishing. Say it explicitly rather than implying it through description.
- Movement: locked-off, slow push in, pull back, handheld tracking, drone pull-away, orbit, whip pan. One movement per shot reads as intentional; two read as a mistake.
- Speed: name it — slow, gentle, at walking pace. "Dynamic" and "cinematic" alone tell the model nothing about how fast the frame should change.
- Keep the subject’s scale stable across the cut if you plan to edit shots together: a slow push in on a close-up is easier to match than an orbit at wide.
The four rejections we see most, and the rewrite
- Asking for a way around the filter instead of the shot you want. Describe the frame — subject, action, camera, light — and let moderation judge the result. Workaround phrasing is rejected even when the underlying idea would have been fine.
- Describing a specific real person as the subject. Describe the appearance instead: age range, build, wardrobe, expression. The model needs a look, not an identity.
- Writing an explicitly graphic moment. Implied action carries the same narrative weight and renders: the impact off-frame, the reaction shot, the aftermath.
- Any suggestive framing involving minors, in the prompt or in a reference image. There is no rewrite for this one — the request is refused and the account is what gets looked at.
Settings that change the result as much as the words
- Duration: 4 to 30 seconds on Seedance 2.5. Write the shot for the length you picked — a 15-second prompt on a 4-second render is a different film.
- Resolution: 480p and 720p are the tiers we sell. Block out at 480p, deliver at 720p; the prompt carries over unchanged.
- Generate audio: leave it on when the clip needs its own soundscape, off when you will lay music over it.
- Reference audio is the lip-sync lever — supply the track when the beat or the delivery has to land on a specific moment.
What gets a prompt refused, and how refunds work when a task fails, is in the moderation guide at /docs/seedance-moderation; the model page at /seedance2.5 has the full capability sheet and a wall of real clips with their prompts.