Seedance 2.5 reads a shot list, not a wish

Most video models take a sentence and improvise. Seedance 2.5 does better than that when you give it structure, and worse than most when you do not - a long paragraph of adjectives gets you a long paragraph's worth of guessing.

Think of it as briefing a producer who will shoot exactly what is on the page and invent whatever is missing. Everything below is about leaving less to invent.

The shape of a prompt that works

Four parts, in this order. You do not need all four every time, but this is the order the model follows.

1. The asset map. If you attached anything, say what each item is and what it is for, by its upload number.

Image 1 = the singer. Image 2 = the venue. Image 3 = the jacket.

Number them by upload order and bind every one. Do not rely on a name written inside an image - that is a reliable way to get two of the same character.

2. One sentence of summary. Subject, place, event, genre, camera. One line.

A singer walks out onto an empty club stage at 3am, handheld, grainy 16mm.

3. The plot, in order. This is where the detail goes. A short timeline beats a paragraph every time.

4. Continuity notes. Anything that must not change between shots: the camera height, the light, the weather, the ambience.

Timestamps are the 2.5 lever

This is the biggest change from earlier Seedance tiers, and the one most people never use.

Work in whole seconds, and make the ranges continuous with no gaps.

0-3s:  wide, slow push in. She stops at the mic, looks up.
3-7s:  medium, static. She adjusts the stand. Hum of the PA.
7-10s: close, slight handheld drift. She breathes in.

Three rules that save a render:

• No gaps. 0-3s then 5-9s leaves two seconds the model fills with whatever it likes.
• Match the plot to the time. Too little in a range and it improvises; too much and it either adds cuts you did not ask for or silently drops a beat.
• Do not use timestamps for fast, repeated actions. "Shakes their head three times a second" is not a timing instruction, it is a request for a mess.

You can also write a moment (at the 5-second mark...) or a relative one (after 3 seconds...). For a transition, give both when and how.

Camera language works by name

You can write camera direction plainly and it will be followed: shot sizes, moves (push in, pull out, pan, track, orbit, tilt), angles (low, overhead, first person), and most named techniques - long take, dolly zoom, aerial, FPV, handheld, speed ramp, bullet time.

For anything niche, add the plain-language version in the same breath: rack focus - focus shifts from the branches in front to her face behind them. The explanation costs you eight words and removes the guess.

Actions and faces: general, then specific

Describe movement in general terms and spend your detail on the two or three moments that matter. "Several sets of high knees, then a somersault" lands better than a move-by-move list, and repeating the same action in a prompt tends to produce it once.

For expressions, write a plain descriptive sentence rather than an idiom. And go one notch milder than you mean: intense emotion words are the main cause of the glowing-eye artefact below.

Audio is free here, so direct it

Seedance 2.5 generates sound, and the sounded rate is identical to the silent rate - a speaking clip costs exactly what a mute one does. That is unusual and it is why talking characters get rendered on this family.

Because it is on, you have to direct it. Put an audio policy at the top of the prompt and repeat it at the end:

Audio: room tone and footsteps only. No BGM. No subtitles.

For dialogue, use the form Character's line (tone): {what they say}. Do not add a separate note about tone after the line, and do not repeat the words - both of those make on-screen subtitles more likely, not less.

What Video Studio gives you for Seedance 2.5

Checked against the live model configuration on November 14, 2026.

Attaching a first or last frame forces the shape to adaptive. This is not our choice - the provider rejects every other aspect ratio once a frame is attached, so we switch it for you rather than letting the render fail. If the exact shape matters more than the frame does, drop the frame and pick the ratio.

When you do use both frames, give them the same aspect ratio as each other, or the last one gets stretched to match the first.

What it costs

89 credits per second at 720p, 40 credits per second at 480p. So a 5-second 720p clip is 445 credits and a 30-second one is 2,670. Rendering at 480p is not a small saving - it is less than half.

When it goes wrong

What the model can do that AetherWave does not

Worth saying plainly so you do not spend credits chasing it. Seedance 2.5 as ByteDance ships it also supports editing an existing video, extending one forwards or backwards, MOV output, and far more reference assets than nine. None of those are wired up on AetherWave today. Asking for them in a prompt will not produce them; you will get a normal render that ignored the request.

Everything else on this page works, and most of it works on the other Seedance tiers too - the asset map, the one-line summary, continuous whole-second ranges, continuity notes, positive phrasing and the audio policy are general craft, not 2.5 features.

This page is our own summary of ByteDance's published prompting guidance for Seedance 2.5, rewritten for what the model does on AetherWave. Where the two disagree, the table above is what your render will actually do.