The Seedance Prompt Stays Between Fifty and Two Hundred Words

Fifty words felt like an insult the first time I read the rule. Then I counted one of my prompts and it was 340. The official docs push a five-part structure: Subject, Action, Camera, Style, and an optional Timeline for multi-beat clips. The whole prompt lives between 50 and 200 words. The measured sweet spot is 50 to 70. That is the number worth taping to the monitor, because the model is not short of attention, it is short of priorities, and every extra clause splits one. Write all five parts. Then cut until the cut hurts. The clip comes back better on the other side of that cut.

Subject is who or what is on screen, and physical specificity matters: age, clothing, distinguishing features, posture. Action is what they do, where one strong verb beats five vague ones. The chapter's own worked example runs 52 words and starts like this. A weathered fisherman in a yellow raincoat hauls a net off the side of a wooden skiff at dawn. Yellow raincoat. Wooden skiff. Dawn. Inside a fifty-word window, weathered is doing real work, and so is the coat, because the model has to spend its attention somewhere and you would rather choose where. Write the subject the way a casting note reads, then give it one verb with somewhere to go.

Camera is where most prompts go quiet, and the docs are blunt that it is the single biggest quality lever. Use cinematography vocabulary rather than the camera shows: dolly in, dolly out, pan left, pan right, tilt up, tilt down, tracking shot, rack focus, crane up, arc shot, shallow depth of field, handheld, static locked-off, low-angle, overhead. The warning underneath the list is the part people skip. Do not stack camera moves. A 360 degree rotation while zooming in as the subject runs toward camera reliably produces garbage. One dominant move per shot, chosen like you are paying for it. The frame that survives is the one the camera commits to, not the one it tours.

Style is genre, era, film stock, colour grade, and a lighting reference, which is the same short list a colourist would ask you for. Cold desaturated grade. Anamorphic lens flare. Natural morning light. Documentary realism. Two or three of those and the render has a world to sit in. The fifth part, Timeline, is optional and only earns its place on multi-beat clips: 0 to 2 seconds close-up on hands, 2 to 4 seconds pull back to reveal. Use it when the shot has to change inside itself, and skip it for a single held moment, because a timeline on a five-second clip is just a second prompt arguing with the first.

The model is not short of attention, it is short of priorities, and every extra clause splits one.

Seedance 2.0 launched in March 2026 and sits at number one on the Artificial Analysis Video Arena, ahead of Kling 3.0, Veo 3 and Runway Gen-4.5. It takes text, images, audio and video at once, and three tiers matter. 2.0 Pro handles hero shots, ads, human actors and narrative, up to 15 seconds at 1080p, roughly 0.24 dollars a second, and it is the only tier that should see a face. 2.0 Fast covers stylized, product, landscape and abstract work at about 0.022 dollars a second on Atlas Cloud, which is 91 per cent cheaper. 1.0 Pro is the legacy tier still on many wrappers: 10 seconds at 1080p, 0.50 to 0.62 dollars a generation.

Four doors. Volcengine for direct access, though some tiers still want a Chinese account. fal.ai for the cleanest API. Replicate for a friendly playground. Atlas Cloud for the cheapest Fast tier. Every host exposes the same parameters: resolution, duration, aspect ratio including 9:16 for TikTok, seed lock, and reference image plus audio on 2.0 Pro only. Fill the negative prompt with text, watermark, blurry, extra fingers, warped face. Draft at 50 to 70 words, first pass on Fast at 480p and five seconds, lock the seed, then re-render at 1080p on Pro. Then the failures. Face consistency across generations stays weak. On-screen text is unreliable. Complex physics breaks. Copyrighted prompts are filtered. And the probability that Fast plus a human face returns a distorted face is not zero. Most people price it at zero until the render lands.