Doubao Seedance 2.5 Prompt Guide: Task Locking and Structured Prompts

Seedance 2.5 pushes video generation into "long narrative × strong referencing × precise editing × multilingual": 30-second single takes, up to 50 reference assets per request, and native dialogue in 10+ languages. For prompt writers, the most important change is not those numbers but the new "locked / unlocked" task system— if you don't internalize this taxonomy, parameter errors and "the output ignores me" will eat half your debugging time. This post distills the official prompt guide into an actionable checklist.

First, tell locked from unlocked

2.5 splits tasks by whether the asset becomes part of the output video's timeline (2.0 has no such distinction):

Edit tasks are triggered by keywords (add/delete/modify/replace…), extension by "extend forward/backward, continue"; asset roles are set with content.role: first_frame / last_frame / reference_image / reference_video / reference_audio.

Engineering limits on input assets

• Hard caps: ≤30 images (up to 4K), ≤10 video clips (≤30s combined), ≤10 audio clips (≤30s combined).
• 1–8 subject images and 1–5 subject video/audio clips work best; beyond that, stability drops and you end up rerolling.
• With more than 5 subjects prefer single views; split multiple angles into separate images instead of packing them into one.
• Grid storyboards work best within 15 panels; stick figures or line art are recommended, and avoid writing much text on the storyboard.
• Keep video-edit inputs within 20 seconds.

Structured prompts: the four-part format

The officially recommended organization, written in order:

1. Asset references: number assets in upload order and bind them to roles — "img1-2 are character 1, matched with audio 1; image 3 is the scene reference". Do not write names inside the images and make the model guess — that is how duplicate characters happen.
2. One-sentence overview: subject + place + event + genre/style + any special camera work.
3. Segmented plot: split by timestamps or "Shot N", and for each segment write the frame, camera move, action, dialogue, and sound.
4. Closing globals: camera setup, environment, sound, and mood that persist throughout.

Realistic nature-documentary style, cinematic true-to-life lighting. On a warm afternoon forest slope, a chubby panda cub tumbles down the hill.

0s-3s: The panda cub lies on the grassy slope and slowly rolls sideways down it, clumsy in its movements, flattening blades of grass; sunlight filters through the trees from the upper left in dappled patches.
3s-8s: The panda rolls to a stop at the lower right of the frame, shifts from its side onto its belly, round face toward the camera, lifting and lowering its head slightly, <soft whimpering sounds>.

Low-angle medium-wide shot with a slight handheld feel, gently tracking the panda; foreground grass blurred, subject sharp; natural ambient sound, warm and realistic overall. No subtitles.

Timestamp rules

• Use 1-second granularity; keep the timeline continuous — no gaps like "0-3s… 5-6s".
• All three forms work: ranges (0-3s… 3-7s), points in time (a fast leftward pan transition at 5s), and relative time (3 seconds later…).
• Too little content in a segment and the model improvises; too much and it over-cuts — keep each segment's information load reasonable.
• Do not use timestamps to control frequency ("shakes the head 3 times in one second" is unreliable).

Negative control and camera language

• Subtitles: "no subtitles", "do not add dialogue captions".
• Audio can be controlled at a fine grain: "no bgm, generate only ambient and action sounds", "no sound at all".
• Standard camera language works as-is (shot size, push/pull/pan/track, low angle, first person), and popular moves are recognized too (one-take, dolly zoom, FPV, bullet time).
• Explain niche terms inline: "rack focus — the focal point shifts smoothly, foreground trees blur while the person in the background sharpens".
• Spell out transitions with trigger and method: "a fast leftward pan transition at 5s (left wipe + natural dissolve)".

Differences from 2.0 at a glance

Run the prompt

For regular text-to-video, drop the structured prompt into prompt and submit an async task through the OpenAI-compatible endpoint (submission is free; you are billed by duration only on success):

curl https://zerofa.ai/v1/videos \
  -H "Authorization: Bearer sk-zerofa-xxx" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "doubao-seedance-2-5",
    "prompt": "(use the four-part structure above)",
    "duration": 10,
    "resolution": "720P",
    "aspect_ratio": "16:9",
    "enable_audio": true
  }'
# → {"id":"<task-id>","status":"submitted"}, then poll GET /v1/videos/<task-id>

When you need content.role (first/last frame, multi-asset referencing) or other native parameters, use the Volcano Ark /ark native passthrough — the request body matches the official docs exactly, with every advanced parameter preserved. Endpoint details are in the video generation API docs; the model and live pricing are on the Doubao Seedance 2.5 model page. For the older model's prompt style, see the Seedance 2.0 prompt guide.

FAQ