Video to Seedance Prompt
Video to prompt reverse-engineers an existing clip into a shot-by-shot Seedance prompt: subject, scene, camera, lighting and per-shot timing, plus the first and last frame extracted as images you can feed straight into image-to-video.
How it works
- 1Upload a clip. The browser pulls the first, middle and last frame locally; only those frames and the video itself are sent.
- 2A vision model reads the video with per-frame timestamps and returns a structured breakdown — subjects, scene, and one entry per shot.
- 3That breakdown is rendered into two prompts, one for Seedance 2.5 and one for Seedance 2.0.
- 4Open it in the creator with the first and last frame already attached, and generate.
Why two prompt versions
Seedance 2.0 does not respond to timestamps — it only responds to shot numbers. Handing it a 2.5 prompt silently discards the segmentation, so each version is rendered separately rather than trimmed from the other.
| Seedance 2.5 | Seedance 2.0 | |
|---|---|---|
| Shot segmentation | Integer-second timestamps (0s-3s:) | Shot numbers (Shot 1:) |
| Aspect ratio | Any ratio in [0.4, 2.5] | Six fixed ratios |
| Max length | 30 seconds | 15 seconds |
Limits and cost
- Formats
- MP4, MOV, WebM
- Maximum size
- 50 MB
- Maximum length
- 60 seconds
- Cost
- 3 credits, plus 1 for clips over 30 seconds
- Reference frames
- First, middle and last, uploaded with the result
- Eligibility
- Accounts that have topped up
What it does not recover
- Dialogue and sound effects. This pass reads frames only, so any line of dialogue would be invented rather than transcribed.
- An exact copy. A prompt is a lossy description of a video; the result matches subject, staging, camera and pacing, not every pixel.
- Negative constraints. Seedance only supports negative control for subtitles and audio, so the output uses positive description throughout.
Questions
- What does video to prompt actually return?
- Two prompts and a set of frames. The Seedance 2.5 version segments the clip with integer-second timestamps; the Seedance 2.0 version uses shot numbers instead, because 2.0 does not respond to timestamps. Alongside them you get the first, middle and last frame as images.
- Can it recover the dialogue in a video?
- No. The analysis reads sampled frames, not the audio track, so dialogue and sound-effect fields are deliberately left empty. Reading speech off lip movement in still frames is not recoverable, and filling those fields would mean inventing them.
- How close will the generated video be to the original?
- Close in subject, staging, camera movement, lighting and pacing — not pixel-identical. Carrying the extracted first and last frame into image-to-video pins the opening and closing images, which is what makes the result actually resemble the source rather than merely describe it.
- What does it cost?
- 3 credits per extraction, plus 1 credit for clips longer than 30 seconds. Credits are pay-as-you-go with no subscription, and a failed extraction is not charged.
- Which video files are supported?
- MP4, MOV and WebM, up to 50 MB and 60 seconds.
- Who can use it?
- Accounts that have topped up. Each extraction calls a paid vision model, so it is not offered on unfunded accounts.
