doubao-seedance-2.5. Try a prompt on a cheap 480p clip first, and only then render in 1080p.
The tips below are based on BytePlus’s official prompt guides for Seedance 2.0 and Seedance 2.5, and on our own test of prompts and dialogue in Russian.
The prompt formula
The “advanced formula” from the official Seedance 2.0 guide:subject + action + scene + lighting and color + camera movement + style + image quality + constraints
Which language to write in
Russian isn’t on the official list of Seedance 2.5 languages. The list has 11: Chinese, English, Spanish, Indonesian, Malay, Thai, Arabic, Portuguese, Vietnamese, Japanese and Korean (BytePlus, November 2026). We tested how the models handle Russian (November 2026, one 4-second 480p clip per case):- You can write the prompt in Russian. On
doubao-seedance-2.0-miniwe compared the same prompt in Russian and in English and saw no difference in the picture. We haven’t tested long, complex scenes in Russian — for those, English is the safer choice. - For dialogue in Russian, choose
doubao-seedance-2.5. In our test,doubao-seedance-2.0-minigarbled some words of the Russian line. - State the language of the line:
says in Russian {…}(orговорит по-русски {…}if the whole prompt is in Russian). The official 2.0 guide recommends marking all dialogue that isn’t in Chinese or English this way. Keep all dialogue in one clip in the same language.
Storyboards and timestamps
If the clip has more than one event, describe them in order — this helps the model keep them in sequence:- Seedance 2.0 only understands shot numbers and ignores timestamps. Don’t set a duration for each shot.
- Seedance 2.5 understands both shot numbers and whole-second timestamps, for example
0-3 seconds: … 3-7 seconds: … 7-15 seconds: …or[1s-4s] … [4s-8s] …. The intervals must be continuous, with no gaps. Timestamps don’t work for very frequent actions such as “shakes their head three times per second”. - Within each shot, describe in order: camera movement → action and emotion → where the subject is → sound.
Actions and emotions
- Be specific, including speed and degree: not “moves”, but “slowly raises a hand”, “smoothly turns their head”, “slightly lowers their head”.
- Prefer smooth, small movements. The official guide advises avoiding sudden, large actions such as sprinting, big jumps and rolls.
- Link actions together: “turns around and, without stopping, raises a hand”.
- Show emotions through visible details instead of naming them:
Camera
Seedance understands standard terms well: wide shot, medium shot, close-up, push-in, pull-out, pan, tracking shot, orbit, static camera, overhead shot, low-angle shot. One camera movement per shot. If you ask for a push-in, a pull-out and a pan at the same time, the picture becomes less stable.Sound, dialogue and subtitles
Seedance 2.0 and 2.5 generate video with sound out of the box. To help the model tell types of information apart, the official 2.0 guide suggests these brackets:- If you don’t need sound, pass
"generate_audio": false(insidemetadatafor/v1/videos) — see the video generation guide. - The model sometimes misspells on-screen text, swapping or dropping letters. The official 2.5 guide suggests two fixes: ask for the word’s letters to appear one by one, in order, or pass the exact text as a reference image.
Constraints: subtitles, logos, watermarks
Sometimes the model adds subtitles, a logo or a watermark on its own. Add constraints at the end of the prompt:- You can’t rule out subtitles completely, but this makes them less frequent — especially in landscape (
16:9), where they appear less often than in portrait. - For Seedance 2.5, the official guide recommends describing what should be in the shot rather than listing what shouldn’t, where possible. Negative constraints work reliably for subtitles and sound (“no subtitles”, “no background music”).
- If you want a specific style (anime, 3D cartoon) but the reference image is realistic, state the style explicitly — otherwise the clip may drift toward realism.
First frame and references
- Image-to-video (first frame). Don’t describe what’s in the picture — the model can see it. Describe what should happen: movement, camera, sound.
- Multiple references. Refer to each subject the same way throughout the prompt and say where it comes from: “the woman from image 1”, “the scene style from image 2”. Don’t max out the reference limit: the official guide recommends sticking to 4–5 references, which makes it easier for the model to tell what matters.
- Photos and videos of ordinary people are fine; celebrities and film or game characters are not. Seedance 2.0 mini and fast have a few caveats — see Seedance: real person.
10 ready-made prompts
Swap in your own subject and adjust the details. All examples are for text-to-video, without references. We tested prompts 5 and 6 exactly as written; the rest are built on the same formula. Prompts 6 and 10 are meant fordoubao-seedance-2.5.
1. Product on a table (ad)
1. Product on a table (ad)
2. Food close-up
2. Food close-up
3. Nature
3. Nature
4. City at night
4. City at night
5. Cat on a windowsill (tested)
5. Cat on a windowsill (tested)
6. A line in Russian (tested, doubao-seedance-2.5)
6. A line in Russian (tested, doubao-seedance-2.5)
7. Cartoon character
7. Cartoon character
8. Anime
8. Anime
9. Text in the shot
9. Text in the shot
10. Timestamps (doubao-seedance-2.5, 8 seconds)
10. Timestamps (doubao-seedance-2.5, 8 seconds)
How to avoid overpaying for test runs
- Test on short clips: 480p, 4 seconds. That’s about $0.17 with `doubao-seedance-2.0-mini` and about $0.52 with
doubao-seedance-2.5at the standard price. Current prices are on the pricing page. Test timestamps and Russian dialogue ondoubao-seedance-2.5right away: 2.0 doesn’t support timestamps, and in our testdoubao-seedance-2.0-minigot some Russian words wrong. - Use draft mode. It’s available only for
doubao-seedance-2.5and only in the ModelArk format: render a 480p draft first, then turn the one you like straight into 1080p, keeping the same prompt, references andseed. See “Draft mode” in the video generation guide. - Change one thing at a time — camera, lighting or action — so you know what worked.
- Failed tasks aren’t billed: if generation fails, the pre-charge is refunded to your balance.
How to try Seedance via RuAPI
No code: open the “Video generation” page in the console and paste a prompt — see Seedance online in the browser. Through the API:Sign up
Top up
Key
sk-... string.Your first video
"model": "doubao-seedance-2.5", and for prompt 10 also "seconds": "8". To poll the task status and download the video, see the video generation guide.FAQ
Can I write the prompt in Russian?
Can I write the prompt in Russian?
doubao-seedance-2.0-mini and doubao-seedance-2.5 followed Russian prompts accurately. We haven’t tested long, complex scenes in Russian — for those, English is the safer choice.Can a character speak Russian?
Can a character speak Russian?
doubao-seedance-2.5: in our test the line was spoken accurately. With doubao-seedance-2.0-mini, the speech was recognizably Russian, but some words got garbled. Put the line in curly braces and state the language: says in Russian {…}.Why did subtitles or a watermark appear in the video?
Why did subtitles or a watermark appear in the video?
Why was the task rejected with SensitiveContentDetected or PolicyViolation?
Why was the task rejected with SensitiveContentDetected or PolicyViolation?
Is there a subscription?
Is there a subscription?