Seedance 2.5: ByteDance's video model — 30 seconds, 50 references, timestamps

What is Seedance 2.5

Seedance 2.5 is a video generation model from ByteDance Seed, the same team that makes the Seedream image models. ByteDance released it on July 31, 2026. It builds on Seedance 2.0 and keeps the same architecture: picture and sound are generated together, and you can feed it text, images, video and audio at once. For the whole family, see the Seedance overview.

The headline change is length. Seedance 2.5 produces up to 30 seconds in a single generation instead of 15, and it builds several connected shots inside one clip: setup, development, turn and resolution. You can keep extending a finished clip round after round, which is why ByteDance talks about videos “several minutes long.”

The model runs in ByteDance’s Chinese apps Jimeng and Doubao Pro and in its developer cloud. The API model ID is doubao-seedance-2-5-260628.

What Seedance 2.5 can do

30 seconds with sound in one pass. Set any length from 4 to 30 seconds, or choose auto and the model picks a whole number of seconds to fit the story. Sound is generated with the picture: lip-synced speech, sound effects and music.

Up to 50 references per request. 30 images at up to 4K, 10 videos and 10 audio clips. Each video or audio file can run 2–30 seconds, but each type is capped at 30 seconds in total. A reference can define a character, a voice, motion, style, a location or a storyboard.

Timestamps in the prompt. Write “0–5s: …, 5–10s: …” and the model maps action, camera and pacing to those windows. Seedance 2.0 ignored timestamps and only followed shot numbers.

Editing finished clips. Add, replace or remove an object, a character, the background, subtitles, music or a voice with a plain instruction: “replace the character in @Video1 with the one from @Image1.” The edit keeps the source clip’s length and aspect ratio.

Extending forward and backward. “Extend @Video1 backward: the hero falls from the sky…” and the model generates a new opening that matches the picture and sound. For edits and extensions, ByteDance recommends MOV output, which keeps color, brightness and audio more consistent.

Animation from a 3D blockout. You can feed it a “clay” video: a scene built from grey primitives with no textures. The model takes motion, paths, camera and lighting from it, and the look from your prompt and images.

480p drafts. Draft mode first renders a cheap 480p preview so you can check the story, shots and motion. The final render reuses the same prompt, seed and references, and the preview is billed separately at the 480p rate.

Speech in 11 languages. Chinese, English, Spanish, Indonesian, Malay, Thai, Arabic, Portuguese, Vietnamese, Japanese and Korean.

Seedance 2.5 vs. 2.0: what changed

	Seedance 2.0	Seedance 2.5
Released	Feb 12, 2026	Jul 31, 2026
Length per generation	4–15 s	4–30 s or auto
Resolution	480p, 720p, 1080p, 4K	480p, 720p, 1080p (10-bit)
Aspect ratio	6 fixed options	6 fixed options or follow the input, 0.4 to 2.5
References	9 images, 3 videos, 3 audio	30 images, 10 videos, 10 audio
Timestamps in the prompt	no, shot numbers only	yes, whole seconds
Multi-view character sheet in one image	not recommended	supported
480p draft	no	yes
Output	MP4	MP4 and MOV
Faster variants	Fast, Mini	none

The two versions have a different look: the same prompt and reference produce visibly different results in 2.5 and 2.0. If a project started on 2.0, ByteDance suggests extending the 2.0 clips with 2.5 and exporting MOV, which preserves the style better than generating from scratch.

There’s also Seedance 2.5 4K with native 4K and 480p drafts. As of October 2026, it’s limited to selected film studios — see our Seedance 2.5 4K news story.

What Seedance 2.5 is good for

• A short film or ad of 20–30 seconds in one piece — no stitching several generations together, no lighting jumps between them.
• A series with one character. Three angles of the character plus a voice sample, and the character stays recognizable from clip to clip.
• Dubbing a clip into another language. An edit replaces the speech and the lip movement and leaves everything else as it was.
• Previs and animation from a 3D blockout. An animator blocks the motion with simple primitives, and Seedance 2.5 renders it in the target style.
• Fixing real footage. Remove subtitles or a watermark, swap the background, or add an object at a specific timestamp.

ByteDance gives practical limits: results are most stable with 1–8 characters from images and 1–5 characters from video or voice. At 9–12 characters, consistency starts to slip and you’ll need retries. With more than five characters, upload each angle as a separate image rather than one sheet.

Seedance 2.5 vs. alternatives

	Seedance 2.5	Kling 3.0	Veo 3.1	Vidu Q3
Developer	ByteDance Seed	Kuaishou	Google DeepMind	ShengShu
Length per generation	4–30 s	3–15 s	4, 6, 8 s	1–16 s
Max resolution	1080p, 10-bit	4K	4K	1080p
References	30 images, 10 videos, 10 audio	up to 7 images or elements, 1 video	up to 3 images	up to 7 images
Editing existing video	yes, with timestamps	yes, in Omni	extension only	extension only
Sound	yes	yes	yes	yes
In AI2Face	yes	yes	yes	yes

Seedance 2.5 wins on single-clip length and reference count. For 4K and tight physics in short shots, go with Veo 3.1 or Kling 3.0. For anime and comic dramas with several characters, look at Vidu Q3.

How to write prompts for Seedance 2.5

ByteDance recommends thinking like a director and writing the prompt in four layers:

1. Reference bindings. Name every file by upload order and role: “@Image1 is the heroine, @Video1 sets the camera move, @Audio1 is her voice.”
2. A one-line summary. Subject + location + event + genre + camera.
3. The plot over time. Timestamps (“0–3s”, “3–8s”) or “Shot 1, Shot 2” — and for each segment, the picture, camera, action, dialogue and sound.
4. Recurring details. Camera angle, lighting, mood, ambient sound — anything that stays constant through the clip.

Describe what you want rather than what you don’t: “an empty promenade,” not “no people.” The exceptions are subtitles and music — the model does understand “no subtitles” and “no background music.”

If the clip has to follow a storyboard exactly, upload the frames as separate images and start the prompt with “Use Images 1 to 7 in order as keyframes.” A storyboard combined into one sheet only sets the overall plot: ByteDance recommends no more than 15 panels, simple line art and no text on the sheet.

A prompt for that storyboard:

Storyboard @Image1, heroine @Image2. Moody short film, 35 mm, cold blue dusk light, film grain.
Shot 1 (0-5s): wide shot of a lighthouse on a rocky cliff, waves crash below, wind howls.
Shot 2 (5-10s): the heroine in a hooded raincoat climbs a spiral staircase, her steps echo.
Shot 3 (10-13s): close-up, her hand turns a round brass door handle.
Shot 4 (13-20s): she enters the lantern room and switches on the huge lamp, a low hum rises.
Shot 5 (20-25s): over-the-shoulder view of a small ship fighting the storm through the window.
Shot 6 (25-30s): wide shot, the lighthouse beam sweeps across the sea, the ship turns toward it.
Sound: wind, rain, distant thunder, no music. No subtitles.

Fixes for common failures, from ByteDance’s official guide:

• Misspelled on-screen text. Spell the word out letter by letter, or supply the text as an image reference.
• Glowing eyes. The model sometimes renders strong emotions like “furious” as glowing pupils. Use neutral wording (“surprised”) and add “normal human eyes.”
• Music despite “no music.” List every variant — music, background music, score, melody — and repeat the constraint at the start and end of the prompt.
• Subtitles despite the ban. Don’t repeat dialogue words in the description, and use the “Character’s line (emotion): text” format. The reference video must be subtitle-free too.
• Fingerprint-like textures on grass and foliage. This happens when a reference image is larger than the output resolution. Downscale it to the video size.
• Bubbling or echo in the audio. Remove water, waves and echo from the prompt unless the story needs them.
• A complex task falls apart. Split it in two: generate from references first, then edit the result.

How to use Seedance 2.5 in AI2Face

Seedance 2.5 is in the AI2Face generator next to Veo 3.1, Kling 3.0 and Gemini Omni, so it’s easy to compare one prompt across several models.

1. Sign up — new accounts get starter credits.
2. Open the generator, switch to video and pick Seedance 2.5.
3. Upload your references and write the prompt using the structure above. Start with a 5–8 second clip: testing an idea that way costs a fraction of a 30-second render.
4. The credit cost is shown before you start and depends on length and resolution. Credits don’t expire.

You can create reference images — a character, a location, a product — in the same generator, for example with Nano Banana Pro. Every image on this page was made that way.

Limitations

• No 4K. The ceiling is 1080p. Native 4K is only in the closed Seedance 2.5 4K version.
• Fixed aspect ratio for edits and extensions. The model follows the source clip’s aspect ratio, and for edits its length too. An edited clip can come out 0.3–0.4 seconds shorter than the source.
• Complex physics and crowds. ByteDance itself says the plausibility of complex motion and scenes with several interacting characters still needs work.
• Text errors. The model mixes up letters in on-screen text: “E” turns into “A” or “I”, and similar-looking characters get swapped in Chinese.
• Real people. Portraits of real people are allowed as references only with confirmed rights to use them.
• Languages. Outside the 11 official languages, prompts and speech are less reliable.

Bottom line

Seedance 2.5 is the pick when you need one continuous 20–30 second clip with a story, a consistent character and sound: neither Kling 3.0 nor Veo 3.1 produces that much in one generation or accepts 50 references. It’s also handy for editing finished footage at a specific timestamp. If 4K or flawless physics in a short shot matters more, use Veo 3.1. If budget is tight and 15 seconds is enough, look at Seedance 2.0 and its Fast variant. Start with a 5-second draft in AI2Face and stretch it to 30 seconds once the prompt works.