Seedance 2.5 Tutorial: How to Use It, Step by Step (2026)

To use Seedance 2.5, you write a prompt that describes subject, action, camera, light and sound, optionally attach reference images (up to 30), a reference video or reference audio, pick a duration between 4 and 30 seconds and an aspect ratio, and generate — the whole clip comes out of one generation, soundtrack included. This Seedance 2.5 tutorial walks through every mode: text-to-video, image-to-video with references, the 30-second single shot, multi-shot prompting inside one clip, dialogue notation, and editing or extending an existing video.

Everything below was written against Seedance 2.5 as it runs in the Framiq studio, where it is the default video model. Where a capability is only reported by ByteDance or the press and not something you can use in Framiq, we say so.

What's new in Seedance 2.5

ByteDance announced Seedance 2.5 on 23 June 2026 at its Volcano Engine FORCE event. The headline changes, as reported:

• Up to 30 seconds in a single shot, roughly double Seedance 2.0's 15 seconds.
• Many more references: reports mention around 50 multimodal references in total (30 images, 10 videos, 10 audio clips).
• Region-level editing and extension of existing video.
• Native audio — dialogue, ambience, music and effects generated with the picture — with 10+ languages reported.
• 3D blockout camera control has also been reported; it is not something you can use in Framiq today.

Resolution reports conflict (some mention 1080p or 4K). In Framiq, Seedance 2.5 runs at up to 720p, in 9:16, 16:9, 1:1 and other common ratios.

What matters in practice: longer clips mean you can tell a complete beat — setup, action, payoff — without stitching, and 30 reference images mean the model sees a face from enough angles to keep it the same person. That is why Seedance 2.5 is the strongest model for consistent AI characters that we offer.

Seedance 2.5 settings explained

Setting	Range in Framiq	What it changes
Duration	4–30 s	Length of the single clip. Longer = more API cost.
Resolution	up to 720p	720p for delivery, lower for drafts.
Aspect ratio	9:16, 16:9, 1:1…	9:16 for TikTok/Reels/Shorts, 16:9 for YouTube.
Reference images	up to 30	Characters, products, locations, style.
Reference video	1 clip	Edit or extend an existing video.
Reference audio	optional	A voice or music the clip should use.
Start / end frame	optional	Pin the first and/or last frame to an image.
Audio	native	Soundtrack generated with the video.

A note on cost: price scales with duration and resolution. As a ballpark, a 30-second 720p Seedance 2.5 clip costs about $7 of API cost. Attaching a reference video costs more than plain text-to-video, because the input clip is billed too. Exact per-model prices are on the pricing page.

Step 1: Text-to-video with Seedance 2.5

Open the Video tab in the studio, make sure Seedance 2.5 is selected, and write your prompt. The structure that works best is subject → action → camera → light → style → sound:

A young woman in a yellow raincoat runs across a wet zebra crossing in Tokyo at night,
laughing, holding a paper bag of oranges. Handheld tracking shot from the side, then the
camera stops and she runs out of frame. Neon reflections on the asphalt, light rain,
shallow depth of field, 35mm film look. Sound: rain on the street, distant traffic,
her footsteps splashing.

Tips for text-to-video:

• One clear action per beat. "Runs, laughs, drops a bag, picks it up, waves" in 6 seconds becomes mush. In 20 seconds it can work.
• Name the camera move. "Slow dolly in", "static wide shot", "handheld follow", "overhead drone pull-back". Without it the model picks for you.
• Describe light like a cinematographer. "Golden hour backlight", "single tungsten practical", "overcast soft light".
• Short on purpose. If you only have an idea ("cat chef making pancakes"), the AI prompt generator button expands it into a full prompt you can edit.

For a broader primer, see our AI video prompting guide.

Step 2: Image-to-video with references

References are where Seedance 2.5 pulls ahead. You can attach up to 30 images, and you should tell the model what each is for.

Start frame: if you want the video to begin exactly on a picture (a product shot, a generated image), set it as the start frame. Add an end frame if you need the shot to land somewhere specific — useful for transitions and loops.

Reference images: use them for identity and look rather than for the exact first frame. A good character set is 6–15 photos: face front, three-quarter, profile, full body, a couple of expressions, consistent lighting. Then write the prompt as if the person were already known:

The woman from the reference photos sits at a café terrace in Lisbon, sips an espresso,
then looks up at the camera and smiles. Medium close-up, slow push in. Late afternoon sun,
warm tones. Keep her face, hair and freckles exactly as in the references.
Sound: café chatter, a spoon on porcelain.

Characters with @Name: in Framiq, create a character once (from photos, or import photos from an Instagram, TikTok or X profile), optionally run the 3D scan — clean studio views of the face, then full body, then profiles and back, each stage approved by you — and tag it in any prompt with @Name. The model then receives that character's references automatically, so the same face shows up shot after shot. Only use your own face or the face of someone who has consented; never a celebrity.

@Lina walks into a sunlit bakery, picks a croissant from the counter and turns to the camera.
Wide shot, then cut to a close-up of her face as she takes a bite. Soft morning light.

Mixed references work too: a character, a product photo and a location photo in one generation. Say which is which: "the bottle from the product photo", "the kitchen from the location photo".

Step 3: The 30-second single shot

Thirty seconds in one generation is the big change. To use it well:

1. Write a timeline, not a sentence. A 30-second prompt with one line of description gives you 30 seconds of the same thing. Break it into beats.
2. Give the camera somewhere to go. A long take needs motion with intention: "starts wide, slowly pushes in, ends on her hands".
3. Keep the cast small. One or two characters stay consistent much more easily than five.

[0:00–0:08] Wide shot of an empty mountain road at dawn, mist in the valley.
A vintage red motorbike appears at the far bend.
[0:08–0:18] The camera tracks alongside the rider, a man in a brown leather jacket,
wind in his scarf, sun rising behind the peaks.
[0:18–0:30] He stops at a lookout, takes off his helmet, breathes out and looks at the view.
Slow push in on his face. Warm golden light.
Sound: engine growl fading to wind, birds, a soft acoustic guitar (music).

Most short-form content — a TikTok story, a UGC ad, a teaser — fits in 30 seconds, so you rarely need anything longer. When you do, Framiq's Extra Long mode (Pro plan, experimental) chains several Seedance clips with continuity and stitches them into one video of up to 5 minutes; see long AI videos.

Step 4: Multi-shot prompting inside one clip

You can ask for cuts inside a single generation. The model generally follows explicit shot labels with timings. Name each shot, its framing and what happens:

Shot 1 (0–5s): Close-up of a coffee machine pouring espresso, steam rising.
Shot 2 (5–10s): Medium shot, a man in a white shirt takes the cup and walks to the window.
Shot 3 (10–15s): Over-the-shoulder shot, he looks at a rainy city skyline.
Shot 4 (15–20s): Close-up of his face, he smiles slightly.
Same man in every shot, same white shirt, same apartment. Moody blue morning light.
Sound: espresso machine hiss, rain on the window, soft piano (music).

What helps:

• Restate continuity at the end ("same man, same shirt, same apartment").
• Vary the framing (wide → medium → close-up) so the cuts read as editing, not glitches.
• Four to six shots in 30 seconds is a comfortable maximum. More and each shot becomes a flash.

Step 5: Dialogue and audio notation

Seedance 2.5 generates sound natively. Its prompt notation, as used by BytePlus:

Notation	Meaning	Example
{…}	Spoken dialogue	{We're late again.}
(…)	Music	(upbeat lo-fi beat)
<…>	Sound effect	<door slams>

Always attribute a line and give a delivery word:

Kitchen at night, two flatmates. Medium two-shot, static camera.
Maya whispers: {Did you eat my cheesecake?}
Tom, calmly, without looking up: {Define "eat".}
Maya shouts: {TOM!}
<fridge door closes> (quiet jazz playing from a radio)

Automatic voices in Framiq: if your characters are linked to a cloned voice (made in the Voice tab — Eleven v4 clones from about 10 seconds of clean speech), write @Maya whispers: {…} and turn on the Voices toggle. Framiq voices each line with that character's voice and hands it to Seedance as reference audio; ambience, music and effects are still generated around it. It costs extra and is off by default. A line with no identifiable speaker is not voiced, so always name who speaks.

Step 6: Edit and extend an existing video

Edit: attach an existing video to Seedance 2.5 as a reference and describe the change. The model keeps the motion and composition and re-renders what you asked for:

Same video, same movement and camera. Change her outfit to a black leather jacket
and the background to a snowy street at night. Keep her face unchanged.

Good edits are specific: outfit, background, lighting, style, one object. Asking for everything at once drifts away from the source. More in our guide to editing a video with AI.

Extend: on any Seedance video in your gallery, click Extend. The clip is used as reference and the shot continues. Describe what happens next, not what already happened:

Continue the shot: she reaches the end of the street, turns left and the camera follows
her into a small lit bookshop.

Remember that a reference video costs more than text-to-video, because the input clip is billed as well.

Common Seedance 2.5 mistakes and fixes

Problem	Likely cause	Fix
Face drifts between shots	Too few or inconsistent refs	Use 6–15 clear photos or a scanned @character; restate "keep her face exactly as in the references"
30 s clip feels static	One-line prompt	Write timed beats with camera moves
Cuts look like glitches	Same framing every shot	Alternate wide / medium / close-up
Wrong person speaks	Unattributed dialogue	Name says: {…} plus a delivery word
Music drowns dialogue	Vague sound brief	Put music in (…) and say "quiet", "in the background"
Edit changed too much	Several changes at once	One or two changes per edit, say "same movement and camera"
Request refused	Content policy	Rephrase; avoid real people who have not consented, celebrities and explicit content

Publishing your Seedance clips

Export in the ratio your platform wants — 9:16 for TikTok, Reels and Shorts. TikTok and Instagram require you to label realistic AI-generated content (TikTok has an "AI-generated" toggle when you post); do it, it protects your account. For format ideas see AI video for social media and our AI TikTok trends roundup.

FAQ

Is Seedance 2.5 free?

Seedance 2.5 is a paid API model, but a free Framiq account comes with free credits to start, enough to try it. On the Free plan you pay the standard rate; on Pro (30/month) you pay at API cost. Failed generations are refunded.

How long can a Seedance 2.5 video be?

Up to 30 seconds per clip in one generation (4–30 s in Framiq). For longer videos, Framiq's experimental Extra Long mode chains clips into one video of up to 5 minutes on the Pro plan.

What resolution does Seedance 2.5 output?

Reports about the model mention higher resolutions, but in Framiq it runs at up to 720p.

How many reference images does Seedance 2.5 accept?

Up to 30 reference images in Framiq, plus a reference video and reference audio.

Can Seedance 2.5 generate dialogue and sound?

Yes. Write dialogue in {…} with the speaker and a delivery word, music in (…) and sound effects in <…>. With the Voices toggle, Framiq can voice lines with your characters' cloned voices.

Seedance 2.5 or Kling for my project?

Seedance 2.5 for long single shots, many references and character consistency; Kling 3.0 for some photorealistic short shots. See our Seedance vs Kling comparison and the Seedance 2.5 model page.

Try it

The fastest way to learn Seedance 2.5 is to run the same prompt twice with one change — a camera move, a light, a delivery word — and compare. Open the Framiq studio, pick Seedance 2.5 in the Video tab, paste one of the prompts above and adjust it to your own scene.