Seedance 2.5 AI Video Generator

ByteDance's audio-video model, running on MaxH3. Write one prompt and get a 30-second shot sequence with sound already in it — then keep extending, or go back and change one second of it.

Optional
Optional
200

What Seedance 2.5 can do

One-take 30 seconds

Seedance 2.5 generates a full sequence in a single pass. That's long enough for a shot to establish, develop, turn, and land. A singer leaves the dressing room, walks the corridor, meets her dancers, and steps into the arena, all inside one continuous take. Not one moment stretched out. A sequence.It also stays clean on the way out. Tell it no subtitles and no background music and it holds to that, so what lands in your downloads folder doesn't need a stripping pass before you can use it.

Video extension

When one pass isn't the whole story, keep going. The character stays the same character, the room stays the same room, and the pacing carries across the seam. Stack a few rounds and you have minutes of footage that never needed a timeline, a splice, or a transition fix.

Native audio

Sound isn't a second job here. Dialogue, room tone, and effects are produced in the same pass as the frames, so a line lands on the mouth that speaks it and a footfall lands on the foot that makes it. What you download is finished, not a silent draft waiting on a sound pass.

Multimodal references

Most models take a handful of files. Seedance 2.5 takes 30 images, 10 video clips, and 10 audio files at once — and each one can be given a job. In a concert scene, one image fixes the hall, another fixes the pianist, a third fixes the vocalist, and a spread of them fixes the orchestra and the seating behind it. Every assignment holds through the take, faces and voices alike. Start from your own material on the reference to video page.

INPUT

Eight references, one generation

OUTPUT VIDEO

0:25

Clay render control

Hand Seedance 2.5 an untextured 3D pass and it reads the staging out of the geometry: where the camera travels, where the subject stands, how the pose changes across the move. It then lights the scene against that same geometry, so source direction, colour temperature, intensity, and shadow fall the way the space says they should. You set the layout, the model handles the look.

INPUT

Clay pass — staging and motion

Look reference — material and light

Prompt

Reference @Clay 1 for camera movement, composition, shot size, spatial relationships, part positions, model structure, assembly order, and motion paths. Reference @Image 1 for material, lighting, colour, reflections, and mood. Render the clay pass as a high-end photoreal car assembly sequence.

OUTPUT VIDEO

0:07

Green screen replace

This isn't chroma key. Change the background and the subject is re-rendered for the new space rather than cut out and pasted onto it — clothing moves with the wind that scene has, footing adjusts to the ground underneath, and the light falls the way that space would fall on someone standing in it. Name the stretch and name the setting: outdoor drill for the first four seconds, locker room for the six after that.

Timestamp & camera edits

Sometimes the acting is right and the shots aren't. Leave every beat where it is and rewrite only the shot plan, stretch by stretch: a low pass through the opening seconds, a push to the hands through the middle, a rise to overhead for the turn, handheld for the last look. Nothing in front of the lens changes — only where the lens is, and when.

GETTING STARTED

How to use Seedance 2.5 on MaxH3

1
Choose your starting point

A written prompt, a still, or a set of reference files. Starting from a still instead of a description? Head to image to video and upload the frame the shot opens on.

2
Describe the sequence

Write it the way you'd brief a camera operator. Give time ranges when the beats matter: what happens early, what happens at the turn, where it ends.

3
Generate, then refine

Extend it if the story continues, or open the clip and edit the one stretch that isn't right.

Ready to create your first video?

Who Seedance 2.5 is for

  • Ad and brand teams — Build a 30-second spot in one generation, with the product, the packaging, and the talent locked by reference files so they read identically from first frame to last.
  • Directors and previz artists — Turn a clay render into a lit, moving shot. The blocking you built in 3D survives into the output, so what you show a client is the shot you planned, not an approximation of it.
  • Product and e-commerce teams — Shoot the scene once, then change what happens inside a chosen stretch of it. The lighting, the set, and the camera stay put.
  • Social and content creators — Get a finished clip with sound already in it, and keep the same character consistent across a run of posts by carrying the same reference set between generations.
  • Educators and explainer producers — Show a historical scene or a physical process across a full 30 seconds, long enough for cause and effect to actually play out.
  • Simulation and training teams — Generate footage of conditions that are expensive or unsafe to film, using a 3D pass to fix the layout and the camera.

Seedance 2.5 on X: What People Are Making

Seedance 2.5 on YouTube: Reviews and Walkthroughs

One Skincare Product, Multiple AI Ad Variations | Seedance 2.5

Ima Studio

Night Shift | Seedance 2.5 AI Short Film

The Imaginarium

I Made a Hollywood-Style Cinematic Film with Seedance 2.5!

GenAI with Keera

Testing ByteDance’s Seedance 2.5 — Full AI-Generated Short Film

Hawro Lak

FAQs

Seedance 2.5 is ByteDance's audio-video generation model, released in July 2026. It produces up to 30 seconds of footage in a single pass with sound generated alongside the picture, and accepts image, video, and audio files as reference material.

Yes. Audio is produced jointly with the frames rather than added in a second step, which is why dialogue sits on the mouth that speaks it and effects land with the action that makes them.

Thirty images, ten video clips, and ten audio files in one generation. What matters more than the count is that each file can be given a job in the prompt — this face, that voice, this location, that camera behaviour — so a scene with a dozen subjects keeps every one of them straight.

An untextured 3D pass used as a staging guide. Seedance 2.5 reads camera path, subject placement, and motion out of the geometry, then renders the finished look on top — including lighting that follows the source directions the geometry implies. Useful when the blocking has to be exact.

Anything that needs a shot to run long and hold together: ad spots where a product looks identical from first frame to last, short film scenes carried by a single take, product walkthroughs built off a 3D pass, and explainers that show a process rather than describe it.

Not necessarily. Reference files do a lot of the describing for you — instead of writing a paragraph about a face, a location, or a camera move, hand the model the image or clip and let it work from that. The text only has to carry what happens and in what order. If you're starting from nothing but an idea, the text to video page has prompt examples worth borrowing; if you already have a still, image to video covers that starting point too.

Thirty seconds, one prompt, one take

Bring a script, a still, or a stack of references. Seedance 2.5 gives you the whole sequence with sound in it — and lets you change one second of it afterward.