Seedance 2.5 API

Seedance 2.5 is ByteDance's new flagship video model, and its headline is length. Where most video generators still think in five- and ten-second shots, Seedance 2.5 produces up to 30 seconds of video with synchronised sound in a single pass, enough for a scene with a beginning, a turn and an ending rather than a single moving postcard. ByteDance launched it on 31 July 2026, five weeks after previewing it at its Volcano Engine FORCE conference in Beijing.

Length is only half the story. The model also takes far more reference material than before: up to 30 images, 10 video clips and 10 audio clips in one request, 50 files in total. That turns a prompt from a description into a production pack, with character sheets, a location, a camera move to copy and a voice to match. ByteDance pairs this with finer editing: changes pinned to a specific moment in the timeline, green-screen background swaps and camera-angle changes, so a clip with one wrong detail no longer has to be thrown away.

ByteDance describes the goal as moving "from merely generating a clip to completing a creative work." On Unicorn API the model offers text-to-video, image-to-video, first-and-last-frame control and every reference slot, at up to 1080p.

ByteDance Seed and the road from Seedance 2.0

Seedance comes from ByteDance Seed, the company's foundation-model research group. It ships in ByteDance's Jimeng (Dreamina) and Doubao apps, and as Dreamina Seedance through the BytePlus ModelArk API abroad and Volcano Engine in China. The line started with Seedance 1.0 in June 2025 and moved through Seedance 1.5 Pro before the release that made the name famous.

That release was Seedance 2.0, launched on 12 February 2026. It introduced what ByteDance calls a unified multimodal audio-video joint generation architecture: one model that reads text, images, audio and video and writes picture and sound together, for clips of 4 to 15 seconds at 480p or 720p. It accepted up to nine images, three videos and three audio clips as references, and it topped public preference leaderboards within weeks.

It also drew controversy. Clips featuring famous actors and franchise characters spread widely, Disney sent a cease-and-desist letter the day after launch, and Paramount accused ByteDance of infringement. ByteDance's rules on real people tightened: using a real person's portrait as a reference requires identity verification or legal authorisation.

Seedance 2.5 is built on that 2.0 foundation. ByteDance says the new release "centers on foundational generation and reference-based generation," with breakthroughs in long-form storytelling, multimodal reference and editing.

Like every Seedance release, it is closed: no public weights, no self-hosting licence, no parameter count and no technical report for 2.5. The 2.0 paper describes a unified, large-scale architecture for joint audio-video generation, and that is as far as published detail goes. Access is through ByteDance's apps and API, and through partners such as Unicorn.

What changed in Seedance 2.5

Thirty seconds in one take

Seedance 2.0 stopped at 15 seconds. Seedance 2.5 doubles that, and the extra time matters more than the number suggests. A 15-second limit forces anything longer to be stitched from separate generations, and every seam is a chance for a face, a jacket or the light to change. Inside one 30-second pass the model organises several connected shots into an arc of setup, development, turning point and resolution, with the same characters and environment across the cuts.

For longer pieces ByteDance supports multiple rounds of extension and says character, environment and pacing stay consistent across several minutes of output.

Fifty references, including sound

The reference limits jump from 9 images, 3 videos and 3 audio clips to 30 images, 10 videos and 10 audio clips. ByteDance says the model reads composition, style, characters and props across all of them and can preserve the look and voice of several characters at once. In practice that means a two-hander with each actor's face, wardrobe and voice pinned down, or a product film with packshots from every angle.

Blockouts, motion and light

Three kinds of reference get specific upgrades. Clay renders, the untextured 3D blockouts that previs artists use to fix camera position and staging, now control the spatial layout of a shot more tightly. Motion references let a clip copy an action or a camera path from existing footage. And lighting control, ByteDance says, now follows physical rules more faithfully, so light and shadow behave consistently as subjects and camera move.

Editing without starting over

The most practical change is editing. Seedance 2.5 offers timestamp-level control, so an instruction can target a specific stretch of the timeline in both picture and audio. Green-screen editing replaces a background while keeping the subject's reaction to the new environment plausible, including how clothes move and how light falls on them. Camera-perspective edits change the angle of an existing shot, and reference-based edits swap or adjust elements while preserving the subject. The point is to keep a take whose acting, expression or lighting is right and fix only what is wrong.

Sharper picture and sound

Underneath the new features, ByteDance reports gains in image, audio and motion quality, naming object textures, skin and eye detail, lighting and colour saturation specifically.

Seedance 2.5 specs on Unicorn

In image-to-video mode the first frame sets the look and the adaptive aspect ratio follows it. An optional final frame turns a clip into a guided transition between two images. Audio references cannot be combined with first-and-last-frame generation. Editing and extension work through the video reference slots: attach the source clip and describe the change or the continuation.

Two more switches are worth knowing. web_search lets the model look up the prompt's subject first, useful for topical scenes or real products, and return_last_frame hands back the final frame so it can seed the next clip in a longer sequence.

Early results on public arenas

ByteDance's launch post contains no benchmark figures, so the first independent signal came from blind-vote arenas. About two weeks after launch, Arena's video leaderboard placed Seedance 2.5 first in its video editing category with 1,411 points, 23 ahead of the next model. On Design Arena it ranked first in multi-image-to-video, a test that hands models several reference images and judges how well they combine them while keeping subjects, style and continuity, with an Elo of 1,400. Early arena scores move as votes accumulate, but both categories reward exactly what ByteDance built 2.5 to do.

Early testers broadly agreed. The 30-second window drew the most praise because it removes the stitching step rather than just polishing the picture, and the same subject carries across cuts inside one generation. In a small three-scenario comparison against Seedance 2.0, testers found 2.5 better on cinematic movement and character consistency while 2.0 rendered a mechanical hinge motion more cleanly.

Where Seedance 2.5 is strong

Scenes, not shots. A 30-second window lets one prompt cover an establishing shot, an exchange of dialogue and a reaction, with sound running continuously underneath. Short drama, explainer segments and narrative ads fit in a single generation.

Reference-heavy production. With 50 slots, a team can feed the model a full character sheet per actor, a location, props, a reference camera move and voice samples. Brand and product work benefits most, because the model has more angles to hold a product's shape and finish.

Previs-driven cinematography. Directors who block scenes in 3D can hand over a clay render and get a rendered shot with the same staging and camera, instead of describing a dolly move in words and hoping.

Fixing instead of rerolling. Green-screen swaps, timestamped edits and angle changes keep a good performance and change the problem around it.

Limits to know before you build on it

A failed long take is a big redo. Thirty-second clips are powerful, but when one goes wrong the whole scene has to be rendered again or repaired with an edit. Settle blocking on short drafts before committing to full length.

Complex physics and crowds. ByteDance's own launch post flags the physical plausibility of complex motion and the stability of multi-subject interaction as areas that need work. Fights, sports and scenes with several people touching or passing objects are where errors show.

Drift over long clips. More seconds mean more chances for a face, a silhouette, clothing or a product's proportions to wander, and testers reported products changing shape while rotating. References reduce this; they do not remove it.

Real people are restricted. References showing real people may be refused or require authorisation, and testers reported moderation rejecting human reference images in advertising and spokesperson workflows. Plan around AI-generated or properly licensed faces.

No 4K. Output tops out at 1080p, so large-screen deliverables still need upscaling.

Some inputs do not mix. Audio references cannot be combined with first-and-last-frame generation, so a guided transition between two images has to take its sound from the prompt instead of a voice sample.

Prompting tips for Seedance 2.5

Write a shot list, not a caption. Seedance 2.5 responds to structure: say what happens first, what changes and how the scene ends, and give each cut its own sentence with a camera instruction. Put dialogue in quotation marks and say who speaks. When you use references, number them and give each a job.

A 20-second dialogue scene from text:

A small ramen stall at night in the rain. Wide shot: an old cook wipes the counter as a young courier ducks under the awning, soaked. Cut to medium shot: the cook slides a steaming bowl across without a word. The courier says, "I can't pay tonight." The cook replies, "You paid last winter." Close-up on the courier's face as he smiles. Rain, sizzling broth, distant traffic.

A reference-driven product film:

The perfume bottle from image 1 stands on the black stone from image 2. Copy the slow orbit camera move from video 1. Morning light from the left catches the glass and throws a gold caustic onto the stone. Halfway through, petals from image 3 drift into frame. End on a close-up of the cap. Soft piano, no voice.

A green-screen edit using a source clip:

Keep the dancer from video 1 and her movement exactly as they are. Replace the green background with a windswept salt flat at sunset. Her dress and hair react to the wind; warm backlight rims her silhouette.

Draft at 480p and 5 to 10 seconds while you work out blocking, then render the keeper at full length and 720p or 1080p.

Seedance 2.5 against other video models on Unicorn

Seedance 2.0 and 2.0 Fast. Seedance 2.0 and Seedance 2.0 Fast share the same foundation, and run up to 15 seconds at 480p or 720p. For short clips with a handful of references they remain the sensible default; step up to 2.5 when you need 30 seconds, more than nine images, audio references or 1080p.

Kling 3.0. Kling 3.0 is Kuaishou's flagship, with multi-shot prompting and native sound. It is the natural second opinion for cinematic shots, especially when a clip does not need Seedance's length or reference depth.

HappyHorse 1.1. Alibaba's HappyHorse 1.1 Reference To Video is another multi-reference model, worth testing for character-consistency jobs that fit in a short clip.

OmniHuman 1.5. For a single talking presenter driven by a voice track, ByteDance's OmniHuman 1.5 is purpose-built for the job and simpler to drive than a full scene model.

The short version: Seedance 2.5 is the pick when one generation has to carry a whole scene, or when the brief arrives as a folder of references rather than a sentence.

Getting started with Seedance 2.5

Open Seedance 2.5 in the Playground, type a scene or drop in a first frame, and set duration and resolution. For code, create a key on the API keys page and submit a job to /v1/jobs with model set to bytedance/seedance-2-5, a prompt, and optionally mode, image, end_image, duration, resolution, aspect_ratio and the reference arrays. The API tab on this page has complete request examples, and the docs explain how jobs work.

Verdict

Seedance 2.5 moves AI video from the shot to the scene. Thirty seconds in one take, fifty references including voices, clay-render staging and targeted edits add up to the most production-minded model ByteDance has released, and its early arena wins in editing and multi-image generation show the gains are visible to viewers. A failed 30-second take is a big redo, it still struggles with complex physics and crowded interactions, and long clips give drift more room to appear. But for anyone who has spent an afternoon stitching 10-second clips into something coherent, it removes the most tedious step in the process.