Seedance 2.5 Video Generator
Thirty seconds in one continuous take, with sound generated alongside the picture. Below is what the spec sheets leave out: what a clip really costs to run, what breaks, and how to prompt around it. Zebracat turns those takes into a fully edited, publish-ready video of up to 5 minutes.
.svg.webp){kind=logo}
.svg)
.svg)
.svg)
.webp)
.svg)
.webp)
.webp)
.webp)
.svg)
The model at a glance
Videos generated with this model
Every clip below was generated in Zebracat with the exact prompt shown. No cherry-picking, no post-production.
What is this model?
Seedance 2.5 is ByteDance's video generation model, released in August 2026. The headline is length: 30 seconds in one continuous generation instead of 15, extendable twice on top of that, with sound produced in the same pass and up to 50 reference files in a single request. What the marketing pages leave out is the rest of the sheet. Output is locked to 24 fps and there is no frame-rate parameter. Resolution runs 480p, 720p and 1080p, and 1080p comes back 10-bit HEVC. There is no seed on 2.5, so there is no such thing as a reproducible re-roll: run the same prompt twice and you get two different videos. Photographs of real people are refused as references outright. Audio-only input, unusually, is accepted.
Nothing here asks you to choose a model. Write the script or hand over an idea, and the scenes that must run unbroken, or must keep one person looking like one person, get assigned to Seedance with the bracket syntax below already written for you. What lands in the editor afterwards is a video with voice, captions and music, sized for the channel you are publishing to.
What thirty seconds actually costs
Every site reselling Seedance hides this behind its own credits, so here is the arithmetic. ByteDance bills the model on tokens: (input seconds + output seconds) x width x height x 24, divided by 1024. Cost therefore scales linearly with both length and pixels, with no volume discount and no per-call overhead anywhere in the chain. At ByteDance's published rates a single 30-second 16:9 generation works out to roughly $3 at 480p, $7 at 720p and $17 at 1080p. A 5-second test is exactly one sixth of that. Resellers sit between the same price and double it for the identical model.
Three things follow, and they shape how Zebracat uses the model. A 30-second take is not a slightly larger 5-second take, it is six of them. There is no partial regeneration, so one bad second costs the entire clip, and with no seed you cannot re-roll the good version back. And passing a reference video is cheaper than it looks: the reference-video token rate is around 40 percent lower, so a 30-second clip built from a short reference can bill less than the same clip generated from text alone. Long takes are worth buying where continuity genuinely carries the scene, and wasteful everywhere else, which is the whole argument for routing scene by scene rather than picking one model for a video.
What our test runs show
- It deletes transitions and jumps to the finished state. Ask to see a character from behind and it can arrive at the rear view without ever performing the turn. The same failure shows up as a lean-in-to-whisper coming back as a kiss, on every re-roll, because the model recognises the destination and skips the movement. Seedance 2.0 handled this unprompted. Write how the scene gets there, not just where it ends.
- Time budgets fail in both directions. Too many actions inside a window returns something that looks fast-forwarded; too much time on a small action returns unintentional slow motion. The ratio that holds is roughly fifteen words of action per four seconds, and no more than three or four beats in any fifteen.
- An untagged reference still gets used. Upload an image without saying what it controls and the model will use it anyway, however it likes. Every reference needs one job and an explicit exclusion, and the exclusion does more work than the instruction.
- Fifty references is a ceiling, not a target. Conflicting references get resolved by averaging, and averaging is what produces the soft, generic look people blame on AI in general. Four to six well-labelled inputs beat twenty.
- Transformation shots are a regression. Anything assembling, morphing or reconfiguring came out cleaner on Seedance 2.0, where parts read as parts. On 2.5 they merge. Those beats do not go to this model.
- The sound effects are better than the speech. Mechanical and environmental audio is genuinely strong. Generated dialogue drifts phonetically, turning a clean line into a nearly-right one. We take the picture from Seedance and the voice from Zebracat.
What this model is best at, and where it falls short
Best at
- Thirty seconds without a cut. Continuity holds across the whole take, which removes the stitch points where multi-clip sequences usually give themselves away.
- Keeping one person one person. Identity, wardrobe and lighting survive a long shot when references are labelled properly.
- Ambient and mechanical sound. Engines, weather, rooms, machinery. This is the strongest part of the model and it arrives with the picture.
- Direction on a clock. Second-level timestamps let a single generation carry three beats and a camera move rather than one static idea.
- Building from a reference clip. Motion, framing and cinematic language transfer from a short video, and the reference path bills at a lower rate than pure text-to-video.
- Fixing a shot instead of rebuilding it. Backgrounds, camera angles and individual elements can be edited after the fact, including green-screen replacement.
Not good at
- Fast motion. Sprints, fights and fast hand work warp mid-move, and this is the most consistently reported weakness of the model.
- Transitions between states. It will jump to the finished pose rather than perform the movement unless you write the movement itself.
- Transformations and assembly. Parts merge instead of moving. Seedance 2.0 does this better.
- Words in frame. Text rendering is poor and did not improve over 2.0. Titles and captions belong in the editor.
- Repeatability. No seed parameter means you cannot reproduce a generation you liked, and no partial regeneration means one bad second costs the full clip.
- Complex physics and several subjects interacting. ByteDance names both in its own release notes.
- Photographs of real people. Uploading a real face as a reference is refused.
- 4K. The model tops out at 1080p. Anything sold as 4K Seedance is an upscale applied afterwards.
Which version of this model you get in Zebracat
Seedance 2.5 (default in Zebracat)
The current model, and the one worth paying for when a shot has to run long. 4 to 30 seconds in a single generation, two rounds of extension, native audio, up to 50 references, timestamp-level direction, white-model and green-screen control.
Seedance 2.0, and why people still argue for it
Half the length at 15 seconds, 12 reference files instead of 50, and around a third cheaper per second. It is not simply obsolete. Experienced users report it handling transformation and assembly shots more cleanly than 2.5, and a recurring criticism of 2.5 is that the work done to support 30-second takes cost the model something elsewhere. Where a scene is short and the motion is complicated, 2.0 is often the better spend. Routing treats that as a per-scene question rather than a loyalty test.
Seedance 2.0 Mini and Fast
Cheaper, quicker cuts of 2.0 for drafts and for b-roll that plays under a voiceover. Lower detail, shorter clips.
Seedance 1.5 Pro and 1.0
The original line. Short clips, no audio-video joint generation. Listed for completeness.
A note on "Seedance 4K"
ByteDance's own API documents exactly three output modes for 2.5: 480p, 720p and 1080p, the last of them 10-bit HEVC, all at a fixed 24 fps. There is no 4K mode and no frame-rate setting. Several of the pages ranking for this model advertise 4K anyway, and at least one advertises a 180-second beta and 60 fps that appear in no ByteDance document. Upscaling a 1080p render to 4K afterwards is a legitimate thing to do; calling it a 4K model is not.
This model inside a full video pipeline, not a download button
Most Seedance sites end at the download button. A 30-second take is a scene, not a video, and the work between those two things is the part that takes the time.
What the long take is for
Zebracat reads your script and looks for the scenes that break when they are stitched: a product demonstration, a walk-and-talk, anything where a face or a wardrobe has to survive from first frame to last. Those go to Seedance. Sound-led shots and product close-ups go to Veo 3, emotional dialogue to Kling 3.0, stills to Nano Banana. Each prompt is rewritten for the model that runs it, references labelled and timestamps included.
What happens around it
Voiceover in 170+ languages or your cloned voice, captions synced word by word, music, and an editor where any single scene can be replaced or rerun on a different model without disturbing the others. Videos run up to 5 minutes and export at 9:16, 16:9 or 1:1, or go onto TikTok, YouTube and Instagram on a schedule.
How to prompt this model
Seedance takes direction in a syntax of its own, and because there is no seed and no partial regeneration, a prompt that is nearly right is a prompt you pay for twice. Zebracat writes these for you. They are documented here because they are the difference between a take you keep and a take you rerun.
The syntax
- Dialogue goes in curly braces. {We open at six.} Name the language first for anything not in English.
- Sound effects go in angle brackets, written as physical events rather than moods. <a kitchen scale clicking as something is set on it> beats "kitchen sounds".
- Music goes in round brackets, on-screen text in corner brackets. (slow piano, no percussion) and 【Chapter One】, though the second is best avoided entirely given how the model renders letters.
- References are numbered in upload order and addressed inline. @Image1, @Video1, @Audio1 through ByteDance and most hosts. A few resellers document square brackets instead, so check before assuming.
The rules that decide the take
- Write the movement, not the destination. "She turns around" is an instruction. "Now seen from behind" is an invitation to skip the turn and morph into the finished pose.
- Size each beat to what happens in it. Around fifteen words of action per four seconds, three or four beats per fifteen. Overfill a window and the model fast-forwards; underfill one and it stretches into slow motion.
- Front-load and back-load. The opening and closing of a prompt carry more weight than the middle. References and camera go first, prohibitions go last.
- Give every reference one job and one exclusion. @Image1 defines her face and hair. Do not use the background from @Image1. Anything you leave untagged still gets used, on the model's terms rather than yours.
- Say what the shot does when it cannot obey. On product work, the escape hatch matters: if a shot would require the label to change, the shot changes instead. Without that line the model redraws the product.
- Restate anything critical in three places. A rule that appears once, such as whether a person speaks on camera or in voiceover, holds about half the time. Repeat it in the style block, in every beat and in the constraints.
- Ask for the flaws. Seedance defaults to clean and steady, so handheld realism has to be named: the frame drifting, focus hunting, exposure fighting the light and losing.
- Write numbers as numbers. Counts drift first. Four chairs stays four chairs only if you say four.
- Bridge cuts with sound. When a sequence spans several generations, end one and open the next on the same audio cue and the same framing. The edit joins on the sound rather than on a jump.
This model vs other AI video models
Spec-level comparison of the models you can run in Zebracat. Every model here is selectable in the same editor.
1,000+ 5-Star Reviews from Creators, Marketers & Makers
.webp)
.webp)
.webp)
.webp)
.webp)
.webp)
.webp)
.webp)
.webp)
.webp)
.webp)
.webp)
Questions about this model, answered
What resolution does Seedance 2.5 actually output?
480p, 720p or 1080p, at a fixed 24 fps, with 1080p encoded as 10-bit HEVC. Those three modes are what ByteDance's API documents and prices. There is no 4K mode and no frame-rate setting, so a page offering 4K Seedance is upscaling the render afterwards.
What does one Seedance 2.5 clip cost to run?
Billing is per token, calculated from duration times pixels times 24 fps, which makes cost linear in both length and resolution. At ByteDance's published rates a 30-second 16:9 generation is roughly $3 at 480p, $7 at 720p and $17 at 1080p, and a 5-second test is one sixth of that. Resellers charge between the same and double. Inside Zebracat it is 40 credits per clip.
Can I get the same video twice?
No. Seedance 2.5 exposes no seed parameter, so identical prompts produce different videos and a generation you liked cannot be recovered. It is the strongest argument for locking style and framing on short test runs before committing to a long take.
Can I upload a photo of a real person as a reference?
No. Photographs of real people are refused as reference inputs. The supported route is to build a character from generated stills first, then use those as your reference set.
How long can one take be, and what happens after 30 seconds?
One generation runs 4 to 30 seconds and can be extended twice. In Zebracat that is a single scene. Several takes are cut to your script into a finished video of up to 5 minutes with voiceover, captions and music.
How many references should I actually use?
The ceiling is 50, made up of 30 images, 10 video clips and 10 audio clips. The useful number is far lower. References compete, and the model settles disagreements by averaging them, which is what produces a generic result. Four to six, each with a stated job and a stated exclusion, outperforms twenty.
Is Seedance 2.5 better than Seedance 2.0?
For length and continuity, clearly. For everything else it is genuinely contested. 2.0 is cheaper per second and handles transformation and assembly shots more cleanly, and a real share of experienced users still prefer it for short, motion-heavy work. Treat it as a choice per scene, not per project.
How good is the dialogue and lip sync?
Serviceable rather than excellent. The sound effects are the strong half of the audio: engines, weather, rooms and machinery are convincing. Generated speech drifts phonetically, which is why Zebracat takes the picture from Seedance and generates the voiceover separately, so one voice runs the whole video.
How long does a generation take?
Minutes, not seconds, and a 30-second take costs more of them than a short one. Times swing noticeably with load, which is why scenes render alongside each other rather than in a queue.
Can I use Seedance 2.5 videos commercially?
Yes. Work generated on a paid Zebracat plan carries full commercial rights, ads and client delivery included.
Is the Seedance 2.5 video generator free?
No. Seedance 2.5 is a premium model and needs a paid plan. Script writing, AI voices and the lower-tier image models are open on the free plan without a card, so you can walk the workflow first.
Why run Seedance 2.5 in Zebracat instead of on a single-model site?
Because a 30-second take is the most expensive thing on the menu and most scenes do not need one. A single-model site sells the same answer to every shot. Zebracat spends Seedance where continuity carries the scene, spends something cheaper everywhere else, and hands back the finished video rather than the footage.
Other AI video models in Zebracat
One editor, every model. Zebracat routes each scene to the model that fits, or you pick one yourself.