Seedance 2.5

What is Seedance 2.5?

Seedance 2.5 is ByteDance's newest video generation model, and the headline change is duration: it produces up to 30 seconds of video in a single generation, double what Seedance 2.0 allowed. It runs in three modes — text-to-video, image-to-video and reference-to-video — generates synchronized audio natively, and accepts reference sets of up to 30 images, 10 videos and 10 audio files at once.

That duration jump sounds incremental and is not. Almost every AI video workflow in production today is really a stitching workflow: generate five seconds, generate another five, hope the character's face and the lighting survive the cut, assemble in an editor. Seedance 2.5 removes the seams for anything under half a minute. One generation, one continuous take, one consistent audio bed.

Resolution runs 480p, 720p or 1080p, and the rate climbs steeply with it — 1080p costs two and a half times 720p, because the provider meters by pixels rather than by seconds. For a long take that adds up fast, so it is often cheaper to finish at 720p and run the result through our video upscaler than to pay the 1080p rate for the full 30 seconds.

What Changed From Seedance 2.0

Seedance 2.0 is still on the platform and still the right pick for plenty of jobs. Three differences decide which one you should reach for:

Read that pricing row carefully, because it cuts both ways. At 720p, Seedance 2.5 is nearly twice the per-second cost of 2.0 — a 10-second 720p clip is 600 credits against 350. But at 480p it is actually cheaper than 2.0 at 30 credits per second, which makes it the best drafting model in the family. Iterate at 480p until the prompt is right, then run the keeper at 720p, and reserve 1080p for shots that genuinely need to hold up on a big screen.

Pricing & Specifications

Seedance 2.5 bills per second of output at 30 credits per second at 480p, 60 at 720p and 150 at 1080p. The tiers are not evenly spaced because the provider prices by pixels: a second of 1080p carries more than twice the pixels of 720p and costs more per pixel on top. There is no minimum charge beyond the 4-second floor and no per-generation fee.

One billing detail that catches people out: reference videos are charged too. Seedance 2.5 meters the runtime of any reference clip you attach alongside the output, because the model reads that footage as input. Attaching a video does drop the per-second rate to 36 credits at 720p, though, so a 10-second generation with 8 seconds of reference video is billed as 18 seconds at the lower rate — 648 credits rather than 600 for the clip on its own. Reference images and reference audio are not metered.

Supported aspect ratios are 21:9, 16:9, 4:3, 1:1, 3:4 and 9:16, plus an “auto” option that lets the model choose based on your prompt and references.

Six things thirty seconds unlocks

Duration limits shape what people even attempt. Here is work that was awkward or impossible at 15 seconds and is straightforward now.

1. A complete ad spot, not a fragment of one

Thirty seconds is the standard broadcast and pre-roll slot, and it is no accident that it is also the length agencies brief against. Previously you built a 30-second spot from six generations and prayed the grade matched. Now the whole thing arrives as one file with a continuous audio bed. If you need structure — hook, demonstration, benefit, call to action — write those as four timed beats in the prompt and the model will pace them.

2. Process and instruction video

Anything that teaches a sequence — a recipe, an assembly, a technique, a repair — falls apart when cut into five-second chunks, because the whole value is watching one continuous action. A 30-second unbroken shot of hands completing a task, with synchronized foley, is genuinely useful content rather than a stylish loop.

3. Property and space walkthroughs

Real estate, hospitality, retail and venue marketing all live on continuous movement through space. Viewers read a cut as a hidden edit, so a stitched walkthrough of a room actively undermines trust in what it shows. One unbroken 30-second glide through a floor plan reads as a real place.

4. A full verse for music and lyric video

A verse or a chorus is typically 15 to 30 seconds. Generate against the beat structure in your prompt, attach the audio as a reference so the visual rhythm keys off the real track, and you get a section you can drop into a longer edit without fighting continuity at every bar.

5. Episodic content with a consistent cast

The 30-image reference set matters more than it sounds for serial content. Build a reference library for a character — face from multiple angles, wardrobe, lighting, environment — and reuse it across every episode. Nine images was enough to suggest a character; thirty is enough to pin one down.

6. Establishing shots and B-roll beds

Editors do not want five-second B-roll; they want thirty seconds they can trim anywhere. Long atmospheric plates — a city waking up, weather crossing a landscape, a factory floor running — are cheap to generate at 480p, and give an edit room actual coverage to work with.

Prompting for long-form

The mistake almost everyone makes on their first 30-second generation is reusing a five-second prompt. A single-sentence description gives the model one idea and 30 seconds to fill, and what comes back drifts, loops or slows to nothing. Long generations need structure.

Write beats, with timings

Break the runtime into three to five beats and give each one a duration, a camera move and an action: “0–8s: slow push in on the workbench as hands lay out tools. 8–18s: camera rises and tracks left following the assembly. 18–30s: pull back to a wide as the finished piece is lifted into the light.” This is the highest-leverage change you can make.

State what must not change

Over 30 seconds there is far more room for drift than over five. Spell out the constants — subject appearance, wardrobe, time of day, colour palette, lens character, whether the audio bed is continuous. A line like “same woman in the red wool coat throughout, overcast daylight, muted palette, no cuts” prevents the most common failure.

Say “single continuous shot” if that is what you want

Left to its own devices, the model will sometimes introduce cuts across a long generation. If you want an unbroken take, ask for one in those words. If you want cuts, describe them as cuts and the model will place them at your beats.

Describe the audio

Native audio is generated from your prompt. Naming the sonic elements — “grinder, chatter, door chime, no music” — produces a markedly better bed than leaving it to inference.

Draft at 480p

At 30 credits per second, 480p is the cheapest way to test whether a long prompt holds together. Prompt structure, pacing and continuity all read at 480p; only fine detail does not. Lock the prompt, then regenerate the keeper at 720p.

How It Compares

Seedance 2.5 does not replace anything on the platform. It owns one axis — length and reference depth — and gives up ground on others.

vs. Seedance 2.0: Use 2.0 for short 720p clips, where its flat 35 credits per second beats 2.5's 60. Use 2.5 when you need more than 15 seconds, a reference set larger than 9 images, or a cheap 480p draft.

vs. VEO 3.1: VEO leads on resolution and on naturalistic, documentary-feeling footage, and it is the better choice for a polished 8-second hero shot. It cannot give you 30 continuous seconds, and it has nothing comparable to a 50-file reference set.

vs. Sora 2: Sora 2 remains the value pick for exploring ideas in short clips. Once a project needs long takes or tight visual consistency against references, that cost advantage stops being the deciding factor.

vs. Kling and the O3 family: Kling is stronger on stylized and anime-leaning work and offers motion-control editing that Seedance does not. They complement each other — Kling for stylized shorts, Seedance 2.5 for long-form realism.

Our full video model comparison goes deeper across every model on the platform, and the Seedance 2.0 guide covers the earlier generation in detail.

Frequently Asked Questions