Seedance Explained: Versions, Access and Limits

It is a text-to-video and image-to-video model family built by ByteDance's Seed research group, sold to end users through apps rather than as a standalone product. ByteDance's own page for the family describes output that "can create 1080p videos with smooth motion, rich details, and cinematic aesthetics" and, more interestingly, native support for narrative sequences: it "maintains consistency in the main subject, visual style, and atmosphere across shot transitions" (ByteDance Seed (2026), read 7 September 2026).

Multi-shot consistency is the part worth paying attention to. Most video models are asked for one continuous take; a model that can cut between two shots and keep the same character and light is doing something that used to require a careful pipeline of reference frames.

The naming is confusing and worth getting straight before anything else. The research page still leads with the 1.0 generation, while the versions people can actually run in most products today are 2.0 and 2.5, and the consumer apps use their own product names on top.

People ask how Seedance 2.0 works before they ask what changed: it takes a text prompt or a still image and returns a clip, with the framing and length fixed before the render starts. Seedance 2.5 keeps that shape and moves the limits. The two releases pull in opposite directions, and which one you want depends on whether you need length and framing or you need pixels. These figures come from our own model pages, read on 7 September 2026.

So the newer release is not a straight upgrade. It buys duration, framing freedom and start-and-end frame control at the cost of output resolution, which matters if the clip is going anywhere near a large screen. For a 21:9 establishing shot that runs twenty seconds, take 2.5 and accept 720p. For a short hero clip that has to hold up at 1080p, take the older release. Both pages are linked at the 2.0 model page and the 2.5 model page.

No. Dreamina is an app; the model is an engine inside it. That distinction explains most of the confusion, because people meet the app first and assume the app is the technology, then search for Dreamina AI expecting to find a model page.

Dreamina is operated by CapCut and describes itself as "an all-in-one AI creative platform powered by advanced image and video generation models". Its own model picker names both current releases of the ByteDance video family alongside Sora 2 and Veo 3.1, with Seedream 5.0 Pro, Seedream 5.0 Lite and Nano Banana Pro on the image side (Dreamina (2026), read 7 September 2026).

That model list tells you something useful: the apps are becoming multiplexers. Choosing an app no longer chooses your model, and the interesting question moved from "which product" to "which engine, for this shot".

Through a workbench that has licensed it. If you are searching for a Seedance AI video generator, that is what you are looking for: a product with the engine in its model picker. In ours it sits in the same catalogue as the other video engines and is chosen in the composer before you type anything.

1. 1

   Open the video workbench

   The video generator opens on the Generator tab with a model picker at the top of the composer.

2. 2

   Pick the version, not just the family

   The two releases have different ratio and duration menus, so the choice changes what the panel offers you next.

3. 3

   Set ratio, duration and resolution

   Vertical for social, 21:9 if the newer release is selected and the shot is meant to feel wide.

4. 4

   Write the prompt, then check the price

   The Generate button quotes the clip in credits before the render starts. Video costs an order of magnitude more than a still image, so read it.

5. 5

   Collect the clip from My Video

   Finished renders queue there with their settings attached, which is how you keep track of what produced what.

For effect-led clips rather than open text-to-video, the AI VFX generator wraps the same class of engine in a single-purpose tool, which is usually faster than prompting the effect from scratch.

Each of the big three is optimised around a different constraint, and the honest comparison is per shot rather than overall.

• Length and framing: the newer ByteDance release is the most flexible of the group in the workbench, with one-second duration steps up to 30s and ratios from 21:9 to 9:16.
• Audio: Google states that Veo generates video with sound, including "sound effects, ambient noise, and even dialogue", and lists 1080p and 4K output (Google DeepMind (2026), read 7 September 2026). Nothing in the ByteDance family competes on that.
• Motion realism and control: Kling's own site leads with its 3.0 series, including an Omni variant (Kling (2026)); in our workbench that tier is the one people reach for when a shot needs start and end frames.
• Templates and volume: MiniMax markets its H3 model through Hailuo with a heavy template library (Hailuo (2026)), which is a different way of buying the same outcome.
• General-purpose editing suites: Runway and Pika sit further up the stack, selling the workflow around the model rather than the model itself (Runway (2026), Pika (2026)).

Test them on your own brief before believing any ranking, including this one. A model that wins on a slow interior shot can fall apart on a fast pan, and every published leaderboard is measuring something that may not be your shot.

Six parts, in this order, and no more than about forty words. The family responds better to a clear subject and one camera instruction than to a paragraph of adjectives.

Subject and what it is doing, then setting, then one camera move, then lighting, then style, then what must stay stable across the shot.

• Name one camera move per clip. Two moves in one prompt produce a drift that no amount of rewording fixes.
• Describe the light rather than the mood: low sun from the left beats cinematic.
• Say what has to stay consistent if you are cutting shots together, because the multi-shot capability is instruction-led, not automatic.
• Keep the negative instructions short. Long exclusion lists tend to reintroduce the thing you excluded.

Our full prompt guide covers per-model differences in more depth and is linked at the end of this post rather than repeated here.

Video is priced per clip and the price scales with length and resolution, so the discipline is to decide the shot before you buy it. Our free plan carries 40 credits a month, and paid plans start at $14 a month for 4,500 credits (LazyKiwi pricing (2026), read 7 September 2026).

Two numbers from our own workbench data (August 2026) put that in context. A free five-second template built on this engine was run by 15 external accounts in the window, and 22 of 39 external video jobs came out at 9:16. Short and vertical is what people actually generate, which is worth knowing before you plan a 21:9 project around a credit budget.

One warning on cost control: a rejected clip costs the same as a good one. Storyboard in stills first, at a fraction of the credits, and only move to video once the framing is settled.

By use case, from the runs and the pages above.

• Social vertical clips of five to fifteen seconds: yes, and the newer release is the one to pick for the duration control.
• Wide cinematic framing on a budget: yes at 21:9, as long as 720p is acceptable for the destination.
• Anything that needs synchronised sound: no. Use an engine that generates audio, or add sound in an editor.
• 1080p hero shots: the older release, or a different engine entirely.
• Multi-shot narrative sequences: promising, and the capability ByteDance itself leads with, but write the consistency instructions explicitly.

Pick the engine per shot. The people getting good results are not loyal to one model, they are loyal to one brief.

— Noah Berger

One caveat about this page itself. Model interest decays quickly and version numbers move every few weeks, so this post sits on a short refresh cycle: the figures above carry the date they were read, and when a page changes we re-read it rather than leaving a stale table behind. If you are reading this months later, check the two model pages before trusting the table.