Seedance 2.5

What it is

Seedance 2.5 is a video generation model from ByteDance, released in July 2026; Wikipedia records the release date as 31 July 2026.^[1] ByteDance describes it as a next-generation audio-video joint generation model and the successor to Seedance 2.0.^[2]

"Audio-video joint generation" means the model is designed to produce picture and sound as one output rather than making a silent clip first and adding sound afterwards. That matters because the two halves of a clip usually have to agree: an object falling, a door closing or a person speaking has to look and sound like a single event. Building both in one model is a way of getting that agreement, instead of hoping a separately generated soundtrack lines up afterwards.

Where it comes from

Seedance 2.5 continues a line of ByteDance video models, taking the place of Seedance 2.0 as the newer generation.^[2] ByteDance's official page for the model offers three entry points — Try now, Get API and Compare now — alongside a curated showcase of example outputs, so that people can watch what the model produces before deciding whether to use it.^[2]

Main features

The most visible change is length. The model generates up to 30 seconds of video in a single pass, and a generation can be extended twice to build longer stories; a beta long-video mode reaches 3 minutes.^[2][1] Longer single takes are useful for scenes that need continuity, such as a conversation or a slow camera move, which tend to break apart when they are stitched from many short clips.

The model also takes input from many directions at once. It accepts up to 50 multimodal references and reads reference videos more precisely, including framing and cinematic language.^[1][2] In practice this means a user can point at existing material — images, clips, styles — and expect the output to follow the shot sizes, angles and visual conventions of what was supplied, rather than only borrowing the general subject.

Editing is the other area ByteDance emphasises. Seedance 2.5 supports local editing, changing specific parts of a scene without regenerating the whole clip, along with controls such as green-screen editing, professional camera movement and performance blocking.^[1][2] Working locally is cheaper and more predictable than rerolling an entire shot because one detail is wrong, and it lets a clip be adjusted after the rest of it has already been approved. ByteDance says editing in Seedance 2.5 is more reliable and responds to a wider range of audio and visual editing requests.^[2]

The model is available through an API and is also offered by third-party API providers.^[2][1]

What ByteDance claims

ByteDance's own account is qualitative. The company says Seedance 2.5 has smoother, more consistent motion and more realistic visuals, and that it goes beyond motion transfer in reference videos into creative interpretation — following not just what a reference does, but what it is going for.^[2] No metrics are published alongside these claims, so they cannot be checked against a benchmark the way a reported score could; for now they describe the direction of the work rather than a measured result.

Where it sits

Generation models are usually judged on three things: how long a usable clip they can make, how precisely they can be directed, and how much of an existing shot can be repaired without starting over. Seedance 2.5 addresses all three — minutes-long output, dozens of references, and local edits — and packages them behind an API so that other products can build on it.^[1][2] The absence of published numbers makes comparison with rival systems difficult, and the length limits still mean that anything feature-length has to be assembled from parts.