ByteDance releases Seedance 2.5

Background

Seedance is the name of ByteDance's video generation model family. Seedance 2.5 is a version in that line. ByteDance released it in July 2026, and the date is given as 31 July.^[1] Video generation models create moving footage from written instructions and reference material. Seedance 2.5 is described as a next-generation audio-video joint generation model.^[2]

A video is a sequence of many frames, so keeping a long clip consistent from start to finish is harder than making a short one. Extending a clip in steps and editing only the part that needs changing are common ways to manage that difficulty. These are general techniques in video generation, and they show what a user can do with the controls.

What happened

Joint audio-video generation means the model produces sound and picture in one pass, so the two are created together rather than adding sound to finished footage afterwards.^[2]

Key details

In a single pass the model generates up to 30 seconds of video, and that clip can be extended twice.^[2][1] A beta long-video mode reaches 3 minutes, which allows much longer sequences than a single pass.^[1] The model accepts up to 50 multimodal references. A multimodal reference is an input of a different type, such as text, an image, an audio file or a video clip, that guides what the output should look like.^[1]

The model also supports local editing. A user can change part of a scene without regenerating the whole clip, which avoids redoing work that was already correct.^[1]

Impact

Seedance 2.5 is offered through an API.^[2] Developers can call the model from their own software and workflows, rather than using only a standalone application. Its longer clips, many reference inputs and partial editing are all features that work on the finished video rather than on a single generated clip.