Seedance 2.5

All models
Video modelByteDance

The hotly-anticipated next generation of the world's #1-ranked AI video model.

Overview

Seedance 2.5 is the most anticipated release in AI video — the expected next generation of ByteDance's Seedance line, which in its Seedance 2.0 version became the #1-ranked video generator in the world, topping the Artificial Analysis Video Arena for both text-to-video and image-to-video in early 2026. ByteDance has not published official Seedance 2.5 specifications as of mid-2026, but industry reporting, leaks, and the clear 1.0 → 1.5 → 2.0 trajectory point to a major leap forward.

Reports and leaks suggest Seedance 2.5 will push well past the current ceiling: up to 4K resolution (versus 720p/1080p today), near-real-time generation, clips longer than 15 seconds, persistent characters that stay consistent across sessions without re-uploading references, and expanded multimodal inputs — many more reference images, audio, and video clips in a single pass.

What is certain is the foundation Seedance 2.5 builds on — and you can use it on Hedra today through Seedance 2.0: native, synchronized audio generated jointly with the video, accurate multi-language lip-sync, multi-shot storytelling that holds characters and style across cuts, precise camera control, and a reference mode for grounding generations in your own images, video, and audio. Seedance 2.5 is expected to extend every one of these — and the moment it ships, Hedra will have it.

A note on transparency: Seedance 2.5 is not yet released, and the capabilities, resolutions, and durations described here are our best estimates based on the public data available so far. They are not confirmed by ByteDance and may change at launch.

Seedance examples

Similar models

Prompt tips

These proven Seedance 2.0 techniques are expected to carry directly into Seedance 2.5:

  • Chunk your prompts: Break the scene into 2–3 second beats, each specifying shot direction, camera movement, and subject action, rather than one broad description.

  • Tag your references: When uploading multiple references, assign clear roles (character, background, motion style) so the model knows how to use each.

  • Audio-driven pacing: Provide a music or voiceover reference and the model will match cuts, pacing, and transitions to the beat.

  • Lock characters with grids: Feed a multi-angle image grid of your subject as the primary reference to hold character consistency across a multi-shot sequence.