Seedance 2.5 AI Video Model

Overview

Seedance 2.5 is ByteDance's next-generation video model. It generates up to 30 seconds of native single-shot video, takes up to 50 multimodal references, edits and extends clips you already have, and adds native audio with support for 10+ languages, building on the Seedance 2.0 line rather than replacing it. It is available on Morphic now.

What Seedance 2.5 brings

The headline is length: up to 30 seconds of native single-shot video, well past the few-second clips most models produce today, which opens the door to full ad spots, short drama, and longer previsualization in a single pass. Alongside that, Seedance 2.5 takes up to 50 multimodal reference inputs per generation (30 images, 10 video clips, and 10 audio), far more than Seedance 2.0's 9 images and 3 audio/video clips. It also adds an audio-only reference that can drive pacing and lip-sync, and native support for 10+ languages. In-clip editing gets two new controls: white-model control, which blocks a shot with untextured 3D geometry before the model lights it, and green-screen editing, which swaps a background while the subject holds.

Editing and extending a clip

Seedance 2.5 works on footage you already have, not just on a blank prompt. Both routes run from the same reference-led generation, and each is a task type you can pin so the model does not read your intent as a fresh render.

An edit is region-level: attach the clip, name the change, and the rest of the frame holds. That covers swapping a subject, adding or removing an object, redrawing part of the frame, and replacing the music or voice. Because the output is anchored to the source, it inherits the clip's aspect ratio and length rather than letting you set them, and the clip you attach has to run 4 to 30 seconds.

An extend adds footage instead of changing it. Seedance 2.5 can continue a clip past its last frame, generate the moment before its first frame, or build a transition that bridges two clips, holding the character, rhythm, and style across the join. An extend keeps the source's aspect ratio, but unlike an edit it lets you set how many seconds of new footage to add. The practical split is simple: reach for edit when the shot is right and one element is wrong, and extend when the shot is right and too short.

Native 1080p output

Seedance 2.5 renders natively at 1080p, a full 1920x1080 frame generated directly rather than upscaled from a smaller render, up from the 720p it shipped with at launch. That makes it a full-HD delivery model: clips land ready for social, web, and ad pipelines without a separate upscale step, with cleaner edges and texture than a 720p render stretched to fit. The rest of the toolkit carries over, so the full 30-second single take, native audio, and up to 50 references all come through at 1080p in one pass.

Quality and realism

Seedance 2.5 also pushes visual quality closer to real footage. Lighting follows physical behavior, with more natural eye catchlights and consistent light direction and falloff. Camera moves (pushes, pulls, pans, and tilts) hold steady across a full 30-second long take, without obvious shake or frame skipping. Performances read as continuous rather than keyframed, with believable micro-expressions, and the overall live-action feel is stronger, with a reduced uncanny-valley effect in skin texture and hair.

Where Seedance 2.5 sits

Seedance 2.5 sits at the top of the Seedance line, above Seedance 2.0 and Seedance 2.0 4K, and above the lighter Seedance 2.0 Mini and Fast tiers. It keeps the Seedance 2.0 toolkit, multimodal input, native audio, and AI camera control, and extends it with longer clips and more references. ByteDance may follow with Pro or Turbo style variants, as it has with earlier releases.

Seedance 2.5 and Morphic

Seedance 2.5 is available on Morphic now, alongside the rest of the Seedance line and other video models like Kling and Veo. Pick it in the prompt bar, write your shot as a brief, attach any references, and generate. Every take lands on the Canvas, so you can run the same brief through Seedance 2.5 and another model and compare before committing.