Seedance 2.0 R2V (Reference-to-Video)

Overview

This page runs ByteDance's Seedance 2.0, a joint audio and video model, with a reference video as the required input. From reference clips it can take subject, camera movement, action and style; from images, character looks and composition; from audio, voice timbre, melody or spoken lines.

The same inputs handle edits and extensions. An edit names what to replace, add or remove in [Video 1], and anything the prompt leaves out stays as it was. An extension continues a clip forward or backward, or bridges two or three clips with new transitions.

Choosing a variant: within its 2.0 series, ByteDance recommends Seedance 2.0 for the highest generation quality, and in this family only it renders 4K. Pick it over Fast R2V when a reference-driven clip needs 1080p, 4K or the best quality the series offers. Seedance 2.0 Fast R2V takes identical reference inputs on a lighter model capped at 720p, and Seedance 1.5 Pro and Seedance 1.0 Pro need no reference clip.

What you can build

• Motion transfer: keep a source clip's pour, spin or gait and swap the subject and setting.
• Subject swaps in existing footage, such as replacing the animal in [Video 1] with the one in [Image 1].
• 4K masters for film or HDR work, using the 10-bit output.

Inputs

• Reference videos (required): 1 to 3 MP4 files, each at least 2 seconds long. At most 15 seconds of footage is used in total, so a long clip is cut short and a later one can be left out.
• An optional prompt that refers to files by type and number, like [Audio 1].
• Up to 9 reference images. The first and last frame slots are ignored here, since ByteDance does not mix frame mode with reference videos. To set an opening or closing shot, name a reference image as the first or last frame in the prompt.
• Up to 3 WAV or MP3 audio clips of 2 to 15 seconds, used only together with an image input.
• Resolution 480p, 720p (default), 1080p or 4K. Aspect ratio adaptive (default) or fixed, from 21:9 to 9:16. Duration 4 to 15 seconds, 10 by default.
• A seed (0 for random), plus switches for watermark (off) and generated audio (on).
• Prompt enhancement (on by default) rewrites the prompt in brand-free terms so content filters reject fewer runs.

Outputs

The result is a 24 fps MP4 in 8-bit color up to 1080p, with mono sound when audio is on. 4K (3840×2160 at 16:9) is 10-bit H.265/HEVC, which some browsers cannot play; ByteDance lists VLC and mpv among players that can.

Recommended settings

• Write longer clips as a Shot 1, Shot 2, Shot 3 storyboard in event order; ByteDance says exact timestamps are followed unreliably.
• For edits and extensions, write [Video 1] directly; "reference [Video 1]" can turn the job into a reference task.
• Prefer slow, continuous movement. ByteDance advises against sprints, big jumps and rolls.
• Define each subject by two or three stable traits, such as clothing or hairstyle, and reuse one label for it.

Limitations

• ByteDance supports seeds on its 1.x models only, so a seed may not repeat a result here.
• Extended clips can jump or briefly roll back where new footage meets the original.

Safety & compliance

ByteDance does not accept reference images or videos showing real human faces for the 2.0 series, so such uploads can be rejected. Reference audio with a real person's voice needs that person's consent.