ByteDance / Seedance 2.0 Reference

## Overview Seedance 2.0 R2V is a ByteDance video generation model built for reference-driven creation. It uses a unified audio-video generation setup, so it can generate video and synced audio together. You provide reference videos, plus optional images and audio, and it produces a new clip that follows the motion, camera language, and style cues you describe. This is useful when you need consistent movement and shot pacing, but you still want a brand-new scene. ## What you can build - Motion-guided ads where your output follows the pacing of an existing product clip - Storyboard-driven short scenes, including multi-shot cuts inside one clip - Style transfer from a reference reel to a new setting, subject, and lighting - Beat-synced clips by guiding visuals with a short audio reference - Variations of the same action with different wardrobes, props, or environments ## Inputs - 1 to 3 reference videos in MP4 format. Each reference clip should be 2 to 15 seconds. Some deployments also enforce tight resolution and file-size caps for reference video. - An optional text prompt that describes the scene, action, camera movement, and style. You can refer to your uploaded assets as [Video 1], [Image 1], and so on. - An optional first-frame image to set the opening frame for image-to-video style control. - An optional last-frame image to steer how the clip ends. This requires a first-frame image. - Optional reference images (up to 9) to lock style, characters, props, or environment details. Refer to them in the prompt as [Image 1], [Image 2], etc. - Optional reference audio clips (WAV or MP3, up to 3). Audio reference typically requires at least one reference image or a first or last frame image. - A choice of output resolution (commonly 480p or 720p; some integrations also offer 1080p output). - A choice of output aspect ratio. You can set a fixed ratio like 16:9 or 9:16, or let the model adapt based on your inputs. - A choice of output duration in seconds. - A toggle to include generated audio that matches the visuals. - An optional watermark toggle, when offered by the hosting platform. - An optional numeric seed to reduce randomness. Setting the seed to 0 typically means “random”. ## Outputs The model returns a generated MP4 video file. - The video can include a synced audio track when audio generation is enabled. - Audio can include elements like ambient sound, sound effects, and speech-like voiceover when prompted. - The response typically includes the downloadable video plus basic file metadata (such as file name and size), and the seed used. ## Limitations - Output length is short-form. Many public deployments generate up to 15 seconds per run. - Reference inputs are capped. Common limits are up to 3 videos, up to 9 images, and up to 3 audio clips, with a total cap across all uploaded assets. - Reference video constraints can be strict. Some providers require MP4 or MOV, a combined reference duration window, and specific resolution ranges. - If your reference media is heavily compressed, low light, shaky, or full of fast cuts, the output may drift or add artifacts. - Some deployments block or reject inputs that contain human faces, even if the request is otherwise benign. - A fixed seed can improve repeatability, but results may still vary slightly across runs. ## Safety & compliance - Don’t upload images or videos that include human faces if your deployment forbids them. - Don’t try to generate content that imitates real people or public figures. - Respect copyright and trademark rights. Don’t prompt the model to recreate protected characters, logos, or scenes. - Avoid disallowed content, including sexual content, graphic violence, and harassment.