ByteDance / seedance 2.5 reference

## Overview Seedance 2.5 Reference is a video generation model by ByteDance that builds a new clip from your reference assets. You feed it reference videos, images, and optional audio, then describe what to keep and what to change. The model copies motion, pacing, and composition cues from your references while it re-renders new subjects and scene details. It outputs a single finished video file, with optional audio that matches the visuals. ## What you can build - Product swap ads that keep the same camera move from a reference clip - Motion transfer, like reusing a pour, dance, or gesture trajectory in a new scene - Style-controlled shots that match a brand look using a reference image set - Video edits that preserve timing while changing props, wardrobe, or background - Music or narration-timed clips guided by short audio references - Short story beats with consistent characters anchored by reference images ## Inputs - One to ten reference video clips in MP4 or MOV format, each 2–30 seconds long, with a 30-second total budget across all clips - One to thirty reference images that set identity, wardrobe, props, environments, or style; some requests with human faces may be rejected - One to ten reference audio clips in WAV or MP3 format, each 2–30 seconds long, with a 30-second total budget across all audio - An optional natural-language prompt that explains the scene and calls out which reference to follow, using bracketed tags like [Video 1], [Image 1], and [Audio 1] - A required output resolution choice between 480p and 720p - A required output duration from 4 to 30 seconds - An optional aspect ratio choice, or an adaptive mode that lets the model pick; if you include a reference video, the output keeps that video’s aspect ratio - A required choice to include synchronized audio in the output, or generate a silent clip - An optional output container choice between MP4 and MOV - An optional prompt rewrite step that removes brand names and risky audio cues to reduce content filter failures - An optional watermark toggle for the exported video ## Outputs - One generated video file in MP4 or MOV format at the requested resolution, aspect ratio, and duration - If audio is enabled, the video includes a synchronized audio track that can contain dialogue, sound effects, and ambience - A generation seed value may be returned for repeatable iterations ## Recommended settings - Use MP4 when you need maximum playback compatibility across browsers and social platforms - Use MOV when you plan to color grade, key, composite, or repeatedly extend and re-edit outputs - Turn on the prompt rewrite step when you enable audio and your prompt mentions brands, copyrighted works, or recognizable lyrics - For motion transfer, keep your most important reference clip first, then describe exactly what must stay unchanged ## Limitations - Output length tops out at 30 seconds per generation - Reference video clips must be short, and their combined duration cannot exceed 30 seconds - Reference audio clips must be short, and their combined duration cannot exceed 30 seconds - Output resolution is limited to 480p or 720p on this model surface - When you supply a reference video, the output aspect ratio follows that video, even if you request another ratio - Providers may reject prompts or references that contain human faces, especially for identity-like edits - Low-quality inputs can reduce reference fidelity, including heavy compression, motion blur, low light, shaky footage, watermarks, and inconsistent frame rates ## Safety & compliance - Upload only reference media you own or have rights to use, including music and voice recordings - Get clear consent before using a real person’s face or voice as a reference - Avoid generating deceptive content, impersonation, or misleading “real event” footage - Expect content filters on disallowed content and on prompts that request copyrighted characters, logos, or protected audio - Don’t attempt to bypass safety systems or moderation checks