Seedance 2.5

Overview

The reference-to-video pipeline of Seedance 2.5 is designed for video generation tasks that need source materials to constrain visual consistency. Developers submit a prompt along with at least one reference video, and may also attach image and audio references, so the model can reference the materials in the order @image1, @video1, @audio1 and generate a 4- to 30-second clip at 480p to 1080p. It suits advertising, short-drama, and brand content teams that need to lock down the appearance of characters, products, and scenes across serialized creation, and is also suited to narrative segment production that requires synchronized sound effects and dialogue.

Core Highlights

• Multimodal materials jointly constrain the visuals — supports up to 30 images, 10 videos, and 10 audio clips as simultaneous references, used to lock down the appearance of products, characters, and scenes.
• Prompts reference materials in order — in the prompt, refer to reference items as @image1, @video1, @audio1 in sequence, clearly mapping materials to visual goals.
• Controllable duration and aspect ratio — output duration can be set from 4 to 30 seconds or left to the model to decide automatically; aspect ratio supports 16:9, 9:16, 21:9, and adaptive following the source material.
• Synchronized audio and video generation — by default generates dialogue, sound effects, and background audio synchronized with the visuals, and can be disabled for pure visual output.

Capabilities

Generation driven by reference video: input at least one reference video, and the model generates a new segment based on it rather than extending or editing the original; suited to productions that need to reuse existing shot language and visual texture.

Mixed image and video references: up to 30 images and 10 videos can be submitted in combination to separately constrain characters, products, and scenes, reducing appearance drift across serial content.

Audio references and audio generation: can attach up to 10 wav or mp3 audio references (must be submitted together with at least one reference video), and choose whether to generate synchronized dialogue, sound effects, and background sound.

Resolution and output format options: output supports three tiers—480p, 720p, and 1080p—and formats can be mp4 or mov for professional post-production; billing is by output seconds.

Web search to supplement information: web_search can be enabled so the model retrieves time-sensitive information involved in the prompt before generation, suited to content that needs to cite current facts.

Use Cases

• Advertising and brand teams producing product clips — use product images and live-action reference videos to constrain appearance and generate 4- to 30-second ad segments for campaign assets and social media distribution.
• Short-drama teams maintaining character consistency — use character images and existing clips as references to batch-generate continuous shots of the same character for storyboard assembly and later editing.
• Post-production teams previewing shot plans — submit scene reference videos and composition images to quickly obtain preview clips close to final quality, for client approval and pre-shoot communication.
• Content teams producing audio-video synchronized segments — provide voiceover or music references and enable audio generation to produce clips with synchronized dialogue and sound effects, used directly as final output or for further mixing.

Limitations and Selection Advice

A single generation is at most 30 seconds; longer narratives need to be split into multiple segments generated separately and then stitched together. Reference materials have explicit constraints: videos must be 2 to 30 seconds and total no more than 30 seconds, audio total duration likewise no more than 30 seconds, and individual file size and pixel count also have lower and upper bounds; non-compliant materials will fail directly. This pipeline is for reference-guided generation and does not support video editing or extending the original clip; multi-material referencing also means higher prompt-writing and parameter-tuning costs. For lightweight, single-shot needs without consistency requirements, consider the text-to-video pipeline in the same series.