HiAPI
Este artigo ainda não está disponível no seu idioma; mostramos o original.
Blogue

Guide

Seedance 2.5 Flash Reference-to-Video: Reference Images, Audio and How the Bill Is Counted

Nov 17, 2026 9 min read
Stacked sheets receding into shallow depth of field, cream and graphite with one terracotta sheet as accent.
Modelo
seedance-2.5-flash
Nesta página

What this model does

seedance-2.5-flash/reference-to-video is ByteDance's reference-driven video model. You supply reference images and reference audio, address them in order with @image1 and @audio1 in the prompt, and the model generates a new clip. Images pin down what a person, product or place looks like; audio sets the rhythm, the music or a voice-over.

Constraints on reference material

  • Images: up to 10, jpeg/png/webp/bmp/tiff/gif, under 30 MB each, 300–6,000 px per side, aspect ratio 0.4–2.5. SVG, HEIC and HEIF are not supported.
  • Audio: up to 5 clips, wav or mp3, under 15 MB, 2–30 seconds each, 30 seconds combined. Audio can be used on its own, without images.

At least one of the two kinds must be present, or the request is invalid.

How billing works

This route is billed on the seconds of video it delivers. Reference images and audio do not add billable seconds.

  • 480p: $0.0679/s — about $0.34 for a 5-second clip.
  • 720p: $0.1525/s — about $0.76 for a 5-second clip.

A fixed duration from 4 to 30 is billed as requested. With -1 (automatic), you pay for the delivered clip's length rounded to the nearest second, with a 4-second minimum. If you need a predictable bill, pass a fixed duration.

Request parameters

  • prompt (required): how to use the material, addressing it in order as @image1 and @audio1. Up to 5,000 characters.
  • reference_image_urls / reference_audio_urls: arrays of URLs for the reference images and audio.
  • duration: an integer from 4 to 30, or -1 for automatic. Defaults to 5.
  • resolution: 480p or 720p. Defaults to 480p.
  • aspect_ratio: 16:9, 4:3, 1:1, 3:4, 9:16, 21:9, or adaptive to follow the references. Defaults to adaptive.
  • omni_reference_task_type: always reference (the default, so you can omit it).
  • output_format: mp4 (default) or mov. generate_audio: defaults to true and produces synchronized dialogue, sound effects and background audio.

Writing the prompt

Number the materials in the order you supply them, then say what each one contributes and what the final shot should look like. For example:

The product from @image1 rotates slowly on a wooden table in front of the
morning window from @image2. The camera pushes in from a medium shot to a
close-up, cutting on the beat of @audio1. Keep the metallic highlights intact.

With several images, say which one is the subject and which sets the scene or style; that is far more reliable than "use these images". With audio, say which part of the action should land on which part of the track.

Common problems

  • Rejected for missing material: send at least one reference image or audio clip.
  • An image will not upload: check the format, that each side is 300–6,000 px, and that the aspect ratio is between 0.4 and 2.5.
  • Audio is rejected: each clip must be 2–30 seconds and under 15 MB, and all clips together no more than 30 seconds.
  • The charge is not what you expected: check whether you used -1. Automatic duration is billed on the delivered clip.

Before you ship

Run one request with a single image at 480p for 5 seconds to confirm the integration and the charge. Then run one with audio or several images and check that the clip uses each item the way the prompt numbers it. Keep both as a baseline for later prompt or material changes. The model page has the current parameters and prices.

Artigos relacionados

Ver todos os artigos →

Comece a criar com artefactos de IA persistentes

Obtenha a sua chave de API e chame modelos de imagem, vídeo e áudio através de uma única API baseada em tasks.

Ler a documentação da API