Seedance 2.5 takes 50 reference items in one request: 30 images, 10 videos and 10 audio clips. The total is the part everyone quotes. The part that decides success is how the budget is shaped: how long each clip runs, which media types stand alone, and the one category the model refuses. The how-to guide covers the browser route.
Where the 50 comes from
The ceiling is new with 2.5. The 2.0 series tops out at 15 items and makes audio and video share a 15-second runtime.
| Reference budget | Seedance 2.5 | Seedance 2.0 series |
|---|---|---|
| Images | 30 | 9 |
| Video clips | 10 | 3 |
| Audio clips | 10 | 3 |
| Audio and video length | 30 seconds | 15 seconds |
| Items in one request | 50 | 15 |
Three rows moved, and output did not: the reference ceiling and the 4-to-30-second output window match for unrelated reasons. The generation comparison covers the rest.
The images: 30, each within 4K
The prompt guide gives the allowance as 最多 30 张 4K 分辨率以内的图片, up to 30 images each within 4K. That is the entire published specification: a count and a resolution ceiling. No file-size cap appears beside it, so upload validation that rejects an oversized image is your policy, not the vendor’s.
Two softer numbers are recommendations rather than refusals. As subject references, 1 to 8 subjects work well; 9 to 12 is possible but stability drops and you may need to reroll. As references for a video edit, 1 to 5 images work well and 6 to 8 is less stable. Note that the first counts subjects, not files, so one image holding four objects is four subjects.
The videos: a 30-second tape budget
Up to 10 clips may total 30 seconds of runtime, a budget across the whole request rather than per clip. The per-clip floor depends on the task: outside an edit a clip runs 2 to 30 seconds, while a video edit raises the floor to 4. A three-second clip that is fine as a motion reference is too short to edit.
The file has published limits too: aspect ratio between 0.4 and 2.5, shorter side between 300 and 6000 pixels, total pixel count between 407,696 and 8,295,044, single file under 200 MB, frame rate between 24 and 60, container mp4 or mov. Those come from the create-task reference page.
The audio: 10 clips, and audio alone
Audio matches video on quantity, at 10 clips and 30 seconds total. The difference is that 2.5 is the first generation that accepts audio on its own.
The capability table marks audio reference supported on 2.5 and 需搭配图片/视频 on every 2.0 variant, so older models reject audio arriving without an image or video beside it. On 2.5 a lone clip is valid, which Volcengine calls 纯音频参考生成视频. Audio arrives three ways: a public URL, a Base64 string in the form data:audio/wav;base64, or a registered identifier written as asset://. The API guide shows where that entry sits in the request.
Real person faces are refused
This constraint stops projects rather than slowing them. Volcengine states that the Seedance 2.5 and Seedance 2.0 series models do not support uploading a reference image or video containing a real human face. A portrait photograph of a colleague is a rejected input.
Three official routes remain. Trust a model-generated result as the input material, meaning have Seedance produce a face first and reference that output. Use a preset virtual human. Or use an authorised real-person asset, which is what the asset:// identifier exists for.
What is not published is the verification workflow: no stated turnaround, no list of accepted identity documents, no claim that any route passes. A project that depends on a named person’s likeness is a process question for Volcengine before any code exists.
Worked examples: two reference prompts
Both prompts are the vendor’s own, quoted in full so they can be pasted into a request.
Example one: one image, six videos
The fruit-cookie advert from the 2.5 tutorial sends one image as reference_image and six clips as reference_video, with omni_reference_task_type set to reference.
明亮多彩的广告片风格，果味饼干为主角，包含草莓、苹果、葡萄、橙子四种口味，草莓味参考@图像1，饼干与对应水果以强秩序感的几何阵列排布，整体画面干净、高级、节奏强。开场水果快速建立视觉聚焦，参考@视频1的构图，音乐重拍切入。随后不同口味饼干整齐排列，切特写，参考@视频2的动态和运镜。高潮段一块饼干被折断，瞬间进入慢动作，果味夹心爆开，碎屑飞溅，果汁感与颗粒冲击被放大展示，参考@视频3的冲击感。横向阵列，形成节奏抛物感，参考@视频4的运动，突出秩序美感与产品丰富度。随后迅速回到快节奏剪辑。结尾英文文字 One bite of crispness, a heart full of delight 快速分词切换入画，配合强节奏文字运动与产品定格，参考@视频5，最终品牌感收束，饼干和水果向四周发散，参考@视频6画面充满年轻、活力、好吃、想分享的广告氛围。The division of labour is the point: one image carries product identity, six clips carry camera behaviour.
Example two: five images, five roles
The second prompt opens with the technique itself: five images named and given a job before the story starts.
素材绑定：故事板分镜@图片1，卧室@图片2，李天@图片3，李倩@图片4，《快快乐乐》书籍@图片5。
镜头一：【全景固定镜头，平视三分构图】冬夜落雪房间，落地窗前，男子双手插裤兜侧立望向窗外飞雪，少女站在身侧静静看向男子，氛围安静克制，雪花持续落在玻璃窗。
镜头二：【中景过肩镜头】少女后背作为前景，男子转头温和看向少女，少女微微低头沉默，窗外落雪持续。
镜头三：【中近景，对角线构图】男子捧着书籍《快快乐乐》缓缓递出，少女抬手接过书本。
镜头四：【少女面部近景，中心构图】少女把书本抱紧在胸口，眼眶泛红，泪珠缓缓滑落，神色感伤。
镜头五：【男子面部近景，斜向构图】男子温柔浅笑，静静注视落泪少女，眼底带着怅然。
镜头六：【全景固定镜头】少女转身缓步走出画面，窗前只剩男子独自伫立，双手插兜望向漫天风雪，房间空旷寂静。Two habits transfer. Naming each image by role stops the model averaging five pictures into one composite, the standard failure when references disagree. And the binding line sits first, so the model meets the roles before the story.
What the vendor has not published
Not published on the official pages read for this review: a single-file size cap for reference images; an accepted image format list beside the 30-image count; any size or format limit for audio clips; any stability percentage, since the only wording is 稳定性下降、可能需要抽卡; and the identity-verification workflow behind the real-person rule. The counts can be planned against; the file-level rules cannot, so treat them as your policy.
Frequently asked questions
How many images, videos and audio clips fit in one Seedance 2.5 request? Fifty: 30 images, 10 videos and 10 audio clips. The 2.0 series and its fast and mini variants take 15.
Can I send audio with no image or video at all? Yes on Seedance 2.5, which the capability table lists as supported for audio reference. Every 2.0 variant needs an image or video alongside it.
Can I use a photo of a real person as a reference? No. The 2.5 and 2.0 series models do not accept reference images or videos containing a real human face. Use a model-generated subject, a preset virtual human, or an authorised asset.
Is there a file-size limit for a reference image? The vendor has not published one. The documented ceiling is resolution, at 4K per image, and the reference-video limit is a separate 200 MB per file.
Every number here comes from the vendor’s pages, and the gaps are named rather than filled.