Seedance 2.5 accepts up to 30 images, 10 video clips, and 10 audio clips in one generation—about 50 total multimodal references when all three categories are filled. The Seedance 2.5 reference limits apply separately to each input type. Cite attached materials in your prompt with labels such as [Image1], [Video1], and [Audio1].
Last verified: 2026-10-23
How many references does Seedance 2.5 accept?
The published limits describe how many reference materials you can supply in a single generation request. Images, video clips, and audio clips have separate allowances:
| Input type | Limit | How to cite in prompt |
|---|---|---|
| Images | Up to 30 images | [Image1], [Image2], and so on |
| Video | Up to 10 video clips | [Video1], [Video2], and so on |
| Audio | Up to 10 audio clips | [Audio1], [Audio2], and so on |
| Combined multimodal references | Up to 50 across the three categories | Use the matching label for each attached reference |
ByteDance’s announcement describes these inputs as reference materials accepted “in a single pass.” The product page lists the same combined allowance, and the Replicate readme repeats the image, video, and audio limits.
These are maximum counts, not a required upload checklist. You can plan a request around fewer materials. A scene with one subject and a simple action may need only a small reference set; a more complex scene may benefit from several distinct visual and audio instructions.
For a broader introduction before preparing your assets, see what is Seedance 2.5?.
Does the total mean any combination of 50 files?
The documented allowance is 30 images plus 10 video clips plus 10 audio clips. It should not be interpreted as a pool of interchangeable slots.
For example, 30 images and 10 video clips remain within their respective count limits. Adding another image would exceed the image allowance, even if you had attached no audio. Unused audio capacity does not establish permission to upload extra images.
This distinction matters when preparing a storyboard or collecting character references. Count each category separately before uploading, rather than checking only the total number of files.
The reference count also does not establish file-size limits, supported formats, or maximum clip lengths. Those are separate upload requirements. Check the interface or API you are using for those details; the published 30/10/10 specification alone does not answer them.
What does a multimodal reference contribute?
A multimodal reference set combines different kinds of source material within one request. The prompt should explain which aspect of each attachment matters.
As a planning approach, assign each reference a specific job:
- Images: identify appearance details, clothing, objects, locations, or composition you want the prompt to describe.
- Video clips: identify an action, movement pattern, or camera movement you want to reference.
- Audio clips: identify the sound, delivery, or timing characteristics you want to guide the scene.
These are suggested ways to organize instructions, not guarantees that every detail will transfer exactly. Merely attaching a file leaves room for ambiguity about its intended role.
For instance, a video may contain both a useful camera movement and an unwanted background. Specify that the camera movement is the relevant feature, and describe the intended background separately. The reference-to-video guide provides a related workflow for turning source materials into a generation brief.
How should you cite Seedance 2.5 references in a prompt?
Use the reference labels to connect each instruction to an attachment. The documented notation is [Image1], [Video1], and [Audio1], with numbering identifying individual references.
A short reference map can make the prompt easier to write and review:
| Reference | Intended role | Example instruction |
|---|---|---|
[Image1] |
Main subject | Use the person’s appearance and jacket from [Image1]. |
[Image2] |
Location | Set the scene in the room shown in [Image2]. |
[Video1] |
Camera movement | Reference the slow forward camera movement in [Video1]. |
[Audio1] |
Sound atmosphere | Use [Audio1] as the reference for the room’s ambient sound. |
Here is an illustrative prompt for those four attachments:
Create a scene of the person from [Image1] standing beside a window in the room from [Image2]. Preserve the person’s jacket and hairstyle. The person turns toward the window and pauses. Use the slow forward camera movement from [Video1], keeping the subject visible throughout. Use [Audio1] as the ambient sound reference. Keep the lighting soft and consistent.
This is a suggested prompt structure, not a tested output claim. Before generating, check that each label points to the intended attachment and that the requested actions are compatible.
Should you use all 30 image references?
Use additional images when they contribute distinct information. Repeated or conflicting references can make your instructions harder to interpret, even when the upload count is valid.
For example, several images might show different views of a product. Explain which features should remain consistent. If two images show different clothing or settings, state which one governs the final scene.
A practical preparation sequence is:
- Define the subject, action, location, and camera behavior.
- Select references that support those decisions.
- Assign each attachment a clear role.
- Remove references whose purpose you cannot explain.
- Check the image, video, and audio counts separately.
If your task begins with a still image, read the image-to-video guide. If you are planning a transition between defined opening and closing compositions, see the first-and-last-frame guide.
For scenes with several people, label each person’s references explicitly and describe who performs each action. The multi-character prompting guide covers that related planning problem.
Do reference limits determine output resolution?
Reference capacity and output resolution are separate settings. The Seedance 2.5 product page lists 480p / 720p / 1080p output options. Attaching more references does not, by itself, select a higher output resolution.
Availability can also differ by implementation. As of the verification date, the Replicate API lists 480p and 720p, while the product page lists the three options above. Check the generation interface you are using before planning a specific output size.
For an initial trial, the offer is free credits on sign-up, enough for one video. Prepare a focused reference set and a clear prompt so that the first generation answers a useful creative question.
FAQ
How many references can Seedance 2.5 use in one generation?
Up to 30 images, 10 video clips, and 10 audio clips, for a combined maximum of 50 references. Each category has its own limit.
Can I use more than 30 images if I upload no audio?
The documented image maximum remains 30. The specification does not describe unused video or audio allowances as transferable image slots.
How do I name image, video, and audio references?
Use labels such as [Image1], [Video1], and [Audio1]. Connect each label to a clear instruction about the attached material’s intended role.
Must I upload images, video, and audio together?
The published figures are upper limits, not a requirement to fill every category. Choose the reference types relevant to your scene and check the available input controls.
Does adding more references guarantee a better result?
No quality guarantee follows from the reference count. Clear roles and compatible instructions make your request easier to interpret; filling every available slot is not itself a quality strategy.
Sources
- Official announcement: https://seed.bytedance.com/en/blog/one-take-creation-flexible-referencing-introducing-seedance-2-5
- Product specifications: https://seedance2video.io/seedance-2-5
- Model readme mirroring the official reference limits: https://replicate.com/bytedance/seedance-2.5
