Seedance 2.5 Prompt Guide: Multimodal Video Prompts

@ references are inline tags — @Image1, @Video1, @Audio1 — that tell Seedance 2.5 exactly how each uploaded file should be used. Instead of hoping the model guesses, you assign roles: @Image1 for character look, @Video1 for camera motion, @Audio1 for music rhythm. Each tag maps to one slot in your reference list, so a 30-second render can keep character, blocking, and pacing separate in one brief.

Up to fifty mixed references on paid packs — images, video clips, audio beds, and text snippets. You do not need to fill every slot; a focused 15–25 reference brief with clear roles usually outperforms dumping fifty similar assets. Prioritize files that change look, motion, or sound most — hero character plates, one camera-reference clip, and one audio bed before decorative extras.

Open with subject and action, then setting, camera move, lighting, and mood in plain language. State duration (up to 30 seconds) and aspect ratio (16:9, 9:16, 1:1, or 4:3). Close with explicit @ tags that name each reference role — for example: "@Image1 for character look · @Video1 for handheld camera rhythm · @Audio1 for beat-synced cuts." The generator shows credit cost before you submit; tighten the brief rather than stacking vague adjectives.

Use First/Last Frame when you have a starting image (and optionally an ending frame) and mainly need text to describe motion between them — product reveals, simple image-to-video, or locked compositions. Use Universal Reference when the job mixes several asset types: multiple images, reference video for choreography or camera work, audio for rhythm, or a 3D blockout for spatial layout. Universal Reference is the default for multimodal Seedance 2.5 briefs on this guide page.

Upload the source clip, note its @Video label, and say what to borrow in plain language — for example: "Reference @Video1 for the fighting choreography" or "Follow @Video1's camera movements and whip-pan transitions." You can also replace subjects while keeping motion: "Replace the woman in @Video1 with @Image1." Be explicit about edit vs reference; "extend @Video1 by 5 seconds" continues the clip, while "use @Video1 as motion reference only" generates new footage with similar movement.

Yes. Upload a music bed or voice track, tag it @Audio1 (or @Audio2, and so on), and describe how visuals should follow the sound — beat-synced cuts, lip-sync to dialog, or ambient mood. Combine with @Video references when rhythm and camera grammar both matter. Native audio is generated with the clip; attaching @Audio tells the model which external track should drive timing and energy across the full 30-second render.

Export a low-poly, untextured OBJ or FBX from Blender, Cinema 4D, Maya, or SketchUp, upload it as an image/blockout reference, and tag it in the prompt — for example: "Use @Image1 as the 3D blockout layout." Describe materials, lighting, and camera path in text; the blockout only defines geometry and framing. Pair with @Video1 if you need a specific dolly or orbit path through the space. Clean room separation and openings read best on camera.

Upload the clip, reference it as @Video1, and state the extension length in the prompt — for example: "Extend @Video1 by 5 seconds" with generation duration set to match the new segment. For longer chains, reuse the same @Image character and style references across segments so identity and lighting stay aligned. Review each extension before chaining again; very long chains can drift on face, wardrobe, or color grade between segments.

Vague reference roles ("use the video") instead of explicit @ tags; confusing @Image3 with @Image4 in long briefs; uploading similar assets into the same role; forgetting to set duration when extending video; and describing ten scene changes without enough reference support for a 30-second single pass. Fix by naming each file's job, capping scene count to what your references can hold, and re-reading @ mentions before you hit Generate.

Yes — text-to-video works with no uploads. You get the fastest brief setup, but less control over exact character identity, camera grammar, and audio timing than a multimodal job. Add @Image and @Video references when a client plate, storyboard frame, or reference clip already exists. Image-to-video and Universal Reference modes are where Seedance 2.5's 30-second, 50-reference workflow pays off for branded or narrative work.