How to use Dreamina Seedance 2.0 (multimodal AI video)

Most video models take one kind of input — a source image, or a text prompt. Seedance 2.0 accepts image, video, audio, and text references together in a single call, and lets your prompt point at each one by name: @Image1, @Video1, @Audio1. The model pulls the specific feature you're asking for from each reference — a person's likeness from an image, an action from a video, a voice from an audio clip — rather than blending everything into an average.

That reference-and-point syntax is the core skill this model rewards. A prompt that says "the girl from @Image1" and "use @Video1 as the action reference" tells the model exactly which pixels or motion to reproduce from which source, the same way HappyHorse reference-to-video numbers its images — except here the references can be video and audio too, and the model reproduces textures, timbres, camera moves, and visual-effect styles from them, not just appearance.

Point the model at an existing video with a reference and describe exactly what should change — a subject replacement, an object-level edit, or an inpainting fix. "Replace all dog food packaging in @Video1 with the packaging shown in @Image1" is a complete instruction: it names the source video, the element to replace, and the reference to replace it with.

Precision comes from constraining the instruction. State plainly what must NOT change — camera movement, lighting, people, pacing — so the model treats everything else as fixed and only touches the named element. The source examples for this model explicitly spell out what stays untouched, which is worth copying: over-specifying what shouldn't move protects the rest of the shot from drifting.

Extension takes a finished clip and grows it — continuing the action forward from where the source video ends, generating a preceding scene that leads into it, or interpolating between two separate clips so the transition between them is smooth rather than a hard cut.

Describe the extension as a continuous transformation rather than a jump cut: "continue @Video1 with a smooth transition to a leaf's first-person perspective... while the polar bear remains unchanged" tells the model both what changes (the perspective and environment) and what must hold steady (the bear), which is what keeps an extended sequence feeling like one shot instead of two spliced together.

These are three different jobs and the model handles each differently. Full generation (multimodal reference-to-video) builds a new scene from your references. Style transfer applies a look — a color grade, an art style — across a whole clip. Local replacement swaps one specific element while leaving everything else untouched. Word your instruction to match: name the whole scene for generation, the look itself for style transfer, or the specific object and its replacement for local replacement.

Mixing jobs in one instruction — asking for a new scene and a style change and an object swap all at once — is harder for the model to satisfy cleanly than running them as separate, focused passes.

A referenced feature doesn't show up: the prompt named the reference (@Image1, @Video1) without describing what to take from it. Add a plain-language description the first time you introduce each reference.

The edit changed more than intended: the instruction didn't state what should stay fixed. Explicitly name the camera movement, lighting, people, and pacing that must remain unchanged, the way the model's own local-replacement examples do.

An extension feels like two clips stitched together rather than one: the prompt described a jump between states instead of a continuous transformation. Describe the change as something that happens gradually, and name what stays constant across it.

Resolution option isn't available: the fast tier caps at 720P and mini at 720P as well — move to the standard tier if you need 1080P or 4K.

Costs are adding up faster than expected: check which tier you're running iterations on. Mini is roughly half the price of standard, and fast excludes-video-input pricing is cheaper still — move trial-and-error to the cheapest tier that still shows you what you need to see.

Reach for Seedance 2.0 when your shot genuinely needs more than one kind of source material — a person's likeness from a photo plus an action from a video plus a voice from audio, in one generation. If you only have one image and want it animated, a simpler image-to-video tool is less to configure for the same result.

The video editing and extension modes are worth knowing even if you generated your source clip elsewhere: you don't need to have generated a video with Seedance to edit or extend it with Seedance afterward, which makes it a useful second-pass tool for fixing or growing a clip from any source.