Video to Video is how you recreate a winning ad's structure without starting from scratch.
When to use this mode
You found an ad or clip whose structure works, and you want yours to follow the same flow
You want to keep a video's camera movement, pacing, or transitions but change what's in it
You want to extend an existing clip, or swap one element for another
How it works
You upload a reference video, then @‑mention it in your prompt and tell the model what to copy and what to change. The model follows the reference's camera movement, editing rhythm, choreography, transitions, or VFX while you swap in your own subject.
This mode is at its best on Seedance 2.0, the only model that takes video references directly. You can trim and pan to select the exact portion of the reference clip you want before tagging it.
Steps
Be in video mode on a model that supports video references (Seedance 2.0)
Upload your reference video
(Optional) Upload a product image or other asset to swap in
In your prompt,
@‑mention the reference video and state plainly that your video should follow its formatSet your settings and click Generate
A worked example
Say you find a video of a whiskey glass sliding in from offscreen and stopping dead‑center in frame. You want the exact same motion, but with a kids' water bottle in a different setting
Attach the whiskey‑glass video as your reference video
Write a prompt like:
Follow @Video1's camera movement and timing exactly: an object slides in from the left and stops centered in frame. Replace the whiskey glass with a colorful kids' water bottle. Change the setting to a bright kitchen counter. Keep the same pacing and final framing.
Click Generate. The output keeps the slide‑in motion and the centered stop, but swaps the object and the scene
The @‑mention patterns
Give one job per asset. Be explicit about which video supplies the camera, which image is the new subject, and which audio sets the mood.
Goal | Prompt pattern |
Copy the camera work |
|
Borrow the motion or choreography |
|
Swap a subject |
|
Extend a clip |
|
Add a soundtrack |
|
Extending a clip
When you extend, set the generation duration to the new segment only, not the combined runtime. If you're adding 5 seconds to a clip, set duration to 5 seconds, not the total length.
Recreating an ad's structure
For a full ad recreation, give the agent a sample ad as the reference. It learns the structure, pacing, and style, then rebuilds that flow using your product and creator images. This is the fastest way to turn a proven format into a version that features your brand.
Get the best result
One job per asset. "Reference @Video1 for camera movement", don't ask one clip to do three things
State what to keep and what to change, explicitly. The model copies what you point at and changes what you name
Protect your references in priority order when you hit the asset limit: motion reference first, then subject consistency, then mood
Use "one continuous take, no cuts" if you want a single unbroken shot
Trim the reference to just the portion whose motion you actually want
Common mistakes
Asking one asset to do everything. Separate the camera reference from the subject from the audio
Being vague about the swap. Name exactly what changes — "replace the whiskey glass with a kids' bottle" — and what stays
Setting the wrong duration on an extend. Use the length of the new segment, not the combined total
Over‑constraining. Copy the structure, but don't pile on so many changes that the reference stops helping
Credits
The cost is shown next to the Generate button before you commit, and scales with model × duration × resolution, plus the references used. Failed generations are refunded automatically. Test with one output before generating variations.