Finding signal on Twitter is more difficult than it used to be. We curate the best tweets on topics like AI, startups, and product development every weekday so you can focus on what matters.
Originally published by @runwayml on X. Tech Twitter preserves the original source alongside this readable edition.
Before you dive in
• Master Seedance 2.0 in Runway: combine text, image, video and audio inputs simultaneously for advanced video generation with multi-reference flexibility and...
Best for builders who want practical takeaways. 6 min read.
Most video models take a single input and generate from there. Seedance 2.0 accepts up to 12 files across four input types — text, image, video and audio — simultaneously. That means you can lock a character with an image, borrow motion from a clip, set the mood with audio and direct the scene with text, all in a single generation.
This guide leads with examples and use cases, then breaks down input types and prompt strategies so you can dive in right away or read through for the full picture.
Prompt techniques & use cases
Multi-shot
Generate multi-shot sequences. We recommend using 10 or 15 second durations for the best results.
You can simply prompt for a “multi-shot sequence,” or describe each shot for more control.
Continuous shot
Animate a storyboard image
Upload a reference image with multiple panels to instantly bring it to life. These images can be created in Runway by using the Ad Concepter and Story Panels Apps.
Apply camera motion from a video to an image
Tip: Rather than supply individual images of your character or object at different angles, use a single character/object image for 360 consistency without consuming multiple reference slots.
Overview of creating with Seedance 2.0
Seedance 2.0 has three core creation modes, each suited to different starting points and goals.
Multi-Reference gives you the most flexibility. Combine images, video clips and audio files to direct different dimensions of your generation simultaneously — locking a character, borrowing motion or setting a mood or pace from sound.
Image to Video (including Start/End Frames) uses one or more images to anchor your generation. Drop in a starting frame, an ending frame or both. Your prompt handles what happens in between.
Text to Video generates entirely from your prompt. Describe what you want through text and the model builds it from scratch.
Understanding prompt balance
We doomscroll, you upskill
Get the 10 tweets shaping how builders think today.
Newsletter
We doomscroll, you upskill
Get the 10 tweets shaping how builders think today.
No matter which mode you're working in, great results come down to a balance: control vs. creative freedom.
Control is how precisely you define what you want through text, images, video or audio.
Creative freedom is the space you leave for the model to work.
Constrain too much and the result can fall apart. Leave too much open and it may not match your intent. The best prompts sit somewhere in the middle.
Multi-References
When working with multiple references, think about the role each one is playing and include that in your prompt:
Apply the motion of Video 1 to Image 1.
Replace the character in Video 1 with the character from Image 1.
Similarly, you may also wish to emphasize what should or should not change between references:
The character from Image 1 speaks the dialogue from Audio 1 with the same emotional delivery.
The number of ways you can combine and apply these inputs is part of what makes Seedance 2.0 so flexible. The sections below break down how each input type works and what you can do with it.
Reference video
A reference video can conceptually be used in two distinct ways, and your prompt should reflect which one you're doing:
Transfer: You want to borrow motion, pacing or camera language from the clip and apply it to something new. Let the video do that work — prompt only for what should change (subject, environment, style).
Edit: You want to change, extend or build on the clip. Prompt for what's new or different; the model keeps everything else intact.
Below are different simple examples of how you might leverage video inputs:
[Transfer — Apply motion] Apply the character motion from Video 1 to the character in Image 1
[Edit — Extend video] Continue Video 1 from the last frame
[Edit — Precede video] Create a sequence preceding Video 1 that ends on the first frame of the reference
[Edit — Interpolate 2 videos] Using Video 1 as the beginning and Video 2 as the end, generate a natural transition between the two references
Reference image
A reference image can be used in two high-level ways:
Inspiration: The model borrows the mood, palette or aesthetic of the image but interprets it freely. Good for style direction without strict constraints.
Transfer: A specific element (a character, object or composition) is carried directly into the output. The model treats it as a visual anchor, not a loose reference.
Below are prompt examples that demonstrate both usage methods:
[Inspiration — Restyle] Restyle Video 1 using the color palette of Image 1
[Transfer — Object placement] Add the soda from Image 1 to the table in Video 1
Reference audio
Use audio to drive speech, rhythm, tone, or atmosphere.
[Editing pace] A multi-shot sequence of Image 1 cut to the beat of Audio 1
[Speech] The character from Image 1 speaks the dialogue from Audio 1 with the same emotional delivery
Start/End Frames
Start/End frames use an image to specify the exact start and (optional) end point of your video.
Your image handles composition, color and visual identity. Your prompt picks up from there.
Start frame only: Describe what changes — motion, atmosphere, how the scene evolves.
Both frames: Describe the transition between the two images, not the images themselves.
Text to Video
No reference image means your text prompt sets the scene entirely. Cover subject, action, environment, lighting, camera movement and style. Think of it like a shot brief.
To learn more about prompting, see our Introduction to Prompting guide.
Troubleshooting
Given the realism of the model, Seedance 2.0 has heavy restrictions around using realistic humans in the inputs. Inputs images or videos that contain a realistic human, especially those that show a face, are likely to be blocked by the provider.
Avoid input images or videos that contain a realistic human
Try using obscured or blurred faces to allow the model to fill in the blanks for your character
Try using an unrealistic style, such as an animated or 3D rendered style for inputs
Looking to learn the latest?
Our full Creating with Seedance 2.0 guide is updated regularly and covers everything in this post in more detail, with additional examples and tips for getting the most out of each creation mode.
