Bring a whole scene to life, not just a single moving image. Seedance 2.5 is built for longer story-driven clips, coordinated image, video, and audio references, and smoother transitions between shots. ByteDance describes single-pass generations of up to 30 seconds, with options to extend a story across multiple rounds. It’s integrated directly into CawCut’s [Video Generation] node, where you can choose Seedance 2.5, select Text, Image References, Frames, or Omni References mode, add your inputs, and run it as part of your workflow.
1. Seedance 2.5 Strengths
| Strength | Application |
|---|---|
| Long-form, multi-shot storytelling: Generates clips up to 30 seconds in one pass and supports extending a story through multiple rounds. | Create a complete short narrative with a clear setup, development, and payoff, or let a sequence of connected shots unfold as one coherent scene. |
| Rich multimodal references: Can draw on images, video, and audio together, and interpret their subjects, style, composition, motion, and sound. | Bring a specific creative vision to life by combining a subject's appearance, a reference for camera or body movement, and sound or rhythm that shapes the scene. |
| Stronger continuity across shots and extensions: Aims to keep key subjects, environments, narrative pacing, and audiovisual style more consistent as a story develops. | Tell a story that follows the same recognizable character, product, or setting across several shots, scene changes, or continued footage. |
| More natural motion and smoother scene transitions: Improves motion quality, camera movement, and transitions between shots. | Create dynamic sequences—such as a character moving through a space or a camera following an action—with movement and transitions that feel connected rather than abrupt. |
| Joint audio-video generation and more precise timing: Generates sound with the video and supports more targeted direction for actions and edits by time. | Make a performance or scene where dialogue, sound effects, music, and visual action land together at the intended moments. |
2. Choose a Mode and Inputs
Choose the mode in the Video Generation composer based on what should guide the clip:
| Mode | Use it when | What to provide |
|---|---|---|
| [Text] | You want to create a scene from a written description. | A focused prompt describing the subject, action, setting, and desired shot. |
| [Image References] | A still image should guide the appearance, composition, or subject. | Relevant images, plus instructions for what to preserve and what should move. |
| [Frames] | The opening composition, ending composition, or transition between them matters. | A required first frame, an optional last frame, and a prompt describing the change between them. |
| [Omni References] | Multiple media types should guide the same generation. | Image, character, video, or audio references, with each one's role explained in the prompt. |
For more about the Video Generation node and how to connect it in a workflow, see Video Generation. To make a sequence of planned scenes, you can prepare a visual plan with the Image Storyboard guide.
For a ready-to-use product campaign path, explore CawCut Ad Video and its step-by-step creation guide.
3. Prepare Inputs and References
Both the prompt and references work best when each has a clear job and does not conflict with the others. Before running the node:
- State the subject and one main action first, then add the setting, visual style, camera direction, and timing that matter to the shot.
- Keep the sequence focused. If the clip has multiple beats, describe them in order and make sure they can fit the selected duration.
- Describe observable results rather than vague qualities; for example, specify “slow camera push-in” instead of only “cinematic.”
- Use compatible directions. Avoid asking for a locked camera and a rapid tracking move in the same shot, or for several unrelated actions at once.
- Use a clear image or clip where the subject and important details are visible.
- Tell the model which reference guides the character, product, setting, style, motion, framing, or audio, and what each reference should contribute.
- Describe what should stay consistent, such as clothing, colors, object shape, or location.
- For a motion or video reference, name the action or camera behavior to use. Explain whether other visual details should come from a different reference.
- For an audio reference, say whether it should guide speech, rhythm, timing, or general sound.
- Keep only references that contribute to the shot. Similar or contradictory assets can make it less clear which detail to follow.
In Omni References mode, add image, character, video, and audio references to guide different parts of the scene. You can also connect compatible upstream nodes. For node connections, see Understanding Nodes. For character-specific reference sheets, see Keep Characters Consistent in AI-Generated Videos; for location layout and shot-to-shot continuity, see Keep Scenes Consistent With Storyboards.
4. CawCut Workflow and Settings
- Open a workflow and add a [Video Generation] node, or select one already on the canvas.
- Choose [Text], [Image References], [Frames], or [Omni References], based on the inputs you need.
- Select Seedance 2.5 from the model menu.
- Add images, character references, video, or audio for the selected mode, and connect compatible upstream nodes where useful. Add the prompt in the composer or connect a text input.
- Review output settings, including duration, aspect ratio, resolution, and audio. Check the credit estimate if the node displays one; Credits explains CawCut credits.
- Run the node, preview the generated clip, and check the subject, movement, scene transitions, sound, and important reference details before using it downstream.
To explore examples of connected creative pipelines, visit CawCut Workflow. If you prefer generating through an AI-agent chat, see the cawcut-generate skill guide.
5. Iterate and Troubleshoot
If the clip misses the intent, diagnose the largest mismatch and make a targeted change:
| Issue | Try this |
|---|---|
| The subject looks different from the reference | Use a clearer reference, describe the distinguishing features, and remove other images that suggest a different appearance. |
| The action is unclear or incomplete | Reduce the prompt to one main action and make its order explicit. Avoid giving the subject several simultaneous tasks. |
| The camera moves in the wrong way | State one camera move and whether the subject, camera, or both should move. Remove conflicting framing instructions. |
| Several scenes feel disconnected | Simplify the sequence, identify the transition between beats, or split the idea into separate clips that you can assemble later. |
| Audio does not match the scene | Specify the intended speech, ambience, or music and its timing. Avoid combining incompatible audio directions. |
| Fine details are unstable | Use a larger, clearer reference, keep the shot less busy, and review the result at full size. Treat tiny labels and exact text as details to verify. |
Change one or two things per generation when possible. That makes it easier to tell whether the prompt, reference, mode, or setting improved the result. If you are building a larger workflow, connect the video output to the next compatible node rather than repeating earlier steps. To join generated clips or add transitions between them, see Video and Transition Nodes.