Why Does Your Video Get Weirder with Every Edit? A Seedance 2.5 Prompt Guide
With Seedance 2.5, you can change part of an existing video or add new footage before or after it, without regenerating the whole clip for a single edit. One distinction causes more confusion than most: the difference between a reference video and a source video. How do you tell the model which clip to borrow from and which one to preserve?
A reference video supplies information about a character, motion, or style. A source video is the clip you want to keep largely intact. Choose the wrong mode, and even a detailed prompt may produce the wrong result. You might ask for a wardrobe change, only to get a new character, room, and camera move because the model treated the original clip as a reference. The outfit description was not the problem; the task was set up as reference generation instead of editing.
Three modes: reference generation, editing, and extension
Start with what you want the video to do: create a new clip from elements of an existing one, change part of a source video, or add footage along its timeline. That choice determines the mode.
| What you want to do | What should be preserved | Mode |
|---|---|---|
| Use a character, motion, style, or sound to create a new video | The original visuals and timeline do not need to be preserved | Reference |
| Keep an existing video but replace an outfit, product, background, or short segment | Everything in the source video that is not explicitly selected for editing | Video editing |
| Add an opening, continue an ending, or connect two existing clips | The source video, with new content added along the timeline | Video extension |
The same clip can serve two different purposes. Used as a reference, it supplies only the details named in the prompt. Used as the source, its visuals and timeline should stay as close to the original as possible. The file type is the same; your intended use determines the mode.
From there, the prompt depends on the job. For reference generation, specify what to borrow from each asset. For editing, name the change and what must stay untouched. For extension, say where to continue and what happens in the new segment.
Reference generation
Choose reference generation when you want to reuse a character, action, shot, or sound but do not need to preserve a finished clip. Every uploaded image, video, and audio file acts as a reference. Seedance uses only the details named in the prompt rather than preserving any one video’s visuals or timeline.
| Reference type | What the model can borrow |
|---|---|
| Subject reference | The appearance and voice of a person, product, prop, environment, or virtual character |
| Motion reference | Motion, expression, camera movement, creative direction, and effects |
| Clay render reference | Motion paths, camera movement, and spatial relationships from a clay-render video, combined with subject or environment assets for rendering |
| Style reference | Visual style, color palette, materials, and grain |
| Audio reference | Music, melody, dialogue, vocal tone, and ambient sound |
| Grid storyboard | Shot order, subjects, actions, and narrative progression |
| Keyframe reference | Visual changes between a first frame, last frame, or multiple keyframes |
| Automatic assembly | Combine multiple images and video clips into a short video, then add text, stickers, or transitions |
One asset can contain a character, motion, camera work, and sound. Instead of writing only “refer to @video1,” name exactly what to borrow. This matters even more as you add assets. For example:
Use @image1 for the main character’s appearance, @video1 for motion and camera movement, and @audio1 for background music.
Here, each asset has one job. If you only write “refer to @image1, @video1, and @audio1,” Seedance has to decide what each one contributes. That makes mixed-up roles more likely.
Reference asset limits
Seedance 2.5 accepts up to 50 reference assets per generation, within separate limits: 30 images, 10 video clips, and 10 audio clips. Video and audio are counted separately, with a 30-second total for each category.
| Asset | Quantity | Size and duration | Formats |
|---|---|---|---|
| Images | 0–30 | Up to 4K and 30 MB per image | JPEG, PNG, WEBP, BMP, TIFF, GIF, HEIC, HEIF |
| Video | 0–10 clips | 480p–4K; up to 200 MB per clip; 2–30 seconds per clip; 30 seconds total | MP4, MOV |
| Audio | 0–10 clips | Up to 15 MB per clip; 2–30 seconds per clip; 30 seconds total | MP3, WAV |
Use only the references you need
You rarely need the full upload limit. When using video or audio as a subject reference, try to stay below five subject assets and keep each clip around 5–10 seconds. A clip that is too short may not capture the character or motion clearly; one that is too long can bury the useful details in unrelated content.
For image-based subject references, aim for eight subjects or fewer. With more than five subjects, single-view references tend to be more stable. If you need multiple angles, use separate images rather than squeezing every view into one.
If you need to cut assets, keep the main characters first, then essential products or props. Add environment and overall style references only if they still help.
Video editing
Choose video editing when the source is already close to what you want and only one part needs to change. Here, @video1 is the clip to edit, not a style reference. Seedance uses its visuals, pacing, and camera movement as the base, then applies the changes named in the prompt.
Write an editing prompt in three parts
Video editing:edit request + @video1 (source video) + optional @imageN / @audioN
- Name @video1 as the source video.
- Describe what should change, the intended result, and when the edit applies. Use @imageN to provide a specific appearance or @audioN to replace sound.
- Add a preservation rule such as “keep everything else unchanged” so the edit does not spill into the composition, motion, or camera work.
Treat the prompt like a change request: identify the segment or object, add any reference assets, then say what must stay untouched. There is no need to describe characters, environment, lighting, or camera movement that are already clear in the source video.
What video editing can change
| Editing method | Best for | Prompt example |
|---|---|---|
| Text-guided editing | Replacing a scene or changing a local area, with optional timestamps | Replace the scene from 0:02–0:05 in @video1 with a medieval stone castle platform. Keep the character movement and pacing unchanged. |
| Reference-image editing | Replacing a person, outfit, product, or another object that must keep a consistent appearance | Replace the outfit of the man in dark clothing in @video1 with the outfit from @image2. Keep everything else unchanged. |
| Add a subject | Adding a character, prop, or effect | Keep the characters, environment, and composition in @video1 unchanged. Add an energy bow to the character’s hands. |
| Remove a subject | Removing production equipment, watermarks, or selected objects and filling the empty area | Remove the drone and foreground track from @video1, reconstruct the removed areas naturally, and keep everything else unchanged. |
| Audio editing | Replacing background music, adding sound effects, or changing a voice | Keep the visuals and camera movement in @video1 unchanged. Replace the background music with @audio1. |
“Keep everything else unchanged” helps limit the edit, but it is not specific enough on its own. Name who or what should change, and specify the location or time range. Then use the preservation rule to mark the boundary.
Video extension
Choose video extension when the source itself is fine but you want footage before it, after its final frame, or between two clips. The original video stays in place while Seedance adds to the timeline, using its visuals, pacing, and style as the basis for the new segment.
Video extension:extension request + @video1 (source video) + optional @imageN / @audioN
| Extension type | Where new content is generated | Prompt example |
|---|---|---|
| Extend the ending | Generate new footage after the final frame of the source video | Continue @video1 for 6 seconds from its final frame: the main character walks out of view, the sky darkens, and the streetlights turn on one by one. Connect the visuals and audio seamlessly. |
| Add an opening | Generate earlier footage that ends at the first frame of the source video | Add 4 seconds before the first frame of @video1: the main character walks into view from a distance. Match the visual style of the first frame. |
| Fill between videos | Generate a transition between two existing clips | Generate a 3-second transition between @video1 and @video2 so the visuals, color, and camera movement connect naturally. |
For an extension prompt, say where the new segment starts, how long it runs, and what happens. Add an image or audio reference only if the new segment introduces a person, product, or sound missing from the source video. After each @mention, say what that asset is for.
In Artflo, upload one video, hover over it, and choose the extend backward or extend forward control. To connect two clips, upload both and select the transition control between them. Artflo fills in the appropriate prompt structure automatically.
What Seedance keeps from the source video
Editing and extension both build on a source video, so Seedance inherits its aspect ratio automatically. You do not need to set it again. Duration works differently: an edit generally stays close to the source length, while an extension lets you choose how many seconds to add.
| Mode | Aspect ratio | Duration |
|---|---|---|
| Video editing | Matches the source video; editing does not recrop the frame | Usually matches the source; model processing may cause small differences |
| Video extension | Matches the source video, keeping the new segment in the same frame | Set the length of the new segment in the prompt |
You do not need to repeat the orientation or aspect ratio in the prompt. Video editing is not the right tool for changing the finished clip’s length. If you need more footage, use video extension and state how many seconds to add.
Before you generate
Seedance 2.5 prompts can be short when asset roles and edit boundaries are specific. If the result drifts, check which clip is the source, what each reference contributes, and whether the edit target, time range, or extension length is explicit. If any of those details are ambiguous, the model has to guess how the assets relate to one another.