You give Seedance 2.5 a careful prompt and get back a clip where the character's face has drifted, two of them appear at once, or they keep talking after the line ends. Most of the time the model isn't the problem. The prompt is asking for too much, in the wrong order.
This guide gives you a simple structure for Seedance 2.5 prompts, with real prompts and the clips they produced. The examples are scenes from Garage Dragon Surprise, a short film made in EasyVid's Director Mode, and each one says which video model generated it.
The short version
A good Seedance 2.5 prompt covers six things, in this order:
| Part | What to write | Example |
|---|---|---|
| Style | One sentence on the look | "Stylized 3D animation render, soft polished lighting" |
| Who and where | Point at your reference images | "@Lena ... @Family_Garage", or "the girl from image 1" |
| Action | One or two actions, in the order they happen | "stumbles back from the bench, turns and runs at the camera" |
| Camera | Height, framing, and one move | "Close-up of @Lena, quickly zooming out" |
| Sound | The sounds you want, and no music | "Footsteps, a crackle of flame, NO MUSIC" |
| Spoken line | Last, in quotes (EasyVid adds it) | says: "There's a dragon in the garage!" |
1. Point at your references, don't describe them again
Seedance 2.5 takes reference images, and that's what keeps a character the same from clip to clip. Once a character has a reference, the prompt doesn't need to re-describe their face or clothes. Point at the image instead: "the girl from image 1", "the driveway from image 2". Re-describing the character in different words gives the model two versions to choose between.
If the model adds a second copy of a character, say so plainly. One of the scenes below includes "Do not create two people."
2. One or two actions per clip
A clip of five to eight seconds has room for one action and a reaction, not a whole sequence. Write the actions in the order they happen, using simple verbs: runs, stops, points, says. If you need more, make it two scenes. That also gives you a cut where you want one, instead of wherever the model decides.
3. Give the camera one job
Say where the camera is and what it does: "close-up", "low camera at the dragon's height", "slow push in". Pick one move per shot. A prompt that asks for a dolly, a pan, and a crane in five seconds usually gets a camera that wanders.
4. Ask for the sound you want, and say no music
Seedance 2.5 generates sound with the picture, so write the sounds into the prompt: footsteps, a crackle of flame, a door. Also say "no music". A music bed baked into each clip fights with the soundtrack of the finished video, and you can't remove it afterwards.
5. Put the spoken line last
Put the character's line at the very end of the prompt, in quotes. A video model tends to speak whatever text comes after a line as part of it, so any instruction you add after the words can end up said out loud. Put everything else first, then the line.
How the examples are written
In EasyVid you tag characters, places, and objects with @names instead of numbering images. EasyVid sends each tagged asset's reference image with the scene, adds the scene's spoken lines at the end of the prompt with each character's recorded voice, and, for Seedance, asks for no music. So the prompts below hold only the picture: who, where, what happens, and the camera.
Example 1: a reaction shot with a cut (Seedance 2.5)
Prompt
Start with a close-up of @Marcus's face, using the exact image provided: @Frame_3. He is shocked to see his lawnmower burning. Cut to him on the right and @Lena on the left. Without another word, he drops his sponge into the bucket and runs toward the burning lawnmower. Do not create two people. Quick cut, no crossfade.
Why it works: "Cut to" splits the clip into two shots inside one generation, each with its own framing. Left and right placement keeps the two characters apart, "Do not create two people" stops a duplicate Marcus, and @Frame_3 (a still saved from an earlier clip) pins the opening shot.
Example 2: two characters, three lines (MiniMax H3 Max)
Prompt
Close-up of @Lena, quickly zooming out as she runs from the workbench up to @Marcus and stands to his left. She looks at him desperately and points back at @Baby_Dragon on the workbench. @Marcus stands beside the car with a dripping sponge in his hand, eyebrows raised, and dismisses her panic. Behind them, @Baby_Dragon jumps down from the workbench and shoots a burst of flame at the lawnmower, setting it on fire. @Driveway:left
Why it works: the camera move comes first, each character has a side of the frame, and the background action (the dragon setting the fire) is written as its own sentence so it doesn't steal the foreground.
Example 3: action and a sound beat (MiniMax H3 Max)
Prompt
@Lena screams as @Baby_Dragon climbs out of @Dragon_Egg:cracked on the workbench on the left and sneezes a small jet of fire. @Lena stumbles back from the bench, turns toward the camera and runs at it with her arms up, shrieking. The final frame is filled with her scared face. @Family_Garage
Why it works: the actions are in the order they happen, and the last sentence says exactly how the clip should end, which sets up the next scene's cut.
Common problems and fixes
| Problem | Fix |
|---|---|
| The character's face changes | Add or improve the character's reference image, and stop re-describing them in the prompt |
| Two copies of one character | Say where each person is (left, right) and add "Do not create two people" |
| The character keeps talking after the line | Move the line to the very end of the prompt |
| Music under the dialogue | Add "NO MUSIC" at the start |
| The camera drifts | Ask for one camera move, or none |
| Too much happens, and badly | Split it into two scenes |
Using these prompts in EasyVid
In EasyVid's Director Mode you don't have to write any of this from scratch. EasyVid writes the script, makes a reference image for each character and location, voices every line, and writes a video prompt for each scene, so characters speak their lines on camera. Pick Seedance 2.5 as a scene's model and edit its prompt using the structure above when a shot needs fixing. See the Seedance 2.5 page for specs, pricing, and plans, or read what changed from Seedance 2.0.