Seedance 2.5 is ByteDance's latest video model, announced on 31 July 2026. It makes clips of up to 30 seconds and takes images, video and audio as references, up to 50 in total. Most of the disappointment people have with it comes from the prompt, not the model.

Where you can use it

Dreamina and CapCut (ByteDance's own), and Higgsfield on its Plus plan and above. Dreamina lists the UK as a market. CapCut's announcement names Europe and does not name the UK, so check your account before you pay. There is no official ByteDance prompt guide yet, so what follows comes from the companies hosting the model.

The formula

Subject and action first. Then, only if they change the shot: setting, visual style, camera, sound. Keep length and resolution in the settings, not in the prompt.

Seven things to put in a brief

1. Subject

2. Action

3. Setting

4. Visual direction

5. Camera

6. Sound

7. References

Use your references properly

Tag them (@image1, @video1, @audio1) and give each one a job: this is the face, this is the outfit, this is the camera move. Say what to copy and what to ignore. Don't expect it to read text written inside an image.

Longer clips

Split into stages, one change per stage, with a clear end state. About three or four beats per 15 seconds works. One moment per time range, no overlaps.

Sound

Write it into the prompt: music in brackets, sound effects in angle brackets, dialogue in curly brackets. For dialogue in English, name the language and accent first.

Common mistakes

• Vague prompts

• Too many events in one clip

• Re-describing your character every time, which breaks continuity

• Long lists of things you don't want

• Asking for an emotion with nothing visible to show it

Work cheaply

Draft at 480p, change one thing at a time, and only render the shots you like at 1080p. Timestamps set the pace but are not frame-exact, and exact signs or subtitles are better added afterwards.

Sources: CapCut newsroom · TechNode · fal.ai guide · Luma guide · PixVerse guide