Seedance 2.5 is ByteDance's latest AI video model. It can generate up to 30 seconds of video in one pass, with sound built in at the same time, and it lets you feed it reference images, clips, and audio so it sticks to a specific character, style, or camera move. That's a lot of power, and it's easy to either underuse it (typing one vague line and hoping) or overload it (dumping in 40 references and a giant paragraph). Here's how to get good, consistent results on BLOCKLORE without needing a professional production workflow.
1. Describe one shot at a time
Write your prompt the way you'd describe a single shot to a camera operator: who's in it, what they do, where they are, and how the camera moves. Avoid cramming multiple scenes or ideas into one prompt; the model handles a focused shot much better than a jumbled one.
Example: "Medium shot of a woman in a yellow raincoat walking down a rainy city street at night. She stops, looks up at a neon sign, and smiles. Camera slowly tracks alongside her."
2. Use a handful of references, tagged clearly
Seedance 2.5 accepts references (images, short video clips, audio) and lets you tag each one with @Image1, @Video1, @Audio1, and so on in the prompt text. You don't need to fill every slot - 3 to 5 well-chosen references (a face, an outfit, a location) usually give more stable results than a large, unlabeled pile. The important part is naming what each one is for, since the model won't reliably guess on its own.
"@Image1 shows the character's face and jacket, keep these consistent. @Image2 is the coffee shop interior, use it for the setting."
"Use @Image1 for the character's face. Use @Image2 for his leather jacket. Use @Image3 for the lighting and color grade."
3. Don't default to the full 30 seconds
A longer clip sounds appealing, but the more time the model has to fill, the more chances there are for details to drift - a face changing slightly, an object disappearing. Start with a shorter clip (5-10 seconds) for anything with a lot of detail or a specific look you want to protect, and only reach for the full 30 seconds when the action genuinely needs that much room, like a scene with a clear beginning, middle, and end.
A 30-second video is a long time for an AI to guess what you want. If your prompt is just "A man walks down the street," the model might have him walk for five seconds and then start doing backflips.
Example prompt: "Use @Image1 for the main character. 16:9 widescreen, smooth cinematic tracking shot. 0-10s: The character walks down a dark, rainy alleyway, illuminated by neon signs. 11-20s: He stops, turns toward the camera, and pulls a map from his pocket. 21-30s: The camera slowly pushes in for a tight close-up on his face as he looks confused. Audio: raindrops, heavy footsteps, no music. Strictly preserve the face from @Image1."
4. Write for the sound, not just the picture
Because audio is generated together with the video, describe what you want to hear, not just what you want to see. If you leave sound out of the prompt, you're leaving it up to chance. Mention dialogue, ambient noise, music, or explicitly say "no music" if you want a quiet, natural scene.
"Audio: rain on the pavement, distant traffic, footsteps on wet concrete, no music."
"Sync the camera movements to the rhythm in @Audio1. When the heavy bass drops at 0:15, execute a fast whip-pan to the right."
5. Build long scenes by extending, not by overloading one prompt
If you're trying to tell a longer story, don't try to script all of it into a single request. Generate a first clip, then use it as the basis for an extension - a follow-up generation that continues from the previous one, keeping the same characters and setting. This keeps each individual prompt focused and gives you a checkpoint to review before continuing.
Extension prompt: "Continue from the previous clip. Same character, same street, same visual style. She now crosses the street and enters the coffee shop."
6. Fix small mistakes with editing, not a full re-roll
If a clip is mostly right but one detail is off - a wrong prop, a background object, a distorted hand - you don't need to regenerate the whole thing from scratch. Seedance 2.5 supports edits that target a specific part of an existing clip while leaving the rest untouched. In your prompt, reference the original clip, explicitly say what should stay the same, and describe only the change you want.
"Edit the previous video. Keep the character, camera movement, and background unchanged. Replace only the cup on the table with a glass of water."
For time-specific fixes, you can also break the clip into time ranges in your prompt (like "0-5s" and "5-10s") and describe what should happen or change in each range, rather than editing the whole clip at once.
"Keep 0-20s completely unchanged. From 21-30s, change the character's action so he looks at his phone instead of the sky."
7. Save what works
When a generation nails a character's look or a scene's lighting, save that still or clip. You can reuse it as a reference in future prompts to keep things consistent across a project, instead of re-describing everything from scratch each time.
Quick recap
Keep prompts focused on one shot, use a few well-labeled references instead of many, don't over-use the max clip length, describe your audio, extend rather than cram, edit instead of regenerating, and save your best results as future references. Like most AI tools, Seedance 2.5 rewards iteration - expect to refine a shot over a couple of tries rather than nailing it on the first attempt.
