The Complete Seedance 2.5 Guide: Cinematic AI Video Step by Step

Seedance 2.5 landed and the thing everyone fixates on is the 30 second generation. That is the headline, but it is not the part that decides whether your video looks like a commercial or looks like AI. The part that decides it happens before you open the model at all.

This is the exact process my team and I used to make a sun cream ad that went viral, start to finish, so you can run the same thing for your own brand or your own client.

What actually changed in 2.5

Four things matter, as of September 2026:

• You can generate up to 30 seconds in one go, and you can one shot a prompt rather than stitching clips.

• You can upload 50 references in the prompt bar. Not just images. Video references and audio references too, so you can hold a specific voice consistent across generations.

• More control means more compute, which means more cost. A 30 second generation runs $40 to $60 right now, which is criminal, and we spent thousands in a week making three client projects.

• You can feed it 3D camera motion built in something like Blender, and it will adhere to it. Your whole pre production can now be nailed down before you generate a frame.

Worth saying plainly: the viral ad was not one 30 second generation. We built it scene by scene, designed each location, character and object, then stitched it in our edit. One shot generations work, but scene by scene is what gives you control.

Start in a Google Doc, not in the model

This is where almost everyone crumbles. The ideation phase. If your foundations are weak the project falls over later and you will never quite work out why.

So I open a plain doc and write the core message first. Ours was: we are promoting a sun cream brand. Then a list of what could be funny about that. Rubbing sun cream into your mate's back makes some guys uncomfortable, so let us build the joke there, and let us put a character in who is far too up for it.

That is it. A message and a list. Ten minutes of thinking that saves you a hundred dollars of generations.

Write it like a director, not like a prompt

Once the idea is blocked out, I dictate the scene as though I am on set. I start by naming the camera shot, because that makes me picture it. Medium shot. Close up. Wide establishing.

Then I describe what happens. It starts with my friend taking a sip of beer while I apply sun cream to my arms. We hear beach ambience, seagulls, general conversation, the beach bar behind us.

Close your eyes when you do this. Put yourself in the location. What would you hear, what would you see, what is moving in the background. Seedance will not add background life on its own, and without it your scene looks like a ghost town.

Then the acting. This is the bit that separates a good generation from a great one. I do not write "he laughs." I write that he spits out the beer in his mouth and laughs obnoxiously for half a second. The timing matters. If you just say he laughs obnoxiously, you might get two seconds of it and your pacing dies.

Give the accents. Give the emotional state. I look disappointed and I plead. My script for a 30 second project runs about a page and a half.

Build your references before you generate

Two reference types do the heavy lifting.

Character sheets. Front, side, back, front close up, top down, height, tattoos, skin texture, eye colour, colour palette. Build them from a single reference photo of your character. Use GPT Image 2 as the image model for this, because it gives you the most realistic skin texture. Nano Banana Pro gives you that plasticky look on faces, so I avoid it for characters.

Our first character was fine but he did not feel right. Tried him with a hat. Still wrong. Ended up on surfy beach blonde hair and that was the one. Expect to iterate. Find design references on Pinterest, then remake them your own way, because you do not own the copyright on someone else's design.

The location reference. This one is underrated and it is the most important image in your project, because it locks the colour grade for everything that follows. Do not accept a flat overcast plate. Go and find a cinematic frame you actually love the grading of, upload it, and design your location inside that grade.

Ours went from a lifeless hut to a proper beach bar by adding a shelf of alcohol, fairy lights, more seating, more tables, more of a community feel. Refine until the vibe is right, because everything downstream inherits it.

Hand the whole thing to Claude

Now I paste the script into Claude with the instruction to write it as a 30 second Seedance prompt, and I give it the context: the character sheets, the location, the product. Context is the whole game here. Without it Claude writes generic description and your generations drift.

What comes back is time coded, shot by shot, with the camera called out and the acting nuances preserved. My key phrases survive intact, so "obnoxious half laugh for a second, then flat and certain" is still sitting in the prompt.

Then paste into Seedance, tag each character reference to its name in the prompt, tag the location, set 30 seconds and 16 by 9, and run it.

One practical note. Character limits matter. Some platforms cap you at 5,000 characters, which is nowhere near enough for a detailed 30 second prompt. Use a platform that gives you room, or your acting direction gets cut.

Run it three times and mine the best

Do not expect perfection out of the first generation. That is gambling.

Run one. Watch the acting performance. Is it what you asked for. If not, go back to Claude, say which beat worked and which did not, refine, run again. Once you are happy with the prompt, run it three times and take the best moments from each, then cut them together.

That is how the viral one was built, and it is the single habit that lifts amateur work to client grade.

The edit is smaller than you think

Our timeline had very few layers. The colour comes out of the model good enough that we did not regrade from scratch. We added subtitles, trimmed the awkward extra second of silence that generations tend to leave, and dropped a music track under one beat to land the joke.

To finish it properly, add your own subtle grade and a little film grain. That is what stops it reading as something that came straight out of a video model.

What this costs, honestly

At $40 to $60 per 30 second generation, and a workflow that wants three generations per scene, this is not a cheap way to work. Budget for it, price it into the job, and test on an unlimited plan while you are learning so you are not burning money on experiments.

How I keep every shot looking like one camera

Everything above is the method. The thing that makes my generations look consistent shot to shot is a custom prompt structure I have built and refined, the global look block that opens every prompt with the same camera body, lens, focal length, depth of field, halation and grain, so that every cut in a sequence feels shot on one camera. That structure, and the Seedance skill that writes it automatically, is the piece I keep for members and buyers. The workflow above works without it. It just takes you more attempts to get there.

If you want the skill on its own, you can get the Seedance 2.5 prompt skill here. It writes the full structure for you, every time.

Common questions

How long can a Seedance 2.5 video be?

Up to 30 seconds in a single generation as of September 2026, and you can one shot the whole prompt. For a finished piece I still recommend building scene by scene and stitching, because you get far more control over each shot.

How much does Seedance 2.5 cost to run?

Roughly $40 to $60 for a 30 second generation at current pricing. Unlimited plans exist on some platforms and are the sensible way to learn without burning through budget.

Why do my AI videos still look like AI?

Usually three reasons. No moving life in the background, so the scene reads as empty. Acting direction that is too vague, so the timing is off. And no location reference, so the colour grade is flat and inconsistent between shots.

What image model should I use for character sheets?

GPT Image 2 for people, because the skin texture holds up. Nano Banana Pro is excellent for many things but tends toward a plastic finish on faces.

Want the prompt structure and the rest of the workflow? Join the community.