Be Good to Me

DENIZ AKKABAK · SEEDANCE 2.5 TUTORIAL · @DENIZ.AKKABAK

MADE IN INVIDEO · invideo.io/i/DENIZ-ig

Be Good to Me

One unbroken 30-second dragon-rider POV — a single Seedance 2.5 prompt that stages the camera, the dialogue and the sound design against the same clock.

A TUTORIAL BY DENIZ AKKABAK · @DENIZ.AKKABAK

MODEL SEEDANCE 2.5 IN INVIDEO · FORMAT 9:16 · 30-SECOND SINGLE TAKE · GENRE PHOTOREAL FANTASY ACTION

Most AI videos fall apart the moment you ask for more than one thing to happen. This one asks for six — a glance at the cliff edge, a three-stride leap into freefall, a wave-strike rescue, a slalom through sea stacks, a colossal whale arcing over the flight path, and a calm glide out — and gets all of them in a single unbroken take, because the prompt never asks the model to improvise. It hands Seedance 2.5 a staged shooting script: every stage owns a slice of the clock and ends on a declared end state, and inside each slice the camera physics, the rider's lines and the sound effects are all told exactly what to do. The model isn't generating a video; it's executing a shot list. That is the whole trick, and this guide walks through how to run it inside invideo.

THE PIPELINE AT A GLANCE

1. 01

   Set the stage in invideo

   Open Seedance 2.5 inside invideo, lock the format to a vertical 9:16 canvas and a 30-second duration, and attach your dragon-rider still as the reference the prompt will swear to preserve.

2. 02

   Hand over the shooting script

   Paste the full staged prompt exactly as written — format contract, six timed stages, audio script and negative prompt — and generate. One pass, one take, no editing required.

STEP 01 · SETUP — INVIDEO

Setting Up Seedance 2.5 Inside invideo

Everything happens in one place. Sign in through the referral link, open the video generator, and pick Seedance 2.5 from the model catalog — it is the model doing the heavy lifting here, because it is the one that can hold timed dialogue and layered sound design inside a single 30-second generation. Three settings matter before you touch the prompt: the aspect ratio goes to 9:16 vertical (this is a phone-first, feed-native ride), the duration goes to the full 30 seconds (the FORMAT line and the six stages are written against that exact runtime — a shorter clip would amputate the ending), and audio stays on, because a third of this prompt is an audio script.

Then attach the reference image as @Image 1. The REFERENCES line is blunt about its job — “@Image 1 defines everything” — and itemizes what it must supply: the rider's chest POV, the gloved hands on cables, the harness clamps, the dragon's grey-green scales and head, the cliff. “Preserve exactly.” The reference is not decoration; it is the casting sheet, the location scout and the color grade in one file.

REFERENCE INPUT — THE WORLD SEEDANCE MUST PRESERVE

STEP 02 · GENERATE — SEEDANCE 2.5

The Call Sheet Prompt

This is the entire production, delivered as one document. Read it like a call sheet rather than a description: FORMAT and STYLE set the contract (30 seconds, 9:16, one continuous photoreal shot at real speed), REFERENCES binds everything to @Image 1, then six numbered stages each own a slice of the clock — and each one closes with a declared “End state” and its own camera instruction. Below the stages sit the AUDIO script (effects inventory in angle brackets, four timed English lines in braces, and the pointed “No music”), a CONSISTENCY lock, and a negative prompt that reads like a list of everything AI video usually gets wrong. Copy it exactly; the stage clock is the choreography.

OUTPUT

STEP 03

Anatomy — Why the Dragon Obeys

Seven load-bearing decisions keep a 30-second stunt from collapsing into soup. Each one is a specific line in the prompt doing a specific job:

1. 1

   The contract

   “FORMAT: 30s / 9:16 / photoreal POV flight / single continuous shot, no cuts” — the whole deal in one line, stated before anything else, with STYLE adding “real physics, real time speed only”. Seedance treats the opening lines as format law; all six stages are interpreted inside that law instead of as excuses to cut.

2. 2

   The reference lock

   “@Image 1 defines everything” — then an itemized list: chest POV, gloved hands on cables, harness clamps, grey-green scales, the cliff. “Preserve exactly.” Naming the elements one by one turns a vague style wish into a checklist the model can actually honor across all 30 seconds.

3. 3

   The stage spine with end states

   Six stages, each owning a range like “[Stage 3 - 9-14s]” — and each closing on a declared checkpoint: “End state: skimming the swell, wet lens.” The end states are the real discipline: the model always knows what frame it must arrive at before the next stage begins, so the ride can never drift off course.

4. 4

   The consistency lock

   “Locked first person POV, camera never leaves the rider's chest, hands and cables always in frame. Exactly 1 dragon, 1 whale.” The hands-and-cables clause is the POV insurance policy, and the blunt creature count is what stops Seedance from multiplying whales the moment one looks good.

5. 5

   Physics as camera language

   The prompt never says “shaky cam”; it assigns each stage its own physics — “shake on strides only, then wind trembling”, “hard g-force compression” at the wave strike, “rolling lean into banks” between the stacks — and lets droplets and spray accumulate on the lens. Motion is described as consequence, so the camera behaves like a body, not a drone.

6. 6

   Sound with no safety net

   “No music.” The AUDIO line builds the entire soundtrack from the world itself — an effects inventory in angle brackets (“3 footfalls before the leap only, then none”, breach eruption, deep whale call) plus four timed English lines in braces, from “Ready, my friend?” at 1s to “Straight and safe now” at 27s. Stripping the score forces the physics to carry the drama — which is exactly why it reads as real.

7. 7

   The insurance

   The NEGATIVE PROMPT is a checklist of every way AI flight footage usually betrays itself: “no slow motion, no speed ramps ... no weightless motion, no rigid flight path, no footsteps after the leap”. Each ban answers a specific failure mode — the footsteps clause alone kills the classic bug of ghost footfall audio continuing into the air.

THE RESULT

Output — One Ride, Zero Cuts

Three beats from the finished take: the eye-contact opening at the cliff edge, Stage 5's colossal whale arcing right-to-left over the flight path exactly as scripted, and the calm glide that holds the final frame. Audit your own generation against the prompt's checkpoints: is the take truly unbroken with no speed ramps, do the hands and cables ever leave frame, is the lens wet from the wave strike onward, do the footfalls stop at the leap, does the whale stay singular and cross overhead rather than beside, and does the ride land on “hands relaxed, horizon level”? Every one of those checks maps to a specific line — an end state, the consistency lock, or a negative-prompt ban — which is what makes this prompt auditable instead of just hopeful.

© 2026 Deniz Akkabak — all prompts and direction by Deniz Akkabak · Instagram @deniz.akkabak · Made in invideo — invideo.io/i/DENIZ-ig