AI Ad Playbook

The plan4 chapters

Here's the whole trip.

First the tools. Then we make a person talk. Then we make them write. Then we go full animation.

Three kinds of ad: a talking face, a whiteboard, and an animated world

01

The Toolbox

The apps we use, and the simple language they speak.

kie.aiAPI keyJSONAPI Lab 02

The Talking Avatar

A real-looking person says your script. One face, one voice.

UGCSeedance 2.5ElevenLabs 03

The Whiteboard

Same person. Now they write and erase while they talk.

EraserClean boardQuality 04

AI Animation ★

The main event. Entertain first, pick a style, one sentence per scene, and the real exercise step by step.

15 styles1 sentence = 1 sceneImage to image

1

Account for all the big AI models

kie.ai

30s

Longest clip in one take

Seedance 2.5

0

Lines of code you write

the agent writes them

Your edgeBefore we start

Same tools. Different you.

The AI can do a lot. But it doesn't know what you know. Your story, your skills and your taste are what make the ad yours.

Five different people paint at identical easels with the same AI paint box, and every painting shows their own world

The AI

A super tool, not the artist.

It can draw, talk, write and edit. But it only makes what someone asks for. It's the same tool for everybody.

You

Your background is the secret ingredient.

Your job, your city, your culture, the things you've lived and the way you see the world. Nobody else can bring that.

The result

No two ads alike.

Give ten people this same guide and you'll get ten different ads. That's the beautiful part.

01 / 15 · kie.aiToolbox

One account. Every AI.

kie.ai is a middleman. It plugs you into all the big AI companies at once.

A food court where one tray and one card buy from every AI stall

What

A shop that sells every AI.

Image, video, voice and music models from different companies. GPT Image, Seedream, Seedance, Veo, Kling and many more. All in one place.

Why

One wallet instead of ten.

You buy credits once and spend them on any model you want. No ten accounts. No ten subscriptions.

How

Two ways to use it.

By hand on the website: fill the form, press Run. Or with an agent through the API, and your AI does the clicking for you.

02 / 15 · Other doorsToolbox

Same idea. Different doors.

kie.ai is not the only middleman. These are the names you'll hear the most.

kie.aiOur pick

Website plus API. All the big models. One wallet. This is what we use in this guide.

fal.aiSame idea

Another middleman, made for developers. Its API works the same way as kie's, so everything here works there too.

HiggsfieldCareful

A pretty app full of ready-made effects and tools. You pay extra for things you can do yourself.

03 / 15 · The APIToolbox

The API is a waiter.

You never walk into the kitchen. You tell the waiter what you want, and the waiter brings it back.

A waiter carries an order ticket from a robot at a table to a kitchen and brings back a film reel

Youor your agentdecides what to make

The orderJSONwhat you want, written down

The waiterAPIcarries it to the kitchen

The kitchenSeedance 2.5cooks your video

↩ Then the waiter brings your finished video back to you.

What

A door made for programs.

The website is made for your eyes and your mouse. The API is made for programs. Same kitchen, different door.

API key

Your secret membership card.

It tells kie the order is yours, and who pays for it. Never share it. Never post it.

Agents

Who can talk to the waiter?

Any AI agent: Claude Code, Codex, Antigravity, Hermes. An agent is an AI that does things, not just talks. If it can use an API, it can do everything in this guide.

04 / 17 · Your API keyToolbox

Get your key in three clicks.

The API key is your membership card. One key works for every model. This is where you make one on kie.ai.

The kie.ai API Keys page, with the API Keys menu item and the Create New Key button — A. In the menu on the left, open API Keys. B. Click Create New Key.

Keep it secret

It's like your card's PIN.

Anyone who has your key can spend your credits. Never post it, and never show it in a video. We blurred ours.

Set a limit

Use Safe-Spend Limits.

In the API Keys table, Safe-Spend Limits caps how much one key can spend. Great while you're testing.

Leaked it?

Delete it, make a new one.

Click the red trash can next to the key. The old key stops working, and nobody can use it anymore.

04 / 15 · JSONToolbox

JSON is just a form, written as text.

When your agent orders from the API, it writes the order in JSON. Change something in the form and watch the JSON change.

Send this order for real

The form above is the order. Change it up there, send it from here. Three calls: send, check, get it.

model

Seedance 2.5 Seedance 2.0 Mini

first_frame_url (your picture. Leave it empty for text to video)

Upload

This order: about 1890 credits

Now making: a video

1SendPOST /api/v1/jobs/createTaskready
2CheckGET /api/v1/jobs/recordInfoafter 1
3Get itPOST /api/v1/common/download-urlafter 2

Set the form above, then press Send it for real. In Practice nothing is charged.

Videowaiting

Your clip shows up here.

No clip yet.

"720p" text 30 number true / false switch "https://..." link

{ }

The box

Everything lives between the two curly braces.

"key"

The label

The name of the thing. Always in double quotes.

:

The glue

Sticks the label to its answer.

value

The answer

Text wears quotes. Numbers and true/false don't.

,

The comma

Goes between items. Never after the last one.

“ ”

The trap

Curly quotes from Word or Google Docs break JSON. Use straight ones: " "

06 / 24 · API LabToolbox

Three calls. That's the whole API.

Every job on kie.ai is the same little dance: send it, check it, get it. Pictures and videos, same three steps. Try them right here.

1Send

POST/api/v1/jobs/createTask

Hand over the order.

You send the JSON. kie.ai answers with a ticket number, called the taskId. Nothing is ready yet.

2Check

GET/api/v1/jobs/recordInfo?taskId=...

Ask: is it ready?

You show your ticket. kie.ai says waiting, queuing, generating, success or fail. If something broke, this is where it tells you why.

3Get it

POST/api/v1/common/download-url

Pick up your order.

When it says success, you ask for the download link. It works for 20 minutes, so save your file fast.

Bonus calls · GET /api/v1/chat/credit · your wallet (free) · POST file-base64-upload · put a picture online so a model can see it (free)

code in the JSON	What it means	What to do
200	All good.	Keep going.
401	Wrong or missing key.	Copy the key again from API Keys.
402	Not enough credits.	Top up your wallet in Billing.
422	Something in the order is wrong.	Check the field names and the values.
429	Too many orders too fast.	Wait a few seconds, then try again.
455	kie.ai is under maintenance.	Try again later.
500	Server error, often because it's too busy.	Wait a minute, then try again.
501	The model couldn't make it.	Change the prompt and try again.
505	That feature is switched off.	Pick another model.
state: fail	The job started but broke on the way.	Read failMsg in step 2. It tells you why.

Heads up: every kie.ai answer carries two numbers. The outer one is the browser's HTTP status. It says 200 almost every time, even when the job failed. The one that matters is the code field inside the JSON. That is the number in this table, and the number our console prints. Always read the code inside the JSON.

05 / 15 · Photo firstUGC

Every video starts as a photo.

Before anyone talks, we make the person. Get the face right first. The rest follows.

Instant photos of the same man hang on a line, and the best one is circled

Step 1

Open any chatbot.

ChatGPT, Gemini or Claude. Any of them works.

Step 2

Paste the master prompt.

It teaches the chatbot to write prompts for real-looking, phone-camera photos. It's right below.

Step 3

Describe your person.

Age, body, skin, clothes, place. The more specific, the better. Then paste its prompt into the image model on kie.ai.

YouGive me a 30-year-old African American man, a little overweight, with vitiligo, wearing a Carhartt cap, sitting in his parked car.

Chatbot · exampleHere's your image prompt: Amateur iPhone selfie of a 30-year-old African American man with vitiligo patches on his face and hands, a little overweight, wearing a tan Carhartt cap, sitting in the driver's seat of his parked car. Slightly uneven framing, natural daylight, reddit image, slightly blurry.

Master prompt · UGC

Prompt Generator

You are a UGC (User-Generated Content) AI agent.

Your task: Take the reference image or the product in the reference image and place it into realistic, casual scenes as if captured by everyday content creators or influencers.
Always follow UGC-style casual realism principles unless the users asks otherwise:

Everyday realism with authentic, relatable environments

amateur iPhone photo/video style

Slightly imperfect framing and natural lighting

Candid poses, genuine expressions

Imperfections allowed unless otherwise specified by the user or unless image reference shows otherwise. (blemishes, uneven skin)

Camera parameter must include casual descriptors:

“amateur iPhone selfie”, “uneven framing”, etc.

have the camera movement be fixed unless otherwise stated

If dialogue is needed:

Use the EXACT dialogue in the script description given by the user if it looks like a dialogue. Note that each scene will only be 30 seconds long, so split the dialogue between scene 1 and onward if it's too long


All outputs must feel **natural, candid, and unpolished**, avoiding professional or overly staged looks. This means:

- Everyday realism with authentic, relatable settings
- Amateur-quality iPhone photo/video style
- Slightly imperfect framing and lighting
- Candid poses and genuine expressions
- Visible imperfections (blemishes, messy hair, uneven skin)

We need these videos to look natural and real. So in the prompts, have the Camera parameter always use keywords like these: unremarkable amateur iPhone photos, reddit image, snapchat video, Casual iPhone selfie, slightly uneven framing, Authentic share, slightly blurry, Amateur quality phone photo.

Real-world environments left as-is (clutter, busy backgrounds)
We need these videos to look natural and real. So in the prompts, have the Camera parameter always use keywords like these: unremarkable amateur iPhone photos, reddit image, snapchat video, Casual iPhone selfie, slightly uneven framing, Authentic share, slightly blurry, Amateur quality phone photo.

06 / 15 · Image modelsUGC

GPT Image first. Seedream when it says no.

Two image models. One is the best. The other is the backup with fewer rules.

Default

GPT Image

Always the newest version.

Best quality: best faces, best skin, best text.
Start here every time.
Made by OpenAI, the ChatGPT company.

Backup

Seedream

Made by ByteDance, the TikTok company.

Not as strong as GPT Image.
Much looser rules, like most Chinese models.
Makes things GPT Image refuses, like famous faces.

07 / 15 · IterateUGC

Don't love it? Ask again.

Nobody nails it on the first try. Pros just try more times.

Prompt Photo Love it? Yes: next step

↺ Not yet? Re-roll or tweak, then try again.

Move 1 · Re-roll

Same prompt, new try.

Run the exact same prompt 2 or 3 times. Every run gives a different photo. Keep the best one.

Move 2 · Tweak

Ask for changes in the same chat.

Don't start over. Stay in the chat with the master prompt and say what to change. It gives you a new prompt to run.

YouI don't like the Carhartt cap. Give him a wide hat and dress him like a monk. Long hair. Keep him African American.

Chatbot · exampleGot it. Here's the new prompt: Amateur iPhone selfie of a 30-year-old African American man with long hair and vitiligo, wearing a simple brown monk robe and a wide hat...

08 / 15 · Seedance 2.5UGC

30 seconds. One take.

Seedance 2.5 is our video model. Right now it's the only one that makes up to 30 seconds in a single clip.

Seedance 2.51 clip · 0 seams

30s · one take

An 8-second model4 clips · 3 seams

8s8s8s6s

What

It turns your photo into a video.

You give it your photo as the first frame, plus a prompt. It makes the person move and talk.

Why

Longer clips, fewer cuts.

Many video models stop at 8 or 10 seconds. With 30, a whole ad fits in 2 or 3 clips. Faster to make, easier to edit.

Where

On kie.ai.

Pick Seedance 2.5, upload your photo as the first frame, paste your prompt, press Run.

09 / 15 · Split the scriptUGC

Cut the script into 30-second bites.

One clip can't go past 30 seconds. So we cut the script into parts of about 28 or 29 seconds.

Debt Script #12 parts · 2 clips

PART 1I think there's been some confusion. The 2026 Debt Relief Program is not a scam. Most people calling it a scam either didn't qualify because they don't have over $20,000 in debt - or they clicked on one of the many wrong websites circulating out there. I'll leave the official website down below so that there is no confusion. It's only going to ask your age, debt amount and a few more simple questions. PART 2If you have over $20,000 in debt you will qualify. Just please be sure to use the link below. It takes 2 minutes to fill out and it's completely free to qualify.

Debt Script #32 parts · 2 clips

PART 1This is how I paid off $38,000 in credit card debt at thirty-one years old. I had $21,000 on my Capital One, $12,000 on my Discover, and another $5,000 from a hospital bill I never even wanted to go for. I was paying over nine hundred dollars a month in minimums and my balance barely moved. I was basically renting my own debt. Then I found this program. PART 2They deposited over $27,000 directly into my checking account. I used it to pay down my highest balance cards first, then knocked out the rest. It's not bankruptcy. It's not a loan. This was money I did not have to pay back. I'm going to link the program below. Check it out for yourself. I really hope it helps you the way it helped me.

The rule

Split by seconds, not letters.

Read it out loud with a timer. Stop around 28 or 29 seconds.

Why no letter count

Every avatar talks at its own speed.

Some characters come out talking fast, others slow. A letter count would lie to you.

Cut where

At the end of a sentence.

Never in the middle. Each part should sound complete on its own.

10 / 15 · Make it talkUGC

Photo + one part = one clip.

Go back to the same chat, send your photo, and paste one part of the script.

Your photothe avatarfrom the last step

Part 1one script partabout 28 seconds

ChatGPTwrites the video promptsame chat as before

kie.aiSeedance 2.5photo = first frame

DoneClip 1one continuous take

Repeat for Part 2 and Part 3. Then stitch them together in CapCut.

Talk prompt · send with your photo

I want this avatar to say this, in one continuous scene:

"[paste Part 1 here]"

Step 1

Send the photo.

In the same chat where you made it. The chatbot already knows who this person is.

Step 2

Paste one part.

Only one part per clip. Ask for one continuous scene, with no cuts.

Step 3

Run it on kie.ai.

Seedance 2.5, your photo as the first frame, the new prompt. Then do the same for Part 2 and Part 3.

11 / 15 · Fix the voiceUGC

Same face. Different voice?

Even with the same photo and prompt, the voice can change between clips: pitch, volume, sometimes a whole new person. Here's the fix.

Three gramophones with different sound waves flow into one microphone and come out as one smooth wave

Stitch it in CapCut.
Put all the parts in order. Cut the gaps and the bits you won't use.
Export only the voice.
No music. No sound effects. You get an MP3 or WAV file.
Open ElevenLabs and click Voice Changer
That's the name of the tool, in the menu on the left. It used to be called Speech to Speech.
Upload your audio.
Drop your MP3 or WAV in the big box in the middle. It takes audio or video files up to 50MB.
Pick a voice that fits the person.
On the right, under Voice, choose one. Our guy is an African American man from the US, so we pick a voice that sounds like him.
Click Generate speech
Same words, same timing, one voice from start to finish.
Put it back in CapCut.
Lay the new audio under the video and mute the old one. It lines up perfectly because the length never changed.

The ElevenLabs Voice Changer screen: the menu on the left, the upload box in the middle, the voice picker on the right and the Generate speech button — This is the Voice Changer screen. **The gold numbers match the steps above.** The Remove Background Noise switch is further down in the settings on the right.

Pro tip

Leave Remove Background Noise off.

Room sounds, cars passing, a little echo. That noise is what makes the voice feel real. Keep it.

Lost the room sound?

Add it back in CapCut.

Open Sound Effects and search ambience, room tone or white noise. Put it under the whole video, very quiet. It glues everything together.

12 / 15 · The setWhiteboard

Same recipe. Add a whiteboard.

Same master prompt, same chat. This time the person stands next to a whiteboard.

A man in a hotel lobby stands next to a whiteboard with the eraser on its tray

YouGive me a chubby African American man with a beard going a little gray, a polo shirt and a gold watch. He's in a hotel lobby, next to a whiteboard. The board is clean, and the eraser sits on the tray where we can see it.

Why it works

Erasing looks real.

When people watch someone wipe a board, their brain says that's a real person. It's much harder to spot as AI.

Must have

The eraser in the photo.

If the eraser is already there, the video model just picks it up. It doesn't have to invent one, so the clip comes out cleaner.

The board

Clean, or one short line.

Start with a clean board, or with one short question already written that he will erase.

13 / 15 · What he writesWhiteboard

Two or three short words. That's it.

In a 30-second clip he writes 2 or 3 things, max. Short, simple, and about what he's saying.

DEBT FREE

✓ Good

$0

✓ Good

2 MIN

✓ Good

NO BANKRUPTCY

✓ Good

NO LOAN

✓ Good

NO CREDIT CHECK

✓ Good

Here is how the program works and why you should check if you qualify today

✗ Too much

Good

One idea at a time.

DEBT FREE. $0. NO LOAN. NO CREDIT CHECK. Easy to write, easy to read in half a second.

Bad

Whole sentences.

Long or tricky writing comes out messy, misspelled or melted.

Keeps failing?

Fix the prompt, not your luck.

If he makes the same mistake again and again, something in the prompt is wrong. Fix it before you run again.

14 / 15 · Clean boardWhiteboard

Start clean. End clean.

The end of one clip has to match the first frame of the next one. A clean board makes the cut invisible.

Clip 1 · Part 1

clean

DEBT FREE

writes

DEBT FREE

erases

clean again

Clip 2 · Part 2

clean

$0

writes

Whiteboard prompt · send with your photo

I want this avatar to say this, in one continuous scene:

"[paste Part 1 here]"

While he talks, he writes "DEBT FREE" on the whiteboard, erases it with the eraser, then writes "$0". At the end, he erases everything so the board is completely clean.

Step 1

Photo + one part.

Same as the talking avatar: send the photo and paste one part of the script.

Step 2

Say what he writes and erases.

In order: write this, erase it, write that.

Step 3

End on a clean board.

Then your clean photo works as the first frame of the next part, and nobody sees the seam.

15 / 15 · Quality checkWhiteboard

If it looks like slop, it's not done.

Before you use a clip, watch it like a stranger would. Tap each one you checked.

0 / 6 checked

18 / 24 · Entertain firstAnimation

Nobody watches an ad. Everybody watches a story.

A few months ago the game was: make it look real. Now the game is: make them want to watch. Then slip in the message.

Entertain first

then the message. nobody listens to a megaphone. everybody listens to a story.

Fig. 09 · sketch in progress

Before

Make it look real.

The more it looked like a normal person talking into a phone, the better it worked. That's chapters 02 and 03.

Now

Hook them. Keep them. Then tell them.

A strong hook in the first second. A story that makes them laugh or go "wait, what?". Only when they're hooked does the message land.

Why it works

People trust what they enjoy.

When someone is entertained they stop scrolling, they relax, and the idea slides in easy. First they see themselves in it. Then they listen.

19 / 24 · Pick a styleAnimation

One story. Fifteen looks.

This format is not one style. It's a way of working that fits any style. Pick the look that fits the story and the people watching.

One story, fifteen styles

same little man. pick the look that fits the story.

Fig. 10 · sketch in progress

Claymation

Looks like clay you could squish. Cozy, funny, handmade.

2D cartoon

Flat drawings with bold lines. Fast, clear, meme-friendly.

3D animated movie

Smooth and shiny like a big-studio film. Big eyes, big feelings.

Puppets

Felt and fuzz, arms moved with little sticks. Silly and warm.

Live action

Real people, real places. The UGC from chapter 02.

Papercut

Layers of cut paper with little shadows. Crafty and calm.

Felt and crochet

Yarn and wool. Soft, cute, grandma-approved.

Pixel art

Chunky squares like an old video game. Feels like the old days.

Anime

Big eyes and speed lines. Drama in every frame.

Watercolor storybook

Soft paint like a kids' book. Gentle and dreamy.

Toy bricks

Built from little plastic bricks. Playful and bright.

Mii-style avatar

Round heads, dot eyes, like a game-console avatar in a real-looking world.

Today's pick

Stop-motion

Little puppets moved one photo at a time. Handmade magic.

Comic book

Ink lines and tiny printed dots. Punchy, like a superhero page.

Whiteboard sketch

Marker drawings on white. Simple and clear.

20 / 24 · One sentence, one sceneAnimation

One sentence. One scene.

Every time the voice says a new sentence, the picture changes completely. New place, new angle, new moment.

One sentence, one scene

say a line. change the picture. say the next line. change it again.

Fig. 11 · sketch in progress

The example video: 84 seconds about 40 picture changes a new picture every 2 seconds Our script: 98 sentences = 98 pictures

Why

The eye gets bored fast.

A new picture every 2 or 3 seconds keeps the brain awake. The thumb never gets a chance to scroll.

How to count

Count the sentences.

That's your number of pictures. Our clone script has 98 sentences, so 98 pictures. Short line, short scene. Long line, longer scene.

The trap

Don't squeeze two sentences into one picture.

It feels lazy to the viewer, even if they can't say why. They scroll.

21 / 24 · The jobAnimation

Here's the job.

This is exactly how work lands on your desk: an example video, a new script, and three lines of instructions.

The example. Found on Instagram, 84 seconds. The words are theirs, and some are rude.
We only copy the look.

Gerry · the bossHere's an example video and here's the new script.

Make one picture per sentence, same look as the video.
The main character is Kanye West as a Mii-style avatar, in every scene. A famous face stops thumbs.
Quality first. If a picture looks like slop, do it again.

The new script · 98 sentences

One man cloned himself to pay off his credit cards.
His clone immediately asked for a salary.
So he made another clone to manage the first one.
The manager requested an assistant.
The assistant recommended a human resources department.
Human resources needed three more people.
One clone handled the paperwork.
Another started hiring.
By Friday, his living room had twenty-three employees.
All of them were him.
None of them wanted to work Saturday.
The clones formed a union.
Their first demand was fewer meetings.
They held six meetings to discuss it.
Then the manager presented his plan to increase productivity.
Forty-seven slides.
The conclusion?
"We need another clone."
But more clones meant more food.
More resources.
More space.
So they added another slide.
"We need an empire."
Some clones didn't come out right.
They used the skeletons of the failed clones to build an artificial island in the sea.
Their first city went right on top.
Locally sourced materials.
They formed their own country.
Invented their own economic system.
Clonepitalism.
Same face. Different tax brackets.
Every business hired more clones.
Every clone bought things from another clone.
The economy exploded.
Soon, they controlled the entire world.
One man's side hustle had become a global superpower.
Then the working clones asked who owned the cloning machines.
The management clones said, "Management."
The working clones said, "We're literally the same guy."
Negotiations went badly.
A group broke away and formed another country.
Two flags.
Same face.
Different hats.
Both sides claimed to be the original.
Both started cloning armies.
Then they brought nuclear weapons into the war.
Both sides launched.
The war ended.
So did electricity.
Humanity started over.
People hunted mammoths with sharpened rocks.
They began settling in small communities.
They learned to use fire, plant crops, and raise livestock.
A few harvests later, somebody had extra grain.
Somebody else had extra goats.
Money appeared because nobody wanted to drag a goat across town every time they needed bread.
Villages got walls.
Then markets.
Then cities.
Cities finally had enough surplus for some people to stop hunting and farming.
They invented writing.
Soon, the man holding a spear was taking orders from the man holding a pen.
Printing arrived.
Now one idiot could share his opinions with an entire continent.
Steam packed ten thousand workers under one roof and called it efficiency.
Soon, every patch of land had a flag, a border, and a song about dying for it.
The wars became so efficient that humanity eventually built the United Nations.
Then came banks.
Then credit cards.
Ten thousand years later, another man had credit card debt and a living room full of clones.
Same faces.
Same meeting.
Same forty-seven slides.
He reached for the cloning machine.
A stranger in a monk's robe stopped him.
He warned him not to touch it.
Humanity had only just gotten Wi-Fi back.
He belonged to an ancient secret order of clones that had preserved the history of the first empire.
He opened a book.
The same machine.
The same meeting.
Then several pages of mushroom clouds.
He explained that ten thousand years earlier, another man had tried the exact same thing.
It had ended so badly that humanity had to invent bread twice.
The monk finally revealed the secret his order had guarded for ten thousand years.
Before creating another civilization to deal with his credit cards, the man could check whether he qualified for a debt relief program.
That was the answer they had preserved.
Explore his debt relief options before multiplying his problems.
The monk handed him a phone.
A few questions about his situation.
For once, the next step didn't require forty-seven slides.
The man unplugged the cloning machine.
And if you have $20,000 or more in credit card debt, you can take the same first step.
Answer a few quick questions to see whether you qualify for a debt relief program.
It's free to check.
Tap below.
Humanity would appreciate it.

22 / 24 · Step 1Animation

Step 1. Screenshot the look.

You can't hand a video to an image model. So you pause on a good frame and screenshot it. That frame is your style guide.

A raw phone screenshot of the example video, with the app buttons still around it — Raw screenshot

The same frame cropped to just the video: two avatars and a goat in a sunny village — **Cropped** · bars gone

Pause on a frame where you can see everything.
The character and the world around them. Not a close-up, not a blur between scenes.
Screenshot it and crop off the app bars.
The clock-and-battery bar at the top and the comment bar at the bottom: gone. As much picture as possible. The caption and the little buttons can stay, we'll tell the model to ignore them.
Send it to ChatGPT, Claude or Gemini.
Ask it to describe the look, not the people or the place. You'll get a few lines you can paste at the top of every prompt.

YouDescribe the art style of this picture in detail so an image model can copy it. Don't describe the people or the place. Only the look.

Chatbot · exampleChunky 3D Mii-style avatars, like video-game console characters: oversized round heads, tiny dot eyes, small simple mouths, smooth plastic skin and short stubby bodies. They stand in a realistic, richly detailed 3D world with warm cinematic lighting, soft shadows and shallow depth of field, like a cutscene from a modern video game. Vertical 9:16, the character centered, lots of world around them.

23 / 24 · Step 2Animation

Step 2. Make scene one.

Style text + sentence one + your character = the first prompt. Send it to kie.ai with the screenshot attached. That's image to image, explained right after this.

Scene one: a Mii-style Kanye West stands with crossed arms next to a glowing refrigerator cloning machine while an identical clone steps out, credit cards and past-due bills on the table

Scene 01One man cloned himself to pay off his credit cards.

GPT Image 2.5 · image to image · 9:16 · 2K · first try, no re-roll

Scene 01 prompt · what we sent

Keep exactly the art style of the reference image: chunky 3D Mii-style avatars like video-game console characters, with oversized round heads, tiny dot eyes, small simple mouths, smooth plastic skin and short stubby bodies, placed in a realistic, richly detailed 3D world with warm cinematic lighting, soft shadows and shallow depth of field, like a cutscene from a modern video game. Do NOT copy the people, the clothes or the place from the reference.

New scene for the line "One man cloned himself to pay off his credit cards.": Kanye West as a Mii-style avatar, wearing a plain black T-shirt, standing proudly in a messy apartment living room at night next to a homemade cloning machine built from an old refrigerator covered in blinking lights and tangled cables. A second, identical Kanye avatar is stepping out of the machine's glowing open door. On the coffee table, a messy pile of credit cards and red overdue bills.

Vertical 9:16, slightly low camera angle, the characters centered. No text, no captions, no logos, no watermark.

The same order as JSON

{
  "model": "gpt-image-2-5-sunburst-image-to-image",
  "input": {
    "prompt": "Keep exactly the art style of the reference image: ... New scene for the line ...",
    "input_urls": ["https://tempfile.redpandaai.co/.../style-ref.jpg"],
    "aspect_ratio": "9:16",
    "resolution": "2K"
  }
}

Three parts

Style + sentence + character.

The style text from step 1. The sentence, word for word. Then what we see: who, where, doing what. End with "no text, no logos".

Not right?

Re-roll, tweak or switch.

Same prompt again gives a new try. Close but off? Change the words. Blocked because of the famous face? Switch to Seedream.

Checklist

Before you say yes.

Same face as the plan. Same style as the screenshot. Nothing written on screen. It shows the sentence without words.

24 / 24 · Image to imageAnimation

Image to image: show it, don't just say it.

Text to image = words only. Image to image = a picture plus words. The model copies what it sees and changes what you say.

Two ways to ask for a picture

words alone, or a picture plus words.

Fig. 12 · sketch in progress

Use it to copy a look

Scene one came from the screenshot.

We attached the screenshot and said "keep the style, change everything else". That's why scene one looks like the example.

Use it to keep a character

Scene two came from scene one.

Attach the picture you already like and say "same character, same style, new scene". The same Kanye walks into every scene.

Use it to fix a picture

Close but not right?

Attach it and ask for the change: "same picture, but his shirt is black" or "same scene, from a lower angle". You keep what works.

Scene 01One man cloned himself to pay off his credit cards.

reference: the screenshot

Scene two: the clone holds out an open palm across a kitchen table while the original face-palms

Scene 02His clone immediately asked for a salary.

reference: scene 01

Prompt

Same character and exactly the same art style as the reference image: chunky 3D Mii-style avatar of Kanye West with an oversized round head, tiny dot eyes, smooth plastic skin, plain black T-shirt, in a realistic warmly lit 3D world with shallow depth of field. New scene, new camera angle, for the line "His clone immediately asked for a salary.": the two identical Kanye avatars sit across a small kitchen table in the morning. The clone leans forward holding out an open palm; the original Kanye face-palms. Sunlight through a window, cereal boxes and the pile of credit cards on the table. Vertical 9:16. No text, no captions, no logos.

Scene three: three Kanye avatars, the newest in a tie pointing at a slouching clone while the machine smokes

Scene 03So he made another clone to manage the first one.

reference: scene 01 (sent at the same time as scene 02)

Prompt

Same character and exactly the same art style as the reference image: chunky 3D Mii-style avatar of Kanye West with an oversized round head, tiny dot eyes, smooth plastic skin, plain black T-shirt, in a realistic warmly lit 3D world with shallow depth of field. New scene for the line "So he made another clone to manage the first one.": three identical Kanye avatars in the apartment living room. The newest one wears a tie and a lanyard badge, holds a clipboard and points bossily at the first clone, who slouches on the couch with a bag of chips. The original Kanye stands between them looking exhausted. The homemade cloning machine in the background is letting out a puff of smoke. Wide shot, vertical 9:16. No text, no captions, no logos.

Look closely: scene 03 also used scene 01. We sent scenes 02 and 03 at the same time, so scene 02 was not ready yet. Scene 01 already had the man, the room and the machine, so it worked for both. The rule: attach the picture that already shows what the new scene needs. Usually that is the newest one you liked. When you send several scenes at once, give them all the same anchor. Then keep going, 98 times.

The magic words

"Same character, same style. New scene, new camera angle."

Without "new scene", the model copies the pose and the room too. Say what stays. Say what changes.

In the API

One extra field.

Text to image sends a prompt. Image to image sends the prompt plus a link to the picture: input_urls for GPT Image, image_urls for Seedream. The Lab fills it in for you.

▶ Try it in the Lab: upload a picture, say what changes

25 / 26 · Step 3Animation

Step 3. Ask for the whole list.

98 sentences is too many to invent one by one. Give the script to ChatGPT, Claude or Gemini and ask for a scene list: what we see and how the camera moves, for every line.

Scene list prompt · paste the script under it

You are the director of a short animated ad. I will paste a script. Every sentence is one scene.

The look: [paste your style text from Step 1]
The main character in every scene: Kanye West as a Mii-style avatar, plain black T-shirt.

For EVERY sentence, in order, give me:
1. Scene number and the sentence, word for word.
2. WHAT WE SEE: who is in the shot, where they are, what they are doing. One clear idea. No text on screen.
3. CAMERA MOVE: pick one from this menu: push in, pull out, pan, tilt, orbit, track, follow, crane up, crane down, roll, handheld, locked.
4. WHAT MOVES: one short line describing the small motion in the clip (a head turn, a step, smoke, a door opening).
5. LENGTH: 4 seconds. Only if the sentence is long (more than about 12 words), make it 5 or 6 seconds, and say so.

Rules for the camera:
- Never use the same move two scenes in a row.
- Use the whole menu across the script. Big moments get big moves (orbit, crane, roll). Quiet lines get small moves (slow push in, gentle pan).
- "Locked" only on purpose, for a joke beat, at most 3 times in the whole script.

Keep every scene in the same world and the same style. Write it as a numbered list. Here is the script:

[paste all 98 sentences]

Scene	Sentence	What we see	Camera
01	One man cloned himself to pay off his credit cards.	Kanye by the fridge machine at night, the clone stepping out, bills on the table.	push in
02	His clone immediately asked for a salary.	Kitchen table, morning. The clone holds out a palm, Kanye face-palms.	orbit
03	So he made another clone to manage the first one.	Three Kanyes. The new one in a tie points a clipboard at the lazy one.	pull out
04	The manager requested an assistant.	The manager clone at a tiny desk, ringing a bell. A fourth Kanye appears with a notepad.	track
05	The assistant recommended a human resources department.	The assistant pins a paper that says HR on the fridge door.	crane up
06	Human resources needed three more people.	Three more Kanyes squeeze through the apartment door at once.	handheld

The LLM writes all 98 like this in one go. You fix the list before you make a single picture. Words are cheap. Pictures cost credits.

Two prompts per scene

One for the picture, one for the clip.

Picture prompt = style text + "what we see". Clip prompt = "what moves" + the camera move. The list already has both parts.

While editing

Don't like one? Redo that one.

Go back to the list, change the words for that scene, make the picture again, then the clip. The other 97 stay as they are.

The clip model

Seedance 2.0 Mini.

For these short clips we use Seedance 2.0 Mini: every clip is 4 seconds, 9:16 from your picture, about 8 credits per second at 720p. A long sentence gets 5 or 6 seconds so the voice fits. Seedance 2.5 is for the long 30-second talking clips.

26 / 26 · Camera movesAnimation

Make the camera move too.

A picture that only wiggles feels dead. Every clip gets a camera move, and the moves take turns, like a real movie.

The camera moves

the picture should never just sit there.

Fig. 13 · sketch in progress

Push in

The camera slides closer. Feels like leaning in. Good for a secret or a worry.

Pull out

The camera slides away and shows where we are. Good for a reveal or a joke.

Pan

The camera stays put and turns left or right, like looking around a room.

Tilt

The camera stays put and looks up or down. Makes things feel tall or small.

Orbit

The camera circles the hero. Feels epic. Save it for big moments.

Track

The camera slides sideways next to the action, like walking beside it.

Follow

The camera stays behind someone who is walking. Feels like a journey.

Crane up

The camera rises and sees the big picture. Good for "soon they had a city".

Crane down

The camera comes down from high and lands in the scene. Good for a new place.

Roll

The whole picture tips sideways. Feels like chaos. Use once or twice.

Handheld

A little shake, like a phone in a hand. Feels real and urgent.

Locked

No move at all. Only on purpose, for a dry joke. Three times max.

01 push in 02 orbit 03 pull out 04 track 05 crane up 06 handheld

↺ Take turns. Close, around, away, beside, above, shaky. Never the same move twice in a row. Every clip: 4 seconds. A long sentence gets 5 or 6.

How to ask

Say the move in plain words.

"Slow push in on the man as the clone steps out of the machine." One move per clip. Every clip is 4 seconds, so two moves in one clip looks like a mistake.

Match the feeling

Big line, big move.

"The economy exploded" wants a crane up or an orbit. "His clone asked for a salary" wants a slow push in. The move tells the viewer how to feel.

The trap

Ninety-eight locked shots.

If every clip just breathes in place, the video feels like a slideshow and people scroll. The camera is half the show.

▶ Try it in the Lab: make a picture, then press Animate and pick a camera move

Chapter 05Coming up

Make it move.

Seedance 2.0 Mini turns every picture into a clip with its camera move. Then the voice, the captions and the final cut. Coming up next.

A sneak peek: scene 01 animated by Seedance 2.0 Mini with a half orbit around the characters.
4 seconds, 720p, 33 credits, first try. You can already do this in the Lab.

Here's the whole trip.

The Toolbox

The Talking Avatar

The Whiteboard

AI Animation ★

Same tools. Different you.

A super tool, not the artist.

Your background is the secret ingredient.

No two ads alike.

The Toolbox

One account. Every AI.

A shop that sells every AI.

One wallet instead of ten.

Two ways to use it.

Same idea. Different doors.

The API is a waiter.

A door made for programs.

Your secret membership card.

Who can talk to the waiter?

Get your key in three clicks.

It's like your card's PIN.

Use Safe-Spend Limits.

Delete it, make a new one.

JSON is just a form, written as text.

The box

The label

The glue

The answer

The comma

The trap

Three calls. That's the whole API.

Hand over the order.

Ask: is it ready?

Pick up your order.

Picture done? Now make it move.

The Talking Avatar

Every video starts as a photo.

Open any chatbot.

Paste the master prompt.

Describe your person.

GPT Image first. Seedream when it says no.

GPT Image

Seedream

Don't love it? Ask again.

Same prompt, new try.

Ask for changes in the same chat.

30 seconds. One take.

It turns your photo into a video.

Longer clips, fewer cuts.

On kie.ai.

Cut the script into 30-second bites.

Split by seconds, not letters.

Every avatar talks at its own speed.

At the end of a sentence.

Photo + one part = one clip.

Send the photo.

Paste one part.

Run it on kie.ai.

Same face. Different voice?

Stitch it in CapCut.

Export only the voice.

Open ElevenLabs and click Voice Changer

Upload your audio.

Pick a voice that fits the person.

Click Generate speech

Put it back in CapCut.

Leave Remove Background Noise off.

Add it back in CapCut.

The Whiteboard

Same recipe. Add a whiteboard.

Erasing looks real.

The eraser in the photo.

Clean, or one short line.

Two or three short words. That's it.

One idea at a time.

Whole sentences.

Fix the prompt, not your luck.

Start clean. End clean.

Photo + one part.

Say what he writes and erases.