Create a social media carousel with AI

Three square visuals that read in sequence, cover, tip, call to action, written by GPT Image 2. The art direction is wired to all three. The full workflow is visible without an account.

Interactive demo

Real workflow, real results. Pan and zoom freely.

Use this workflow

Real workflow: one Art direction node wired to three Image nodes on GPT Image 2, one prompt per visual. The three tiles of the demo carousel.

At a glance

What the workflow produces: 3 images in 1:1
Canvas structure: 8 nodes wired by 6 connections
Models wired in: GPT Image 2
Cost of one full run: about 24 credits, roughly $0.29

A carousel is recognised before it is read. Before the first word, the eye sees three tiles that belong together: same colours, same typography, same margins. That is what this workflow builds. The Art direction node carries the house style once, it is wired to the three Image nodes, and each visual prompt only deals with its own text and illustration.

The model is GPT Image 2, and that choice is not neutral: it is the text INSIDE the image that makes or breaks a carousel, and it is the most reliable in the catalogue at writing a short sentence without deforming it, accents included. The demo shows it on three French visuals, « 3 erreurs qui gâchent votre café », « 1. L'eau bouillante », « Venez goûter la différence ».

Two rules are enough to get the text right. Give it in quotes after the word "exactly", and give little of it: one headline and one sentence per visual. The rest, the palette, the flat areas, the paper texture, lives in the art direction. The three visuals cost twenty-nine cents, twenty-four credits.

How it works

1
Write your house style in the Art direction node
Palette, typography, background, margins, illustration style. Once, for all three. That node is what makes the set consistent, not a sentence copied into three prompts.
2
One idea per visual, not three
A cover that hooks, content that delivers, a call to action that closes. A carousel that puts two ideas on one visual gets read for neither.
3
Give the text in quotes
"Large text, in English, exactly: …". That formula is what stops the model rewording, shortening or translating your headline. Then check the accents on the render.
4
Generate all three, read them, fix one
"Generate all" produces the whole carousel. If a letter falls badly, change that text and rerun that node: the other two keep their render and the house style does not move.

Variations and use cases

The same canvas, tuned for different needs: each variation is two or three lines of prompt away.

The tips carousel

The format that gets saved the most: a numbered promise on the cover, one tip per visual, a call to action at the end. It is the one in the demo, and it works for any trade.

Announcing something new

The product on the cover, what it changes in the middle, where to buy it at the end. The art direction picks up your brand colours, and the series is recognisable from one post to the next.

One number per visual

A report, a review, a study: pull out three numbers and give one per visual, very large. It is the most shared carousel, because it reads without stopping.

Before, after, how

Three visuals telling one transformation. The consistency of the house style is what makes it read as a single story rather than three posts placed side by side.

The weekly series

Keep the canvas, change the texts every week. The art direction does not move, so the account has a visual identity without anyone rebuilding it.

The carousel, then the thumbnail

A carousel cover and a YouTube thumbnail obey the same rules: few words, very large, one subject. The neighbouring workflow handles the 16:9.

Common mistakes

Copying the house style into all three prompts

It works until the day you change a colour: you then have to change it three times, and one miss breaks the set. The Art direction node exists precisely for that.

Putting a paragraph in the image

Beyond one or two sentences the letters deform and reading falls apart. Long text goes in the post caption, not in the visual.

Forgetting the word "exactly"

Without it the model rewords, shortens or translates your headline. With it, it copies. One word changes the whole result.

Three unrelated ideas on three visuals

A carousel is not three posts. It needs a promise on the cover and a progression: without it, nobody swipes to the third.

Recommended models

Frequently asked questions

Does the model really write the text without mistakes?

On a short sentence given in quotes, yes, and the three demo visuals show it in French with their accents. On a paragraph, no: no image model holds that length, split it across visuals.

Can I make more than three visuals?

Yes. Duplicate a Text node and an Image node, wire the same art direction, and the fourth visual arrives in the same style. Count eight credits more per visual.

Which format for Instagram and LinkedIn?

The square of the demo works everywhere, it is the safest format from one platform to the next. The node also offers 4:3 and 3:4 if you publish on a single network.

How much does a carousel cost?

Twenty-four credits for the three visuals, eight each. The price shows on each node before you click, and redoing a single visual costs only that one.

Can I use my real brand?

Yes. Put its name in the visual's text and its colours in the art direction. For a pixel-perfect logo, add it afterwards in your layout tool: an image model always redraws it slightly.

How big are the visuals?

About 1250 pixels a side, well above what the networks ask for. For print, the Enhance node scales them up without reinventing the text.

Can I do it from a phone?

Yes, and it is the natural use: you write the three texts on the mobile canvas and publish right after. The create on mobile guide explains the gestures.

Can the model write in other languages?

Yes, in all fourteen on the site. Give the sentence in the language you want and state it in the prompt, as the demo does with "in French".

Create a social media carousel with AI

Use this workflow

Free account, no card required. The workflow opens pre-filled in your canvas.