AI Carousel Maker for Instagram: How to Build a Slide Series People Actually Swipe Through

A carousel lives or dies on one thing: every slide needs to read like part of the same document, not a folder of loosely related pictures. This is exactly where an AI carousel maker for Instagram trips up. Each generation is independent as far as the model is concerned, and it has no memory of the slide that came before.

Different carousel types need different build approaches, so let's go through them one at a time.

The step-by-step carousel

This is the format you see most: one action per slide, moving from the starting point to the result.

What the images need is a matching angle and the same scene, with only the state of the object changing. That's the hardest thing to pull off, because the model rebuilds the frame from scratch every single time.

Here's what actually works: generate the first slide until you're happy with it, then feed it back in as a reference for the rest and change one word in the prompt at a time. Not the whole sentence, just the specific word that controls what's different.

Add the captions on top afterward in any editor. Asking the model to render step numbers directly inside the image is almost always a bad idea. The numbers drift and end up wrong.

The roundup carousel

Five products, five ideas, five examples. The slides aren't connected by a story, but they still need to feel like one visual set.

Here consistency is easier to pull off than in the step-by-step format. Consistency comes from the color palette and the lighting, not from matching composition. Lock down the background and lighting description in your prompt and change only the object.

If these are real products you sell, generate off their actual photos. Otherwise the carousel will show something close but not quite right, and the customer gets a different item than what they saw. For product shots, Nano Banana and Seedream 5.0 are solid picks.

The carousel with text built into the image

Here the text is part of the design, not a caption layered on top afterward.

Your model options narrow fast here. Legibility still varies from one model to the next, especially once the text gets longer or more stylized. Ideogram 3.0 and GPT Image 2 handle it well, while plenty of others produce letters that look fine from a distance and fall apart the moment you zoom in. We put this to the test across a dozen models.

Honestly, I almost always add the text by hand afterward anyway. It's faster, more reliable, and I end up with my own font instead of whatever the model decided to invent.

What goes on the first and last slide

The first slide decides whether anyone keeps swiping at all. It has to work without the caption underneath, one glance should be enough to tell what the post is about.

The last slide is where you ask for the action. Not "thanks for watching," something specific: save this, visit the link, send a message.

The middle slides can be a little weaker, people skim through them fast anyway. If your generation budget is tight, spend your best attempts on the first and last slide.

Where generation falls apart

Three failures show up again and again.

Background texture that shifts between slides. The eye catches this instantly even when the colors technically match. A reference image fixes it.

Object scale that jumps around. The product is large on one slide and small on another. Describe the distance to the subject in words and don't touch that part of the prompt between slides.

Time of day that changes without warning. Left on its own, the model will happily give you morning light on one frame and evening light on the next. Spell out the lighting explicitly every time.

How I actually build an eight-slide carousel when the generator still works one slide at a time

Here's my process for an eight-slide set.

First, I sketch out on paper what goes on each slide. Skip this step and generation turns into endless trial and error.

Then I generate the first slide until it's genuinely right. More than half my attempts go here, and that's normal.

How many attempts that takes varies every time for me. Sometimes it's ten, sometimes forty. I haven't found a pattern.

The remaining seven get generated using that first slide as a reference, changing as few words as possible each time.

Last step: assemble everything, add the text, and look at all eight slides in a row on my phone. Not a big screen. A carousel lives in the feed on a small one.

Eight slides built off a single reference takes about forty minutes, picking the best shots included. Doing one slide per day across a week takes twice as long and looks noticeably less consistent.

FAQ

How do I keep carousel slides in the same style? Generate the first slide, then use it as a reference for the rest, changing as few words in the prompt as possible. Keep the description of lighting, palette, and background fixed throughout.

Can I ask the AI to render the text directly on the slides? You can, but legibility still varies from one model to the next, so it's worth testing before you commit to one. It's usually safer to add captions on top afterward in an editor: you control the font, and the letters won't come out warped.

How many slides is the right number? Six to eight works well for a step-by-step carousel, five to seven for a roundup. Past ten, people rarely scroll all the way through, so there's not much point spending generation budget on slides nobody reaches.

What should I generate first? The first slide, and don't be stingy with attempts on it. It decides whether anyone opens the carousel at all, while the middle gets a quick scan at best.

What aspect ratio should I use? Vertical 4:5 takes up more of the screen in the feed. A 1:1 square is the safer bet if the slides might get reused elsewhere. Either way, keep every slide in the same ratio, otherwise the image jumps as people swipe.

Sources

1. Flami: putting a dozen AI image models to the test for in-image text
2. Flami: AI image generation for social media
3. Flami: templates for a product series