AI Instagram carousel post generator: same size, same style

An AI Instagram carousel post generator has to make slides that match. Generate every slide at the same aspect ratio, 4:5 portrait or 1:1 square, because Instagram crops all carousel images based on the first one. Then keep one look across the set: repeat the same style description in every prompt and pass an approved slide to the model as a reference image.

With Sume, each slide is one POST /v1/images call to a model that lists 4:5. Sume facts come from the Image API docs and current API code; Instagram's numbers come from Meta's content publishing guide and IG User media reference. All were read on 2026-09-27.

What size should Instagram carousel slides be?

Pick one shape for the whole post. For images published through Instagram's API, Meta documents these rules; 4:5 is the tallest shape they allow, and 1:1 is the carousel default:

How do I generate carousel slides with an AI image API?

Make one request per slide, because each slide shows something different. n gives you several takes of the same slide to pick from, not a sequence. The Image API docs allow up to 10 images per request with lower ceilings per model, and current code caps most models at 4.

• Pin a model that lists 4:5 (or 1:1) and send the same aspect_ratio on every slide. A model only accepts the values its catalog descriptors list; the aspect ratio guide shows each model's list.
• Ask for output_format: "jpeg" if you publish through Instagram's API, which takes JPEG only.
• Keep one style block, such as palette, light, and framing, word for word in every prompt, and change only what the slide shows.

curl -X POST "https://api.sume.com/v1/images" \
  -H "Authorization: Bearer $SUME_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "openai/gpt-image-2.5",
    "prompt": "Slide 1 of 5, cover: a ceramic mug on a linen table. Style: warm film photo, soft window light, cream and terracotta palette.",
    "aspect_ratio": "4:5",
    "output_format": "jpeg",
    "n": 4
  }'

How do I keep every slide in the same style?

Reuse the slide you approved as a reference. Send its public HTTPS URL in input_references on every later request, and say in the prompt what stays and what changes, as the request body below does for slide two.

• ChatGPT Image 2.5 takes up to 16 reference images. Models whose input_references range is 0 to 0 are text-to-image only and reject references.
• Localhost, private-network, and non-HTTPS reference URLs are rejected before submission.
• There is no seed to lock a look: seed is in the schema but not served, so sending it returns 400 unsupported_parameter.
• Each slide is still a new generation, so compare the set before you post. More on this: consistent characters across images.

{
  "model": "openai/gpt-image-2.5",
  "prompt": "Slide 2 of 5: the same mug, held in two hands. Keep the style, palette, light, and framing of the reference image.",
  "aspect_ratio": "4:5",
  "output_format": "jpeg",
  "input_references": [
    { "type": "image_url", "image_url": { "url": "https://example.com/carousel/slide-1.jpg" } }
  ]
}

Can the generator output an exact pixel size?

Not every size. On the ChatGPT Image models, custom sizes need both edges in multiples of 16, which rules out 1080×1350, and current code serves their aspect_ratio: "4:5" at 1024×1280. That width is inside the 320 to 1440 range in Meta's docs. If you need exact pixels, resize the downloaded file yourself; the aspect ratio guide covers custom sizes on other models.

What does a carousel cost, and what are the limits?

• Every image in n is billed, plus a 5.5% agent fee by default, so five slides with four takes each is 20 billed images. ChatGPT Image 2.5 is priced by tokens, so its size and quality set the price of each one. A failed generation is not billed; AI image generator API cost lists every model's price.
• Slow configurations, a large n among them, can answer 202 with a job instead of 200 with the images. Poll GET /v1/jobs/{id}/status, then fetch GET /v1/jobs/{id}/result.
• Result URLs are Sume-hosted and signed, so download the slides you keep.
• Check any text the model draws letter by letter before you publish; AI images with text covers that case.