Can GPT-Image-2 Return Multiple Images? Use the Image API n Parameter

Yes—when you use the official OpenAI Image API, set the n parameter to the number of final GPT-Image-2 candidates you want. The default is one image. A request with n: 4 is meant to return four separate final image payloads in result.data, not one four-panel collage.

That answer applies to the Image API. It does not automatically describe the ChatGPT UI, a Responses API image-generation tool call, or a gateway that happens to use a similar model name. Those routes may offer a different control—or no direct count control at all.

The direct API answer

OpenAI's Image Generation guide documents n for generating multiple images in one Image API request. Treat the resulting array as a set of candidates: save them separately, check how many actually arrived, and select before spending on more runs.

Generate and save four final images

This Node example requests four 1024px square WebP candidates and writes each returned payload to a distinct file. It deliberately does not assume a particular response count before checking the array.

javascript
import OpenAI from "openai";
import { writeFile } from "fs/promises";

const openai = new OpenAI();

const result = await openai.images.generate({
  model: "gpt-image-2",
  prompt: "A studio product photo of a matte black desk lamp on a pale gray background, centered with soft side light. No text, logos, or people.",
  n: 4,
  size: "1024x1024",
  quality: "low",
  output_format: "webp",
});

if (!result.data?.length) {
  throw new Error("The request completed without final image data.");
}

await Promise.all(
  result.data.map((item, index) => {
    if (!item.b64_json) {
      throw new Error(`Image ${index + 1} has no base64 payload.`);
    }
    return writeFile(`lamp-candidate-${index + 1}.webp`, Buffer.from(item.b64_json, "base64"));
  })
);

console.log(`Saved ${result.data.length} final candidate(s).`);

The API's default output is base64 image data; output_format: "webp" makes the saved extension match the requested format. If your SDK or wrapper uses a different parameter name or returns URLs, consult that route's own documentation instead of copying this example unchanged.

What n does—and does not—promise

Use a batch when the candidates will be evaluated under the same acceptance rule—for example, four product-shot compositions with the same subject, aspect ratio, and brand restrictions. Do not batch unrelated jobs just because they can share an API call; separate prompts make failures, cost attribution, and review harder to understand.

Do not confuse final candidates with streaming previews

Streaming has a different purpose. The official guide lets you ask for partial_images while an image is still rendering. It may send zero to three intermediate preview events before the final result, and a fast render can send fewer previews than requested. Those events help make a UI feel responsive. They are not the same thing as requesting n: 3 final images.

The practical distinction is simple:

• Want three images you can compare or persist? Use n: 3 in the Image API generation call. Want to show progress while one image renders? Use stream: true with partial-images.
• Want three scenes inside one poster? Ask for a three-panel layout in the prompt, then expect one final file.

Pick the correct surface first

This matters when debugging a result that has only one image. First record the endpoint, SDK method, model field, and raw response shape. If the code never calls the Image API generations endpoint with n set, changing the English prompt will not add a batch parameter. If a provider accepts n but returns one item, treat that as a provider-contract issue until it reproduces against the official route.

Batch safely: save, select, then refine

Multiple outputs are useful only if you have a selection rule. Before submitting a batch, define the few observable checks that matter: correct subject, readable text, no prohibited logo, correct crop, or a required layout. Save the prompt version and requested settings alongside each filename. Then choose the strongest candidate before raising quality, making a targeted edit, or sending another batch.

Each requested final image can consume generation capacity and create cost. Start with a small, reviewable candidate set and lower quality for draft composition checks. Once one direction passes, make the next request narrower: preserve the winning composition, change one detail, or request the final quality. Do not blindly rerun a large batch because one candidate missed the mark; that obscures what improved and increases review work.

For production automation, also handle partial success. Count the returned items, store the request identifier or your own correlation ID, and mark the job incomplete if the count does not meet your downstream requirement. A response with valid image data is not proof that every candidate satisfies your content, brand, or safety acceptance rule.

Related next step

• GPT-Image-2 API Guide
• GPT-Image-2 API Pricing
• GPT-Image-2 API Pricing can help you estimate a batch before it enters a production queue.

FAQ

Can GPT-Image-2 return four separate images in one API request?

Yes. With the official Image API, use model: "gpt-image-2" and n: 4, then save each item in the returned image data array. The default is one image when n is not supplied.

Are partial_images the same as multiple final images?

No. partial_images are intermediate stream previews for one rendering operation. Use n when you need several final candidates to compare or store.

Why did I get a collage instead of several files?

The prompt may have requested a multi-panel composition, or your route may have produced one image by design. A collage is one final output. Check that you are calling the official Image API generations endpoint and that n is an actual request parameter rather than plain prompt text.