text-to-image

Edit images from text - GPT Image 2

Edit or recompose one or more input images from a text prompt using OpenAI's GPT Image 2.5 model.

Key Features:

Accepts 1-16 input images, used together as the visual context
Restyle a single image, or compose a new one from several
Same output controls as text-to-image (resolution, aspect ratio, format, compression, quality)

Best for:

Product and scene variations from an existing shot
Combining a subject, a background and a style reference in one call
Adding or correcting text on an existing image

Output: one edited image per num_images (up to 10), delivered as CDN URLs in the task's generated array. The input images are never modified.

Credits: editing sends your images to the model and their input tokens dominate the cost, so it is priced differently from generation: flat units per output image by quality (low x1, medium x4, high x6), with resolution not affecting the price.

The request is processed asynchronously: the endpoint returns a task id immediately. Retrieve the result by polling the task endpoint or by providing an optional webhook_url.

post/v1/ai/text-to-image/gpt-image-2-5-edit

Request

Base URL: https://api.magnific.com
URL: https://api.magnific.com/v1/ai/text-to-image/gpt-image-2-5-edit
Auth: API key in header x-magnific-api-key

Request body

promptstring required

Text instruction describing the edit to apply to the reference images.

Tips:

Say what should change, not what the image already contains
Several references are used together as one visual context
Text inside the image can be corrected or replaced verbatim

Examples:

Simple: "Put the product on a marble surface"
Detailed: "Replace the background with a sunlit studio, keep the product and its shadow, change the label text to ORGANIC"

reference_imagesstring[] required

The images to edit, in order. GPT Image 2.5 uses them together as the visual context for the prompt: it can restyle one image, or compose a new image from several.

With the default aspect_ratio: auto, the output keeps the aspect ratio of these images.

Each image may be up to 20 MiB, and one request may carry up to 64 MiB of images in total.

webhook_urlstring uri

Optional callback URL that will receive asynchronous notifications whenever the task changes status. The payload sent to this URL is the same as the corresponding GET endpoint response, but without the data field.

num_imagesinteger

How many images to generate for this prompt. Each image is billed separately.

variant'flare' | 'sunburst'

Which model of the GPT Image 2.5 line renders the request. Both share every other parameter, the quality scale and the price.

flare: optimised for latency (default)
sunburst: optimised for editing precision

resolution'1k' | '2k' | '4k'

Resolution tier of the generated image. Combined with aspect_ratio it selects the exact pixel size.

With aspect_ratio: auto, only 1k leaves the size to the model; 2k and 4k render a square at that tier.

Credits: the tier multiplies the price set by quality — 1k x1, 2k x2, 4k x3. On the edit endpoint it does not affect the price.

Shape of the generated image, expressed as *_width_height (e.g. square_1_1, widescreen_16_9).

auto (the default) lets the model choose the shape. When the request carries reference_images, the output keeps their aspect ratio, including ratios that are not in this list — which is what makes auto the right choice for editing a picture whose proportions must be preserved.

quality'low' | 'medium' | 'high' | 'xhigh' | 'max'

Render quality tier. Higher tiers spend more provider compute; lower tiers are faster.

low: fastest, for drafts and iteration
medium: balanced (default)
high: high detail and prompt adherence
xhigh, max: the extended tiers of GPT Image 2.5, for the most demanding renders

Credits: low, medium and high are billed on their own tier. xhigh and max are currently billed at the high price.

output_format'png' | 'jpeg' | 'webp'

The file format of the generated image.

png: lossless, supports transparency. Best for logos, text and sharp edges
jpeg: compressed, no transparency. Smallest files for photographic content
webp: compressed, supports transparency

output_compression only applies to jpeg and webp. A jpeg output cannot carry an alpha channel, so it is rejected when background is transparent.

background'transparent' | 'opaque' | 'auto'

How the image background is produced.

transparent: render without a background. Requires an output_format with an alpha channel (png or webp)
opaque: always render a filled background
auto: let the model decide

When omitted, the provider default applies.

output_compressioninteger

Compression level for the generated image, where 100 is the least compressed. Only applies when output_format is jpeg or webp; ignored for png.

moderation'auto' | 'low'

Strictness of the model provider's own content filter.

auto: the provider's standard filtering (default)
low: less restrictive filtering

This is the model provider's filter and is applied in addition to Freepik's platform moderation, which cannot be disabled.

Example request

{
  "prompt": "Put the product on a marble surface",
  "reference_images": [
    "https://example.com/product.jpg"
  ],
  "webhook_url": "https://www.example.com/webhook",
  "num_images": 1,
  "variant": "flare",
  "resolution": "1k",
  "aspect_ratio": "auto",
  "quality": "medium",
  "output_format": "png",
  "background": "auto",
  "output_compression": 80,
  "moderation": "auto"
}

Response

OK - The request has succeeded and the GPT Image 2.5 Edit process has started.

Example response

{
  "data": {
    "generated": [
      "https://openapi-generator.tech",
      "https://openapi-generator.tech"
    ],
    "task_id": "046b6c7f-0b8a-43b9-b35d-6489e6daee91",
    "status": "CREATED"
  }
}

Changes

No recorded changes to this operation across all 1 revision of this API.