Transparent backgrounds arrive in the GPT-Image-2 API

The GPT-Image-2 API can now generate transparent-background PNG and WebP files in preview for websites, mockups, presentations, and campaigns.

A product image, icon, or illustration generated against a white background generally requires another step before it can be used on a website, in a presentation, or across a campaign. The subject must be cut out, its edges corrected, and any fine or translucent areas inspected. GPT-Image-2 can now generate files with transparent backgrounds directly through OpenAI’s API.

The feature remains in preview. To activate it, a request must use `background="transparent"` with the `gpt-image-2` model. OpenAI’s official documentation states that the output must be saved as PNG, the default format, or WebP. JPEG does not support this option.

The resulting file contains an actual alpha channel that controls the opacity of different areas within the image. The subject can therefore be placed against a light, dark, colored, or illustrated background without retaining a visible rectangle around it. The same asset can then move between an online store, an advertisement, a presentation, a website mockup, or a catalog.

The difference from a background-removal tool lies in when transparency is created. A cutout service analyzes an existing image to separate the subject from its surroundings. GPT-Image-2 generates the subject and its alpha channel within the same file. This approach may preserve some difficult edges more effectively, but it does not guarantee a production-ready result in every case.

OpenAI demonstrates the feature using clear and frosted glass objects, translucent ribbons, thin reeds, plant fibers, and fine cords. These elements regularly cause problems for cutout tools, which may remove part of the subject, create halos, or make transparent materials appear opaque. OpenAI’s practical guide shows how to verify that an alpha channel is actually present before integrating the file.

However, the preview was not accompanied by independent comparisons or edge-accuracy measurements. Hair, fur, glass, smoke, soft shadows, and partially transparent objects will still require inspection. A file can technically contain an alpha channel while retaining fringes, uneven cutouts, or incorrectly interpreted opacity.

The prompt also remains decisive. OpenAI says its instructions can take precedence over the transparency setting. If the request describes a studio, backdrop, background color, or complete scene, the model may generate that environment despite the use of `background="transparent"`.

The company therefore recommends requesting an isolated subject that is fully visible and has no backdrop, frame, stand, or cast shadow, while explicitly stating that a genuinely transparent background is required. This wording becomes particularly important when the object itself contains transparent components that need to remain visible.

The feature retains GPT-Image-2’s other output controls. Developers can select the dimensions, choose low, medium, or high quality, and specify the file format. The documentation lists common formats such as 1024 × 1024, 1024 × 1536, and 1536 × 1024, while noting that GPT-Image-2 supports more flexible dimensions when they meet the service’s requirements.

The use cases presented by OpenAI begin with e-commerce. A brand can generate a product once and place it against several seasonal backgrounds without repeating the generation. This method does not guarantee consistency between products created separately, however. Shapes, proportions, labels, and brand details must still be checked against their references.

Transparency can also be used to create icons, stickers, ornaments, interface elements, or graphics intended for design templates. For print-on-demand products, the same design can be applied to several garments or objects, provided that colors, margins, and the printer’s technical requirements are checked beforehand.

OpenAI also presents generated charts for use in presentations. A diagram without a background can sit on a dark slide or gradient without introducing a white frame. This use requires an additional precaution: a generated image remains a graphic rendering rather than a visualization calculated directly from source data.

The documentation recommends checking labels, values, and proportions before professional use. When strict numerical accuracy is required, the chart should be produced with a deterministic tool. Image generation can instead be reserved for illustrations and decorative elements.

The announcement currently concerns the API. OpenAI does not say whether the same setting will be exposed in this form within ChatGPT. It has also provided no timeline for the feature to leave preview or commitment regarding the stability of its behavior.

Native transparency can therefore remove one possible operation from the production pipeline, but not the final quality check. For simple, isolated elements, it may avoid a separate cutout step. Complex materials, very fine edges, and brand assets should still be tested against several backgrounds before publication or production.