gpt-image-2 transparent backgrounds: skip the cutout step

Every catalog pipeline has the same annoying step in the middle: get the subject onto a clean background so it can sit on a white PDP, a colored promo tile, and an email header without three separate shoots. OpenAI's gpt-image-2 now returns transparent backgrounds directly, which removes that step for generated and edited assets.

What actually happened

Per the OpenAI API changelog on August 20, gpt-image-2 and the pinned gpt-image-2-2026-04-21 snapshot support transparent backgrounds in preview, in both the Images API and the Responses API image generation tool.

The mechanics are simple. Set background to transparent and output_format to png (the default) or webp. JPEG has no alpha channel and won't work. The background parameter accepts transparent, opaque, or auto — and auto is the default, so if you don't set it, the model decides and you get inconsistent output across a batch.

OpenAI's image prompting guidance is blunt about the prompt: ask explicitly for an isolated subject on a fully transparent background, with no scenery, no solid backdrop, no checkerboard, and no stray shadows. The model will happily paint a checkerboard pattern as an image if you're sloppy about it.

Why transparent product imagery matters for your business

One asset, every placement. A transparent PNG composites onto a seasonal background, a size-comparison graphic, a bundle tile, or an ad unit without a designer touching it. That's the actual savings — not the generation, the reuse.

Preview means pin the snapshot. Preview features change. Call gpt-image-2-2026-04-21 rather than the floating alias, and regenerate a reference set before you move off it. Your catalog is not the place to discover a silent behavior change.

Set background explicitly on every call. auto is the default. In a batch job, a default that "decides" is a batch with two kinds of output and a person cleaning it up later.

Know which images you're allowed to synthesize. This is a marketing-asset tool, not a substitute for a photo of the thing you're shipping. Generating the product itself misrepresents what the buyer receives, and marketplaces are already forcing the question — Amazon requires sellers to tag AI-generated imagery. Composite real product photography onto generated scenes; don't generate the product.

Sources: OpenAI API changelog, OpenAI image generation guide.