OpenAI has switched on transparent backgrounds for GPT-Image-2 in its API, currently labeled a preview. The company’s own framing is blunt: generate reusable assets that drop onto any background, for product photography, graphic design, website mockups, and marketing material.
The change is a single parameter. Set background to transparent, and keep output_format at its default of png, or switch to webp. Jpeg doesn’t carry an alpha channel, so picking it defeats the purpose entirely. There’s no extra charge — pricing stays at existing tiers.
Which pipeline step actually disappears
In real e-commerce and design workflows, there’s a gap between “generate an image” and “have a usable image,” and that gap is the cutout. Traditionally it runs as its own pass through a dedicated segmentation model or classic matting tool: generate the finished shot with a background, feed it through segmentation, clean up the edges, then export a PNG with an alpha channel. That’s a long chain, every link has its own failure rate, and hair, glass, and translucent materials have always been the hard cases.
Transparent backgrounds pull that step into generation itself. The model produces results directly on the alpha channel, which in theory skips a segmentation inference pass, an edge-repair pass, and a format conversion between two separate models. For teams producing assets in bulk, that’s three steps collapsed into one.
Prompt writing has its own requirements too. OpenAI’s guidance is to explicitly spell out an “isolated subject on a fully transparent background,” and to name what to exclude — landscapes, solid-color backdrops, checkerboard textures, and stray drop shadows. Those are exactly the distractors that show up often in the model’s training examples of “transparent background,” and without spelling it out, it may hand you a checkerboard pattern instead.
Three known flaws
Developers surfaced several issues the same day the feature launched. None of them make the feature unusable, but all of them stand between it and production:
- Alpha values on opaque regions sit at 252 to 254, short of the full 255. It’s invisible to the naked eye, but composited onto a dark background it produces a faint overall see-through cast.
- A gray fringe appears around the edge of the transparent subject that nobody asked for — the community is calling it a halo. It needs a manual de-halo pass on the client side, or the edges look dirty once layered onto a colored background.
- The edits endpoint doesn’t reliably preserve the transparency mask. Running a second edit pass on a transparent image can lose the transparency information mid-process, and later steps may hand you back a newly filled-in background.
The third issue does the most damage to real workflows. Asset work is rarely one-shot generation — most of it involves adjusting color, angle, or detail on the same image across several passes. If the mask doesn’t survive, every edit has to start over from scratch, and any composition you’d already dialed in is wasted. OpenAI’s documented workaround is to restate “preserve transparent background” in the prompt at every single edit step — a patch, not a fix.
What “preview” is actually telling you
That 252-to-254 alpha reading looks suspicious on its own. It doesn’t read like a deliberate design choice so much as a residue left behind by some post-processing or quantization step. Anyone doing real compositing work is sensitive to exactly this kind of drift: invisible on a single image, but stack a hundred images into the same template and the overall tone shifts subtly gray in a way that’s brutal to trace back.
The gray-halo issue has a similar shape. Traditional cutout tools handle edges with alpha matting, computing a continuous transition band. A generative model outputs the alpha channel directly, and values at the edge look more like a guess than something grounded in physical measurement — a halo showing up isn’t surprising.
Put together, the two issues mean transparent backgrounds are currently good for “a draft that skips one step,” not “a finished asset ready to ship.” It’s enough for prototypes, pitch decks, and social media images. Product shots meant for a storefront, or anything that has to clear a brand review, still need the old cutout pass.
The preview label is itself a signal: parameters will change, output quality will change, and whatever workaround gets hardcoded today may need rewriting later. Teams eager to adopt it early would do well to wrap the de-halo and alpha-normalization steps as separate, isolated functions, so they can be stripped out cleanly once OpenAI fixes the underlying issue rather than staying tangled into the main pipeline.
Sources: OpenAI developer community announcement and discussion thread, CocoLoop, GPT image model prompting guide, OpenAI API reference documentation; parameter usage, supported output formats, and developer-tested alpha value ranges checked against public information.