Image Generation API
GPT Image 2 now generates transparent PNGs directly.
On August 21–22, 2026, OpenAI added a preview feature to the GPT Image 2 API that generates images with no background directly. Simply set the background parameter to transparent, and the API returns a PNG or WebP with a real alpha channel already baked in. The extra step of cutting out the background after generation is no longer necessary.
Making "cut it out later"
a thing of the past
Transparent generation itself isn't new. What matters here is that it now works on GPT Image 2.
When teams put AI-generated assets to real use, logos, product cutouts, and icons usually need to ship as background-free PNGs. Until now, that meant generate, then cut out the background with an external tool like remove.bg or Photoshop, then composite it into the design — a three-step routine. Subjects with soft edges, like fine hair or glass, often needed manual touch-ups on top of that.
Here's the twist: transparent generation itself already existed as the background: "transparent" parameter on GPT Image 1, released in 2025. But its successor, GPT Image 2 (model id gpt-image-2, with the dated snapshot gpt-image-2-2026-04-21), shipped in April 2026 without support for that same parameter. Teams doing e-commerce, sticker production, and presentation work kept posting on the OpenAI Developer Community that transparency broke the moment they migrated.
What shipped this week is, in effect, a preview that restores transparent-background support to GPT Image 2. It isn't just a new feature — it's a genuinely necessary capability being backfilled roughly four months after a model migration quietly dropped it.
| Before | From Aug 22, 2026 |
|---|---|
| GPT Image 2 only produced opaque PNGs | Set background:"transparent" directly in the API |
| Remove background externally, then re-composite | Done in a single generation call |
| Edges around hair or glass tended to break down | Alpha channel is built directly during generation |
| Transparency only worked on GPT Image 1 | Transparency works on the current model too |
Not cutting it out —
drawing it transparent from the start
Instead of a post-processing mask, the alpha channel is built into generation itself.
According to the official OpenAI Cookbook, GPT Image 2 doesn't strip the background after generating an image — it builds the alpha channel directly into the generation process. That's said to produce cleaner edges on details that typically trip up standard background-removal tools, like glass reflections or fine strands of hair or fiber. The alpha-capable output formats are PNG and WebP, and the feature spans three endpoints: /v1/images/generations (new generation), /v1/images/edits (editing), and the image_generation tool on /v1/responses.
Just three steps to use it
Add two fields to an existing prompt, and a transparent PNG comes straight back.
Set the parameters
Set background to "transparent" and output_format to "png" or "webp" in the request. Some reports suggest simply writing "with a transparent background" in the prompt text can also trigger it automatically.
Generate
Call it from generation, editing, or the Responses image_generation tool. Worth noting: the edit endpoint doesn't trace the original outline exactly — it effectively redraws the image with transparency built in (more on this below).
Export as-is
The returned PNG or WebP already carries a real alpha channel. Slot it straight into a slide deck or product page without a separate background-removal pass.
E-commerce product shots
For designers, dropping a product shot onto a new campaign background becomes a single request. No more preparing a separate cutout mask or white backing board.
Deck and presentation graphics
Marketers can drop diagrams and icons straight onto a branded slide background. It removes the familiar "AI image with an awkward white box" look that has plagued generated visuals in decks.
Logo and merch artwork
Print-ready assets for stickers or t-shirts arrive with the outline already transparent. It cuts one round trip that marketers used to make to a Photoshop-savvy teammate before placing a merch order.
What happens next
This is currently a preview, and OpenAI hasn't given a timeline for general availability. As with past OpenAI image API features, quality tweaks and pricing adjustments over the coming weeks or months seem likely. Since transparency itself is billed at no extra cost, teams can trial it against their existing image-generation budget without a new line item.
Three recommendations from the editorial desk:
- Teams with an existing Photoshop round-trip should pilot the switch to
background:"transparent"on low-risk assets first — internal decks, social icons — rather than moving production e-commerce imagery over immediately. Check edge quality visually before rolling it out wider. - If you use transparency on the edits endpoint, test with the assumption that the outline won't match the source image exactly. Early preview behavior looks closer to a "redraw" than a precise trace.
- Keep a manual edge-quality check in the loop for high-precision output like merch and print, at least for now. Automating the transparency step doesn't fully remove the need for a quality pass.
From "cut it out later"
to born transparent from the start.
This isn't unqualified good news
A few caveats are worth keeping in mind. First, as noted on the OpenAI Developer Community thread, transparency on the edit endpoint isn't a precise trace of the original subject's outline — it's closer to the model regenerating the image with the background set to transparent. For cases where you need the exact same product photo with just the background removed, a dedicated background-removal tool may still be more accurate.
Second, as a preview feature, the parameter's name or behavior could still change before general availability. Anyone wiring this into a production batch pipeline should build in a check that verifies the returned alpha channel behaves as expected, rather than assuming it always will. Third, transparency support isn't actually new — it already existed on GPT Image 1. This release is better understood as restoring a capability that a model migration had dropped, not as a brand-new feature, and it's worth resisting the urge to oversell its novelty.