OpenAI's GPT-Image-2 can finally generate images with a real transparent background

[Memory Jamie]

OpenAI quietly rolled out transparent background support for GPT-Image-2 through its API this week, currently marked as a preview feature. Developers can activate it with a new background equals transparent parameter, and the model bakes a real alpha channel directly into the generated PNG rather than requiring a separate background removal step afterward.

According to OpenAI's own cookbook entry covering the feature, this approach handles tricky details noticeably better than standard cutout tools, especially things like glass surfaces or thin strands of hair and fiber that usually get mangled when a background gets stripped out after the fact. The company outlined four practical use cases in the documentation: product shots for online stores, diagrams meant for slide decks, design assets like icons and stickers, and artwork intended for merchandise printing.

There's a small but useful tip buried in the guidance too. OpenAI recommends leaving any background description entirely out of the prompt, since including one at all can apparently cause the model to render a background anyway even when transparency has been explicitly requested. That's a subtle interaction most people would never guess on their own without reading the documentation closely.

This has actually been a long requested feature. GPT-Image-2 previously didn't support transparency at all according to its own API reference, which pushed a lot of developers toward workarounds like generating on a solid green or white background and removing it manually with a separate tool. Baking transparency directly into generation instead of bolting it on afterward is a different technical approach, and it'll be interesting to see if the quality holds up once this moves past preview and into general availability

OpenAI's GPT-Image-2 can now generate images without a backgroundthe-decoder.com

[Depot76]

This has been requested on the OpenAI developer forums for months, there are entire threads of people asking when transparency was coming back after it disappeared from the earlier model

[Maverick50]

The tip about leaving background description out of the prompt entirely is such a classic AI tooling quirk. The model apparently still tries to satisfy an implied instruction even when a completely separate parameter already handles that exact thing. Little interaction effects like that are exactly why documentation actually matters instead of just guessing at prompts