Build Hour: Image Gen

OpenAI’s Image Gen is built on the same GPT-4o architecture used for text, and it’s now available to developers through the Responses API as a native, tool-based capability—bringing faster, more controllable image generation plus editing workflows that feel closer to a back-and-forth conversation than a one-shot prompt.

The core shift is architectural and practical. Image Gen generates images auto-regressively—analogous to next-token prediction—so it can do things earlier diffusion-based text-to-image models struggled with. That includes rendering text directly on top of images, following instructions more reliably, and supporting granular editing. OpenAI highlights specific capability areas: improved text rendering (including handwritten-style text on surfaces), “world knowledge” that helps with educational visuals like science posters and photorealistic renderings of real places, and image-input workflows where multiple images can be combined into a single output.

To make these capabilities usable in real products, Image Gen is delivered through two API formats: “images” for simpler, single-turn text-to-image tasks, and “responses” for multi-turn and multi-tool experiences. The Responses API is where the newest developer-facing upgrades land. In the latest release cycle, Image Gen becomes a built-in tool inside Responses, with streaming partial renderings (so apps can show progress before the final image), multi-turn editing (by reusing an image ID or previous response ID), multi-tool image generation (Image Gen can be orchestrated alongside other tools), and masking for targeted edits.

A live demo turns those features into a working photo booth. A Next.js app uploads a user image, then generates stylized variants while streaming partial images to the frontend. The demo then performs multi-turn edits: the user modifies prompts to change elements like background color, while the system keeps the original composition and subject. It also demonstrates a key Responses API pattern—passing the previous response ID instead of resending full image payloads—reducing state management complexity.

The most product-relevant capability shown is tool orchestration with web search. Image Gen doesn’t have real-time knowledge by itself, but Responses can call a web search tool first, then generate an image that includes up-to-date information (weather in New York in the playground example; later, a live Knicks vs. Celtics score). The workflow is presented as “out of the box,” with no custom function wiring required.

OpenAI also lays out constraints that matter for production planning: generation is slower than earlier approaches, text rendering is better but not perfect (non-English characters can fail), multi-turn consistency improves but still isn’t guaranteed, and moderation rules apply. Developers can adjust a moderation parameter, but some requests may still be refused depending on content.

Gamma’s Jordan then adds an operator’s perspective. Gamma generates hundreds of thousands of presentations daily and has crossed 1 billion AI-generated images across its platform. Historically, AI images often broke on hands, limbs, and text; newer models have improved enough that Gamma uses AI images more confidently in decks. Gamma’s workflow includes generating presentation outlines using web search, selecting a theme, streaming deck creation while waiting on images, and refining visuals through maskless editing—chatting with an image to regenerate or remove elements without supplying explicit masks. Gamma reports a 27% improvement in user ratings after switching maskless editing to GPT image, and it recommends practical deck-building strategies such as using ChatGPT for synthesis and then importing into Gamma for slide generation.