ChatGPT Sketch turns your bad drawings into detailed AI images

The Verge reports that OpenAI announced ChatGPT Images 2.5 on Tuesday with a new Sketch path: type @Sketch, draw inside ChatGPT, then instruct the model how to turn that doodle into an image. Reporter Jay Peters tested a rough mouse-drawn cat and got a realistic photo matching instructions; on-image comments let him change eye color without restarting from scratch.

OpenAI says Images 2.5 delivers more natural lighting and richer textures, follows editing instructions better across multiple turns, and cuts image-generation latency by up to 50% versus Images 2.0. Availability covers ChatGPT, ChatGPT Work, and Codex users on desktop, mobile, and web.

The strategic stakes are less about prettier cats than about who owns the multimodal editing loop. As image models converge on quality, the product that wins is the one that makes revision cheap, spatial, and conversational — especially for non-designers who cannot write a perfect prompt but can circle what is wrong.

Signal Desk's read: Sketch is a distribution play disguised as a creative toy. Embedding doodle-to-image inside the same chat box that already holds work context, files, and Codex keeps users from bouncing to Midjourney, Ideogram, or a separate canvas app for the first draft. @Sketch as an invocation is a small UX detail with a large habit implication: image gen becomes another slash-command workflow inside ChatGPT rather than a destination site.

The more important feature may be comments on image regions. Multi-turn editing that actually sticks has been the weak joint in consumer image products — users regenerate, lose composition, and blame the model. If Images 2.5's "across multiple turns" claim holds, OpenAI is competing on controllability, not just sample beauty. Latency cut "up to 50%" versus 2.0 matters here because iterative workflows die when each tweak costs a coffee break.

This is also a Work and Codex signal. Shipping the same image stack to workplace and coding surfaces suggests OpenAI wants diagrams, UI mocks, and annotated screenshots inside agentic work, not only social posts. The likelier read is that Sketch is onboarding for spatial prompting: teach consumers to point and draw now, then reuse that muscle for agent tasks that need visual grounding later.

Competitive pressure rises on anyone whose editor still treats the prompt box as the only steering wheel. Expect rivals to ship in-chat sketch or region-comment parity quickly; the moat, if any, is how well edits compound without drifting identity or layout.

Consumer image generation has largely shifted from novelty one-shots to editing and brand-consistent iteration. OpenAI's prior Images releases competed on fidelity; 2.5's messaging emphasizes lighting, texture, multi-turn instruction following, and speed.