For the past week, ZDNet's David Gewirtz has been poking around OpenAI's new ChatGPT Images 2.0 release, which he claims is a much more consequential release than it first appears. We've all been impressed with AI image generators - they make pretty pictures, do fun tricks, and produce a lot of AI slop. But so far, they've been limited in their understanding of what they're producing. Images 2.0 adds subject-matter intelligence to image generation, meaning ChatGPT can now receive assignments and produce high-value outputs - like, say, redesigning your app's user interface without making it look like a 1998 GeoCities page.
Gewirtz fed two UI designs he's actively working on into ChatGPT Images 2.0 on his $20/month ChatGPT Plus plan. In a few short minutes, ChatGPT returned two redesigned user interfaces, both of which resulted in a whole bunch of design improvements he plans to incorporate into his products. The first was a Mac app he's been vibe coding with Claude Code since January - a project that's taking a while because he only has an hour or two a week to work on it, and it uses macOS's internal AI for image processing and analysis. He uploaded a screenshot and prompted the AI to "Redesign this user interface to make it more attractive and easier to use." At first, he didn't like the alterations - the most obvious change was the loss of the colored buttons, and the AI didn't quite understand that there are viewing options at the bottom of the grid view. But on the other hand, the new mockup included five design notes he plans to incorporate, plus having a mockup makes it much easier to show Claude Code what he wants it to do.
Next up was the UI for the starting page of his security product. His design was clean but quite rudimentary, mostly reflecting the fact that he dislikes coding in CSS. Last fall, in his first agent-based vibe coding project ever, he used OpenAI's Codex in his ChatGPT Plus plan to redesign the UI from fairly ugly to unobjectionable. This time, he pasted the screenshot into ChatGPT but accidentally hit return before he could give it a prompt. ChatGPT decided on its own to analyze the page, deducing that "The biggest issues are weak visual hierarchy, too much gray, a very long intro block, and three lower cards that compete equally for attention even though they are not equally important." It also recommended a "more modern admin aesthetic." Gewirtz then instructed it to "Provide me an image of the redefined interface," and the AI produced a design that invented a logo (mostly because he hadn't provided one) and added several features he liked, including a Quick Setup zone, a Need Help zone, a Configure Privacy/View Docs section, and a Site Status section at the bottom.
Gewirtz feels this capability is as game-changing as ChatGPT was back when it first came out, and as agentic pair programming was when it landed last summer. He submitted two product user interfaces to the AI and received essentially peer-reviewed commentary along with a set of prototype designs. As a solo programmer, this output is invaluable - even if he had a full in-house team with programmers and designers on staff, it would probably have taken a week or so to run this analysis and construct prototypes to review. The payroll expenses alone for that project would have been fairly substantial. But for $20, he was given two very helpful, very constructive, and not-at-all-AI-stupid redesigns that contained actionable nuggets that will make his products better.