Hands-on with GPT IMAGE 2: generating UI shots from full prompts

On October 3, 2026, a post titled "GPT IMAGE 2 in practice: generating interface mockups with complete prompts" appeared on V2EX. The author used a fictional task app called Today to show how one complete prompt per request can describe a single screen and produce static concept images in three forms: a mobile screen, a SaaS dashboard, and a landing page.

The author makes clear that "one prompt" means a full description of one page per request, not a single submission that returns three ready-to-use products. The deliverable is a static concept image: not code, not an editable component library, and not a clickable prototype.

The recommended order of work is to write the state first and deal with layout second. If all three tasks are incomplete, the screen must show three empty circular checkboxes together with a "0 of 3" progress indicator, so that the list and the progress card do not contradict each other. According to the post, the most common failure in interface generation is not an ugly button but mismatched state: a list where all three tasks are unfinished while the progress card shows one completed, or a heading that says Tuesday while the date strip highlights a different day. The main phone example in the article uses September 29 and the 0/3 prompt.

The post also notes that the teaching cover was generated by Codex and is not actual H3Max test output. The phone on that cover shows 1/3 completed and a different demo date, which makes it a visual variant; the small screen, chart, and landing-page copy in the cover and blueprint are not verbatim results of the prompts below. The word "free" in the image belongs only to the fictional Today showcase and does not represent H3Max pricing rules. So far the item ends with this practice write-up, with no further reported developments.