ChatGPT breaks out of identikit poster style with named design aesthetics
A blog post by a developer who posts as "John Hartnup" starts from a common complaint: AI-generated event posters have converged on a single identikit style, illustrated by a widely shared Facebook post and an Independent article on the trend, plus an example from a beer festival poster the author found. The author's point is that the posters are not bad on their own; the problem is that after seeing the same look about 20 times, the repetition itself becomes irritating.
To test whether ChatGPT could do better, the author invented a sample brief for a spring fayre (date, venue, activities like a tombola, a samba band and a craft workshop) and asked for a poster in a "clean, unfussy, bright" style, avoiding pastel or airbrushed art. The first result still felt close to the generic look the author wanted to avoid. Asking ChatGPT for something in "a completely different design aesthetic of your choice," using the first poster as a "what not to do" example, produced something the author judged clearly better simply because it stood out.
Asked to name the style it had used, ChatGPT described it as modernist and Bauhaus influenced, with flat illustration layered on top, and broke that down into named traits: asymmetrical layout, bold sans-serif type, a limited high-contrast palette, and a Swiss-style information-first grid. When the author asked for a list of alternative styles to choose from, ChatGPT returned 15 named options grouped informally, including Bauhaus/Modernist, Swiss Style, Risograph print, cut-paper collage in the manner of Matisse, botanical illustration, brutalist graphic design, 90s rave flyer/acid graphics, Memphis design, Japanese minimal poster, monochrome-plus-accent, wayfinding/signage style, modern icon systems, stamp/letterpress style, and contemporary indie festival posters.
The author then generated posters in several of these styles by name, including Stamp/Letterpress, Japanese Minimal, and Memphis Design, and pushed further into more unusual requests: a poster in the style of Designers Republic (a design agency the article describes as known for record covers since the late 1980s), one imitating a professional designer adding typography to a child's paint drawing, one in the style of a photocopied 1980 punk fanzine, one styled as a 90s drum and bass flyer with early 3D and fractal computer imagery, and one styled as a 1940s poster for a cubist exhibition.
One technical wrinkle the author flags: while generating the Designers Republic-style poster, ChatGPT invented extra poster text on its own ("A day of music making and family fun") because it judged that style would include such text. That invented phrase then stayed in the chat's context and kept reappearing in later posters in the same session, since none of the details were real, the whole brief was invented for the test. The author notes this is avoidable by asking for the desired style from the very first prompt rather than accumulating stray details along the way.
The author's conclusion is that the resulting posters are still recognizable as AI-generated to a careful eye, but are distinctive rather than ugly, and that variety is available simply by naming a style rather than accepting the model's default choice. The post also mentions, without elaborating, that Claude and Gemini can output posters as HTML, PNG or PDF with editable real text and layers rather than a flat image, calling that a separate topic. Inspired by the exercise, the author says he went on to build a separate catalogue of 100 poster styles with ready-to-paste prompts and example images for each.
Key facts
- The author tested whether naming a specific design style to ChatGPT, rather than accepting its default look, avoids the repetitive 'identikit' AI poster style now widely mocked online.
- ChatGPT, asked to name the style of a poster it produced, described it as Bauhaus/Modernist with Swiss Style and geometric-minimalist influences, then listed 15 named alternative styles on request, from Risograph and Memphis Design to Japanese Minimal and 90s rave-flyer aesthetics.
- The author generated posters in styles including Stamp/Letterpress, Japanese Minimal, Memphis, Designers Republic-inspired, and a 1940s cubist-exhibition poster, all from the same invented spring fayre brief.
- While generating a Designers Republic-style poster, ChatGPT invented its own poster text, which then persisted into later posters because it stayed in the chat's context.
- Inspired by the exercise, the author separately compiled a catalogue of 100 poster styles with ready-to-paste prompts and example images for each.
Why it matters
Complaints about identikit AI-generated imagery, the same look regardless of subject, are common, and the usual response is to blame the model. This post argues the repetitive look is a prompting default rather than a hard limit: naming a specific design tradition (Bauhaus, Risograph, Memphis) rather than asking generically for 'a poster' pushes the same model into visibly different, more distinctive output. It is a small but concrete demonstration that variety in generative AI output can be a matter of vocabulary, not model capability.
Who it affects
Anyone who generates flyers, event posters or similar graphics with a chat-based AI tool and has run into the same generic look: community organizers, small event promoters, marketers producing quick graphics, and developers building tools that generate visual assets from text prompts.
How to use it
The technique described is simply to ask ChatGPT to name the design style behind an image it produced, then to ask it for a list of alternative named styles, and to request a poster explicitly in one of those styles (Risograph, Swiss Style, Japanese Minimal, brutalist, and so on) instead of leaving the aesthetic to the model's default. The author also warns that within one chat session, invented details (such as text the model adds unprompted) persist in context and can bleed into later generations unless a new session or explicit correction is used. The author separately built a catalogue of 100 such styles with ready-to-paste prompts, referenced but not detailed in this article.
How solid is it
This is a first-person account of one author's experiment with a single invented brief, not a systematic study: no control group, no quantitative measure of how 'identikit' the outputs were beyond the author's own judgment, and only a handful of style requests are shown. The specific dialogue exchanges and named styles quoted are ChatGPT's own descriptions of what it produced, taken at face value by the author.
Risks and caveats
The author himself concedes that the resulting posters are still recognizable as AI-generated by a careful viewer; the technique changes the style, not the underlying tell. The examples come from a single made-up event brief, so it is unclear how the approach holds up on posters with real, non-negotiable content. The context-persistence issue the author flags, where the model's own invented text keeps recurring in later prompts, is a caution that unwanted artifacts can quietly accumulate across a chat session rather than resetting with each new request.
“The problem with these is not so much that they're bad. They're OK, I don't love them. The problem is that once you've seen that style 20 times it starts to irritate just from the sheer repetition.”
— the author