AI-generated posters don’t have to be horrible
TL;DR Highlight
Naming a specific design style in ChatGPT broke the generic AI poster look
Who Should Read
A developer who makes event or marketing posters with ChatGPT image generation and is tired of every result looking the same
Core Mechanics
- Defaults give the same craft-fayre look
- The author asked for a spring fayre poster and banned pastel, airbrush, and people, but still got the default template feel.
- The problem isn't that it's bad but that seeing the style 20 times becomes irritating.
- Asking for a totally different design gave Bauhaus
- He treated the first result as a "what not to do" and asked again.
- The second attempt came back in a Bauhaus-style geometric look, which he found much better.
- Ask the AI to name the style
- If you don't know a style's name, ask "What is this design style called?"
- Asking for a list of styles produced a 15-item menu including Swiss, Risograph, Brutalist, and Memphis.
- Picking one style at a time gave variety
- He requested Stamp/Letterpress, Japanese Minimal, Memphis, Designers Republic, and more in turn.
- He says they still look AI-made but are distinctive and not ugly.
Community Reactions
- Opinion that only the surface style changes
- Commenters noted every poster keeps the icon+text layout for the tombola and cakes lines, like Windows XP themes.
- Another said icons make no sense on a poster.
- Opinion that the better ones are just minimalist
- Minimalist designs leave little room for personality, so a fair test would compare AI and humans on the same style.
- One pointed out the Japanese Minimal example used cherry blossoms and the flag, the most obvious associations.
- Critique that the 90s DnB poster misses the culture
- A commenter said the wireframe renders are broken, and the smiley logo and community messaging don't fit DnB.
- Another argued the original posters were better because they said what the event was at a glance.
- Experience from a print shop
- A copy-print center worker said AI posters often have the wrong size, leaving white borders when printed.
- They said they spot AI by size, not style.
How to Apply
- Add a style name to your current image prompt
- Append a line like "Risograph" or "Swiss Style" to an existing poster or banner prompt and run it.
- Compare it side by side with the default output.
- Re-ask with the first result as "what not to do"
- After an unsatisfying first result, ask for a completely different aesthetic and call the old one a "what not to do".
- When you like one, ask for its style name and note it down.
- Request a style name in that era's poster format
- Combine the movement with the medium it would have used, like a 1940s cubist exhibition poster.
- Start a fresh chat for new styles so earlier text doesn't leak in.
Try It
# Paste into ChatGPT (with image generation) — the article's method as a prompt
Produce a poster for {{event name}}.
{{date and time}}
{{venue}}
{{list of activities}}
Go for a clean, unfussy, bright layout with a bold striking graphic. Avoid pastel/airbrush/oil style art or images of people.
# After the first result, follow up:
Make another one using a completely different design aesthetic of your choice. Treat the current one as a "what not to do".
# If you like one:
What is this design style called?
Give me a list of styles to choose from.
Now please give me {{chosen style}}.
# Next to try: in a new chat, ask from the start with an art movement plus medium, e.g. 'in the style of a contemporary 40s poster for a cubist exhibition'Terminology
Related Papers
AI is a tool and using it well takes skill
A rebuttal to Hickey: LLMs are tools, and skill in directing and checking them decides the outcome
Reflective Prompt Optimization of a Deep-Research Agent: A Topic-Disjoint, Multi-Seed Protocol for Measuring Transferable Gains and Prompt-Level Overfitting
Reflective prompt optimization raised held-out citation reward from 0.702 to 0.781 on a deep-research agent.
GPT‑6 and Intelligent UI for everyone
GPT-6 reportedly adds generated UI to chat; with no article text, only comment reactions are available.
Base Models Can Reason By Taking a Cue From Training Data
Prefilling just two opening tokens lifted Olmo-3-7B's MATH-500 accuracy from 42% to 78%.
Labels Override Definitions in Jev-Style Typed Decision Models
Renaming options to A/B raised routing accuracy by +0.1511 over meaningful labels.
How Much Prompt Is Enough? A Blackbox Minimization of Few-Shots in LLMs
Delta debugging trimmed few-shot prompts by 65.3% on average with unchanged propositional output