Vibe-imaging: chat-to-image & the Great Ghibilification
In a recent post we discussed vibe-coding, the ability to use a chatbot to write code and build functional development prototypes.
Last week the Internet was turned upside down by the launch of the new ChatGPT 4o Image Generation model. With the flick of a switch, OpenAI gave tens of millions of paid users access to a state-of-the-art image model, right in their chatbot.
Vibe-imaging (working title!) was born.
Under the hood, this new model, along with Google’s soon to be released Gemini Flash Image Generation work in a fundamentally different way than their predecessors. Being true ‘multimodal’ models, when you chat to create an image, the model understands what you mean, beyond you simply typing what you want to see.
Another of ChatGPT’s stand-out features compared to past models is the ability to generate realistic text within images. This sprung wide open a swath of new use cases such as comic strips, infographics, menus and greeting cards, all of which OpenAI were eager to show off on their site.
This got us thinking about some other use cases and we ran some experiments, one of which was;
Text-to-webdesign
With it’s new text rendering abilities plus fantastic prompt adherence, ChatGPT can now create wireframes, and then mockup designs from the wireframe.
Example 1. Beach Clothing brand - wireframe to designs
The prompts were:
"I am doing a new eCommerce site for an Australian fashion brand, can you help me design the web site? Come up with a wireframe first"
"Please create a visual wireframe of the homepage"
"Great, now create a visually stunning design for this web page, the theme is Beach Clothing for men and woman, and beach accessories, in Australia"
"Make it more award winning, less boxy"
"Show it on an iOS screen"
Example 2: Furniture brand - Shopify theme to designs
It can also take a Shopify theme, and show your product range in it, notice how it even updated the text!
Prompt:
"I like this Shopify theme, can you do a mockup using my furniture products on this style page?"
Both of these examples show very rudimentary prompting, and modest quality outputs. I don’t think UX designers will be threatened by this, but they do now have a new tool in their toolkit.
Ai·dentikit
Before OpenAI dropped their image bombshell, we were already experimenting with the recently released (not multimodal) Google Imagen 3 model, which has impressed us with the quality of it’s outputs.
We had an idea we wanted to play with; can you describe yourself with words, and have the image model sketch a portrait of you. This reminded us of a police sketch artist, and hence Ai·dentikit was born! You can try it here.
If you are interested to know how it works, read this explainer post from Raj.
The Great Ghiblification
Last week after ChatGPT released their new image generation, X and LinkedIn were awash with AI generated illustrations in the style of Studio Ghibli.
This great article from Carly Ayres dissects how AI can replicate in seconds what took Ghibli creator Hayao Miyazaki decades to perfect, questioning the legality and also the ethics of this. As Carly puts it: "They gave us an everything generator but everyone is obsessed with making it do the same thing."
Potential risks and issues
With ChatGPT’s new image generation model now in the hands of millions, the potential is exciting, but the challenges are just as significant. Let's unpack...
AI generated images with ChatGPT
🚫 IP infringement? Now just a swoosh away.
Brand logos, trademarks, and iconic designs can be mimicked in seconds. It’s fast, convincing - and potentially unlawful.
🫣 Golly, is that real!?
Hyper-realism makes it harder than ever to separate fact from fiction. Even digitally altered landmarks can look 100% legit.
📸 Copyright breach? “Wasn’t me.”
These models are trained on the internet’s collective memory. That includes copyrighted architecture, like the Opera House, art, and branding - now easily summoned on demand.
🎭 Deepfakes made disturbingly simple.
Photorealistic faces. Public figures in fictional scenarios. The ethical line? Blurrier than ever.
🧑🏽💻 Prompt precision is next-level
If you can describe it, ChatGPT can generate it. That means more control, more creativity - but also more confusion. What’s real, what’s synthetic, and who decides?
Undoubtably, the tech is impressive. But the risks are real too. It feels like we need to move as fast on governance as we are on innovation, with most of that being in our hands. Can we be trusted?
And finally, a touché that our friend Jim Stewart posted on LI:
