The Exploration Machine - AI image generation in design and innovation
It’s simply stunning.
The rate of progress in AI image generation has been breathtaking.
In the images below, you can see an image of a “tranquil Nordic style home office with park view” from Midjourney V1 in February 2022 and V6 in December 2023.
The pace of improvement in under 2 years in image generation has been spectacular.
This pace of development has been both thrilling and anxiety-inducing. In this week's issue, I wanted to explore how we can constructively use image generation in design and innovation.
We’ll dive into five ways of using image generation in design and innovation:
Explore early directions
Visualize novel concepts
Prompt with care
Know the limits
Personalized images
1. Explore early directions
Image generators are the ultimate exploration machine.
In simple terms, they work by:
Learning patterns from millions or billions of images and their descriptions
Generating novel images based on specific user request prompts by understanding their context
By learning patterns in the information in the pixels of digital images and their descriptions, image generation models learn visual patterns beyond the obvious. They can produce truly novel images based on the “rules” of images they’ve seen.
AI image generation shines in the early solution exploration phase of the design process (visualization inspired by Roope Rainisto)
Pulling patterns from billions of images enables us to explore unlimited visual directions. With the right prompting, tools like Midjourney help us with divergent ideation.
Examples of this early exploration include:
Quickly generating dozens of early versions of a logo or icon based on a simple visual idea
Exploring how a change in a variable (color, material, style etc) influences the direction of a visual concept
Getting several examples of a data dashboard UI for early inspiration
In the example below, I created early variations for the logo of an imaginary vertical farm startup.
Logo variations for a vertical farm startup. Midjourney Prompt: a minimal, modern, and flat logo for a vertical farm company. Use elements of nature in vertical layers in a subtle way. Solid white background. --v 6.0 --s 50
I used a simple prompt using the visual idea of vertical nature as a basis. Based on the prompt, Midjourney combines elements of thousands or millions of nature-themed pictures and logos it's seen.
Did it create a perfect logo in its first go? Absolutely not. However, doing several iterations of this exploration is often still immensely helpful to jolt our own creative process.
The key to early exploration is to use image generation as a starting point - not the end destination for creating final outputs. I’ll often iterate with dozens of variations of the initial prompt to get ideas and possible directions. In Midjourney, it’s practical to use the variations feature to create slightly different versions of a promising direction.
Based on the initial exploration with AI, it's time for us to add our own creativity to the mix.
2. Visualize novel concepts
Sometimes, we get to design something that has never existed.
Tools like Midjourney can help us imagine things that have never been done with incredible accuracy.
I imagined a new premium vertical farm concept for high-end restaurants in the example below. I asked Midjourney to use brands from different domains, like Blue Bottle Coffee and Aesop, as inspiration to create something that doesn’t exist.
A new design concept for a vertical farm. Midjourney prompt: premium brand lifestyle photos of a vertical farm used in a high-end restaurant dining area. Vertical farm includes elements of natural wood and solid black metal. Use a style inspired by Blue Bottle Coffee and Aesop. The brand photos should include candid moments, natural light and stylish nordic inspired interior design. --ar 16:9 --style raw --v 6.0 --s 50
Using generative AI to create the first visual mockups can be useful if you're venturing into completely new areas with your project. It can help you communicate your vision and have more informed conversations to drive it forward.
Here are other examples of how image generation could be used to visualize new concepts (prompts in the captions).
Doctors office designed by Apple. Midjourney prompt: Pristine ultra high quality interior architecture photo of a doctors office designed by Apple. Stylish minimal interior, natural material, plenty of natural light. Tasteful touches of embedded modern technology. --ar 16:9 --v 6.0 --s 50 --style raw
A new electric car designed by Muji. Midjourney prompt: Extremely high-quality photograph of the interior of a new electric car designed by Muji. Take inspiration from the natural materials, neutral color palette, and functional simplicity of Muji. Include a big digital display for the driver --style raw --ar 16:9 --v 6.0 --s 50
A new health app on a tablet. Midjourney prompt: a health app dasboard on a tablet device. The app will give users an easily glancable overview of their health data. Take inspiration from companies like Google and Apple. The user interface is uncluttered, beautiful and intuitive. Hands holding the tablet in a home kitchen. --ar 16:9 --s 50 --style raw --v 6.0
3. Prompt with care
Image generation is less forgiving compared to Large Language Models like ChatGPT.
Tools like Midjourney don’t give you any feedback or hints on how to use it. It doesn’t correct your mistakes or suggest potential avenues. Some of its core features require the knowledge of how to use parameters.
Here are some core principles to help you get started:
Don’t overprompt. 3-4 sentences of specific instructions that still leave room for creativity are better than full paragraphs of text. Image generation tools can be easily confused.
Use references to existing styles or brands - to get non-generic results; it’s useful to have a stylistic direction in mind and not leave it up to the AI model to decide. I often include references to styles or brands as inspiration in the prompts. A resource like Midlibrary.io can help you find inspiration. Be mindful not to overdo it and fall into copyright infringement.
Learn to use parameters - Midjourney enables users to influence their image output with parameters.
– style raw reduces the creativity of the model and adheres more litelarily to the prompt. This can produce more authentic results with fewer opinions from Midjourney. I usually prefer this mode.
– s - 50 - 1000 - means how much stylizing Midjourney applies to the images (lower number means less). I typically prefer to keep this number low; it produces a more authentic and less “AI generated” feeling images (see example below)
Aspect ratio (-- ar 16:9 or other) enables you to choose the aspect ratio of the image
4. Know the limits
Image generation can be immensely useful for early exploration and visualizing novel concepts. The current versions of AI image generation also have some significant downsides.
The potential limitations and risks of image generation include:
Mistakes and hallucinations - AI image generations can create weird mistakes - like hands in unusual positions. These visual hallucinations have decreased significantly in the latest versions of Midjourney and other models.
Generating complex visualizations like user interfaces - UI’s combine text and several complex visual elements. Image generators don’t produce perfect UI’s and are best used as early inspiration
Text elements - the latest Midjourney version can create images that include text, but the results are often imperfect
Default AI style - All AI image generators have a built-in visual style. This can help novice prompters achieve passable results, but it also creates bland and unoriginal work. Like discussed above, use the –style and –s parameters in Midjourney together with creative prompting to achieve stand-out results
Misleading and harm - AI-generated images, wrongly used, can create misunderstandings and real harm. It’s best to consider the possible ripple effects of using an AI-generated image and being transparent about it being AI-generated to mitigate any potential risks
Copyright infringements - image generators create new visuals based on their learning data. Typically, these creations are completely novel, but sometimes, they can produce something that draws too much direct inspiration from existing work. One way to ensure that the image is indeed novel is to run it through a Google image search to find any similar pre-existing work.
5. Personalized images
How would your kitchen look in a retro 1960s style? What would your company’s new brand look like based on a few sentences of description?
With the advancement of AI image generation, we’re bound to see these models used for increasingly relevant and complex personalization. Customers will be able to see images and later videos deeply personalized for their needs. All within hopefully thoughtful guardrails set by the company.
We’ve only seen some early exploration using models like Stable Diffusion and DALL-E, but we’ll surely see more innovation in this space in the coming years.
The Faulty Exploration Machine
So, will image generation immediately replace all designers, photographers, and artists?
Regular readers of these articles will guess my answer: not in the immediate future.
It’s an incredibly fast-improving, impressive technology for generating never-before-seen images. It’s also trickier to prompt well than language models, prone to visual mistakes, and easily slips into a bland default mode.
But with the right prompting, it can be immensely helpful in the early stages of the design process. Not as a copying machine but as an exploration machine helping us envision new directions.
With the pace of improvement we covered earlier, the biggest question is how impressive will image generation be in 5 years?
Want to make sure you catch my next article?
This article was shared on my weekly newsletter on how to use generative AI for design and innovation.
