Generative AI for Industrial Design

July 31, 2023

Recently, I spoke about generative AI to a group of designers and researchers at Amazon Web Services. Specifically, I gave insight into my processes and workflows around AI. Around that time I had started using generative AI for industrial design. Similar to what I did with the children’s book, I was using ChatGPT and Midjourney to create product images. I figured this was an opportunity to show my process.

It’s only been about a week since I gave my talk. But in that time, OpenAI released a new feature called custom instructions. I have only started experimenting with it, but it may be a more effective approach for someone serious about using AI for industrial design. It might save some setup time and possibly give better results. Regardless, this approach will work fine for anyone just wanting to create a few interesting products.

This blog post is an adaptation of my talk.

I can already tell. Some of you are wondering what a guy who designs software is doing talking about industrial design. What do I know?

Maybe not much about industrial design. But I do know a bit about combining ChatGPT and Midjourney—let me give a bit of background. At the end of 2022, I wrote a post about using generative AI to write and illustrate a children’s book. That post has been a huge point of discussion and continues to be the top post on my website. And I frequently get questions about it.

However, the questions I get aren’t so much about the children’s book—but about the process. People are curious about the process more than the output.

I’d like to think of this talk as more about a workflow to use generative AI creatively. Today we’re just going to be looking at it through the lens of industrial design.

Why industrial design?

Original prompts: rv childrens toy made of wood, a cute selfie drone with a face and pastel colors, coffee maker with a vaporwave aesthetic, robot lawnmower with a rugged, aggressive styling, a steampunk robot vacuum.

While industrial design is what I’m interested in right now. That’s not what I set out trying to design.

I initially set out to see if I could create a process for designing software interfaces using generative AI. You’d think that basic web layouts would be easier for an AI like Midjourney than photo-realistic images of consumer products. But it turns out that isn’t the case.

So I spent time thinking about alternative design fields that might be interesting to explore—architecture, interior design, fashion—but industrial design seemed like the most interesting one at the time.

To be clear—AI will be coming for product and web design. It’s not far off. I expect it to be just as impressive as Midjourney has been for photography and illustrative work. I have more thoughts about that, but I’ll save it for another time.

Prepare, Learn, Create

We are going to follow a three-step process to turn short phrases into amazing product images. I’ll show you what’s possible, what’s difficult and outline my process for getting there.

Step 1: Prepare

First, we’ll provide ChatGPT with a prompt to prime it. That prompt will introduce that we’re going to be creating image descriptions called prompts and that we’re going to be using these prompts in Midjourney, an AI image generator. Remember ChatGPT does not know about the world before 2021.

Then we’re going to outline some high-level goals. Broadly what would we like the ChatGPT to do? We want it to produce a lot of vivid, detailed and varied descriptions. We want to make sure it includes camera details, viewing angles, moods, lighting material descriptions and more.

The last thing we’re going to do is confirm—we’re going to confirm that it understood the above and let it know what to expect next. If everything works as expected, you should get a response that says “Give me example prompts!” If you get anything else, you should check the prompt and regenerate.

Step 1: Prepare

We are going to be generating descriptions of product images, to be used in an AI image generator called Midjourney. These descriptions are called prompts.

Use the following info as best practices to create Midjourney prompts:

Focus on clear and concise descriptions, with different concepts separated by commas, then follow them with any parameters. Parameters are not separated by commas.

Be specific and vivid: Describe every single aspect of the image, including: Subject, Style, Color, Medium, Composition, Lighting, Shadows, Environment, Perspective, Depth of Field, Textures, Scale and Proportions, Foreground, Midground, Background, Material Properties, Design Properties, Aesthetic Style, Time of Day, Motion or Stillness, Patterns and Repetition, Setting, Reflections or Transparency, Symbolism, Light Source and Direction, Art Techniques or Mediums, Contrasting Elements, Framing or Compositional Techniques, Dominant Color Palette, and any other relevant context.

Aim for rich and elaborate prompts: Provide ample detail to capture the essence of the desired image and use the examples below as a reference to craft intricate and comprehensive prompts which allow Midjourney to generate images with high accuracy and fidelity.

Incorporate relevant camera settings like focal length, aperture, ISO, & shutter speed. Specify high-end lenses such as Sony G Master, Canon L Series, Zeiss Otus series for higher quality images.

Diversify your prompts: Explore various styles, moods, colours, art mediums, and aspect ratios to create a wide range of visually appealing and unique images.

I will continue to provide you with information about Midjourney AI. Simply reply with, “Give me example prompts!” if you understand.

Step 2: Learn

In the second prompt, we want to teach ChatGPT what a good output looks like. We’ll do this by providing quality examples.

Where do you find examples though?

With Midjourney, there are two great options. You can search for product images in the community app, or you can go into a public channel and try and find decent prompts there. The images don’t necessarily have to be product images either—you’re looking for objects with a clear focus. Be mindful of the parameters the image is using. I only want photographs and 16:9 dimensions, so whenever aspect ratio or style parameters appear I change them to –ar 16:9 –style raw.

I know this is a time-consuming step. It does require some additional work, but if there’s any secret sauce in getting great results, this is it. You can copy my examples in the prompt below, but I’d encourage you to research and add a few additional ones.

How many examples will you need? You’ll want three examples at a minimum, but shoot for 10 or more.

After we provide examples, once again, we will have it confirm it understands. If you get anything other than “Example prompts studied!”, you should check the prompt and regenerate the response.

Step 2: Lean from example prompts

Here are great examples of prompts generating product images that you can learn from:

Package Design. product display. Pentawards. vray tracing, collecting and modes of display, simple and elegant style, glass as material. yaka art, smooth surface, vibrant and colorful, industrial themes. clean and simple designs, gamercore. neon cold lighting. volumetric lighting. product photography, masterpiece, Minimalism, industrial design, product design, surrealism, epic detail, realistic rendering, 3 d rendering, oc rendering, clean shadows. 32K –ar 16:9 –v 5.2 –style raw

Black perfume bottle luxury for brand presentation on rocky background, commercial advertising photography, the two sides of the main body is illuminated by contour light, the left side is illuminated by a soft light box 45 degrees, the background is in the evening snow mountains, Rule of Thirds, dark background, top advertising photography, creative, studio lighting, shot with Canon 5d4 and Canon 100mm f2. 8, photography, ultra detailed, realism, 16K –q 2 –s 750 –ar 16:9 –v 5.2 –style raw

Professional commercial photograph of a product, medium close up shot of the product, full focus on the product. Avoid glare on product. Soft warm natural light. Even lighting. Rule of thirds. Minimalist, stylish, sleek, super trendy. Full Focus , In style of commercial adverts for Vogue and Conde Nast magazines. In style of ISO 800, F/ 1. 4, white balance, White background, Nikon D7200 –q 2 –ar 16:9 –v 5.2 –style raw

A jewelry design, sakura – themed ring, gemstones and diamonds, luxury, closeup, product view, trending on artstation, cgsociety, ultra quality, digital art, exquisite hyper details, 4k, Soft illumination, dreamy, fashion, rendering by unreal engine -ar 16:9 –v 5.2 –style raw

A photograph of a robotic vacuum cleaner autonomously navigating a room, keeping floors clean with minimal human intervention, hyper realistic, super detailed, dynamic pose, photography, 8k –q 2 –s 750 –ar 16:9 –v 5.2 –style raw

Professional photo of a modern lawn mower with aerospace and space vehicle inspired colors, materials and features, Dashes of vibrant aqua neon on key features, shot in a dark warehouse with smoke for effect, Cinematic, 35mm lens, f/ 1. 8, accent lighting, global illumination –uplight –s 50 –ar 16:9 –v 5.2 –style raw

Design a wine packaging, brand identity, quartz, black, include historical design accents, august macke, elegance, sophistication, wide angle shot, 1100s, anthropomorphism, analytical cubism, natural lighting, soft lighting from above, golden hour, highly detailed, 8k – -ar 16:9 –v 5.2 –style raw

Top view of a 1970s futuristic white colored car, driving fast, action camera, sunny day, high detailed, chase, holographic projection of the car floating some lightning bolts, dynamic photography, Gopro, a Tesla car in the center of the frame, regionalism, dreamlike environments, virtual reality, flat, low – angle shots, whirring contrivances –s 950 -ar 16:9 –v 5.2 –style raw

Ultra – detailed, best – quality, cutting machining, digital art, milling machines, lathes, drill presses, metal shavings, whirring motors, safety goggles, gloves, protective clothing, precise shaping, desired components, well – lit environment, ample space, organized tools, efficient workflow -ar 16:9 –v 5.2 –style raw

Photography of an organic chair made of mycelium, organic shapes, industrial design, elegant, white sheet background, natural lighting, realistic, photorealistic, kodak portrait 400, 35 mm, analog photography, depth of field, coarse, noise, hyper detailed, insane detail, high texture details, 8k –q 2 –ar 16:9 –v 5.2 –style raw

Next, I will provide additional requirements for creating prompts. Do not generate any prompts until I give you specific input to do so. Simply reply with, “Example prompts studied!” if you understand.

Step 3: Create

In the last two steps, we told ChatGPT what we wanted to do and provided good examples. In this step, Command, we’ll give it specific requirements for creating product images. Most importantly, we will have it take text and generate three prompts based on the best practices and examples outlined in steps one and two.

I have it generate three prompts because that’s the queue limit for my Midjourney plan. There’s no reason you can’t do more.

What else? This is where we’ll outline any guardrails—max word length and Midjourney parameters. I will say with my current approach I sometimes still get prompts wrapped in quotes.

One final time we’re going to confirm ChatGPT knows what we want it to do. Once you hit enter, “Ready to prompt!” should be displayed.

Step 3: Create

You will now receive a text prompt from me.

Once you receive the prompt, do the following:

Create three creative prompts for the Midjourney AI art generator using the best practices mentioned in step 1 and example concepts in step 2.

Do not include explanations in your response. List three prompts with correct syntax without unnecessary words.

Each of the three prompts must contain less than 100 words.

Make sure each prompt ends with –ar 16:9 –v 5.2 –style raw.

The prompt should not end in any punctuation.

The prompt should not be in quotes.

Simply reply with, “Ready to prompt!” if you understand.

Example inputs and outputs

Now we should have a solid system for generating and iterating on prompts. Let’s look at how to make a few products come to life.

Here we’ve entered the prompt A light-up dog harness into ChatGPT.

The output of that gives us a more detailed prompt. It provides the scene, lighting and additional details. I would never have thought to mention LEDs with intricate patterns, but that is a design detail that makes these harnesses distinctive.

In the Midjourney results, you can see how those details make their way into the final images.

What about something more stylistic and less feature-specific? Here we want a robot lawn mower with rugged, aggressive styling.

ChatGPT responds with additional details describing our rugged, aggressive mower—resembling a stealth drone, matte black, neon green accents. It also gives details about lighting. Here you can see the strong backlight from the sun in all of these images.

Finally, let’s try something a bit out there—a coffee maker with a vaporwave aesthetic.

For those that don’t know vaporwave is an 80s-esque aesthetic that features vibrant hues, geometric patterns and nostalgic motifs. It draws influence from the 1980s culture—particularly music, fashion, and art. If you’re still having trouble imagining what vaporwave is, take a look at James White’s Uzicopter blog, which serves as a fascinating vaporwave mood board.

The ChatGPT’s output gives us pastel colors and strong geometric shapes which are heavily incorporated into the Midjourney designs.

Volume is king

Those of you that have worked with generative AI have probably recognized that gems are often found with volume. If you’ve followed the prompts above exactly, you should get three detailed prompts for every short prompt entered.

I like batching my prompts into sets of three. I chose three, because that’s the maximum concurrent jobs Midjourney can run. That doesn’t mean you couldn’t do more though.

What is difficult?

It’s not all rainbows though. For whatever reason there are some difficult things to generate. Holly asked me to generate a stuffed animal—a monkey with rabbit ears. After a dozen times, I started typing straight into Midjourney to see if it could even come up with such a thing. No luck.

Monkey plush toy sporting oversized rabbit ears, rich chocolate brown fur, contrasting pink inner ears, adorable facial features, shot against a vibrant nursery room backdrop, warm afternoon light streaming in from a window, shot with a Zeiss Otus series lens for sharp detail –ar 16:9 –v 5.2 –style raw

I went back to ChatGPT and tried a few more iterations. Eventually, I got this, which fits the description. For whatever reason some products can be difficult to generate.

There are a few other oddities I’ve found—like specifying a specific number of wheels. If you specify that you want a product with three wheels, chances are you’re going to get something that looks like it has four wheels. I’ve found it helps to reference an object that already has a specific number of wheels associated with it—a tricycle for example. It takes some tweaking though.

Resources

Further reading on AI for industrial design.