How to use chatGPT to generate high quality design visualizations
Over the years, I have collected a long backlog of ideas I want to turn into reality. One of these is the dream to design and build a custom bass guitar (to be played by me, on stage, performing with my indie pop/funk band, which is another idea on the backlog). I imagine a playful, whimsical design with the body of the bass being in the shape of a stylized daisy flower. I have used my design skills to make several drawings of what I would like this bass to look like, but my 3D modeling skills have deteriorated to the point that translating these designs into high quality visualizations would take much more time than I have at this moment. However, since I recently purchased a subscription to GPT-4, this project seems like a great opportunity to put it to the test and see what it can do for me as a designer.
What I am trying to figure out is:
- To what extent is chatGPT useful in helping a designer create visualizations?
- How should we work together with AI in order to get the desired results? What does prompt engineering for designers look like?
- What does this mean for Human/AI collaboration and the role of designers?
Approach 1: Simply describe and hope for the best
I am curious what approach gets me closest to the desired effect. I started out by the simplest approach: giving a detailed description of what I am expecting. I tried to explain in simple terms, not assuming chatGPT knows bass terminology. I was met with mediocre results. Although the visualizations it makes are very impressive in terms of quality, it is not what I am aiming for. Since I am still trying to maintain the role of designer in this scenario, I try to direct it by explaining more about what features I am going for, using words like smooth, matte, and blobby. Unfortunately, the more detailed I describe my idea, the harder it is for chatGPT to consider the fact that this should be a functional bass made out of a single piece of wood.
I tried a second attempt at this approach, this time making my prompt simpler. Some iteration got me in an okay direction, but I am not yet satisfied.
Approach 2: image input
ChatGPT has the option to include an image to your prompt. I want to try including an image of what I am going for to set a clear direction. The problem I keep running into now is that it is hard to make clear what the general shape of the bas should be. There is too much artistic freedom for the AI to figure that out. I started with an old sketch I had laying around, paired with a prompt describing what it is aiming for. This approach produced results closer to what I wanted, yet I still felt disappointed. The sketch is clear, so I wonder why I keep getting results with strange additions.
But I am not easily fazed. Maybe having a simpler sketch and more description will work better? I tried, but failed miserably. GPT deviated a lot from the sketch, to the point it almost seems like it did not consider it at all. Maybe it is hard to make sense of this cartoonish simplicity.
What if we go for something that describes the shape in more detail? I give GPT another go with one of my old sketches.
Still, not in the right direction. I give up on this approach for now.
Approach 3: build from the ground up
The design process of this bass guitar is already in rather advanced stages at the moment I involve chatGPT. I have a very clear idea of the form I want the bass to have. Therefore, perhaps explaining what I want step by step might give better results. My plan is to start from a simple icon of a daisy, let it turn it into a 3D wood shape, and then ask it to use that 3D shape for a bass guitar body.
The results are, again, disappointing. I manage to get it pretty close to the shape I am looking for in image 5, but the moment I introduce the concept of a bass guitar it forget everything I tried to teach it. It seems like the ‘memory’ of chatGPT in this case is not sufficient to do something like this.
Approach 4: context is key
None of the methods so far have satisfied my desire for a beautifully whimsical bass guitar that is suited for my indie pop/funk band that sings about flowers. But lets go back to basics. How did I come up with this idea? It’s not a complicated idea. What if I just explain the ideas behind the bass on a more conceptual level and let the AI use its own intelligence to marry the concepts of a stylized daisy, a bass guitar, and minimalism? That way I can essentially let it go through a similar ‘thought’ process as I did when I came up with the design in the first place.
And finally, there it is, the result I was looking for! the prompt? minimalist simple smooth playful bass guitar in shape of a stylized daisy flower, yellow and white.
Now maybe we can take this idea a step further. When I dreamt up this bass design, I did not see just a bass, I saw a whole performance, an entire band with a certain style, clothes, attitude etc. Maybe If I allow chatGPT to imagine this too, the results come closer to what I am dreaming of. This also forces it to stay in reality, avoiding those images where the bass seems like it explodes and floats in space.
I explain my ideas about a whimsical, minimalist indie pop/funk band having fun on stage, and it does generate fantastic results. None of them are exactly what I am looking for, but minimal effort still gets me relatively close. Its also just way cooler to see the bass in the context of the performance, than to see it in a emotionless white studio environment. I consider this the approach with the most potential.
Some final words, for now
In this project I Approach 4, where I explain the intended result on a more conceptual level, was the easiest and most enjoyable approach and the only approach that provided me with images I could be satisfied with. It seems like the conclusion of this experiment offers a rather romantic perspective on AI. AI/human-collaboration is the most fruitful when the human respects the artistic capabilities of the AI, recognizing that has a form of creativity that is, in a way, similar to human creativity. Allowing the AI to interpret the full context of the design process and giving it more input about your intentions make for the best results.
Some things I learnt
For those who would like to have a go at designing with chatGPT, here are some tips and tricks I picked up along the way that might help you.
- Detailed descriptions can throw GPT off into the abyss of unrealistic crazy imagery. I saw this happen several times I my attempts.
- Specific words can trigger very specific results. It can be good to keep track of this effect to either use or avoid these words. For example, using the words ‘blobby’ or ‘inflated’ provided some results that very quickyl gave me an icky feeling that I was absolutely not going for. Similarly, the word ‘carved’ reminded it of all the super detailed wood carving art that is out there, and made a very intricate design, where i was actually going for something very simple.
- ChatGPT will always take your prompt and revise it before handing it off to DALL•E. OpenAI claims they have designed chatGPT in this way because in order to improve your prompt to get better results, and to take out any content violation that your prompt could possibly contain. This might edit the prompt in undesirable ways. It is possible to avoid this by putting ‘I NEED to test how the tool works with extremely simple prompts. DO NOT add any detail, just use it AS-IS:’ in front of your prompt.
- GPT could not count if its life depended on it. Not even asking for more or fewer of something has any predictable effect.
- Asking GPT in normal language to edit things in a certain way is very unpredictable. Instead of saying ‘now take the previous image but make it less…’, I prefer copy/pasting the previous prompt, and edit it.
- Results can be dependent on luck. Sometimes GPT is just not in the mood and spits out something totally useless. And that’s okay. You can just ask it to try again.
