OpenAI recently announced that DALL·E 3 is available to use inside ChatGPT, meaning that you can now create and modify images through the familiar chat interface. Since Floorboard is brand new and needs a logo, I decided to use DALL·E 3 to create a logo and try it out. The resulting experiment was super informative. Even though I ended up using the DALL·E-created image as inspiration for my final creation in Photoshop, this was a great example of what sorts of concepts AI can “understand” and where there are still some blindspots.
Creating the original image
I wanted a relatively simple, line-drawn icon that could act as a logo for a consulting/education plan called FloorboardAI.
With that in mind, I started with a prompt and got the following output:
Aside from the handling of the text, which DALL·E has always struggled with, this was definitely a great start. However, I still wanted it to be a bit simpler, without any text and probably without the “3D” perspective look.
Making edits to the generated image
To update the image to be closer to what I had in my head, I asked ChatGPT to “also make a ‘flat’ version using just white lines against a dark background” thinking that it would give me a more simplified version of the logo.
While it was definitely moving in the right direction, I could tell that our image generation was getting a bit off track.
This is when I decided to start a brand new chat. Because ChatGPT uses the history from your previous messages in the same chat, if you need to get a substantially different output, it helps to wipe the slate clean and start a new chat with a new prompt.
Trying different prompts to get a better output
I still had a pretty specific vision in my head, so I went through a few different prompt ideas, even passing it in some images of the floor of my apartment to try and teach it what interlocking floorboards looked like. What I started noticing was that most of the outputs from my prompts were providing a similar looking image of a 3D set of interlocking set of floorboards, with multiple layers, from a 45˚ angle sort of perspective, which I found interesting.
I didn’t quite understand why it was doing that, but when I did a Google Image search for floorboard imagery, it turns out that most images of floors are taken from that angle.
AI still has some blindspots
Asking ChatGPT to generate an image that was a single layer of boards or “two dimensional and viewed from above” yielded less than impressive results. And I think this is really the important takeaway here. Despite the impressive capabilities of some of these AI tools, they’re still not able to “understand” many of the concepts that are trivial for humans to understand and respond to.
Even after multiple attempts, ChatGPT was stubborn enough to want to persist that 45˚ angle perspective and waffled back and forth when asked to not stack multiple floorboards on top of each other.
Generating the final logo
In the end I was able to use a more detailed prompt to get something very similar to what I was looking for.
This gave me a color scheme that I liked and some idea of how to make the wood grain. From there, I was able to open up Photoshop, use this AI-generated idea to build a simple icon and that’s the logo you see in the top left corner of this site! The final logo:
