Using text-to-image Generative AI to help create our Podcast Logo
When we decided to start this podcast back in April, Millie and I didn’t really know much about the process. Millie had been a guest on a podcast(link) in the past, and I have a couple of friends who run their own - that’s about it. What we did know was, a) that we wanted to spend a year researching and writing a horror screenplay, and b) we wanted to experiment with generative AI tools like ChatGPT to help us along the way. If you’re not sure what generative AI is, here’s a primer to get the gist of it.
We typed “how to start a podcast” into Google and hoped for the best. Aside from a device to record with, most results said that you just need 3 things to be able to publish your first episode: a name that represents you, an image or logo, and theme music. Our tired, screen-addled brains were grateful that we asked the AI overlords to help get us started.
So, how did these freaky internet bots help us create our Logo? Many AI text-to-image generators are free, or have some free functionality, while others (like OpenAI’s Dall-E) require you to purchase credits. I went with Picsart and DeepAI’s free tools, and Canva’s version that’s available with my existing paid subscription. Full disclosure, this was my first time using any of these programs so my first attempt at writing prompts for them was pretty weak.
Some of my starting prompts were super basic, like these:
Inside of a New York City subway car
Spooky New York City skyline
Full moon over the Brooklyn Bridge
To which I selected one of the available style choices (“Playful, “Filmic,” “Ink Printing”), yielding results like this:
These weren’t really what I was looking for, so I started to refine them by adding info about the style, amount of detail, type of light, positioning, etc. For example:
“Full moon at night over the Brooklyn Bridge, centered, in the style of a children’s book illustration, spooky” (below, left) or “Inside of an empty dirty New York City subway car at night, realistic lighting, medium detail” (below, right):
Most of these were warped or alien in some way. Like, any text you specified to be included came out as gibberish language, and anything to do with human - or, in our case, robot - forms were jacked up. Here are some of the robot rejects:
The prompt for both was: “one metal robot hand reaching up out of the dirt in a cemetery at night during a full moon, centered, high detail, realistic lighting.” As you can see, there’s a lot of weird stuff going on here. Extra fingers, half finger, half hand, hand off-center, moon off-center, tiny moon, weird double-crosses on graves - and that’s with some pretty specific prompting!
Canva’s tool was easiest to use because I could refine the prompt quickly after each batch was generated, and see the change until I got something that I liked that was usable. I think the prompt for these two was something like, “Inside of a spooky New York City subway train car in the style of a retro anime” and “New York City subway train car passing by in the style of a comic book”:
I messed with the colors, saturation, filters, and some lil ghost buddies. Then layered the text, added various effects to the fonts, etc until they looked like this:
There are still some things I’d want to change about it, and it’s obviously derivative of the base “Retro Anime” style it was pulling from. But the point was to try and do some things that were out of our element with the help of generative AI and I think this was a good example of collaboration. Here’s one of the robot ones that also came out alright:
Honestly, this took a LOT of time to prompt and re-prompt, select, and edit. Likely way more time and effort than it would have taken a professional Graphic Designer to ideate concepts, get approval, and do edits for something like this.
Text-to-AI generators are fun and easy to use, but yield better results with detailed prompts
The generated images are, by design, derivative of other human artist creations in some way
A great way to get inspiration, or develop a raw concept
You can make the output yours with lots of editing and a human touch
Is not a replacement for the quality and originality you’d get from a real life Graphic Designer, or any other visual artist