AI19: Using ChatGPT to produce creatives

How to create icons and images of the same style to emphasize a style and elevate the entire design's aesthetic

I have spent the last two weeks designing landing pages for our upcoming product, an AI agent for office work.

As a tiny bootstrapped company of two, SK and I often try to do as much as we can ourselves. This allows us to stay lean, move fast, and keep costs low.

But sometimes, that would mean a compromise in the quality. For example, we often use Heroicons icons and Unsplash images because they are nice enough and free.

Heroicons icons are beautiful. But they are all in a single style and are limited.

While the photos on Unsplash are amazing, they are overused. Anyone who needs a photo of a lady working on a laptop has probably used this Unsplash image.

Even if you find a nice photo after scrolling through 200 photos on Unsplash, it is almost impossible to find another photo, say a man working on a laptop, in the same style.

ChatGPT has entered the chat

Using ChatGPT (Plus), I was able to generate various icons in the specific styles and different images with a consistent style.

They might not seem like much on their own. But when I combined them with the other elements on a design, they helped to emphasize the style further and elevate the design aesthetic overall.

How I use ChatGPT & Sora

I use both ChatGPT and Sora to generate images. Both of which are included in the subscription. It isn’t clear how ChatGPT and Sora are different but it seems that:

• They are using the same underlying model to generate images.

• Sora is like a power tool. It has various configurations and allows you to turn your images into videos.

• ChatGPT might generate an image prompt based on your message to help you get a nice image while Sora would take your input as the image prompt.

• Images generated in Sora will be published to the public feed automatically, unless you turn it off under the settings. ChatGPT images should be private.

Here’s how I have been using them:

ChatGPT: Images that I’m not sure how to describe, and icons

I often ask ChatGPT to suggest what a good visual might be for my idea. For icons, once I get my desired style, I can simply add to the chat something like “Now generate (another object) in the same style”.

Sora: Images that I can already roughly visualize in my mind

Honestly, I haven’t explored all of Sora’s features yet. But I have been able to get what I wanted in just a few tries. The Explore tab is a good place to get inspiration for what to generate.

Some of the generated images can still be creepy or odd, such as super short legs or laptop screens facing outwards. But as a whole, the images are so much better than just a year ago and seem as good as Midjourney’s.

Have you been using ChatGPT or other AI tools to generate creatives? I’d love to hear how you have been doing it or if you have any tips to share!

---

Jargon explained

• python-docx: This week I learned about and played with a Python library for reading and editing Microsoft Word documents. Specifically, I was trying to read tables. I hadn’t thought about how complicated it could be; for example, I couldn’t simply read and store the data cell by cell because some cells could be merged. The recommendation is to repeat the merged cell value for each layout-grid cell occupied by the merged cell.

• Encoder and decoder: An encoder turns text or images into tokens (or embeddings) to be fed into the main transformer of the model to generate more tokens. A decoder then turns the generated tokens into text or an image as the output.

---

Interesting links

• Open AI launched Codex, a coding agent: I have been using Windsurf but I’ll be keen to try this once it’s available to Plus users. I want to see how integrated it is with my local development environment because I probably won’t want to be approving PRs for changes. Latent Space shared some suggestions on making the most of Codex by the Codex team.

• Conversational Interfaces: the Good, the Ugly & the Billion-Dollar Opportunity by

  Julie Zhuo
  : I hope years from now, when the hype dust settles, we will not be using AI via a chatbox only. SK and I made several UI explorations ourselves; Pebblely is a good example of how people don’t only want a textbox.

• Building, launching, and scaling ChatGPT Images by

  Gergely Orosz
  : I like the section on how ChatGPT Images work.

---

Recent issues