For years, I’ve used photos from Unsplash and similar platforms to add as featured images to blog posts I’ve written. When sharing a link or when creating an archive page for my posts, I prefer having a visual element to work with. With stock images though, I find myself either spending too much time or not being happy with what I find on short notice. Next to that, you start seeing the images you’ve used across the web.

So, since the image generator vastly improved in ChatGPT’s 4o model, I tried that route: setting up a custom GPT to create images to accompany my blog posts. The results are not great as they still feel a bit too generic in style, but I think they add more personality than the stock images did.

Setup

Here’s how I worked on the prompting. First of all, I uploaded the drawings I made for my daughter’s and son’s birth cards. I asked ChatGPT to distill the style from it.

Then, there were a few specific features to add:

• The ratio is 3:2 as that’s the one of my camera. That means that I can set the template for both photography posts and archives, and those of blogs alike.
• The colour is black-and-white with some grey tones. This makes that I can change the colour scheme of my site without the featured images looking off. I’ve found that adding a duotone filter on the images adds a bit of thematic colour without being too disruptive.
• The background is white. That means I can make it fluid into the background of my site, assuming that I’ll stick to white for a while longer.
• I didn’t want text added. While the model is getting better at text, my blog has enough text in it to not add more text. I don’t think that text is needed either to make a compelling visual.
• Finally, I wanted to add some diversity, and not just white men in the illustrations.

Tweaks

The first results were okay, but there were some challenges.

One of the main challenges was that the GPT didn’t strictly follow the instructions, and this was especially bad in re-rendering the image. So when the first result wasn’t exactly what I wanted, the clarifying prompt would almost always be met with at least one of the strict rules to not be followed: I would almost always get either the dimension, the colour or the no text rule violated. So I gave it instructions and this prompt was added to the GPT. While it looks promising on “paper”, in reality, it’s still very much a hit ‘n miss. Sometimes it works, sometimes it doesn’t:

When re-rendering or updating images, these constraints must be just as strictly followed as in the initial prompt. If the user requests a revision, reaffirm these specifications and ensure they remain intact.

The second challenge was that the GPT’s interpretation of diversity seemed to be mostly focused on race, with a little on gender as well. And in doing so, I would say that in 70% of the cases, a male person of colour emerged from the use of the app. The two problems I had with this were that first of all it seemed to ignore the more autobiographical side of some of my posts — I am a white male — and second, that diversity is more than just race. I again updated the instructions and the GPT returned this as the prompt clarification:

Character diversity must be broadly and thoughtfully interpreted. This includes representation across race, gender, age, body type, and visible ability (e.g., people using wheelchairs, hearing aids, guide dogs). Avoid overrepresentation of any specific group (e.g., only men of color). Match the narrative tone of the blog post—if the post is clearly autobiographical, characters can reflect the author’s identity more directly unless otherwise specified. Diversity should feel authentic and inclusive, not tokenistic.

Use it yourself

I’ve made the custom GPT publicly available. Have fun with it!

Some examples

Here are some examples of illustrations I’ve created for my blog.