102: Revealing the Secret Prompts and Process Behind our AI Images

Apple •Spotify• Pocket Casts •Youtube •Overcast •RSS

What’s up folks. One episode that’s been on the list for a while now is revealing how we do our AI images. 

Most of the comments we get on our social posts aren’t “wow amazing content, love the CDP topics, I learned so much about email deliverability…” It’s usually “that cover art is SO cool, what’s the prompt that you use, what tool are you using for these amazing images?”

So without further ado, let’s go behind the curtain and walk you through the process that we use to repurpose our audio only podcasts into long form blog posts packed with eye popping AI images.

Here’s today’s main takeaway: Ditch your raw transcripts and transform your audio podcast into a visually engaging blog with unique AI-generated images. Ditch lame stock images and learn how to use Midjourney for standout visuals that elevate your content and captivate your audience. But a word to the wise: these AI tools are addictive. Use them at your own peril.

Jump to a Section 👇

How to turn your audio-only podcast transcript into a long form blog post

We’ve been using Otter.ai to transcribe every episode after you’ve edited them. What’s the first step to converting that long not so great raw transcript into the polished blog posts we have on the site today?

Step number 1 involves our trusty friend ChatGPT. Here’s the prompt I start with:

I’ll provide podcast transcripts with [guest name], and I want you to convert each Q&A into a blog passage. Third-person only. No fluff or weird words. Remove ‘ums’ and ‘likes.’ Each passage gets an H2 title and ends with a key takeaway. Ready for the first question and answer?

So I go through our transcript, I copy paste the question we asked and I copy paste the answer from our guest. Here’s an example from our episode with Scott Brinker.

And here’s the ChatGPT response:

So I copy paste that passage into a doc and I repeat the same thing for every question and answer, here’s another example:

And ChatGPT’s output:

Naturally, the output isn’t always perfect. You’re asking ChatGPT to turn your raw transcript filled with ‘ummms’ and ‘likes’ and probably has several mistranscribed words. So it’s always worth going through it and looking for issues. 

I also use ChatGPT to help me write a summary of the episode and suggest variations of titles as well. 

I start with getting a list of all the takeaways from each section and paste them all into Chat GPT and ask it to come up with a summary based on those takeaways.

This is my prompt for the summary:

And this is the output:

And titles is next. So I do the same thing, paste all the takeaways and ask it to suggest a bunch of titles. This is my prompt:

And this is the output:

Getting Started in Midjourney

First step here is creating a Discord account if you don’t have one already. Then you’ll want to create your own server so you can generate images in your own private channels. 

Then you join the Midjourney Beta on their site, that’ll get you invited to the Midjourney Discord server. You can check out the #getting-started channel in there to get some startup instructions. You can start seeing what others are building in any of the #newbie channels. 

But I prefer building in private so you can hit the ‘Show Member List’, click on the Midjourney bot and then add it to the server you just created. 

Since the Pope Francis in a puffer jacket earlier this year blew up their servers and they now require paid plans for all subscribers.

Well worth the $10 to start playing around though. You just need to hit the /subscribe command in Discord and you’ll get a link to sign up. 

One thing we get asked a lot is that our illustrations and images have a crazy likeness to our guests. We’ll reveal the secret sauce behind that. It’s a bit of trial and error but it all starts with another Discord bot. It’s called Picsi.AI by InsightFaceSwap Bot.

It’s a highly realistic portrait creation tool that you can use for free with their Discord invitation, or head over to Patreon to subscribe and access more features and higher usage limits.

The creators are InsightFace.ai, an open source Python library that offers 2D and 3D face analysis. It efficiently brings together top-notch face recognition, detection, and alignment algorithms. It’s designed for performance, both in training and deployment phases. Both research institutes and businesses stand to benefit from using InsightFace. 

Picking your style

Another thing that lots of people comment on is how consistent it is. You can just tell that’s Humans of Martech. Picking your style in Midjourney is one of the most important steps. At first you want to experiment and play around with a variation of style prompts and once you have something you like. Save it. 

There’s probably thousands of styling permutations and combinations that can give you a pretty unique style. You can reference specific styles like cyberpunk, 8-bit, cubism, pop art. You can reference styles from video games like Zelda, GTA or Firewatch. You can also reference famous artists like Picasso, Davinci, Warhol, Kubrick and Tolkien. You can stick to real life and even reference specific cameras and lenses.

I went through a bunch of my favorite ones in my virtual talk at MOPSapaloza all using the same prompt:

/Imagine A person presenting during a virtual conference [style]

Cyberpunk

8-bit

Cubism

pop art

Zelda

GTA

Firewatch

Picasso

Da Vinci

Warhol

Kubrick

Tolkien

Real life 

Nikon AF-S DX Nikkor 35mm

Fujifilm XF 56mm

Adding images to your blog post

JT was worried about revealing the prompt but I already dropped it at MOPSapaloza and on the Martech Podcast. But folks can also take one of our images and using the /Describe command Midjourney can spit out a prompt that’s probably close to what I use. 

I’ve settled on 3 recurring keywords in my prompt:

  • Flat illustration: simple, two-dimensional elements and bright colors. It avoids gradients, shadows, and textures to achieve a clean and straightforward look
  • Modern: characterized by simplicity, function, and clean lines. It avoids excessive ornamentation and often embraces new materials and technologies
  • Geometric: clean lines and basic shapes like squares and triangles. It’s minimal, precise, and often uses a limited color palette for impact

I’ll sometimes remix a prompt and add a few extra keywords:

  • Neo-pop: bold colors, clean lines, polished and slick
  • Bold stencil: strong, clear lines and high-contrast colors, 
  • Minimalist: simple forms, clean lines, and a restricted color palette, sense of calm and clarity

So I like to use ChatGPT to help me come up with cool prompt ideas for Midjourney. It doesn’t know what Midjourney is but it knows what DALL-E is so I’ll ask it to suggest DALL-E prompts based on the paragraph below and I paste each blog post section.

This is my prompt:

Write a DALL-E image prompt based on the paragraph below. Stick to one subject and a background. Start with “Flat illustration, modern, geometric,” and end with “–ar 9:4”. Avoid using words, symbols, icons, or hands. No tightrope or balance scale. Give me 10 variations. Make ’em creative.

Here’s an example using one of the blog post passages from our episode with Scott Brinker.

And the output from ChatGPT:

Here’s the output from Midjourney when we paste “Flat illustration, modern, geometric, a robot holding a sculptor’s tool, reshaping a clay model of a marketing job depicted as a billboard –ar 9:4”

And here’s another example from our other blog post passage from Scott’s episode:

And the prompt ideas output:

Here’s the output from Midjourney when we paste:

“Flat illustration, modern, geometric, an early-stage marketer standing at the base of a mountain that represents ‘AI in marketing’. The mountain is sectioned into different paths each symbolizing a different specialization within marketing. The marketer, with an AI-powered tool in hand, confidently steps onto the path of their chosen specialization. Above them, a superhero cape with the word ‘superpower’ billows in the wind, symbolizing their future advantage in the evolving marketing landscape –ar 9:4”

Making the Cover Art for Humans of Martech

Okay let’s talk about the cover art specifically. How do we generate images that look so much like the actual guests?

Step 1 is uploading an actual image of your guest, usually what works best is a high quality front facing, no glasses headshot.

The command you use after connecting InsightFaceSwap bot is

/saveid [idname] [image]

So you’re saving “Scott” with his headshot. The bot is then associating the facial features from your guest to the id name you’re saving. We’ll be using this to apply that overlay to the Midjourney output. 

Step 2 is pasting 2-3 other headshot images from your guest in Midjourney and copy pasting the URLs for each image in your prompt. Here’s my prompt:

[image link 1] [image link 2] [image link 3] in the style of flat illustration, neo-pop, geometric, bold stencil, minimalist –[aspect ratio]

So using some of Scott Brinker’s headshots, this is the output in Midjourney:

You can always repeat a command if you want to see variations:

So you’ll notice that some of these kind of look like Scott, but most don’t. Because we’re applying real life image features to an illustration, the output isn’t always great, but let’s try to apply Scott’s face to one of these illustrations by right clicking on one of them, selecting apps > INSwapper.

Not bad, but not perfect:

Here’s a couple other attempts:

I didn’t love what I was getting here so I tried to go with another theme. We all know Scott as the Godfather of Martech, maybe we can combine one of these illustrations with Marlon Brando, using this blended prompt:

[illustration image link] [godfather image link] in the style of [your style], flat illustration –[aspect ratio]

Here’s the output:

Some interesting images here, none that look quite like Scott though.

Let’s see if we can apply InsightFaceSwap to one of these:

Not ideal again, you can see the struggles with using an image with glasses here.

Let’s try something different, can we blend Scott’s famous martech maps with one of the illustrations using this prompt:

[illustration image link] [martech map image link] in the style of [your style], flat illustration, bold stencil, minimalist –[aspect ratio]

And here’s the output:

I feel like we’re getting somewhere. Some of these are really neat:

Let’s try applying Scott’s face to one of my favorites:

Not great… 

Instead of having the martech map image link at the start, let’s try adding it to the end of the prompt like this:

[image link 1] [image link 2] [image link 3] in the style of [your style], flat illustration, neo-pop, geometric, bold stencil, minimalist with this background [martech map image link]  –[aspect ratio]

That 3rd image makes me feel like we’re finally getting somewhere. Let’s run a few light variations of that one:

I love the second one, it gives me Bob Saget vibes though let’s try to add Scott’s face to it:

And there you have it folks.

I don’t usually spend this much time on every cover art… but like I said, it can get a little addictive. And while there isn’t a perfect synergy between all our cover art, you can still tell there’s a theme. 

Listen to the full episode 🎧

✌️

—
Intro music by Wowa via Unminus
Cover art created with Midjourney

Find your next episode
Semantic search across every episode transcript. The more context you give, the better the match.
Find my episodes →

Apple •Spotify• Pocket Casts •Youtube •Overcast •RSS

, , , ,

<< Previous episode

Next episode >>

All categories

Monthly archives

See all episodes

Future-proofing the humans behind the tech

Apple •Pocket Casts•Google •Overcast •Spotify •Breaker •Castro •RSS

1 Comment

  1. Pingback: The Role of Front-End Martech Tools in the Age of Generative AI - Marketing Ops

Leave a Reply