So What? Marketing Analytics and Insights Live
airs every Thursday at 1 pm EST.
You can watch on YouTube Live. Be sure to subscribe and follow so you never miss an episode!
In this episode of So What? The Trust Insights weekly livestream, you’ll learn about the current state of AI image generation. You’ll discover how AI image generation tools have evolved over the past two years and how you can use them to create marketing graphics. You’ll see real-world examples of AI-generated images, and you will learn how to write effective prompts to get the results you want with AI image generation.
Watch the video here:
Can’t see anything? Watch it on YouTube here.
In this episode you’ll learn:
- What tools are currently available for AI image generation
- AI image generation within Photoshop
- Ethics and compliance of AI image generation
Transcript:
What follows is an AI-generated transcript. The transcript may contain errors and is not a substitute for listening to the episode.
Katie Robbert: 00:00
Foreign. Well, hey, everyone. Happy Thursday! Welcome to So What?, the Marketing Analytics and Insights live show. I’m Katie, joined by John, and—as you can see—John and I are trying to remember how all the buttons work. Chris is traveling this week. John, how’s it going?
John Wall: 00:37
Good. Yeah. This is so funny. It’s like banging the rust off the machine here as we use it. I was in Keynote today. I haven’t been in Keynote in like seven or eight months. It was ridiculous. Yeah. So, today’s show—I just have to say out the gate—there’s a very Saturday Night Live feeling to this show because the stuff that I’ve got to show today, it’s not that it’s ready to go and it’s a fantastic presentation. That’s right where I want it to be. It’s going on today because it’s 1 o’clock on Thursday, and so we’re going to go with what we’ve got, and this is going to be it.
Katie Robbert: 01:08
Well, and what? We were talking pre-show—when I should have been remembering how to start the show—but we were talking pre-show that I feel like it’s okay that the stuff we do is a little bit more raw and unpolished because we’re learning and growing along with everybody else who’s trying to figure it out. And so this week we are talking about the current status of AI image generation. And so, when we last talked about this two years ago—which is crazy, John—you and I remember we did the episode. Things have come so far, and I personally have not had the time to really test everything. So this is where we’re going to start to explore what the heck is happening with AI image generation. So, where would you like to start?
John Wall: 01:56
Yeah, well, we can still stick to the framework and talk about what we’re trying to do here. So the first one is—let’s just start with the use case here—so as a marketing professional that needs to use graphics, I want to figure out what is the state of these tools now. Now, we’ve checked two years ago. We want to see where they’re at today. And I’m doing this so that we can take advantage of some of this stuff. Like, some of these tools can maybe help out with what we’re doing on a daily basis. And—and I kind of—in general—when we did that first round, the results were so bad that I basically said, yeah, we’re—this is too early—like, forget this, we’re not—I’m just not going to use these things for now. And—and it is funny.
John Wall: 02:35
It’s like two years to within 10 days since we did this last. So, obviously—as everyone knows—Chris is on the road for April. Like, that’s one of the hottest conference seasons, which is why we end up having to run the ship, at least for a week or two while he’s on the road.
Katie Robbert: 02:53
Well, I mean, let’s be clear, I’m still in charge. I’m just not the technical front of the company.
John Wall: 03:00
Right. Well, it’s just he spends so much time in the trenches with this stuff that he can just pop his head out once a week and be like, “Hey, look, here’s all the insane stuff I found,” and that fills the show. Whereas, you and I are actually doing the business things and the guardrails that keep this moving.
Katie Robbert: 03:18
We’re not the technical demo guys.
John Wall: 03:20
Right. Exactly. We’re not. We’re not the mad scientists of the crew here. We’re like the bus driver. So you’re getting like the John Candy and Melissa McCarthy of Trust Insights here holding this together.
Katie Robbert: 03:33
But, when I think about this user story—so—as a marketing professional, I want to see what’s going on with AI image generation so that I can see where it fits into what I’m doing. A lot of companies are going to say, “Well, we have a whole creative team.” Trust Insights does not have a creative team, as evidenced by our website and anything else that we put out there. And so, for a small company like ours where that role is just as much as we need, it hasn’t been a priority in our budget. Finding tools to make some images better than—like—you or I can make images. John, I don’t know about you. I took art for about 10 years, and after 10 years, I was still at day one. Like, I don’t have a creative bone in my body.
Katie Robbert: 04:19
And so, putting images and graphics together, it’s just not a skill set I possess. So, when I see tools like this that can help me, I’m really excited to get the support.
John Wall: 04:34
All right, cool. We’ve got Brian already chiming in. He loves Midjourney for style, but ChatGPT seems to be the best detector. Yes, we have some classic tech stuff we’re going to be showing to everybody here. So, yeah, you’ll get a kick out of this. And, yeah, you bring up a great point, which is that dividing line of—okay—the general public kind of needs these tools to—if you don’t have a graphic designer—well, your options are go take some pictures, go buy something, or throw a prompt together and see what the heck you get out of it. But a lot of what we’re going to show today is too. I call slot machining.
John Wall: 05:08
You know, this idea that you throw something in there and you pull the lever and you kind of have no idea what the heck is going to come out. And, yeah, I might as well. Let me just dive into some of this.
Katie Robbert: 05:19
Let’s get into it.
John Wall: 05:20
Start with the—with the show here. And again, this is going to be great here. See if I can get all this stuff to work here. Okay, so that’s that. I need to share this thing. How the hell do I fire that? Or are you the only one that can fire that? All right, there we go. All right, so that’s running. I’ve got this window. All right, so this is it. This is the exact deck we had used two years ago, AI for image generation. So this is where we started. So our angle back then was we had framed it in terms of podcasting was like, okay, so what are you using for podcasting to create images for the show? And we covered three tools at the time: Stable Diffusion, DALL-E 2, Midjourney 5. And then we also talked about transcripts.
John Wall: 05:58
And that’s become a whole other thing—which is interesting—this idea of using these AI tools to generate transcripts of the shows themselves and how accurate are they? And accuracy was a huge problem back when we ran this. That was something that didn’t come through clean at all, but has really cleaned up a lot. But, yeah, I only had enough time to go down the rabbit hole for images and figure out what is going on this front. As we dig in, we swapped out, ditched the transcripts and decided to go with image tools. And this was just like everything that I always do on this front. I just go to Chris and I say, “Hey, I need the new tools. What should I be digging into? Tell me the good ones.” This was the list he gave.
John Wall: 06:42
He’s like, “All right, ChatGPT, Google JEDI, Google Gemini, Facebook Meta.” And that was his exact quote. “They’re doing some crazy stuff: Fluxone, AI, and Metalama 4.” And so then I—so—and—you know—innocently, I was like, “Oh, well, I’ll just go pull these and go.” But now I’m realized, I realized as I dug in that—like—okay—so I’ve got to go test the old models, and I’ve got a whole new grocery list of new ones. So I’ve, like, more than doubled the work for figuring out what’s going on here. So the adventure begins. I started to jump in, and the original case that we had run was the Simon Sinek interview over on Marketing Over Coffee. And there was a graphic up there. There’s the show notes.
John Wall: 07:25
And the idea was, okay, what can we use for AI tools to just generate some stuff for this? And to give you a little bit more of an idea, too, here’s Simon—handsome guy—that’s him with his book. That’s the Marketing Over Coffee show logo that gets used on the episode. So this is stuff that’s out there on the web that these models are going to pull from. And so we had started with that and—yeah—back at the day, we used Stable Diffusion, which—back then—it was confusing. Like, you couldn’t even find it on the web. You had to go to a secondary website and get there. And it hasn’t gotten any better, too. Now let me. So Stable Diffusion is now owned by Amazon and it’s baked into a bunch of tools. Like, I couldn’t even.
John Wall: 08:05
I don’t think there is an interface where you can just go to and say, “Hey, Stable Diffusion. Crank me out a picture.” Although I could be wrong on that because all these things are all over the place. But so. And this is what it came up with. And this is what we were just dying about—you know—talking about how it just could not get text. Yeah. And—you know—we had—like—drunk Popeye Simon Sinek over here on the top. Right. Like, I don’t know what was up with that. And basically all the interest came up with similar things. And in the bottom right there—like—weird hands, that was still a thing all over the place at that time. So that was where we’re at. Well, just jaw-dropping when I fire things up.
John Wall: 08:43
Now I go to ChatGPT two years later, and just the same prompt: Create an image for Marketing Over Coffee podcast interview with Simon Sinek. And it throws. Yeah. Is that nuts? I was like, “Oh, wow. Okay, so.”
Katie Robbert: 08:58
So, first of all, it’s correct language. It didn’t bungle all the words. And I don’t know Simon, but I’m fairly certain that’s a close enough replica that someone wouldn’t think, “Oh, that’s not him.”
John Wall: 09:14
Yeah, isn’t it? It’s like, bang on. But then it totally kicked it to another level, too, because I was like, “All right, so the game has totally changed. Like, I need to figure out what’s going on.” But then, after it fired the image, this is the next thing it said to me. I said, “Wow, that’s really good.” He said, “I’m glad you liked it. Would you like variations or versions formatted for social media?” And I was like, “Oh, of course I do,” because that was the big beef two years ago, all of them would just crank out—like—”Here’s a 100 by 100 image, or here’s a 300 by.” Like, you had no control over the sizes of them. Right.
John Wall: 09:49
But now for it to just be like, “Okay, yeah, you know, crank me out a LinkedIn one, crank me out an X version, crank me out an Instagram one.” So that was. I was like, “Okay, ChatGPT is like bringing the A game here. This is a whole new thing.” And now, as we dig into the—you know—ChatGPT was not doing images. In fact, I’d have to go back and look, maybe ChatGPT was publicly available so we could get in there.
Katie Robbert: 10:15
But images were—oh, yeah—ChatGPT was publicly available, but at that time two years ago, it wasn’t doing image generation. And so it’s only—and I apologize if I’m misquoting the timelines—but it feels like it was only within the past few months that you could do image generation easily within the ChatGPT interface without being sent somewhere else. And so now it’s baked into part of the opening prompt where it’s like, “What do you want to do today? Or what question?” And image generation is just baked in. It’s not like, “Oh, let me find where image generation is, go to that and find somewhere else.” Like, it’s all in one place now. And I feel like that’s a fairly recent—and by that, I mean within like 2023 feature.
John Wall: 11:07
Yeah. And if I’ve got the lineage right, too, now DALL-E 2 that we had originally tested is not available. You can’t even get to it anymore. And DALL-E 3 is baked into a bunch of tools. And ChatGPT does call DALL-E 3. So this sort of looks like where it goes. But that brings up another point along the way that was just crazy to deal with all of these things, is all of them have different kinds of paywalls and gates, and they’re tied to other people. Like, there were a number of services where it’s connecting to our Trust Insights business account. So I wasn’t really clear on which ones are free and which ones are we getting some tokens because we’re already paying for the Google stuff, right? And getting into the Meta stuff, we’ll talk about that.
John Wall: 11:52
That was a whole different road to go down and get lost on. But—but—so that was ChatGPT—you know—kind of coming out of the gate and really surprising me as far as kind of what can be done and where it could go. So Google Gemini was the next stop on the list: “Where is this going to go and what’s going to happen with this?” And this is where I ended up with that. You know, I was just like, “All right, what have you got?” And it’s like, “Well, sorry, not working today.” And—you know—it sat overnight and still did not even come up with anything. And so it was basically broken. And that was something that I ran into.
John Wall: 12:26
That’s why I was like, “All right, I got to do screenshots of this stuff because I don’t want to be sitting here and having prompts take on me or not going.” So Midjourney. Looking back. So this is another thing. When we’re looking at Midjourney, it didn’t even have a web interface. Like, you had to go to Discord.
Katie Robbert: 12:43
Right. I forgot about that.
John Wall: 12:44
Right. It was totally—you know—blocked from the rest of the world that this was the only way to get over to it. Well, so this time I go in and run and this is. It had done a good job as far as coming up with some stuff. You know, it’s still—the words were messed up—but at least—I don’t know—he looks a little bit more like Simon Sinek. It’s still not him—obviously—but at least it was closer to the mark. But then I got in. I was like, “Oh, okay, no, memberships start at $10 a month.” And I was like, “Okay, ChatGPT is doing fine for free, or at least on the Trust Insights version of whatever the heck paid is.”
John Wall: 13:18
And so I’m not going to cough up 10 bucks to figure out what it can do. So maybe it is better than all the other stuff out there. But I was not about to go down that route. I did have another interesting thing. So, when we were testing that out, Chris had a prompt. He was like, “Hey, fire against these styles of art, you know, see what would happen.” Okay. Gil Elgren does these kinds of 1950s pulp fiction kind of madman.
Katie Robbert: 13:47
Ish.
John Wall: 13:47
Yeah, totally. Mad Men, you know, thing. And so this is what it came up with for Christopher and I drinking coffee in a bar.
Katie Robbert: 13:57
And I totally see the resemblance.
John Wall: 13:58
I know, isn’t it. Yeah, exactly. Like, which is Chris, which is me. I, you know, it’s.
Katie Robbert: 14:02
I mean, you must think it’s like looking in a mirror.
John Wall: 14:07
So the thing was, okay, you know, what’s happened in two years? Where do we go with this? What’s gone? And, interestingly enough, again, as I—we ran this through—it was like, “Oh, okay, you know, same prompt. Here’s Christopher S. Penn, John J. Wall in a bar drinking coffee by Gil Elvgren.” And even this summary that comes back: “Wow, that’s a fun and imaginative prompt. Christopher S. Penn and John J. Wall, known for their work in marketing and podcasting, styled in a scene by Gil Elvgren, famous pinup artist.” And so it even asks, “Hey, do you want us to come up with a general image or do you want to upload a photo?” And so that I did. This was a whole change of actually giving it what it wanted to look like. So.
John Wall: 14:46
And this is a shot from a good 10 years ago.
Katie Robbert: 14:49
I was going to say, look at the babies.
John Wall: 14:52
Yeah, exactly. Baby faces. This is. We were talking with David Meerman Scott a long time ago, and—but it did tank initially, too. It was like, “Okay, a lot of people are creating images, so it’s going to have to sit.” And I don’t think it was overnight, but I did have to come back a couple of hours later. But the final output of that.
Katie Robbert: 15:10
Oh, wow. I mean, looks like neither of you, but it looks more like you than it did two years ago.
John Wall: 15:21
Right, exactly. My thought was it’s way better than anything else that had come out previously. And yet it’s still horrible, huh? Like, it’s definitely not us, but it is a lot. And the—the other kind of interesting thing with that was that it didn’t get as close on the stylistic side of it. Like, these candelabras are definitely not 1950s, you know, going on. But that was at least interesting that—you know—the current tools now are way ahead of what we had two years ago. So—I mean—at least has a passing resemblance to us. Oh, all right. And so then another one was we were talking about show images, right? We were talking about—you know—put something out there. And this was the prompt at the time was cup of coffee and a studio table next to a studio microphone and an iPhone.
John Wall: 16:09
Because that was kind of the—you know—the classic podcast images of the time. And it came up with all these weirdo microphones. You know, it was like patching bizarro cables and photo equipment and stuff. And then—of course—the other joke was every coffee cup must have a latte art on it. You know, you have to have that. But overall, it was pretty close. Well, it was the same deal. We ran that through, and—right out of the gate—they come up with something that’s—you know—pretty much dead on, less complicated, but—you know—none of the weird artifacts or bizarreness. It was a lot closer to.
Katie Robbert: 16:49
This strikes me as I’m giving you exactly what you asked for, which is why when we talk about prompting, we talk about specificity and getting detailed. And so we obviously have—you know—a bunch of different prompt structures, the most recent being Trust Insights AI RAPPEL. So—you know—the PARE framework will walk you through getting really sophisticated. But what you asked for was a cup of coffee on a studio table next to a studio microphone and an iPhone. And I would say—like—95% of what you asked for is there. It’s a little. You can see—like—on the right side of the mic, it’s a little wonky, but—like—that’s only if you really look at the image for a very long time. Otherwise, you’d be like, “Oh, there’s coffee and a mic and a phone. You’re good.”
Katie Robbert: 17:35
It’s exactly what you asked for.
John Wall: 17:38
Yeah, isn’t it? It’s amazing how it just does—again—what it says. And this was—you know—learning from the previous round was, “Hey, your prompts have to be a whole lot better.” Like, I was. All my prompts are obviously garbage because I’m not getting what I want. But the thing was, I did want to use the same prompts for this round so that we could see that gap of where things go. But—yeah—you can if you. You know, that’s the great thing of all. These are just the ugly first shots. You know, if you keep talking to it and moving it along, you can get closer to where you want to go. And—yeah—it can come up with some interesting stuff. Let me think. We also had. Did I get into meme? Oh, Facebook Meta. Yeah.
John Wall: 18:16
Okay, so this is the next. I was ready to dig into this and kind of see where it goes. It gave me just a headache trying to get in the door. It would not let me get in. And it was a matter of—like—it wanted to verify my age. It wanted me to do captchas. It wanted me to log in with my Facebook account. And I kept chipping away. Finally, I had to go out of my browser and go to a browser that was my personal one that I use, Facebook so it could grab my Facebook login, which—for some reason—it couldn’t do. And so it came up with this, which—I mean—what’s your gut on this one here?
Katie Robbert: 18:50
Well, again, I don’t know Simon that well, so I can’t say, “Oh, yeah, that totally looks like him, or that’s him adjacent.” But I don’t. I mean, I don’t look at this image and immediately think, “Oh, there’s weird stuff in there.” But knowing that it’s an AI-generated image, I want to know who the other person is. Like, where did that person come from? Like, you know how this is a little weird? You know how—like—when you have a dream and it—like—puts together people that you’ve literally never seen, and they’re meant to represent people you may have—like—passed in the grocery store, or—like—one time when you were five at the amusement park, you maybe saw a face. That’s what this—you know—strikes me as. It’s like, “Where did that face come from?”
John Wall: 19:29
Yeah, that was the one that got me, too. And—like—why is that person so small? They’re like.
Katie Robbert: 19:35
They’re either like squatting down or it’s a child with a five o’clock shadow, right?
John Wall: 19:40
Because it’s like a regular head, but—like—way down lower than it should be. And, yeah, the Simon thing looks like it. It’s like 95% Simon, but it’s—it’s kind of like. Or did you make an evil clone of Simon? You know, it’s like, it’s definitely not him. And there is a little weirdness with his hand there. He’s got a little—I don’t know what the heck is going on there under his thumb. There’s like a half finger in there or whatever. The other thing that was interesting to me was the idea that the green awning is across the street. Like, that’s. You can totally think of. That’s a Starbucks pole, right? Like, every Starbucks has a green awning, but here it’s across the street. So that was kind of.
Katie Robbert: 20:16
Well, so maybe he’s in the indie coffee shop and the Starbucks is across the street. So it’s like a social commentary against shopping small businesses, right?
John Wall: 20:26
It’s a buy local, shockingly, like Starbucks, but aren’t Starbucks. So that was. So that was the same. I was kind of like, all right, it was a lot of headache to get over here. And the results were a little bit weird. So that’s enough for me on that front. Flux 1A. This one I knew nothing about; this was another one that Chris was like, “Oh, God, dig that up, go check this out.” And as I get—you know—I go to the website and it’s just—you know—the classic—I know—classic—like tech. Right out of the gate, I’m like, “Okay, we have—you know—women and cats and right on the mark.” But then at first I was like, “All right, this—you know—seems to be something that’s different from everything else.” So let’s get started. Let’s fire up.
John Wall: 21:12
And I go to the first screen and they show a sample image and it says, “You know, flex AI image generator. Result. The original prompt was a dark-haired woman in her early 30s playing the piano.” And I’m like, “That’s not a piano. I don’t know what the story of that is.” So that was a little bit weird. So you can imagine I was like, “All right, let’s fire this up. Let’s throw the MOC prompt in there.” And it was like two years back—you know—”Yeah, you know, no dice with the words. We’ve got some coffee cups.”
Katie Robbert: 21:45
You know, I often sprinkle whole beans around my morning coffee just for—you know—aesthetic effect.
John Wall: 21:52
Right. It makes it feel more coffee if there’s a couple of.
Katie Robbert: 21:54
It makes it feel less like coffee shop to have like just beans everywhere, getting under everything.
John Wall: 22:00
And this is another interesting one, too. I’ve noticed on a couple of tools having the watermarks on the image, you know—like—”Hey, we’re taking all this stolen stuff, but now this is ours, we’re defending this. So here’s a watermark.” Because we wouldn’t want people just willy-nilly grabbing images and doing stuff with it on the internet. That would be.
Katie Robbert: 22:18
Well, but you know, it’s the old-school business model where—you know—like Getty Images or something like that. Like, if you want to do it for free, you have to deal with the watermark. But if you want that watermark out of there, know, pay up, isn’t it?
John Wall: 22:34
Yeah, that’s—you know—you’re a marketer when you go out to a restaurant or a hotel somewhere and you’re like, “Wait a minute, that’s a Getty image. I see the water on that.”
Katie Robbert: 22:42
Like, oh, I’ve seen that laptop, coffee cup, notebook before.
John Wall: 22:47
Right, right. That ethnically diverse group huddled around the laptop in the conference room.
Katie Robbert: 22:52
Oh my God.
John Wall: 22:53
Yep, that’s how that.
Katie Robbert: 22:54
All right, so Flux one is stuck a couple years behind.
John Wall: 22:58
Yeah, you’re just like, whatever.
Katie Robbert: 23:00
Yeah, so what? Like, oh, so. And we also—I thought—so wait, Meta Facebook is different from Meta Llama 4, right?
John Wall: 23:11
This is—he just gives me the list and I’m like, “Okay, what the heck are all these things?” So Metalama. And then I realized as I dug in, the idea with Llama is that it runs locally. The whole thing with Llama is you can download it and have it on your computer and go. But I was like, “Okay, so Meta—you know—powered on their servers, didn’t do that well, so should I set it up on my machine?” And I was just like, that got well.
Katie Robbert: 23:37
And I will say we have done episodes on the live stream where Chris has walked through how to set up local machines. So if you want to go to those, find those instructions, go to Trust Insights AI, YouTube, go to our site, what playlist. And we’ve done quite a few episodes on setting up local AI. It’s not easy, once you have it set up. Yes, you can run it fairly straightforward. Like the systems themselves have their interfaces, but it’s the actual setup and knowing that you’re going to be using up memory on your computer in order to do this.
John Wall: 24:11
Yeah. And I was just—you know—it was just a George Bush “not going to happen. Not going to happen. I was not going to build a server and like play around with it and spend time on that.”
Katie Robbert: 24:21
It’s a good opportunity to revisit the user story because—you know—as a marketing professional, I want to understand what’s working in AI image generation so that I can figure out what to use. And if your criteria doesn’t involve setting up a local model, great, get it right out of there then. Don’t even consider that as part of your tech stack because these are things that you need to be able to make work for you, not just what everybody else is using. And that’s something that we talk about a lot is there’s a lot of AI hype right now. I mean, has been. And we all sort of get this feeling like, “Oh, we’re being left behind,” and—you know—I’m not moving fast enough. Really challenge yourself to think what works for me specifically and for you, John.
Katie Robbert: 25:09
Setting up a local model doesn’t work. And that’s okay. That means that it’s not the right tool for you.
John Wall: 25:16
Yes, that was where I was going to dodge that bullet there. It’s. Brian just brought up a point I’d heard about this Midjourney print magazine. I have not checked it out myself though, so I’ll throw that out there if anybody wants to dig into that. The idea that you get a print magazine and it has the images and the prompts so you can get ideas on what kind of prompts that you should be writing and give you some guidance on that front. So thanks for that, Brian. We’ll throw that out there. All right. Know your memes now. There was a—there’s been a lot of stuff with image generation and memes, so we wanted to get that out. The first one was the one that I ran yesterday, which was a reminder that it’s going to be May.
John Wall: 25:55
And I was definitely underwhelmed by this one. When I put that in there, I was expecting something a little more and sinkish.
Katie Robbert: 26:04
So, obviously I know the original meme, but I don’t know what it means in this context. So—I mean—this is all a surprise to me as it is for hopefully a lot of our viewers. Like, this is all new information for me.
John Wall: 26:15
Yeah, I—you know—I figured we’re doing image memes, but—you know—it being May 1st, the—the May meme is on fire today. And so I—you know—dropped a prompt in there, see what kind of image. This one, I think was from Gemini. And—yeah—I was—you know—the original meme is Justin Timberlake, and this was kind of underwhelming, so we blew past that. The next one, I believe, is. This is. Again, it’s like keynote issues here. I can’t see my next slide, so I’m rolling the dice here. But I’m thinking where it’s Muppet images. Are you ready for the Muppets?
Katie Robbert: 26:50
I’m—I am ready. I love Muppets.
John Wall: 26:54
There we go. This is me and you as Muppets, huh? And so the thing with this is I’m still cursed by John Wall from the Washington Wizards, NBA basketball player, I was going to say.
Katie Robbert: 27:06
Is it confusing you with the other John Wall? Because otherwise, I have some concerns about how that turned out.
John Wall: 27:12
Right. Exactly where this is. No. So basically the idea is that—you know—50 billion John Wall basketball football photos are out there, and so that has outweighed everything else—you know—in the world out there. And so you and I have been dragged into the NBA.
Katie Robbert: 27:28
And I was going to say I couldn’t understand the basketball connection, but now it makes more sense.
John Wall: 27:33
Yeah, that’s there. And now I was actually. What do we have? We’re not bad for time. I was thinking about trying to clean this up and run it again. I don’t know. Do you want me to try?
Katie Robbert: 27:44
Yeah, if you—you know—I can sort of talk through what it is that you’re doing if you want to get that set up. I would give it our headshots if you have those handy, because otherwise it thinks that you’re an NBA player, which—I mean—you might have really good skills. And so here I am rambling to give you time to get this set up on the back end. So go ahead and be doing that—you know—and then because by association, it didn’t bother to try to figure out who I was, it just said, “Oh, well, if John Wall is an NBA player, then Katie Robbert must be part of the NBA as well and a commentator.” And so—like—in terms of me as a Muppet, this. If I gave it a picture of me with curly hair and a sunburn, maybe, but otherwise, I don’t think this is—you know—accurate. So I’m seeing your mouse move around a lot, John. Is this something we should skip and come back to?
John Wall: 28:39
Well, no, here, I’ve just thrown in the prompt.
Katie Robbert: 28:42
Oh.
John Wall: 28:43
In fact, all right, if we want to go live here, I have to.
Katie Robbert: 28:45
Yeah, let’s go live.
John Wall: 28:46
Switch this. Now I have to remove that.
Katie Robbert: 28:50
And then let me present a different.
John Wall: 28:51
Screen at a different file.
Katie Robbert: 28:56
But so—you know—again, this goes back to good prompting. And it’s not just about good prompting; it’s about giving enough background information in order to get what it is that you’re actually looking for. So if you. Oh, there’s. There’s us. I think that was from Marketing Profs. Oh, no, this was. Was this. This past year. Yeah, that was 2024. And so it’s—it’s not enough just to give it a prompt. If you want to give it more context and get more accurate results, give it background information. So if—for some reason—you know, you’re going through this whole—you know—”My team has to be Muppets,” maybe give it some images of Muppets, give it a lot of different versions of pictures of people on your team, and then give it the prompt. And that’s going to get you better results overall.
Katie Robbert: 29:49
But just saying, “You know, here’s my name. Turn me into a Muppet,” is not a strong enough prompt. And you can see that’s what happens because—you know—no offense to you, John J. Wall, but John Wall the basketball player is much more famous than you. So clearly that’s what it found first, because you gave it no context to say, “I’m looking for John J. Wall the podcaster, not John Wall the NBA player.”
John Wall: 30:14
All right, so it’s grinding and running. I can smell the smoke here as this is. Oh, wait. It was almost. I thought it was doing something, but.
Katie Robbert: 30:26
I think you have to. That’s our picture. That’s the first.
John Wall: 30:30
The original image.
Katie Robbert: 30:33
Let’s move on, and we’ll come back if it decides to do its thing.
John Wall: 30:36
Yeah, exactly. Let me remove this one here. Stop screen. It’s too bad that I should be able to tab between windows. I should be able to run a different one. I’m coming back to the presentation here.
Katie Robbert: 30:50
What we can do in the post-show, when we put this up on the website and everything, is we can also share some of the pictures that we didn’t get to show. So if the Gemini prompt actually goes through, we can show that version as well in the post-show notes.
John Wall: 31:06
All right, so that was that one. All right. This was the one you had teased in the beginning. Action figure. This is the other meme that’s out there. And it was actually kind of funny. It was over in DALL-E 2, where there’s some discussion of this. “Turn yourself into an action figure.” As I drill down in, here’s the prompt that we had used for this. If you want, I’ll throw the prompt. That’ll be over in the Analytics for Marketers board. You can go over there, get a copy of this, and you can try it out.
Katie Robbert: 31:35
I haven’t tried this, so it’s something that I see everyone else doing. And I was like, “Oh, I should get to that eventually.” And I just haven’t. But so I’m interested to see what happens.
John Wall: 31:44
All right. Yeah. And so I did a first run and we basically filled in a bunch of stuff about you. So the first shot was like,
Katie Robbert: 31:55
Okay. So wearing my leprechaun hoodie, I—yeah—I said hoodie.
John Wall: 32:05
And then I think I said something about St. Patrick’s Day scene. So it’s kind of weird. A green mask. We got the dog putting data on the coffee cup. That was not. I just said coffee cup. And it threw data on there. So that was pretty interesting. I was surprised with that.
Katie Robbert: 32:24
I feel like perhaps. And maybe can you go back a slide to the.
John Wall: 32:30
To the prompt? Yeah, yeah, to the.
Katie Robbert: 32:31
No, to the slide before that. So you see this Ryan Allen. And so it looks a little bit more 3D than the one that it came up with for me, but I want it. What I wanted to see was how detailed the face was or if it was still very cartoonish. Because I look at this, I’m like, “Well, that just kind of looks like a—like—a color form versus an action figure.”
John Wall: 32:55
Yeah. So that’s a great point, because I thought the same thing. I was like, “You know, that really looks cartoony. It barely looks real.” So we go back and fired up another prompt here. I was like, “Okay, let me put—you know—put all my stuff in there.” And the big thing was I switched it to more of a Kenner Star Wars action figure and make the figure more muscular—you know—to see what it could come up.
Katie Robbert: 33:17
Oh, sure. You get all the good details. That’s fine.
John Wall: 33:20
Here’s. I know. Isn’t this. I’m like, “Well, first of all, I didn’t put dwarfism in the prompt anywhere.” I don’t know.
Katie Robbert: 33:31
Well, and your—your action figure is so big and muscular that you’re outside of the packaging for said action figure.
John Wall: 33:39
Is it? And—like—for some reason, there’s no blister pack for the figure.
Katie Robbert: 33:43
No. Well, because you’re so muscular.
John Wall: 33:45
Right.
Katie Robbert: 33:46
Like, you sort of hulked out of it.
John Wall: 33:48
Yeah. So I—you know—if I were to do future prompts, I would dial it back a little bit. It did totally.
Katie Robbert: 33:53
I look like a Polly Pocket. And you’re just hulking out of your. And I—you know.
John Wall: 34:00
For some reason, I have two Thor hammers. I don’t know what that’s about.
Katie Robbert: 34:03
Two Thor hammers. That’s right. Okay. So, yeah, I gotta say, I’m underwhelmed by my action figure. I would not buy her off the shelf.
John Wall: 34:15
This is so funny. The comments have totally blown up here. The Muppet Show theme playing. Yes, absolutely. Craig mentioned coffee, puppy, laptop. “Believe in yourself” has said AI sense of humor. So—yeah—so that was—you know—the action figure meme. I—yeah—I don’t know. It’s a little bit creepy and weird as far as I’m concerned, but yeah.
Katie Robbert: 34:38
I think—you know—you could. I’m—I’m gonna try it and see if I can get a better version. So did you give the prompt a headshot as well for reference?
John Wall: 34:50
For both, yeah. Which is interesting. Yeah.
Katie Robbert: 34:53
Interesting. And so my action figure has decided that there is zero definition to my face.
John Wall: 34:58
Well, no, I think it was just. It took the cartoony a little too serious. And another thing.
Katie Robbert: 35:03
Yeah.
John Wall: 35:03
I didn’t get to do any slot machining on this. Like, it would be fun just to run the exact same thing again and just see how different it is—you know—or where it goes. All right. And—oh, yeah. All right. So yeah, we actually have. I was concerned we’re going to burn through this in 10 minutes, and we haven’t, so that’s great. But I do want to hit a couple of the other things. And Photoshop was the big one in doing show images. This was an interesting one I wanted to do. The episode was on AI clones, and so I said, “Just make me a show image of AI clones.” And it was interesting. This is again—I call this hot-dogging—the idea that you ask for steak, but you get something that’s based on everything out there that’s not as good.
John Wall: 35:43
And so because you’ve got Star Wars with Attack of the Clones and all this Star Wars stuff about clones out there, it actually pulled it towards Star Wars. There’s kind of a C-3PO here and R2-D2, and this is kind of a Vader over here that it’s pulled over in that direction.
Katie Robbert: 35:57
I was going to say, is this meant to be like famous clones throughout history? Because it’s very like. I don’t get it.
John Wall: 36:05
Yeah, I was just—you know—this was totally. Let me just throw it to you and see what the AI generates. And the prompt was—you know—”Create AI clones, do a batch of clones—you know—and just show me what you get.” And this was at the end—for some reason—it spread them over history, which was kind of interesting, but that’s where it got. And then this. I wanted to lead this as a transition to talk about Photoshop because all these other things are freestanding tools. But Photoshop has a bunch of stuff baked in, and there’s other tools that have this image generation baked right into the product. And I wanted to throw this one out here. This is Tom Webster from our interview with him, and I wanted to lead with the result first. What do you think about this, Katie?
John Wall: 36:47
Does this work?
Katie Robbert: 36:50
Again, it’s. I know Tom Webster well enough that I would look at this and not instantly think AI-generated. I would think it looked close enough. Like, I wouldn’t. Because—you know—here’s the thing, like with all of these images, with anything we put out there, the content, the social. As much as we want to think people are like studying it and cherishing it, these are not works of art. People are giving it maybe a two-second glance and then moving on; they’re giving a longer glance if something is apparently wrong with it or sticks out or there’s an anomaly. But there’s nothing really anomalous about this. It looks enough like Tom Webster that I would look at this and go, “Oh, Tom Webster.” And then move on with my day.
John Wall: 37:33
Yeah, it actually is Tom Webster. So this is the headshot that. So you’ve got that now. The big thing is I get these images all the time. I get regular headshots. And this one is better than average for Tom because it’s—it’s got most of them. Some are literally just ahead. And I’m like, “Okay, how the heck do I make just ahead into an 820-wide magazine format shot?” And so this is something that Photoshop does pretty well. If you give them the—I give them the Tom image and then I’m able to just say, “Okay, extend the background based on the original shot.” And this is what it comes up with.
Katie Robbert: 38:06
I’m guessing it does it pretty well because that’s what the software was meant to do in general—you know—so it’s not like you’re bringing it to an AI image generator that doesn’t do this. Like, Photoshop is meant to do this manually or with AI. So I’m not surprised. Like, you can see—like—maybe the background. Like, you could see if someone did—like—a rack focus where it’s like, focused on Tom. The background is meant to be blurry. It looks like it could be a regular photograph. You wouldn’t first immediately go, “Oh, must be AI.”
John Wall: 38:37
Yeah, right. It does cross over enough that—like—the first—your first inclination is not, “Oh, that’s fake.” Like, you’re just like, “Okay, that looks pretty decent” to give a couple of. So this is Mari Smith. You know, same deal, just the shot of her with the chair. And this looks like it could be a hotel lobby or something. And this is believable. Good run with that. And Casi Bruno here. This one was interesting. I had run it and it was—just got too busy—and so I used a little vignetting and—you know—brightened up her original shot so that she’s still focused. But yeah, these are all good. But. And all of these are still—you’re dealing with slot machining. You know, it’ll give you three that come out the door. And usually one of them is horrible.
John Wall: 39:18
You know, it just doesn’t work at all. And you can usually get one or two that are on the mark and same deal. You can refine your prompts and get closer if you want to go. But then—like—for pure image generation—like—this is—it’s still not that solid. We had a show where mariachis came up for some reason. And—like—that’s. Mariachi is a recurring theme. Sangria is a recurring theme on the show. But—you know—when you drill in close. Like those don’t look. Those are like nightmare.
Katie Robbert: 39:45
Like a nightmare.
John Wall: 39:46
Yeah, that’s like a little bit weird. So—so—the kind of pure AI—you know—it’s still not there for getting that together.
Katie Robbert: 39:54
But again, it goes back to the more—and this is true of text-based prompting and image prompting—the more background information you can give it, the more research you’ve done already, the more data you can provide, the better the results are going to be. But asking it to create something from scratch to your point is slot machining.
John Wall: 40:15
And you know, I don’t know about Photoshop in the last three or four months, but you know, the last times—few times I’ve dug in late—it’s just not there with faces. It just cannot make a face that works. If you want an intelligent fill or a background or something pulled out, yeah, it works great. But yeah, I’ve been struggling with that, actually. I had. Oh, well, we should. Let’s just see if our Muppet. Let me see if anything came up on the Muppet side there. Yeah, it’s choked on the Muppet. I don’t know what’s wrong with it. Although, let me download that. Just make sure that. Yeah, no, the download is still us, so it just can’t handle us as Muppets, unfortunately. I think that’s where we’re at on that. But I did want to take a couple seconds too.
John Wall: 40:54
Ethics and compliance have kind of come up. I mean, one thing for me was just rampant theft. You know, people, these machines grinding out stuff that are trademarked images or people and modifying images without consent. You know, teams that could use stuff to change images or do things they—you know—shouldn’t be doing with images. And you as an organizational behavior expert, this was where I wanted to at least get some of your opinions on where you think this stuff is going and what needs to be watched. And—and—where are we at today?
Katie Robbert: 41:26
Well, you know what’s interesting and—you know—this is something that I say with Chris a lot on the podcast and in other contexts is new tech doesn’t solve old problems. If anything, it just highlights existing problems and magnifies them. And so—you know—art theft and modifying images without consent, that’s not an AI problem. It’s just a problem that’s been exacerbated by AI and puts more of a spotlight on it because things are being generated more rapidly. But—you know—I mean, think about—you know—someone—you know—a famous Van Gogh painting—like—you know—Starry Night or—you know—the—like—his Sunflowers or something like that.
Katie Robbert: 42:05
There’s no shortage of people who have made replicas of that and pass them off as originals or close enough to try to make a quick buck off of it—you know—or—you know—there’s—you know—you have the brands that are on like Amazon or Etsy, they’re knockoffs of small businesses that the small business has to go out and find in order to see, “Hey, you’re selling my thing. It’s not proper trademark.” Like this is not—I guess my point is this is not an AI-specific issue. You know, theft and the modifying images. It’s an issue in general with content creation, image generation. AI just makes it happen faster or sometimes harder to spot. And so—you know—it done it goes back to making sure you have a clear user story: “Why am I using image generation in the first place?”
Katie Robbert: 42:58
But then also digging deeper into the five Ps: purpose, people, process, platform performance, and your risk mitigation. So with AI in general—without getting too deep into—you know—the compliance, you need to be willing to accept a certain amount of risk with whatever is generated with AI so that you can say, “AI created this thing, I stand behind it. I’m putting it out into the public now because now it’s not just AI-generated; my name is on it as well as someone who is endorsing this thing—you know—And so with image generation, you’re saying, ‘Okay, AI created this thing, but now my name is under it. So if it’s copyright infringement, if it’s—you know—just a straight up—you know—rip off of something, I’ve done my due diligence and I’m okay with that.'”
Katie Robbert: 43:48
And so the technology—this is going to sound a little convoluted—but the technology is not responsible for not ripping off images. You as the human are responsible for making sure you’re not putting out things that you won’t stand behind, which could be ripped off. So you have to do your due diligence to make sure that’s not happening.
John Wall: 44:09
That sounds good. Let me think. One other upcoming meme, “Believe in yourself,” said, “May the fourth be with you.” That is another one coming down the pipe for Star Wars fans. Yeah, good artist copy, great artist steel. This is true. The only other one was as far as compliance. Is there anything that you’re concerned with as far as the slot machining? You know, the fact that do you make this your process and it’s just—you just say, “Okay, you know, try 10 times and hopefully you get something good,” or—you know—is that just kind of gives me an instant angst when I look at processes like that.
Katie Robbert: 44:46
Well, again, it goes back to making sure that you are being specific enough in your prompting and giving enough background information so that you’re not saying something as simple as, “Generate, you know, a tree with a squirrel,” and then just sort of like crossing your fingers and hoping for the best. You need to say, “I specifically want an oak tree that is—you know—in the late summer sun, in fall in New England with a squirrel that’s gray,” and—you know—like whatever the details are, because if you’re leaving it up to AI to make those decisions, great, you get what you pay for. It’s going to look kind of crappy. If you want something more specific, that’s where you know that built into your process.
Katie Robbert: 45:31
And so again, I keep harping on do your—you know—due diligence, do your requirements ahead of time. This is why. Because if you leave it up to chance and just say, “Oh, this is what AI came up with,” like, it’s going to be crappy.
John Wall: 45:44
All right.
Katie Robbert: 45:45
Yeah.
John Wall: 45:45
When it comes to leaving it to chance, I think that pretty much wraps up that deck. We’ve rolled the dice and got some interesting stuff.
Katie Robbert: 45:51
I mean, we covered a lot. You know, AI image generation has come pretty far since the last time, John, you and I covered this two years ago, as we’ve seen. But it still has a ways to go. And—you know—you as the human are still in charge. You’re the one who’s responsible for giving it clear instruction, giving it background information, and then making sure that what comes out that you’re going to put out there in public is something that you’re comfortable with; that it’s not—you know—a rip off, copyright infringement because you don’t want to deal with lawyers. That is just not something you want to have to do. So make sure that part of your process includes human intervention and QA.
John Wall: 46:32
I like it. That’s all I’ve got. Anything else from you?
Katie Robbert: 46:34
I think that’s it. I think—you know—we have a lot of other stuff we wanted to talk about. Something we want to talk about maybe up in the upcoming weeks is voice generation. I think that we can cover that. But yeah, I think that does it for today. So let’s see if I can remember how to close out the show.
John Wall: 46:57
Thanks for watching today. Be sure to subscribe to our show wherever you’re watching it. For more resources and to learn more, check out the Trust Insights podcast at TrustInsights.ai, TI podcast, and a weekly email newsletter at TrustInsights.ai. Got questions about what you saw in today’s episode? Join our free Analytics for Marketers Slack Group at TrustInsights.ai/analytics-for-marketers. See you next time!
|
Need help with your marketing AI and analytics? |
You might also enjoy: |
|
Get unique data, analysis, and perspectives on analytics, insights, machine learning, marketing, and AI in the weekly Trust Insights newsletter, INBOX INSIGHTS. Subscribe now for free; new issues every Wednesday! |
Want to learn more about data, analytics, and insights? Subscribe to In-Ear Insights, the Trust Insights podcast, with new episodes every Wednesday. |
Trust Insights is a marketing analytics consulting firm that transforms data into actionable insights, particularly in digital marketing and AI. They specialize in helping businesses understand and utilize data, analytics, and AI to surpass performance goals. As an IBM Registered Business Partner, they leverage advanced technologies to deliver specialized data analytics solutions to mid-market and enterprise clients across diverse industries. Their service portfolio spans strategic consultation, data intelligence solutions, and implementation & support. Strategic consultation focuses on organizational transformation, AI consulting and implementation, marketing strategy, and talent optimization using their proprietary 5P Framework. Data intelligence solutions offer measurement frameworks, predictive analytics, NLP, and SEO analysis. Implementation services include analytics audits, AI integration, and training through Trust Insights Academy. Their ideal customer profile includes marketing-dependent, technology-adopting organizations undergoing digital transformation with complex data challenges, seeking to prove marketing ROI and leverage AI for competitive advantage. Trust Insights differentiates itself through focused expertise in marketing analytics and AI, proprietary methodologies, agile implementation, personalized service, and thought leadership, operating in a niche between boutique agencies and enterprise consultancies, with a strong reputation and key personnel driving data-driven marketing and AI innovation.
Newsletter
Podcast
This “So What?” episode sounds super interesting! I remember when AI image generation was a joke. Now it’s actually useful? Gotta check this out to see what’s new and if it can actually help with marketing stuff.