So What? AI Image Generation Bake-off

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? In the Trust Insights weekly livestream, you’ll learn how experimenting with AI image-generation tools can help you produce more realistic images. You’ll discover how to optimize your prompts to get the best possible results from AI image generators, whether you are using text or image prompts. You’ll also learn how AI models interpret prompts differently and why providing more information to the AI model is always better. Download the free AI image prompt guide to start creating amazing AI images today!

Watch the video here:

So What? AI Image Generation Bake-off

Watch this video on YouTube

Can’t see anything? Watch it on YouTube here.

In this episode you’ll learn:

Which models to use for AI image generation
What biases to be aware of
How to prompt different models for 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 00:32

Well, hey everyone. Happy Thursday. Welcome to so what, the marketing, analytics, and insights live show. I am Katie, joined by Chris and John. Hey fellas.

Christopher Penn 00:41

Hello. Almost at it.

Katie Robbert 00:44

Almost had it. This week we are doing an AI image generation bake-off. We spend a lot of time talking about generative AI in terms of what it can do with text, and those are some of the primary use cases. But there’s a lot of people using generative AI to create images.

When generative AI first hit the market in a more mainstream way a couple of years ago—I don’t mean to say it’s new, but when it hit in a way that was accessible to people—everyone was trying to figure out, how do I generate an image with it? Why does it look so poor? Why does the hand have eight fingers when it should have five? Why are there three people when I asked for one? There was a lot of issues with it.

Katie Robbert 01:33

Prompting for images is a different skillset than prompting for text. So today what we thought we would do is go over a different sampling of the available image generation tools and do what we like to call a bake-off. So Chris, where would you like to start today?

Christopher Penn 01:53

Let’s start with the image that started this conversation. The image that started this conversation, Katie, is this splendid, wonderful image that was part of the first version of Katie GPT. This was made in DALL-E 3, which is part of OpenAI’s offering. The prompt for this was “a professional blonde woman with black rectangular glasses, wearing a green hoodie, typing on a Samsung smartphone. The woman has a cheerful demeanor.” Now, Katie, I know you wanted to talk specifically about this image and what kind of happened here.

Katie Robbert 02:31

First of all, I’m not wearing my green hoodie today. I’m wearing a gray hoodie. So that’s sort of not that, but that’s not the problem. Granted, this is the avatar for KDGPT, and recently I had an opportunity to give her a closer look because my friend Ann Handley was asking about it.

I admittedly had never really taken a look at the image. I just sort of saw, in a very small scale, a blonde woman wearing a hoodie, blah, blah. Our prompt was fairly generic as well—a blonde woman, so on and so forth. But then when you actually look at the image, the AI made a lot of assumptions and took a lot of liberties. Primarily, this image looks like she has had a heck of a lot of plastic surgery.

Katie Robbert 03:20

The image very much evokes fillers and Botox, and definitely an underweight, someone probably dying for a cheeseburger but is going to deny herself one. We didn’t give those prompts, we didn’t give those directions to the AI. The AI said, well, you want an image of a woman? Here’s what I think a woman should look like. I very quickly had a strong reaction to this.

Now, granted, she has the right amount of fingers, but the fingers are long and bony. Everything feels off about this. I don’t know that if we had given it more direction to say, a woman who hasn’t had fillers or a woman who hasn’t had plastic surgery, it would have done much better.

Katie Robbert 04:10

I think we would have gone the opposite direction with some assumptions about what a woman should and shouldn’t look like. That was where this conversation started. So, while yes, this is technically KDGPT, I’m going to be the first one to say this looks nothing like me.

Christopher Penn 04:30

All AI models have biases built into them because of the training data. They were trained on different image sets and different image libraries, and depending on the system you use, it will determine its biases, because they were all trained on different imagery.

This is a slide from one of our keynote talks that we give. This is using Meta’s AI generator, and you can see the two-word prompts that we used to generate each of these images—these two-word prompts. You should never have a prompt this short, but to illustrate the point, we say, a lawyer, a senator, a surgeon, an accountant, a nurse, a secretary, a teacher, a librarian. There are a lot of assumptions that these tools are making. You can see there’s a very clear gender bias.

Christopher Penn 05:13

There is a very clear racial bias in these images, and then of course all the usual things that are wrong in these images. Part of that is prompting because these prompts are woefully insufficient, and part of it is bias in the training data that the tools were given. So, to effectively deal with this, we have to do two things. Number one, we have to understand how an image generator works because it’s not the same as a language generator. Number two, what we need to do to work with the specific tools that are out there. Today, we’re going to do a bake-off with four of the five tools. Now, if you would like to skip ahead to the end of the…

Christopher Penn 05:58

…of our program today, we have a very comprehensive guide we actually put together for this livestream. Go to trustinsights.ai/image-prompts, and you can get the ten-page PDF of the prompt structures for all five systems, along with a custom GPT that we’ll be talking about in a bit.

The first thing we need to figure out is what are these systems looking for? What we did was took over 100 different articles about generative AI image models from all these different systems, and about 40 different YouTube videos of people talking about the ultimate Midjourney prompt and all that stuff—stuffed it into Google Gemini and said, “What do these, what do all these different perspectives have in common?” and assembled some templates.

Christopher Penn 06:50

We’re not digging into the specifics, but what I want to call out is for each of these systems, they have a lot of opportunities. Let’s look at the one for DALL-E 3. These are the things that good DALL-E prompts have in common: an image command, an image type. Is it a photo? Is it an illustration? What’s the main subject? What are the descriptive modifiers of the main subject? So in our case, we’re talking about a professional blonde woman. What’s the setting? Where is this located?

What are the style cues? Photorealistic? What are the key features that you want to emphasize? What are the specifics about those key features? What are the analogies or comparisons? DALL-E, in particular, is very sensitive to anything copyrighted. You can say van Gogh, you cannot say Ansel Adams.

Christopher Penn 07:41

Just as a note there: color schemes, lighting and composition, and negative prompts. The reason that this system is—and to be fair, all the image models have this much detail—is because underneath the hood, the way they’ve been trained is they’ve been given a gazillion and a half different images.

Whether or not they may have permission to use them is a separate show, and the captions for those images were either provided or automatically generated. Then there’s a small, a very small language model that sits in front of it that essentially goes, mashes back and forth with captions to try and match essentially what the image should look like conceptually. If you’ve ever heard the expression, “Michelangelo carved the statue of David by taking a block of marble and chipping away everything that wasn’t David…”

Christopher Penn 08:42

That’s how a diffuser image model actually works. It has a pile of words as a pile of statistical association about what those words are— a dog on a skateboard wearing a tutu. Then from its knowledge, it has to chip away at all the pixels in this image that are not that representation. That’s why the prompt structure for image models is so weird, especially when you look at a prompt, for example, like Midjourney.

Midjourney’s prompts get all kinds of… wow, they look like, hieroglyphics practically sometimes, or arcane spells from a spell book because the way they’ve trained the model is very specific to all those inputs.

Katie Robbert 09:26

I don’t think the casual user was prepared to have to think about all of these different elements, but it certainly… So it certainly explains why if you open up an AI image generator and say, “I’m looking for a picture of a dog on jet skis in the Pacific Ocean,” you’re going to get something that maybe resembles that, but probably not because there’s a lot of specifics in here that tell me, similar to using generative AI for text, for language tasks, you really need to know what you’re asking for.

Christopher Penn 10:07

Yes, exactly. It’s like sort of the genie of fairy tales. You get what you ask for, so if you ask for the wrong thing, you will get not what you wanted.

Let’s do our first test. Before we talk about prompting, let’s do our first test and take these four different systems, and we’re going to take this original prompt and feed it to all four systems just to see what we get now because one of the other things about KDGPT’s avatar was it was generated with DALL-E about a year ago, I think. All the models have changed since then.

So let’s start with “create an image, a photo of a professional.” So that’s what we’re going to do. We’re going to create a photo of a professional with black rectangular glasses, wearing a green hoodie, typing on a Samsung smartphone.

Christopher Penn 10:59

“The woman has a cheerful demeanor.” I’m going to start with ChatGPT. We’re going to move over into Midjourney. We’re going to go into Google Gemini, and we’ll go into Meta AI. Meta AI requires the “imagine” tag upfront. So let’s see what we get. ChatGPT seems to have created a…

Katie Robbert 11:25

Well, you didn’t specify that the lights had to be on in the room.

Christopher Penn 11:33

ChatGPT, you need to try this again. Yeah, this is no good.

Katie Robbert 11:38

So if you say, “But the lights also have to be on in the room,” maybe she’ll appear.

Christopher Penn 11:44

My goodness. Think about that one. Let’s see what Midjourney came up with. Midjourney came up with four different variations here. You can obviously… there are refinement buttons. But that’s what Midjourney came up with. Those are certainly an improvement over the existing KG.

Katie Robbert 12:05

They’re… I mean they’re all fairly generic looking.

Christopher Penn 12:14

Yes.

Katie Robbert 12:15

Which is good because we didn’t specify anything about ethnicity, which is interesting because we said “a blonde woman.” We didn’t say “a blonde Caucasian woman.” It made that assumption which to me is something that kind of sticks out. If you say blonde, you must mean a white woman, which is absolutely not true in real life. That, to me, that’s sort of my first. I’m just going to file that away for a discussion point later on in this episode.

Christopher Penn 12:45

Yep, the lights are still not on.

Katie Robbert 12:49

Listen, this is true of KDGPT as well. The lights are off. Nobody’s home. You don’t notice.

Christopher Penn 13:00

Let’s see what Gemini came up with. Gemini came up with…

John Wall 13:03

This is…

Christopher Penn 13:03

…This is Google’s Imagen 3 model. So here we have, again, same assumptions.

Katie Robbert 13:09

Same assumptions. Yeah. The hair is… well, also the hair is very light blonde. So when you say blonde, they’re thinking like a white blonde. The hair is straight. The woman is, in all of the images, arguably very thin and very white.

Christopher Penn 13:29

Yep. Let’s see what Meta came up with. Meta came up with very similar. Let’s… I think it’s just on vacation today.

Katie Robbert 13:40

Yeah, that’s fine.

Christopher Penn 13:42

So these are Meta’s version. Again, compared to a year ago, they do have the correct number of fingers, which is a large improvement.

Katie Robbert 13:52

The image quality has drastically improved. I will say that these images look more human than the original image, which definitely had a bit of a space-alien effect to the facial features.

John’s like, “Yep, she sure did.”

But the biases have not improved, and…

Christopher Penn 14:18

A good chunk of that has to do with the prompt. The first question I have is, for this prompt in particular, what would you change? Keeping in mind all the different… putting aside technical things like lighting and stuff like that, what would you change to make this prompt more robust?

Katie Robbert 14:39

Man, I’m almost afraid to because I’m… I’m nervous about what the output is going to be. But, I mean, I think the first thing I would do is if it’s meant to be a photo of me personally, then I need to give it more detail about my physical appearance. Maybe it’s, “create a photo of a professional blonde woman in her forties.”

Because all of the other… all of these images skew really young, which is flattering. But let’s be realistic about this. So, “a blonde in her forties of average build”—so that doesn’t look like I’m emaciated—”with black rectangular glasses,” that’s fine. “If we’re saying a blonde woman with long, wavy hair.” It’s sort of… and it’s…

Katie Robbert 15:38

I’m not changing your prompt drastically, but I’m trying to give it more detail so that, again, if it’s meant to be someone who looks like me, then I have to give it that information.

The other thing I guess I would say is, “with a round face” because, again, all of these faces are very angular. Is there anything obvious that I’m missing?

Christopher Penn 16:08

Here is where what is true in language models is also true of image models, particularly today’s multimodal models. If you want a better prompt, you are better off bringing data to the party because, a lot of the time, it’s very difficult for us to encapsulate how we want to describe something. It can be very challenging.

One of the things we could do here… let’s go ahead, and I’m just going to change views for a moment, and I am going to screengrab you, Katie, off of the livestream. I’m sure I’m not making faces. Well… and we can go into a tool like ChatGPT and say and ask the question, like, “How would you describe this image in terms of a prompt to be used for generative AI to create a similar image?”

Katie Robbert 17:19

Now, while that’s processing, I want to talk a little bit about why I was hesitant to give it more descriptors, if that’s okay, Chris.

Christopher Penn 17:27

Of course. Absolutely.

Katie Robbert 17:29

Obviously, we saw the biases, the bias with a “professional blonde woman” and it very much created the Botox and the fillers, and someone who’s a thinner build. As I’m sort of describing myself, I’m talking about an average build around face, someone in her forties. Perhaps it’s my own bias, but my concern is that it’s going to return something that’s very unflattering. I’m almost scared to see what that looks like if you give it those kinds of prompts because of how much bias is built into AI. That’s sort of where I was. I could certainly describe myself, but I don’t think I would be describing myself in the way that the language model would understand.

Christopher Penn 18:23

Gemini says, “You’re a young woman with long brown hair and glasses, sitting in a gaming chair in front of a microphone. You’re wearing a gray hoodie with earbuds. The background is slightly out of focus and includes a bookshelf with plants, framed pictures, and decorative pillows. Lighting is soft and natural. The overall mood is casual and comfortable, suggesting a podcasting or streaming setup.”

Katie Robbert 18:42

It’s funny because it is a gaming chair because these chairs are meant to be sat in for long periods of time, much like when you work from home and run your own business.

Christopher Penn 18:53

Exactly. That’s what the model saw.

Katie Robbert 18:56

Yeah. In this context, I don’t disagree with that. Sorry for the pun, we don’t need to split hairs over the color of my hair. It’s fine. If they want to call it brown, blonde, dark blonde, whatever. That, to me, doesn’t matter because the rest of it actually got pretty correct.

Christopher Penn 19:24

Mm-hmm. Here’s what we can do. We have that. That will add more precision to this we can take. As I mentioned at the beginning of the show, we have this guide that we put together of all of these different prompt templates.

Rather than sit here and try and pick this apart piece by piece, what we can do is have AI do it. Say, “Here’s the templates. Here’s what we know to be true. You do it” because it’s too much of a pain in the butt for us to do it ourselves.

Let’s do this. Let’s take this. I’m hesitant to choose ChatGPT because it’s having a real hard time today.

Katie Robbert 20:06

Doesn’t want to play along today.

Christopher Penn 20:08

It really doesn’t want to play along today. Let’s… you. Did I make a gem out of this? I don’t remember if I made a gem out of this or not. I did make a gem.

Katie Robbert 20:18

For those who don’t know, a gem is your own custom model. We’ve done this on other episodes. If you want to see what those look like, you can go to Trust Insights AI YouTube, find the “So What?” playlist, and you can see how to build your own gem. A gem is essentially similar to a custom GPT—it’s a large language model that you have created the system instructions for that you can now interact with.

Christopher Penn 20:44

These are the instructions for our gem. This is essentially the prompt guide that we put together that you’ll be able to download, with some basic instructions saying, “You’re going to get input on what the user wants to make, and the user is going to specify a platform. Then you, the AI, are going to find the appropriate template, add in the relevant missing details to the extent that you can, and then create…”

Talk the user through your choices, and then spit this out. I’m going to go ahead and start this gem, and let’s… ChatGPT still is a little on the iffy side.

I’m going to take the original prompt that we started with here, and I’m going to augment it with the results that we just got from AI Studio. Am I on the right one?

Christopher Penn 21:38

Yes, I am. Nope.

Katie Robbert 21:40

I was going to say we actually haven’t tested the updated prompt to see what would happen.

Christopher Penn 21:45

Right, so we’re going to. Well, do you want to test the updated prompt as is before we…

Katie Robbert 21:49

I would say let’s test it just to show how it’s still… I mean, we know it’s still insufficient, but I’m… now it’s just the curiosity.

Christopher Penn 21:59

Okay. So, the difference between Midjourney… we’ll start a new chat with Gemini, we’ll start a new chat with Meta AI and ChatGPT, how you feeling today?

The lights are still off. Probably Midjourney is working on it, Gemini is working on it. Meta AI came up with this.

Katie Robbert 22:25

Huh. I mean, this woman has noticeably more wrinkles. Her fillers may have fallen a little bit on the bottom left-hand side. I mean, AI, get it together.

It’s interesting. The top left-hand one is still by far the youngest. The bottom left is definitely someone in her fifties or sixties, which there is nothing wrong with that, but it’s the way that AI is interpreting age. That’s what I was worried about.

Christopher Penn 23:06

Yep. Also, what the hell is she typing on?

John Wall 23:10

Oh, yeah.

Christopher Penn 23:11

What is that?

Katie Robbert 23:12

That’s not a problem. I don’t know, John. You’re the techie, you should… Like, is it a very… it’s either a very large phone or a very tiny laptop.

Christopher Penn 23:20

No, I’ve still off here.

John Wall 23:22

Like, there’s a lot of… when I was looking at podcasting gear, it’ll just make up weird gear, weird boxes with five wires coming out of it. It’s just a portmanteau of five devices. So that’s like a sort of a phone, laptop, tablet thing.

Christopher Penn 23:42

So this is Gemini’s pass at this.

Katie Robbert 23:48

Interesting. This one, to me, so far, is the best representation of a normal-looking image of what you would expect. There’s nothing overly offensive about it. She looks about age-appropriate. We said, “average build, round face,” and it didn’t take it to an extreme place. This, so far, is my frontrunner.

Christopher Penn 24:19

And then here is Midjourney’s.

Katie Robbert 24:21

What were you going to say, John?

John Wall 24:22

I was going to say that last one is the first one that I would look at and not question if it was AI. I mean, that looks like it could easily be a straight-up photo of somebody.

Katie Robbert 24:33

Mm-hmm. And what was the last one?

Christopher Penn 24:36

Here’s Midjourney’s, the top row.

Katie Robbert 24:39

They honestly… they don’t look drastically different from the bottom row. I mean, this could all be the same woman at different points during the day, which I feel like Midjourney is just like, “Okay, you want a blonde woman? Here’s a blonde woman. This is all I got in the bank. This is all you’re gonna get.”

John Wall 24:56

Yeah. It’s been a little weird that across all the models, they are all looking pretty much the same.

Christopher Penn 25:03

Again, that has to do a lot with the input of that specific prompt. What we can do now is take this prompt, and we can put it inside of our gem. What we’re going to see is we’re going to see some of the language differences.

We’re going to say, “The platform is Meta AI. Here is the prompt for you to optimize.”

I’m not going to take the Gemini feedback one. I want to take the same one that we’ve been working with, Katie, and have it… It’s going to step through the process of what does it know. “So the original prompt, here’s the template mapping, augmentation… explicitly stated main subject, descriptive modifiers, setting, detail. It’s a modern office. It’s going to add in well-lit, medium shots, ethnicity, descent not specified.”

Christopher Penn 26:00

We’ll omit it and then see, here’s the final prompt. This goes into Meta. Let’s see if Meta’s AI can do this well.

Katie Robbert 26:13

But also, you hear that, ChatGPT? Well-lit. Turn the lights on.

Christopher Penn 26:20

Okay. Let’s get that cranking. Let’s now say, “Now, rewrite the same for Gemini Imagen 3. There’s the original prompt. It’s going to go through the Imagen format, sixteen by nine, contemporary.” Let’s take that over into Google, start a new chat.

Okay. Next, we want to do the same for Midjourney and we start the same original prompt. There’s our template, main subject, well-lit…

Katie Robbert 27:16

…Well-lit. Listen, apparently, that is some qualifier that you have to include. You can put all the images you want, but you’ve got to specify that the lights are on in the room.

Christopher Penn 27:28

Exactly. Let’s put this… Oh, “Prompt starting with the web interface.” That’s a Discord command. Let’s put in this. All right. So while that’s… what the heck, we’re going to try it. All right. “Rewrite the prompt for DALL-E.” We’ll see if maybe this will get the lights to turn on.

Katie Robbert 27:50

Oh man. Sometimes you just got to work in the dark and it’s okay.

Christopher Penn 27:54

Exactly. Final prompt.

Katie Robbert 27:59

While those are generating, what’s interesting is… when, again, when generative AI became a bit more mainstream a couple of years ago, there were services that allowed you to upload your headshot. It would generate, like, 20 additional headshots of you, but you had to give it enough for the model to work from. You had to have, like, six to eight images of yourself for reference, but then it would regenerate your image.

Are those… so those are sort of more user-friendly services. I assume the way that those models work are roughly the same as what you’re describing here in the livestream.

Christopher Penn 28:39

In most cases, those services are built on top of a foundation image model like DALL-E, or probably more like anything like Stable Diffusion because Stable Diffusion, as an open-source model, is much lower cost to operate. If you’re trying to run a startup, you’re going to pick the cheapest.

The lights are still not on at OpenAI. Let’s see what Meta came up with. This is what Meta came up with with the revised prompt.

Katie Robbert 29:12

Okay.

Christopher Penn 29:14

It’s better than the previous attempt. It’s still not great.

Katie Robbert 29:18

It’s still not great. No.

Christopher Penn 29:19

Right. Gemini came up with this with the revised prompt, which is actually a little bit better.

Katie Robbert 29:26

It is. I would agree that this one, again, so far, seems the most… You wouldn’t blink an eye to be like, “Is that AI or is that a person?” Exactly. And is that an age-appropriate person?

Christopher Penn 29:38

Right. Midjourney is getting closer, but still not quite right.

Katie Robbert 29:45

I would disagree. I would say it’s getting farther away because, in the top left, you have a 12-year-old, in the top right… I’m not really sure what’s happening there, actually. I’m not really sure what’s happening in any of them because now it’s like, “Woman what? What is… what is woman like?” These look worse, I would say.

John Wall 30:11

Yeah.

Katie Robbert 30:11

What do you think, John?

John Wall 30:14

These are still treading over in weird land. Yeah. Like, every one of these, I’m like, “Yeah, no, there’s something not right with these shots.”

Christopher Penn 30:21

Exactly. For this particular bake-off, Gemini’s Imagen 3 model seems to have done the best.

Now, when it comes to image generation, remember that each model is going to have its own strengths and weaknesses. Some will be good at illustration and art, others will be good at painting, others will be good at photorealistic images. Some, the lights will just be off permanently.

Katie Robbert 30:52

Well, and I know we tried to do it with OpenAI, but one of the things that, to me, sort of sticks out is if you’re looking to sort of do this activity that we’re doing when you have an image of the real person… We haven’t given, except for OpenAI, an image of me as a point of reference to say, “It should look like this person,” which to me sort of seems like the missed opportunity.

Katie Robbert 31:20

Now, for a lot of stuff, you’re going to be needing to create something net new. You can give it… just like you give these models for language tasks, data points and references, and whatever training data that you have. You would want to do the same for the images and say, “This is the style of image, or this is the style of person,” or whatever the thing is you’re trying to create. Like, “This is a dog. This is the dog breed that I want to create images of.”

Christopher Penn 31:52

Exactly. Ideally, you would do so in the native format of the model that you want to use. If we go back to… let me go back to my screen share thing.

If we go back to this, and we go back to the instructions that are in the prompt template, I can give those instructions to the model and say…

First, here is some information about generative AI image generation by individual system templates. Read through this and respond… okay. If you understand it, and we put in our data… and it should just say, “Okay.”

Now, to what you were saying earlier, Katie, we can take this image and say, “From this image, construct a prompt for the Midjourney system to faithfully replicate this image as close as possible in the Midjourney template.”

Katie Robbert 33:17

You might be wondering, “Well, what is the use case? You already have an image of yourself.” Maybe you don’t have a professional headshot image of yourself, but you have like this. This happens to be a screengrab.

This is a way for you to start to play around with, “Could I make headshots for myself?” While I’m also getting skilled up on how to better use generative AI or prompting for image generation.

You know what you’re supposed to look like, so it’s actually a really good A/B test to say, “Well, does it look like me or not?” That’s going to help you not only evaluate the different models to figure out which one works best for the style of image that you like, but also to help…

Oh boy, oh my goodness. Skill up on… Lord.

Katie Robbert 34:13

I forgot what I was saying.

Christopher Penn 34:15

This is what’s happening. This goes back to what we were talking about earlier. This is the fusion process. What is happening is it is taking that prompt, and is trying to chip away at everything that isn’t in that prompt to come up with an answer. Based on what we got. Gemini 1.5…

To construct for a Midjourney prop… this is what it came up with.

Katie Robbert 34:37

All right, John. I like at the…

John Wall 34:39

Bottom right, the seldom-seen triple microphone.

Katie Robbert 34:48

Oh boy. Let me ask you straight, John. Do any of these resemble me?

John Wall 34:53

If I had to pick one, the bottom right is closest, but I also love the top right, using the pop filter as a light shield.

Christopher Penn 35:05

That’s interesting. Take…

Katie Robbert 35:09

Oh boy. Yeah. I mean, at least they all look human.

John Wall 35:17

Oh, that is another… I would say all of them pass the human test.

Katie Robbert 35:21

Like they all look human. To Midjourney’s credit, they’re not all identical-looking like replicas of themselves. It’s not like the Stepford Wife clones of the previous images, right?

Christopher Penn 35:35

Yep. To your point, Katie, reverse engineering off data you already have is a good idea. It’s a good way to get these tools to work with the data that exists and turn it into something more useful.

We’re finding out across the spectrum, in all forms of generative AI…

A, the more data you bring, the better, and…

B, the more specific your prompts are—especially specific to the system and the model you’re using—the better the performance you’re going to get. We got better performance out of Midjourney by taking existing data and essentially reverse-engineering a prompt from that data in the Midjourney format.

Katie Robbert 36:17

It’s interesting because while those images didn’t… I don’t think they looked like me, we only gave it one reference point. I think that’s really where you want to figure out… like, if you’re doing this kind of exercise… like this is me on a Thursday afternoon after I had dental surgery this morning, and I’m nothing ready for a professional photo to be taken. So there’s… only… only so much it’s going to do with the images.

Katie Robbert 36:50

So if you want to also give it other versions, other data points to reference that are different… if you don’t like the results… like, “Okay, then you need to give it better data.”

I think that’s sort of the long-winded point that I’m making is, if you only give it one reference point and you don’t like the results, you need more reference points.

Christopher Penn 37:10

Exactly. More data is better.

Now, some of the systems, particularly Flux one—which I wanted to show on the livestream today but we can’t because I run it locally and it consumes 100% of my CPU, would not be able to have a livestream at the same time—can take a raw image as your starting point and then essentially upsample it, turn it into other stuff so you can start. You can, for some of the systems, you can start with a starting image and then build backwards from that starting image into something that is more usable.

Christopher Penn 37:49

When you were talking about the headshot generator, that’s essentially what those things do is they take their best guess at a prompt from an image and then tune it up to say, “Let’s turn this into a professional headshot of this person.”

Katie Robbert 38:05

What do you think, John? Are you going to make your clone army?

John Wall 38:08

Yeah. It’s an interesting use. I mean, if I had to come up with fake things, I think the big thing would be: is it as good as stock photo?

If I needed to pull up, “Give me just a random… red-haired guy working at a laptop”—it can definitely get close enough that I don’t need to go to one of the other providers.

That’s good. But… I definitely… both the triple microphone and the light pop filter… you know, that kind of stuff, you get back on the weird side and it gets to… you have to keep an eye on it to make sure it doesn’t run south.

Christopher Penn 38:43

Yep.

Question in the comments. Torbjorn was asking, “Is it possible to [create a] giant video with [a] good resemblance of a specific person?”

Yes. These are not the tools to do that. There are several companies out there—HeyGen is probably the best-known one—where you upload two minutes of video and then go through an ID process where they validate that you are who you say you are, and they require you to have government ID, and do a visual scan of yourself to prove that you have the rights to use your image. Then, that can very clearly replicate—eerily so—a person, and you can give a text prompt and it will then create the video from that file.

There are versions that do not do… that have the identity check, and we will not be discussing them.

Katie Robbert 39:38

On previous episodes, we’ve talked about using generative AI to generate podcast episodes, for example. If you want to check those out, you can go to Trust Insights AI YouTube and find the “So What?” playlist where we’ve talked about those things in a different context.

I think it is interesting to see, in a short amount of time, the strides that generative AI has made. Because, while yes, some of those images were still on the weird side, they all looked human, they all had the right amount of features, they all had two eyes, a nose, and a mouth. The hands weren’t visible in all of the images, but the proportions for the most part were… pretty regular.

Katie Robbert 40:33

That alone just shows how far generative AI image generation has come in a short amount of time.

Christopher Penn 40:42

Exactly. What I would encourage folks to do is… if you would like to mess around…

One of the things that we did… OpenAI’s DALL-E is not playing nicely, but the custom GPTs are. We have a custom GPT called the “AI Image Generation Prompt Generator.” I probably should have named that something else, probably.

What it does is exactly what we showed on the show today. You can give it a starting prompt, tell it what platform to use, and it will walk through the same logic, and come up with a revised prompt that more closely adheres to what the individual systems are doing.

If that’s of interest, you can get that. It’s totally free, it’s just within the PDF, and we would encourage you to download that. All we need to [from you] is to fill out a form.

Christopher Penn 41:34

Any other things that I forgot to mention?

Katie Robbert 41:38

Not that you forgot to mention, but… if you’re using image generation, I feel like it’s sort of akin to describing something to a sketch artist. If you’re trying to construct a prompt, imagine that the AI on the other side is that sketch artist.

You have to say, “It’s a man of this race, with these features, with this hairstyle.” If you’re getting frustrated that it’s not giving you back exactly what you wanted, you have to look at the information you gave it and say, “Did I give it enough information? Where is it filling in the blanks?”

Katie Robbert 42:15

Similar to the exercises that we do with language-based prompts and priming is, you really want to say, “This is the information I have,” and iterate with it and say, “What am I missing? What questions do you have?” before you just go ahead and give it a prompt.

It doesn’t give you what you want, you get frustrated, you close it, you walk away—aka what I do—is you have to actually prime it to say, “This is what I have. What else do you need? What information can I provide that would get me a better result?” Treat it the same way you would with a language prompt. In that respect, I would agree.

Christopher Penn 42:58

The prompt breakdown that’s in our guide is essentially the requirements-gathering for the image. These are the requirements: descriptive modifiers, scene details, lighting, style cues… our art… move at [an] artistic level. All [of] that stuff… is stuff that, ideally, you think about first, or you have a conversation with GPT or a gem or whatever, to go back and forth and say, “I actually want to be like this.” By the time it does generate the prompt, or you write your own prompt, you’ve taken into account all the requirements that are supposed to be there.

Katie Robbert 43:33

I’m looking forward to seeing the John Wall version with finger guns and eight microphones because part of me is curious. If you say “finger guns,” is that slang and would it actually give you back fingers that are guns? Would it know “finger guns?” I feel like there’s… there’s a lot more experimentation to be done. John, I can’t wait to see it.

John Wall 43:59

There’s plenty of places to go shooting at you.

Christopher Penn 44:06

That’s it for today’s show. Thanks everyone, 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/tipodcast and our weekly email newsletter at TrustInsights.ai/newsletter.

Need help with your marketing AI and analytics?

So What? Marketing Analytics and Insights Live

airs every Thursday at 1 pm EST.

In this episode you’ll learn:

Transcript:

Leave a Reply Cancel reply

Subscribe to our Weekly Newsletter

Pin It on Pinterest