Using DALL-E 3 For Inspiration And Production by Beatrix Meszaros


So hello everyone, I'm Beatrix and I've worked in digital advertising and marketing for 15 years already and now I am basically training marketers how to integrate AI tools into their everyday work.

Demonstration of AI in Advertising

And I've prepared with a live creative demo for you today. So look at the screen.

Iconic Advertisements

Here on the screen you can see previous iconic McDonald's Halloween ads. Have you seen any of these? Maybe they are just circulating on the internet.

Just think about them, look at them. What do you think?

How long and how many brilliant minds do you think it takes to bring these ideas to life? We can agree that it takes months. Do you agree with that? Definitely.

And it requires a massive group of creative people for the brand.

Experimenting with AI for Campaign Brainstorming

Around Halloween, I just played with the idea of being responsible for this year's Halloween ad for McDonald's. And my aim was to test Chad GPT's and Dolly's multimodal capabilities. and to see whether I could simplify the entire brainstorming process for the campaign.

So let's move on to ChatGPT Plus and let me show you what I've come up with. While I switch my screen, I'm just really curious to know how many of you have a ChatGPT Plus subscription? So it's the paid one. Could you raise your hand?

Okay. And how many of you are on the free one? Yeah, so yeah, I think the good majority already had. Okay, cool.

Setting Up the Creative Prompt

Okay, so first, I just copied my first prompt here just to save some time. And I have come up with this first prompt. So the prompt is like, act as an advertising creative director with 20 years of experience in the advertising industry. Then I provided context to it.

What I have done, I have downloaded these seven ads simply from the internet. Let me attach these to the chat.

Initiating the AI-Assisted Creative Process

Let me start the generation and I will explain in the meantime while it's generating. I think many of you are already familiar with this technique when you are assigning a specific role to ChatGPT. You are responsible for this year's creative Halloween ad for McDonald's. All Halloween ads must follow a consistent style and visuality.

I attached seven real previous Halloween ads from McDonald's for you to analyze. I want you to do a comprehensive analysis on the images. focusing on image style, color schemes, design, and theme patterns. Based on your analysis, I would like you to come up with five new creative ideas for this year's Halloween at Visuals.

After that, I want you to write dolly prompts, As the previous speaker said, ChatGPT is much better in writing prompts for itself than us. And finally, I would like you to generate key visuals for each concept. Think step by step.

So I have given the entire process to it just to see what I want to achieve with this task. But I prompted it to do it step by step because everything, the outcome is always much better if you do this chain of thoughts type of processes step by step. So let's see what happened.

Analyzing Past Campaign Visuals

So here is the analysis. Can you see this properly or do you want me to zoom in a little bit more? Let's see. This might be better for you, I guess.

OK, so first we are testing its visual and image analysis capabilities. And as you can see, it captures that each ad features a single central item, which is a very important thing if you just look at it. they always had one product in the center of each ad. So one single item.

Then we can see that the items are presented with a minimalist background, often dark or a solid color to highlight the product and its transformation. Absolutely correct. Color scheme. Predominant colors are black, red, and yellow, which align with the McDonald's brand color palette.

Absolutely correct and very important observation. The use of red and yellow is often minimal. Accent colors such as white and variations of orange are used to evoke the Halloween theme. Very nice, I think, design and theme patterns.

And you see, it has structured the analysis already. The theme is consistently Halloween with playful and not too scary twists. which is the actual creative idea there. So it has noticed that there is a creative twist on each ad.

So it's very, very important. The products are depicted in a way that hints at traditional Halloween motifs. Again, super important observation.

Text is witty. Okay, typography, it's usually simple, not much typography we have on the ads, but it has noticed that as well.

And yeah, here is the summary of the analysis. Keeping the central focus on a single McDonald's product, applying a Halloween twist, utilizing a minimalist and dark background, incorporating the McDonald's color palette subtly, employing witty and concise text that plays on Halloween themes, maintaining brand consistency. Do we agree with this analysis? Can we accept that?

Yeah, I agree. I would maybe delete one point from the analysis because it says that employing witty and concise text that plays on Halloween, actually, there aren't much, if we go back to this, there aren't much text on the original ones, but let's accept this. Okay, so now let's say

Generating New Conceptual Ideas

Okay, now generate five extremely creative new conceptual ideas as the next state. And let's see what it comes up with. And normally, I would recommend doing more than five. So I managed to get the best results when I was testing over 20 concepts.

I slightly simplified the process for this demo. So let's see what we have.

First one, we have a McFlurry poltergeist, which is a concept McFlurry cup with the ice cream swirling out in the shape of a ghost with bits of cookie and candy mix-ins forming the eyes and mouth of the ghostly figure. We have a tagline as well, spiritedly delicious. Yeah, that's the thing. Now it's writing a tagline for each of them, which we don't necessarily need.

The original didn't have any. Fright Night. A carton of French fries arranged to look like a creepy hand reaching out of the carton with ketchup at the tips to resemble painted nails or blood.

Beavis burger, a classic McDonald's burger with the cheese melting over the side to look like a witch's head, complete with a burger patty broomstick on the side. Okay, let's say that we don't accept this concept, and before moving on to the next step to write actual daily prompts for this, let's ask it to revise all your creative ideas against your initial analysis and observations and improve them accordingly.

remove text from each concept. We don't need that. And make sure that the background is always a plain dark color.

This is actually a mistake that I have noticed it is making. with this project, because again, if we look at the original ads, the background is always plain dark. And for some reason, it is always suggesting stuff on the background.

But it is a good idea to prompt it to do a revision. OK, now we have improved concepts. Let's say we will accept these.

So we have the McFlurry Phantom Whirl, whatever that is. Okay, Midnight Fry Spell, Enchanted Burger, Nugget apparition, what? Okay, apparitions, okay.

Creating Prompts for AI-Generated Imagery

Okay, now. move on to the next step and write a DALL-E prompt for each. And yeah, as I said, typically it comes back with much better.

Now it's creating an image. Yeah, the beauty of a live demo. But I'm sure, have you experienced troubles with ChatGPT over the weekend?

It was crazy. But I do have a backup. You can see my taps.

So I will show you. Oh, probably this is one of the worst generations I have ever gotten from this. OK, let's stop.

this process, yeah, I just stopped the generation. Write the dally prompts first before generating any image. And I really hope it gets it now.

Oh yeah, this is awful. And one interesting thing, I don't know whether you have, is there anyone who tried to generate any branded content with any AI image generator? Yeah, I see a few heads nodding.

What was your experience, bad? I think only Dali is capable of doing fairly good generations. I will show you really good ones, actually.

It is capable, but that's the only one. I've also tested it against Midjourney, which I will also show you later. Theoretically, Midjourney is also capable of doing branded content, but they are not that good, I think.

Or maybe I just didn't get the prompt right. So now we are getting our prompts. And again, you see, beauty of the live demo, it stopped generating the prompts after two, and now it's creating an image.

Let's see what it is generating for us. Yeah, even worse. Okay, okay.

But yeah, that's the thing, you need to test a lot, but I will show you, you can get it right. So yeah, it generated, what happened? We asked it to write all the DALI prompts before generating any image.

It generated two prompts, one image. Now it is doing prompts again. Yeah, I'm not sure that it is stable now.

Sorry, I'm not sure I could hear you. Yes. I mean, are you asking whether it is possible to put text on the images?

Yeah. Ah, yeah. So it isn't putting text on the images if you are not prompting it to put text on the image in the first place.

But no, what's going on? And it crashed. Okay.

Troubleshooting and Demonstrating Backup Results

Let me move you to another chat, which I have here as a backup to show you what it is capable of doing when it is actually working. Yeah, I was expecting this after the weekend crashes. So let me show you what I have.

I will do just a walkthrough of a previous chat, okay? I think we will be more efficient like this.

So again,

Showcasing the AI's Capabilities

The very beginning was the same, uploaded the seven creatives, used the same initial prompt that we discussed. I got the analysis of the images, which was again mostly correct. Then I prompted it to move to the next step to generate five new ideas based on the observations.

Here are one, two, three, four, and one idea, but yeah, it is what it is. It generated all the prompts. I'm not reading everything out loud, if you don't mind.

I think if we get to the visuals, then just everything will fall into place. So these were the first five ideas. Again, I prompted it to review all the ideas against the initial observations.

And I prompted it that I feel like they are less minimalistic than the original ads. Also, the original ads always have plain, dark backgrounds. I will give you $100 if you come up with better ideas. Are you bribing it?

If not yet, you need to, yeah? Let's start bribing it. Okay, now it has refined all the ideas, but again, we have one, two, three, four, and won again.

Anyway, new prompts we have, and here I positively affirmed it. These sound much better. Positive affirmations work much better with chat GPT as well, not just with humans. In general, I'm not sure that you are using this trick as well, but I highly recommend doing that as well.

So I said, now generate an image for the first concept. Again, we are moving step by step. It always works much better if you work step by step, piece by piece. And again, this was the first generation, and I wasn't impressed at all.

I was like, OK, it should be photorealistic as the original ads. Refine the prompt based on this feedback and generate a new image. All future images have to be photorealistic. And it came back with this one.

Now we have actually at least a nice branding. So even though I think we can all agree that we don't like this image, but It kind of aligns with the originals. But let's move on to the next one.

Now generate an image for the second concept based on the feedback. And it came back with this one. This one is pretty cool, I think, as a conceptual idea. Obviously, none of these would make as a finished concept. But as a conceptual idea, this one is fairly good. What do you think?

This is decent. I started getting happy when I saw this. Then, third concept. I was again like, okay, let's see what else we have here.

Now generate the fourth. And came back with this. What do you think about this? Yeah, pretty good, I think. Again, aligns with the original.

I think it can be a decent idea, conceptual idea. And let's see the fifth one. And again, not that good. So I tried bribing it again.

I was like, I will tip you 100 if you make the last one more creative. Let's see what happens. Not so good. So then I've decided to move on with the second one, you know, with the fries.

You agree? This is nice. This is decent. And the fourth one with the Happy Meal box concept.

And yeah, and I just told it, I asked the seed number. Do you know this seed number trick? in ChatGPT, the seed number is the unique identification number of each image, and you can ask ChatGPT what the seed number of the image is, and you can use that as a reference for future generations, and you can modify the original image that way. So first you ask for the seed number, it gave me the seed number,

the seed number, and then I said add text Happy Halloween to the seed number, which was the Happy Meal box. And it came back with this again. That is the box, but where is the text? I'm like, hmm, again.

Okay, I said there is no text on the image, regenerate it. put the text on the image. Finally. Okay.

And I was just further playing with it. But the rest of the generations were not that good. Yeah, I said that I love the second and the fourth concept, review them against your analysis and make suggestions for further improvements.

I was curious to see whether it can further improve them. It made further suggestions. Theoretically, they sounded correct, but the result got worse. So I think the initial ones were much better. And just to wrap this case study up, because I'm afraid I'm running late, let me show you here.

I've just arranged in a presentation the best results that I managed, literally managed to get out of this process. So these were the best I managed to get.

Conclusion and Final Thoughts

And I truly believe, having a really solid experience in advertising, some of them are decent conceptual ideas. And I managed to get the best results while I was spending, let's say, an hour testing over 20 different concepts.

And I have tested mid-journey as well. I have literally asked ChatGPT to write prompts for mid-journey. I had to tweak the prompts a little bit because the mid-journey prompt structure is different from the DALI.

These were the best results from mid-journey, but I believe that the DALI results are much better than those. Maybe I just wasn't getting the prompts right. Maybe more tweaking with mid-journey could get me there.

I pretty like the coffee cup idea. That's a nice one, I think.

Anyway, that was it for today.

Connecting with the Presenter

And yeah, if you want to keep in touch, you can scan this. You can find me on LinkedIn. Sharing lots of AI content, mostly very similar tutorials and very real life use cases and case studies on my LinkedIn around marketing, because that's my background.

We have told I don't have a tech background. I'm just really trying to help marketers to be more productive in their work with AI tools.

So thank you for your attention.

Done? Mark to tidy your library & unlock trivia