We take it back... Maybe AI images are better than we thought?

We take it back... Maybe AI images are better than we thought? Speaker 1: Alexa, play That Amazon Ads Podcast. Unknown Speaker: Which one would you like to hear? Speaker 1: The best one. Unknown Speaker: Okay, now playing that Amazon ads podcast. These gentlemen are completely changing the game. Speaker 2: After listening to that Amazon ads podcast, my ads are finally profitable. Unknown Speaker: I also heard they're pretty cute. Speaker 1: All right, so we're back with episode 95. And as you may recall, if you're a good listener and a good faithful follower of this show, episode 94, we were bashing Amazon AI and ChatGPT AI for image generation, specifically for Amazon products, product images and lifestyle images. Now, these episodes are just one week apart, but it's important to note we are over a month ago. Did we record that last episode? I believe we recorded it towards somewhere around the middle of February and it is now the end of March. We took a long hiatus because we had a team trip for the AdLabs team in Berlin. And then after that, Andrew and I went to Madrid where we saw a Real Madrid soccer game. That was very awesome. And then after that, I spent two weeks in Morocco and Qatar. And now we're back. And recording the next episode, just boom, dealing with the jet lag and moving on so that we don't skip a beat and don't skip a week. But I'm pulling up the screenshot here so you can see exactly what we were looking at previously. And we're going to try this again because a lot has changed just in the last like five, six weeks. So this was the hammock image that, man, if there's a brand and they're just like watching this episode, like that's my product. They're probably very excited because we're giving you guys some good stuff. We don't know who this is. But this was the image that we just downloaded this image, we uploaded it to Amazon, and we gave it this prompt that said, I don't even, we don't have the prompt saved in the notes, but it was something along the lines of like, put this in the forest, you know, outdoorsy, you know, try to sell this product basically. And the Amazon images that we got were not great. So, Andrew this morning messaged me with some very big improvements that he was seeing in ChatGPT's image generation and we were just experimenting with it for the last hour before we hit record on this episode. Very impressed overall. However, We had to check Amazon's AI generation again to see how have things improved there. So we uploaded the exact same prompt, exact same image. We tried it again and here are the results. Boom. Speaker 2: No updates. Speaker 1: No updates. This was actually the, this is the prompt. So yeah, suspended by sturdy trees, fire pit, winding path to a mountain view. Did not follow that prompt at all. We still have this, you know, magical hammock suspended by nothing, just invisible beams. And we also have the hammock just sitting over a fire because who doesn't want to roast like a pig in their hammock with invisible beams when they're camping. And then this one's pretty funny. It's like someone's backyard. It's like a little koi pond. And then we still have the abstract images as well. Amazon continues to throw in these abstract images that I can't imagine a use case for someone to use abstract images in their lifestyle ads and why Amazon would give that. Yeah, you see the theme is abstract, theme abstract, theme backyard. Maybe we should have assigned a theme. Speaker 2: I find when you do that, it makes it a little weird, but yeah, yeah. Speaker 1: Nothing says exterior. There's not even an option here. So yeah, very, very limited here. So anyways, now we took two other products. So we're now going to look to the good examples, right? So we took this Bloom High Energy Pre-Workout Raspberry Lemonade and we took this little kid's bike. Andrew, are you thinking about buying this for Navy? Why are you looking at this? Speaker 2: I was thinking about it. Yeah, I just looked up toys for one-year-old girls and that was like one of the first thing that came up. Speaker 1: And she just turned one like two weeks ago, right? Speaker 2: Yeah, just last Sunday. Speaker 1: Happy birthday, Navy. So yeah, so Andrew took these images and uploaded them. And then he just sent them to me in Slack. So I'm just going to open up the Slack channel. And you can see Andrew, tell us what was the prompt that you gave for this image? Because this looks Phenomenal. Speaker 2: Yeah. So what I said was make a realistic image showing the attached product image sitting on a counter in a modern kitchen next to an opened container of raspberries and some lemons. Have it sitting next to a cute blender bottle that matches the color scheme. And that's pretty much all I said for that one. Speaker 1: This is, well, first of all, it's a great prompt. And it followed it remarkably well. And this was on what GBT model? What are the settings that people need to do to get this result? Speaker 2: Yeah, so first of all, you got to have a pro subscription. That's step one. Speaker 1: Number two, it's the best $20 a month of any software subscription you'll ever buy. Speaker 2: Don't even like second guess yourself, just buy it. Speaker 1: Maybe AdLabs might be one better software, but keep going. Speaker 2: Yeah, yeah, yeah. So I use the 4.5 model, which is the latest model that they've released. This is where I've found the best results for image generation. And then I think that's pretty much it. I just click the, there's like a little three buttons and you hit create an image and then you add the prompt and copy paste or drag and drop your image and let it go to work. Speaker 1: It's phenomenal. And one thing I observed is, let me come back to this thing. So notice how the angle on the can changed, right? This is a straight shot at it. You can't see any parts of the top of the lid. But in the image that we get here, you're seeing a little bit of the top of the lid because the camera is positioned differently, which means that they're not just taking this actual, what Amazon does is they just take the actual JPEG and slap it on an AI background. GPT is actually re-rendering it to the best of its ability to make it extremely realistic. And you can see the reason why this is so phenomenal is because the lighting's changing. If anyone's familiar with photography, like lighting and angles, all of that's critically important. So you can see how the light's coming from this left side and is casting a shadow on the right side. And that same lighting is going to affect the countertops and the lemons and this shaker bottle as well. That's something that Amazon doesn't Does not do very well where on this hammock, you know, like, well, I'm not really sure where the lighting is coming from on that. Where is the slide coming from? It's like kind of coming from like the bottom or head on. And then, you know, where they put their light sources all over the place. Sometimes it's behind it. Sometimes it's in front of it. So it's very confusing, right? So not the best, but this is great. And also this is, you know, when we did that one episode on optimizing main images, one of the critical things was including the, you know, if there's a product with ingredients, putting those ingredients in the image. And that's great. And I see that this brand already did that. They put the raspberry lemon in the product images. That's how it actually appears here too, right? So that's pretty good. In the past, we've commented on like actually put the berries and the lemons beside it just to really make those ingredients pop so people know what the flavors are just so that they don't click through and then later realize, oh, it has raspberries. I didn't want that. I only like lemonade and then click out. But if they would have seen the raspberries up front, In a more prominent manner, you can protect yourself from some wasted clicks. So this looks really good. And then we're gonna go to the next image here. Speaker 2: Oh, wait, one other thing to add here. That's super important. And what the biggest change that I've noticed in this image generator is, is that it gets the words right. Usually with these image generators, they will Have you know the words all jarbled and it looks like an alien language of some sort this took that image and Perfectly matched the very impressive wording on the actual container We've got high energy pre-workout raspberry lemonade, but it's not quite perfect So if you zoom in there, you can see that there's a little bit of that still showing up, but this is like you're a better commiserate Yeah, whatever that is. But this is getting you like 98% of the way there, right? Speaker 1: No one's going to notice that. Speaker 2: I mean, you could Photoshop and just cover that up real quick and you're golden. This gets you pretty much most of the way there, which it didn't do that before. Speaker 1: Yeah. So then let's go to the second image here. This is crazy, I can't believe that this is not a real person. This is just a phenomenal, like if you're just gonna throw this into your products things or on a lifestyle image, obviously you do wanna like throw in some prompts to make this landscape aspect if you're gonna be doing the, yeah, if you're gonna be using this in a lifestyle image. We're still noticing a couple of issues with this, like very small though. Like that blender bottle looks like it has two mouth caps. So it's a little bit weird, but you know, it's not too weird. That could have just been a weird design that someone had for this blender bottle. But it is also now saying proberry lemonade with two B's. And so some more things get messed up here, probably because the size of that can, maybe the smaller it gets, the harder it is to render the text properly. So very interesting, very, very strong overall for, you know, you can pay a ton of money for a photographer to get these types of images. And you'll, you know, maybe pay several hundred dollars for per product to get like, you know, 10, 15 images, or you can pay $20 a month for unlimited access to this for all of your products. And then just, you know, have an intern on your team, just go through and get all the images for you. Or you could do it yourself. So anything else to add to this one, Andrew? Should we switch to the bike? Speaker 2: No, let's check out that little, little girl's bike. Speaker 1: Yeah. What do you see here? Speaker 2: Yeah, so this one turned out pretty good. The biggest thing is some of the misspellings a little bit. So that one says Syreed. If you look at that, it's like Syresti. Other than that, I mean the classic AI generated image fingers on these people. You can see that the man has like a, you know, he looks like an alien. Speaker 1: Like he's like, yeah, he has, he has a thumb and then a nub. Speaker 2: Yes. So nothing wrong with that. Speaker 1: Women's hands are better, but they do look like they were crushed in a work accident and stitched together, um, by a low budget doctor. Speaker 2: Yeah, very thin, but to an untrained eye that might not get noticed. Speaker 1: She's missing a pinky. Speaker 2: Super prevalently. Yeah. Speaker 1: Her pinky got chopped off. Speaker 2: Otherwise, this is great. I basically prompted it, told it to have parents supporting their kid as she's riding her new bike in the driveway and stuff. Speaker 1: Look at this beautiful looking home. It just smells like success and it's like, if you buy this bike for your kid, you can be successful too. Speaker 2: Exactly. Speaker 1: And that's why Andrew's buying it. Speaker 2: A hundred percent. I already bought it. Speaker 1: Yeah. They do look a little bit cartoony, you know, so you would probably want to try a few more prompts just to like, please make it more realistic. But I think don't use the word hyper-realistic. I was trying to do that before to make it, I thought hyper-realistic meant just like very realistic. It actually means like above realistic. Which is not what you want, I think. Okay, now we also had, oh, I don't have the link to it. I could open the link to it. Let me just grab this here real quick. So one more tab for you. I'll just switch this one out to the, oh shoot. Oh yeah, here we go. So you have this EM's cat food. And what was the problem for this one? Speaker 2: IAMS? IAMS? Unknown Speaker: I would say IAMS. Speaker 2: Yeah, for this one, I basically said create a realistic image for the attached cat food product. I'm imagining an image of a The cat food bag sitting on the ground with some kibble spilled out onto the floor. The cat is eating the kibble off the floor, have a bowl sitting next to it. This image should be in the setting of a modern kitchen. And actually, that's slightly different. I variated it to actually improve this image because if you look at that image, you'll notice that there's a cat getting its head squashed by the bag. So in that one, I was trying to have the bag laying on the floor and like it opened and the cat kind of crawling into it. So I thought that would be kind of cute. So you'd have a cat crawling in trying to get more food out of there while another one's eating. But ChatGPT didn't quite understand what I was going for there and instead just crushed a kitten. Speaker 1: Yeah. I mean, if you just photoshopped out that cat, I mean, the rest of this is good. But yeah, it missed the mark on having that cute cat climbing into the bag thing. Like look at how they made the wrinkles on it though. That's very impressive. And again, like lighting source you can see is over on the left hand side. So all the shadows and everything are very consistent. And yeah, just how it wrinkles the bag a little bit to make it look like it's actually sitting there. Let's go back to this one. This image just looks fake, right? Like it seems like this was their, you know, what they, what the designers send to the manufacturers for printing as concept, you know. And the rest of these images honestly aren't that strong. It's just like, it's just like value props. Veterinarians recommend EMs, nutrition facts, this kind of stuff. You don't ever see a cat. I mean, aside from like the little cat on the bag, which actually I just noticed, They use the same cat breed, that's cool. But yeah, you don't see any cats in their whole thing. Okay, here's one, here's two, but there's no cat food there. It's the same grumpy looking cat. Speaker 2: It's probably the same picture, honestly. But the one we just generated is, in my opinion, gonna be better and could be utilized in a very similar fashion on this listing. Very quickly, within like a minute or two, upgrade that listing with some solid lifestyle imagery. One thing with these ChatGPT generated images lately, at least so far, That I've noticed is it takes a little bit longer. So if you're testing this out, it does take maybe three to five minutes to sometimes upload and render these images out. So it might take a little bit longer, but that's way faster than having a photographer go out and actually take really high quality lifestyle images like this. Speaker 1: And the text on this one is a little bit messed up. Indoor weight and hair and hairball care, chicken and furkey, her life, re-life, uh, indoor weight and hairball care, chicken and turkey recipe, but not bad. Like, I don't think either of us noticed that maybe because we were more caught off guard by the cat's head getting crushed by the back. Yeah. So then, okay. So this next image here, I was trying to get ChatGPT to render that hammock again. This was, I didn't upgrade the model. So this was kind of just, I was using the same models before and it came out the same as it was previously. And then kind of coming back to the hammock, I then switched the model. And I said, you know, create a realistic image of this exact hammock product being used in the beautiful outdoors. Use Yosemite as inspiration. There should be an attractive person somewhere in the image, smiling and having a grand old time. Landscape aspect ratio. So this looks pretty good, except for the fact that the guy does not have legs. But very, very realistic. Teeth a little messed up, but you know, that's okay. Speaker 2: He's having a good time. Speaker 1: Yeah, he's having a good time. We got the Yosemite. It's, it's actually suspended by trees, not nothing. Same thing like I, it's very impressive how it did the shadows like everything like that looks really good. So I did correct this prompt and I said make sure he has legs and make it sunset. And then and then it got a little weird. Not bad. I mean, just not what I had intended. But, you know, the stuff it's trial and error. Get it to work as best you can. But yeah, the you do have to upgrade that model now. Coming back to cat food, here's Andrew's updated image. This one has the Grigham turkey recipe and the number ingredient 15 chicken. You know, you got to do some work to it, but overall, like if this is a lifestyle image on a sponsored brand ad, the image is going to be significantly smaller, right? It's going to be about that size maybe. So no one's going to see that text anyways. Speaker 2: Yeah, you could easily just cover it up in Canva or clone stamp it in Photoshop and it would be totally fine. Speaker 1: This is a much better lifestyle image. It's like the cat seems happy with the food. It's cool that the cat is the same one on the bag. I don't know why. I'm not an artistic director or anything like this, but I would imagine having the same cat as the one that's on the bag is a good idea for whatever reason. There's a little bit of coherence there as opposed to having a black cat on the bag and a white cat eating the food. Maybe that's better for cat diversity if you're very pro that. I currently don't have any opinions on it. And then we had Andrew do one more shot where I was like, hey, try to get a better hammock thing from ChatGPT using the upgraded model because mine didn't work that great. And Andrew got this one, which is pretty good. The hammock is a bit too oversized. It's still a little close to the fire, but very nice. Speaker 2: Yeah, all the image generators struggle with hammocks for some reason. It must be kind of a weird layout or a weird setup for them, but pretty much everything else works great. Supplements, it's like spot on almost. Spikes, it's great. Hammocks are the true challenge. If you're in the hammock category, good luck. Speaker 1: That's the true Turing test these days for images is how good can you do with a hammock, with a hammock prompt. All right, in conclusion, We did do one final test with Grok AI because the AdLabs developers are constantly raving about it. Anytime ChatGPT doesn't work, they're always just like, try Grok, try Grok. And then Grok tends to do a pretty good job. So we were curious. I do not have a paid Grok subscription, but I was curious to see how it would do. So I gave it the image and the prompt, and then it just gave me a big scene description thing. And I was like, I think he didn't want to waste time generating an image. So it was like, would you like me to make this image based on the description or just anything? And honestly, I didn't read that because it was TLDR, but it made it. And not bad, but not great. Very AI-ish. Like this woman, her foot's kind of distorted. Dog's place. Speaker 2: We've got two hammocks kind of hanging out. Speaker 1: Two dogs. So I corrected it. I was like, get the hammock higher off the ground, suspended by trees, add a campfire, make a golden hour. Did not get the hammock higher off the ground. It did suspended by trees. Once again, this man's foot is burning over the fire. These AIs really like to burn people in hammocks. Speaker 2: This one broke. This guy fell through his hammock and is sitting on the ground now. Speaker 1: That guy looks just like a camper. Speaker 2: Looks like a normal human. Speaker 1: Very normal guy. Yeah, really looks the part. Now, this is free. Actually, Andrew, Can you just send me the cat photo thing? I'm going to try actually the blooms. And this will be the final thing we test before we wrap up here. So Grok AI, they have a paid version. I'm not on it. I'm on the free version. So this is great value for completely free. And image generation is not bad. What I do love about AI right now is all of the intense competition. Everyone's racing to do better than the last. So this stuff's going to load for a second. So we're just going to Fast forward. Speaker 2: All right. Yikes. Speaker 1: Not the best. I would argue much worse. Speaker 2: Significantly worse. Speaker 1: Probably better than Amazon. Probably better than Amazon. No, no, not better than Amazon because they're just, yeah, the text here is classic AI stuff. But again, we were having these text issues six weeks ago and they suddenly got better. Now, ChatGPT is relying on DALI, which really just specializes in image generation. I think Grok is making their own AI image generation. I'm not sure if they're partnering with another like AI image company. So obviously, people who specialize in things are going to do better at that one thing. So it's probably a better idea to have like multiple AI specializations brought together in one tool, as opposed to having one tool that's trying to be the master of all. Yeah, this one labeled the blender rather than the product. It did a great job with the raspberries and lemons, completely failed at recreating the product and the text. ChatGPT improved model is the way to go and my concluding statement is that it is slower now but slow is smooth, smooth is fast and I think it creates phenomenal images. That personally, I intend to use for the brands that I'm working with. I think this is great. I'm gonna explore more about this. And it's especially great for when there's like a holiday or a special event and you just want something a little more themed and your client's really slow to talk to their production team to get something up and running. And like, you just need this ad within the next week, then, or you need this image in the next week. You can just generate something, say, give it to the client, can I use this? They're probably gonna be excited and say yes. And yeah, so great for sponsored brand ads, even great for your own product listings. And I would say if you're not starting to use this now, you're going to start falling behind because this is probably going to give you better images than what you already have. And we all know the importance of images. People don't read the product titles or the descriptions anymore. They're expecting to get all the information from the images. So most important thing that you can do for listing these days. Speaker 2: Andrew, listen. AI image generation is very quickly improving the writing is on the wall like this has come so far so fast and I would predict that within the next six months this is going to be dang near perfect to where you could really leverage this to make a big impact on your overall. Business on your Amazon listings on all of your marketing materials is going to be very powerful, very effective. So start using it, start testing it out. You could take what we did today and run that through Photoshop, make a couple tweaks, add some additional wording and graphics, and you're going to have a much better outcome with the overall appeal of your listings and just the professionalism and the quality of the images that you're getting. So it's coming, it's happening fast. Photographers, beware, learn how to use this and better get started with ChatGPT it seems because it's definitely leaps and bounds above everything else that we've seen so far. Speaker 1: True that. So if you guys like this content, make sure you like and subscribe. And also there's ChatGPT Sora, which is the AI video generator. If you would like us to explore that, comment Sora. And if we get One comment saying Sora, we will make a video on that podcast episode on Sora. Thank you guys so much for listening and we'll see you next week on That Amazon Ads Podcast.