43: Way better AI photography for product marketers

This is going to be a different one. In the old days of product photography and marketing, this is how you did it. You had to take a picture of it with lights and you had to have the backdrop and all sorts of complexity and it took a lot of time, a lot of people working on it. Today we're launching a new product and I'm so happy to be able to show you a new model that does this better than any of the other AI picture to video models you've seen. Who are you?

Thanks, Robert. I'm Darius. I'm the founder of nex. We are building content automation tools for brands. We really saw that there were two key problems with folks trying to use existing AI tools to solve this problem. One of them has to do with product consistency and the second has to do with relying on models that were significantly less efficient to run.

Yeah. So what's the breakthrough here? Because there's so many tools hitting every day for creative people to create videos or pictures from photos. Right. You put a photo in it to like cling or something like that and it gives you a video. There's so many of these. What's your breakthrough today?

So we are releasing two things. First is a new image generation foundation model that we've built from the ground up. Something like 6x efficiency versus existing approaches. And the second is a product that allows brands to be able to use that model to create fine tunes with better text and product consistency versus existing approaches.

Yeah. This is a big deal, Right? Because if I'm a marketer or a, a brand manager and I take a picture of something like this, the other models often don't keep the text proper. Certainly when you ask the model to create a business card or something else that doesn't look like this. Right.

100%. And there are a large number of industries where this can be applied. Thinking toys, skincare, pretty much every retail or CPG product where there is a need for brand consistency. Product consistency, that's where we shine. Yeah.

This is not a general consumer tool. So if you're a consumer looking to play around on the weekends with some making some fun movie or something, there's other tools for you. This is aimed at brand managers who are working at brands and need to take really professional photos of their products or of their logos or of their brand works and use them in a number of different contexts. Right.

100%. Why don't I share my screen really quickly here and I can talk you through the key innovations that we made in order for us to make this happen. Would that be okay?

That sounds great.

I'm going to go ahead and share my screen here. Are you able to see my screen?

It's popping up. There it is.

Okay, great. There are two key technical innovations that we made. The first has to do with improving the convergence speed of continuous diffusion models. For reference, what you're seeing here is a spreadsheet of the effective flops that we applied to the training of our Icon 2 foundation model, which as it stands rests here at 3 to the power of 23 flops. This is about 6 times the compute that was applied to the stable diffusion 3.8 billion parameter model and yet cost us something like 6 to 10x less in terms of number of GPUs and in terms of cost. When we talk about the utilization of our approach, we can calculate this by using the OpenAI method, which is just six flops per parameter per token multiplied by three for the forward and the backward passes of the model. And we achieve a peak throughput of about 1050 teraflops for the middle resolution range for this model, which is something like 1050 over 1980, which is about 53% model flops utilization.

Yeah, I'm just going to. You're talking nerdy and all the brand creative people are like, what is he talking about? Can you convert that all to English? In other words, you're getting a lot sharper, a lot better image out of the model for a lot less money. Money. It's what I heard, how I tried.

So the translation is it's no longer cost prohibitive for brands to create product consistent photography with standard marketing budgets. Right. And really the core thing that we are trying to solve for is today the most expensive part of using AI tools is the human that has to sit in front of the computer to do the prompting. And we're trying to develop systems that enable us to get around that altogether. So the last nerdy thing that I wanted to highlight here is that this is the new repository that DeepSeq just released a couple days ago called Deepgem, which is their repository of high performance kernels for hopper level chips. And you can see on this column here, you see how it says 350 teraflops up to 1358 teraflops. Our calculated number here of 1050 rests very squarely in the range of some of the peak throughputs that the Deep SEQ team was able to achieve for their model convergence. And I think that's a testament to the technology that we've been able to build internally here at nex, which is comparable to the performance of top organizations in the world. And that performance is now accessible to brands and built primarily for brands. So there are two more things that I just wanted to get into. First is that when we benchmarked the Icon 2 image model against other models in the space, using the Deneval score, we find that we're able to achieve a performance of 0.84. This is higher than Deep Seq's Janus Pro Playground V3's Model Stable Diffusion 3 Flux Dall E. Pretty much every model that's currently ranked on this benchmark. You can see this is a greater breakdown which we'll share the full chart with you as well as the images we use to calculate this score. We're comparable in terms of most counting tasks for a single objects, two objects and excel in things like position and attribution. So from a benchmark standpoint, this model performs very well against other models in the space, despite its significant efficiency.

Let's see what this means in the real world for a brand creative person, right? Okay, so let's say I'm a brand person at a brand and I have a product. I have my hip water bottle here. You have a customer of yours, first of all. So let's go back up to 30,000ft. What do they need to do to use this model? Do they need one photo? Do they need 20? And do they need to light it properly? Tell me what this really means for a marketer who has to use photos to make Amazon page listings and custom brands for their websites and their videos and their super bowl commercials. But talk to them. What does this mean?

So working with NEXT and getting started with Next is really easy. All it takes is for you to send us a version of your product. What we will go and do is internally we will scan that product into our image generation model and allow you to render new views of that product in different backgrounds, aspect ratios and more. One of the top challenges that we heard from brand cmos is that creating brand consistent content at scale that's personalized to your users frequently is expensive, is challenging, takes a lot of labor, right? We sometimes hear that these brands are spending 1500 plus per photo shoot and that's one off. Right now you're stuck using those same photos for your products throughout the entire year. With our technology, you scan your product into our model once and you're able to use it to generate as much content as you want. The use of the model is unlimited and as an owner, you can now get Significantly more content for basically a one off cost of having to scan this product into our model.

So let's say I'm a brand person, like at Sephora or something like that, and they have a new lipstick line and they probably have 20 different lipsticks that they want to scan. They send them all to you, you scan them all and then what do they get back and what can they do with that lipstick? Then maybe give us an example of what kinds of creativity this unlocks for somebody who wants to put their lipstick in the hands of somebody riding a horse or. I don't know, let's get creative.

So what I can share is that we have been working with partners in skincare and in the beverage industry as well. And being able to get these multiple angles, get the text correct, even being able to do things like render text vertically, that's been something that's been pretty challenging. Another partner that we're working with is a bakery that sells these baked goods. Right. Which traditionally you would just use standard text image prompting to be able to coerce these models to be able to spit out. Whereas for us we're able to put these into different lighting conditions, different angles and more with ease.

And it looks probably a lot better than the kind of croissant that you're going to get out of stable diffusion or something like that, right?

Absolutely. Most importantly, it's not just about the looks, although that's a really important part of it. It's also that it'll match the product that you're trying to render.

Yeah. Why do we need to send the product to you to scan? Why couldn't I just take my iPhone and take some pictures and feed it into a model and do that? What are you doing on the scanning part? That really requires a little bit more diligence than letting somebody do it in their office or their warehouse.

So this is something that we pride ourselves in offering. It's a kind of service that today, effectively, there's very few if any companies that are offering this kind of effectively white glove service for brands. And it allows us to get the highest quality results. There are other companies in the space. There's a lot of folks that are trying to do this product photo for AI, AI product photo type of offering. But they're really often put into three categories. The first is where they don't even do any fine tuning at all and are simply doing background removal. And you can very clearly tell that this is the case because the lighting of the object does not match the lighting of the background. If you try to put that object resting on a cube or on a rock, it's not actually resting on it. You can very clearly see that it's hovering and that there are these artifacts. So that's the first one that's obviously very easy to achieve. You can do that very quickly, but the results are very bad.

Yeah.

The second approach is to is what a lot of companies offer in the space today, which is you can take a handful of photos from your phone or from your existing image library and try to use those images to create this model. But the problem is that for most non technical brand executives, brand CMOs, that image quality that you take from your iPhone is not great. It doesn't inject the right information into the model. It doesn't teach the model about things like angles or it fails for things like doing text. What we've seen is with these two lower cost but lower effectiveness approaches is that these are actually a net negative to use if you are a brand. And it's a net negative because your customers are going to see that you cut corners in doing your brand outreach work and that's going to reflect poorly on your product. So our perspective is that like brands need to get the highest quality imagery that they can. And we are trying to move that frontier of this is quality, this is cost. Today using a real studio might cost you $1,000 plus and the quality is obviously very high. And so how can we move that so that the quality stays the same? Maybe it's a little bit less thanks to the technology still developing, but the cost goes down and that's where we're trying to be. And I think we are one of the first companies to be in that space.

So I'm a photographer and I take a lot of pictures of things, things like this merge cube. You know, I have some flat lights or studio lights hitting the cube and it reflects differently as you move it around. Right. Now that's one thing if you're shooting it with a real camera, you know, my old camera. Right. But it's another thing if you're trying to put this cube into, you know, a scene or a story, or have a model holding it in different lighting conditions. Tell me a little bit about how your model looks at lighting and how it can work with different kinds of lighting that a creative director might be working with in 3D modeling to make a really badass image that fits their creative vision.

So what I'd say is that there are still limitations in our approach and we expect that as the models improve and that as our techniques improve, the quality and the diversity of image outputs are only going to get better. For example, today we sometimes still struggle with things like humans holding onto the objects. And that's just simply by the nature of these models. Where you.

Some of the models that used to put six fingers on the models. Right.

Six fingers, four fingers. That kind of stuff that can be solved by repetition. Right. Just keep trying it until you get something good. We want to avoid that in the future, but sometimes that's just the limitations of where we are today. But what I'd say is that overall, being able to achieve things like the text consistency, the lighting consistency with background objects, even being able to do things like being able to put props in your scene and interacting with your objects, these are all things that even if we were only able to achieve, these, would still be tremendously valuable to brand owners. Yeah.

Now it makes sense that you're shooting something still and solid, like a bottle or a lipstick thing or this merge cube. What about clothing or things that are more fluid or drapes or other kinds of things that have, you know, fabric in them? Even cars have a little bit of fabric here and there. Right.

So we haven't yet tried. So, okay, if you are a brand owner, you know, that has some kind of fabric product or a fashion product, we'd love to work with you and see, test the limitations of our technology effectively. My hypothesis is that we will be able to do very well with things like cloth materials, but might still struggle with things like human renderings, at least in the near term. Over the long term, we are looking to create effectively a database of virtual influencers where these influencers can earn money by licensing their likenesses to us for us to build these digital replicas and then incorporating your product into that model. So you can do things like have that virtual influencer wearing a particular product.

Yeah.

This is at the core of our approach, which is we see these diffusion models and these AI models as fundamentally tools in the photography process, and they enable us to overcome limitations of traditional approaches. I think about putting together a traditional photo shoot. You've got to sign the licensing agreements. You've got to fly that model out. You've got costume, makeup, lighting, sound. You've got all these different things and no sound for static photo shoots. Yeah. And you do that. You do that once, and that could be a $10,000 engagement, $50,000 engagement. We see that the future of this is using AI to help you create content that's more adaptive, content that's matching today's fast paced social media environment and all of that.

Yeah, it's one thing if I'm scanning a little cube or something like that, or a lipstick bottle. What about bigger things, Skis and bikes and maybe even cars. How big can you scan?

So what I say is that we're working with a partner now that is selling like a kind of sports product similar to like skis in size and we're seeing good results already. So I would say that we're pretty good with that. But any larger, again, haven't tried, we will likely have to develop some new technology to be able to do things like vehicles. But in those instances, these car makers, they're not mom and pop brands, right? They have capital to spend and making.

Sure they have 3D imagery of their cars. A lot of the advertising on the super bowl is done in computers. It's not even shot on scene anymore. Right. It's a crazy world for them. But I just wanted to get an idea of where is your limitation of size that you could scan. Could I send a bike to you and scan a bike? If I was a bike company, that.

Kind of thing, I'd say that right now we found a lot of success in working with small to medium sized objects, objects that are around the home, like I said, skincare bottles, drinks and beverages, toys, foods. That has been a great area of success for us.

Okay, and what should a brand manager, we've been talking about price the old way cost thousands of dollars to get a photographer in a studio and get a makeup person if you're going to shoot people and stuff like that. What should a brand be prepared to pay for AI photography campaigns when they want to move from traditional photography to using AI to do their work?

So this is where I'm going to get into very salesy mode here, which is to say that on the high end, right, you've got these photography studios which can be in the thousands of dollars for a single shoot. So each photo will cost you 50 to $100, something like that. And then you've got on the low end, you've got a hundred of these companies sprouting up where you snap some photos on your iPhone and you do background removal or all those things, those guys will charge you maybe 30 bucks a month, 30 to $60 a month, the standard SaaS subscriptions. What we are offering at Next is really a platform that allows you, the brand to handle all of your content. So not just the product photography itself, but the ideation that goes into what product photos should I create the Creation of the content. So not just the landscape images, but formatting them for different platforms, whether it's Instagram, your Shopify store, or your Amazon store, to the distribution of that content, Pinterest page, Pinterest, whatever it is, to the distribution of that content. So we want to automatically be able to distribute that content for you and then finally to optimization, and the next platform handles all of that for you as a brand.

Got it.

We are pricing our platform today at about $500 a month for our starter package. And for our first D brand partners, we're calling them founding partners, we are going to reduce that price significantly.

Yeah, that's good. Where do we go from here? If I'm a brand person working at Sephora, what am I thinking about? Okay, now we have our photo, our product scan. We paid our 500 bucks and we're all happy with the first images. What else can we do with these images? Because I know I just interviewed the found of Luma, who's taking pictures and making videos out of them. Is that on the roadmap? And yeah, let's talk about that before we get into cultural personalization, because where this is really heading is highly personalized imagery for each person who's looking at a Facebook page or something like that. So let's go to personalization. So we have pictures coming back. We. You give us a 3D model. If I'm at Sephora and I send you a lipstick thing and it's, do you give me a 3D model that I. How do I use that on my screen?

Right. We have a very intuitive way to use our product. I think this actually might be a good time for me to go through a little bit of what the product looks like.

Yeah.

So I'll go through this really quickly. There's three things that the next platform can do for you as a brand. First, we can help you organize your existing content. Second, we can help you create new content using our AI models and tools. And third, we can help you edit content that either the AI has created or that you might already have. And so we see a future where, like, real content, whether that's UGC content that you've recorded yourself and AI generated content all work together. So our primary product is this AI agent tool called Marco. And this is where you're going to be spending most of your time. You should really treat Marco as an Internet right on your marketing team, and you can just ask it to perform tasks for you with context. We also have a creative platform, and these are.

Can you give Me some idea of. Can I say, can I put this on inside, you know, a forest? Can I put it on a beach? Right.

Do you just stuff like that or let. Let's take a look. So this is what a Marco session looks like. And you can see very standard chat interface. And the power of Marco comes when you add context. So in context you can add four types of context. You can add a product which is a custom product model, brand guidelines which help guide the model towards the guideline that you specify audiences. So things like a skincare enthusiast or a coffee lover, that's Gen Z or others, and you can really specify that and help the model better target your products. So what I'm going to go here is I'm going to select this pineapple bun, for example. I'm going to select an audience. So let's say like coffee consumers here and click done. And then you said you wanted to see this on like the beach. Sure.

I'm just thinking creatively, I might have an ad campaign about summer. Right. And I want beach scenes or something like that.

Right.

So fun scenes on the dock in the bay or something.

I can literally just submit this. And what's happening in the background is that our model is looking through this audience guideline and saying, what are the prompts that best fit this kind of audience? It's indexing into our custom model and saying, make sure to use this model in your outputs. And it'll also go and create as many assets as you need in order for you to achieve your task. In this case, we're asking it to generate a single image. But if you ask it to do more complex things, help me create a product video that might include generating three different images. This is pretty funny here. This is your pineapple bun on the beach. You can see that it's able. This is this explanation here. Sunlit beach photo with design elements and with brand alignment here. It's able to capture that. What was I trying to say? Oh, yes, there's more. So Marco is not just able to be used to create new content. So this is AI generated content. You can also use it to edit existing content as well. So as an example, I have a video here. Down here you can see this MP4 that I can simply add as context. By the way, this is connected to the next Digital Asset management system as a brand. Another big problem that we've heard is organizing all of your assets across your entire team, from videos to images. It's complicated, it's expensive, and we offer that as well for you Just out of the box. We offer brands like 2 terabytes worth of storage, so you'll probably not run out for a long time. But this is a real video.

Yeah, the scenic brought me back. Didn't fit my creative vision. Is there a way to maybe prompt it to give you 30 different versions so that maybe I go out to lunch and I come back and there's 30 different ones for me to pick through? And is your model learning by that? Right. And getting better at understanding me as a user or whatever?

Absolutely. Both of those things. The answer is yes. So in this case, I can simply ask it to give you another variant or 30 variants if you so wish, and it'll just go and it'll just keep working. And this is what I mean when I say that you should treat Marco as an intern, where Marco is just going to go and do the work for you. It'll also learn. So we capture information from the user in the form of knowledge and we store that for future use. And that's just added to the context of the model.

So if I was a brand manager and I didn't want my baked good shoving up on a beach, I could maybe teach it. Hey, don't ever show it on outdoor scenes. Only show it on indoor scenes or something like that and enforce that at the model. Is that where we're going?

So that is something that would be best handled under the brand guidelines.

Got it.

Where you can add that directly and you can always use that.

When I worked with Eastman Kodak back in the day, a long time ago now, they had a big brand document that, that showed how. What, how the font had to be, what color the font had to be, how big the font the logo had to be in relationship to anything else. Right. There's all these brand guideline that these big companies develop over years to make sure that their brand is always portrayed in the way that they wanted to. Right.

100%. And that's something that we've seen not just in big organizations, but even all the way down to smaller teams too. If you're talking even a team of 10 folks working on a brand, they're very specific about how they want the results to look. Yeah. So the. The last thing that's here is.

Yeah.

And we, you know, we. What am I trying to say now?

The last thing.

Oh, yeah, yeah. The very last thing is that again, the system is not just able to create new content using AI, it's also able to edit existing content as well. So in this case, I've asked it to do something like remove ums from a video, and it'll go in and actually go transcribe the video, select those words and crop it out for you automatically. Okay, so this is an example of a flow, right? A workflow where you could imagine you have some video, you want to perform some edits onto it. Traditionally, you'd have to go to 15 other tools, whether that's Cap Cut, whether that's Canva video or wherever else.

I'm using descript, which does a similar thing, right?

100%. And there's so many of these tools. And we are hoping also that Marco will be able to integrate all of those tasks into one framework. So you can just ask the model to do something for you on your video, whether it's word removal, adding music, stitching things together, including real and AI images, and it'll just go and do that for you. So, really, an omnimodal marketing agent for brands.

That sounds real interesting. It starts to get real interesting when AI is programmatically doing that on your behalf. Hey, we have a Facebook ad. We need an image on the Facebook ad. The user's in India, so it needs to be culturally relevant to India, something like that. And then it comes to your model and starts asking for an image like that. When are we going to start seeing your kind of tooling hooked up that way? So AIs can really make personalized marketing without very many humans involved at all, very soon.

In fact, for us, you might see it in the next couple of weeks. So we're working very hard to be able to integrate exactly that kind of functionality in. We think that humans always need to be in the loop to pull the trigger, so to speak. So there will always be a human performing the approval. But being able to do things like automatically analyze TikTok for recent trends, automatically analyze Instagram for what is the hottest product category that people are talking about today. Being able to look at your existing video database and pull from that, or create new content, create new videos, new images. All of that is going to eventually be done by Marco. And we think that the hardest part has so far been at the model level, where it's just been very challenging and difficult to be able to pull off the construction and development of these kinds of algorithms. But as we move from the model level into the application layer, that unifying these together is going to lead to some really powerful results.

That's cool. These brands already have a lot of 2D photos and videos that they've shot. My friend Scott Jordan, who makes Clothes. He has giga vibes and gigabytes of videos and photos he's made of his different clothing. That's why I was asking about clothing for him. But let's say I'm at Sephora. I know Sephora has huge databases of all their product photography and videos that they've done in all their commercials and all their brochures and all their marketing vehicles that they've had to generate. It sounds like you're both doing a new 3D scan with their latest photos or their latest product, but you're also able to use that older work in some way together.

Absolutely. That this is actually a really key element of my philosophy for how AI is going to support brands. I seriously doubt that we will see a world where it's all AI generated. Right. As a consumer of some of these products, I would certainly not want to see that world. Right. And I think that it's because some of these AI images sometimes lack authenticity. So really the future is going to be, you are a brand owner and what you want to do is tell a story to your consumers. And how you tell that story is going to change. It's going to incorporate images that you might have already taken, images that you are going to take for new product releases or with new brand partnerships or whatever it is. And then where there are gaps, where your, where the real world effectively can't catch up with the pace that the consumers want to consume content, that's where the AI tools can really be useful and, and can deliver a ton of value. Right. Where in today's world, on platforms like TikTok, you almost need something new, like every day. And most brands cannot keep up with that. They cannot afford that. And we think that with our technology, you'll see more and more brands being able to engage in better ways with their customers.

Yeah, we're also moving into a 3D world. You know, it's hard to imagine that.

Hold on a second.

But with things like Apple Vision Pro and other, you know, Quest three, it's starting to happen. In a few years, we're going to have a pair of glasses on that's going to let us see a holodeck, a 3D environment around us. And that holodeck is probably going to be scanned. If I'm going to put a product into a 3D environment like that, you're going to need a really high quality 3D scan so that a consumer can pick up that product, look at it and manipulate it in 3D space. And it has to be properly lit in that space. It sounds like your scans are going to be really perfect for that coming future. Right.

So what I'd say is that our Scans are not 3D. And in fact we've taken like a musk approach to this. You know, where the, the thing that he had is like no lidar pure cameras. Right. And that was a revolutionary idea to not use a 3D scanner and to estimate depth using static cameras. So our approach is actually to use just normal cameras, but with extra data in there and with a special camera technique that allows us to very accurately capture detail, orientation, text and more of the product. And that's really where we are headed. So what I'd say is, yes, can our technology be used in a 3D world? 100%. But we avoid the challenges of dealing with 3D. 3D is just really annoying. Hard for models to operate on, hard for AI. So we prefer to stick in the 2D world and inject as much physical information as we can into it.

The government showed me a long time ago that you can take a 2D photo and extrude it into 3D.

Right. So.

So that's a really interesting approach. And that sounds like something a South by Southwest panel. You know, hey, 2D can do 3D and you probably want to stay with 2D because it's easier. Like the Tesla car, it's. Oh, that's a big idea. What else would you tell brand creators about you or what's going on in the industry right now? Because it seems like so much change is hitting. Just last week we got three new voice models, we got new music models, we got OpenAI shift, a new model, Anthropic cloud shift, the new model. There's so much change hitting everybody right now. So what's going on in your head from that? And maybe talk to the brand people about your position. What's happening?

I agree with you entirely that there is a ton of activity happening in the AI space right now and it's often very difficult to catch up. As a brand owner, you really want to work with the companies that care about your problems and your use cases. That's what next is where, yes, we are developing some really incredible technology to be able to accelerate AI training and diffusion training, but so that we can help you, the brand owners, get higher quality product photography. There's all this talk about AI for video editing, AI agents, which is really confusing. AI agents for. What is it, customer service and all that. Yeah, we are building that technology for you. And if you are an owner of a brand that sells Physical products, small to medium sized physical products. We'd love to work with you. We'd love to give you a demonstration of our technology, a trial of our product and see how we can hopefully not just reduce costs for you on the marketing side, but also help you get a much higher return on your marketing dollars.

Yeah.

And that's what we're really excited for.

Thank you so much for spending a little bit of time with me today. I know it's south by Southwest which is going on like tomorrow or today or yesterday depending on when people see this video. But it's. It happened in early March, in 2025. I went 25 times to south by and and ran events for brand people. They're going to be very fascinated about you, I have a feeling and want to use you because it's. It reduced like you said, this technology costs but also giving much more creative freedom so that somebody who has a creative vision of how their product would be portrayed on TikTok or on Instagram or wherever has a freedom to play. This is really a fun thing to do. It's not like you have to hire a photographer and you get the photos back in five days. No, you can play and just sit at the keyboard and go, what if I put my bottle on this? What if I put my product on this?

The last thing I say is that ironically there are other services that where you would do exactly that. Right. Where you literally sit down on as a brand owner on a computer screen and you can ask the studio photographer to take pictures in different angles and conditions. We are building a very similar user experience but instead of now having to talk to a real physical photographer, it's all happening in the AI verse. It's all happening in digital space. Space. And we think that's going to be the future. Cool.

Thank you so much. Where do we find you? I. I assume you have a website.

We have a website. You can reach me directly at Darius Next Art or you can go to our website and sign up for a demo there.

Thank you so much.

Thanks Robert. Thanks for the opportunity.