Scaling AI Photo Editing to 300M Users with Photoroom’s Matt Rouif | AI Basics with Google Cloud

Foreign.

Everybody. Welcome back to this week in startups. We're continuing our basics series. Why do we do these? Well, so many founders email me DM me all the time with basic questions. So when we see a question come up over and over and over again, we find experts, we bring them together and we do these very short, tight, concise, basic videos and they explain to you everything you need to know about important topics like legal or marketing. And we decided to add AI to the mix. You can see all the basics@thisweekinstartups.com basics. We do this to support founders, just plain and simple. We know they need this information and we are delighted that our friends at Google Cloud have sponsored this season. You can check out their report the future of AI perspectives for startups. It's got insights from 23 different top AI experts. I think a couple of my besties are in there. Like I think Chamath and Freeberg are in there today. Very lucky to have Matt Ruif here. He is the co founder and CEO of PhotoRoom. It's a power space AI photo editing app with 300 million downloads and they process, wait for it, 5 billion images a year, reportedly. Matt, you got 50 million in revenue already and it's what year? When did you start this company? And welcome to the program.

Hi Jason. Thanks. Yeah, we started in 2019 just before COVID actually. So it's been like a six year journey.

Amazing. So you can make images on every language model today. You can jump in and just describe what you want and get something that's, let's call it funny that you could tweet, you know, with Grok images or with Gemini or these other services. You know, you can make some interesting photos for a birthday card or maybe for a presentation. But this is what I find. I don't get exactly what I need. I typically get 60 to 80% of what I need. And so I'm curious twofold, when do I get to 100% of what I need? And why does the world need a standalone company like yours? And I think these two questions are related.

Yeah, it's a good question. Well, I think if you want control. So we started as 2019. It was not the world of Genai yet. And what we brought to the table is getting control. Exactly what you said. And so we started actually for Genai. So you remove the background of product photos, your product photos you can sell on ebay, Shopify, everywhere. And this is actually the first step, like you get 100% control of cutting out the background and so this is what we bring to the table. A lot of the generative AI space is about creativity. Create these dreamy photos and we really focus to help create photo that sell. So we bring trust, trust with your product. We don't change anything in the product you're taking into photo. And we also bring trust to the brand so you can put your exact color, your exact logos and have the power of Genai. So that's why people love photoroom so much. Like you get the benefit of Genai, but also the control of the traditional AI image editor.

And this is super important for professionals. You know, if we're screwing around and we want to, you know, send stuff to our group chat and make people laugh, it doesn't need to be perfect. But if you're working for a very, you know, if you're working for Coca Cola, like this logo, whether it's in Mexico City or it's in the Philippines, it's got to be perfect everywhere around the world. So I suspect you probably have global companies that have global design standards and then local design standards under those.

Yes, yeah, exactly. So we started from a mobile app, but we expanded and we serve some of the biggest e commerce in the world. Doordash is using us, Amazon is using us. Because the control we give, like you really have exactly your logo, exactly your colors. It's not only the logo of Coca Cola, it's exactly, you need the red of Coca Cola. You need all these controls to appear for your brand. So it's really this approach.

So yeah, and these businesses, even Doordash is becoming international. People don't know that, but they're, they're going global now and they're buying some companies everywhere. And this is super important. If you're Uber eats or your DoorDash, you want to have some common experiences and you want to enable your partners to make consistent images that don't break the design and build trust. And when you have broken design, it lowers trust in products. It's a super important thing for founders to understand. So maybe you could show us the product and then talk to us about, you know, how you go to market with a product like this and find your customers. So sort of two questions, maybe just a demo. And then when you're ready, explain to us how you got those first couple of dozen customers and then how you.

Scaled it to add actually on top of the doordash analogy, what's key for them? And there were a few scandals, like a problem with that is that you don't like, let's Take a photo of a salad. You don't want peanuts to be added to your salad. With generative AI hallucinations, you don't want to have a lawsuit because of that. And that's why our approach is quite powerful for them and very important. So yeah, here's a little demo in the Photroom website. We also have a photoroom. Well, we started with the app too. So we really have everything you want here. And I drag and drop a pair of red sneakers in PhotoRome. So as you see, we're the best at removing background from that you can create. Well, we would suggest like to make it accessible. What we really do is make it easy to create anything from that. We want to build trust because our audience is the non designers. If you're selling some sneakers, if you're a restaurant, you don't have all the skills. You don't have 10,000 hours on Photoshop. So it needs to be easy. And so I can create, I can use this AI background that is totally generated. And here you can see my sneakers. They're all the initial pixels but at the background it's all AI generated. But it really sheds light on your beautiful product. And if you look back you can even add a virtual try on. So here I can use my pair of sneakers and put them together with a model that I can choose.

Wow.

Can remove the background fidelity on this by the way.

For people who are watching, the fidelity on this is just extraordinary. These Nike high tops, I guess they're Air Jordans. I'm taking a guess, not sure which model. But you know, when you attach them to a model here or put them on a background, they look indistinguishable from an actual photo shoot.

Exactly. And for a small business a photo shoot is like 10k. Like a cheap photo shoot is like 5, 10k a day. You need at least 3, 4 people. I was able to create a prompt and something that looks good for me in a matter of seconds. These drawings look amazing now with my model and I can reuse the same model for multiple products I'm selling. One of the features that works super well with our like we have bigger brands is the idea of creating, you know, a batch so you have consistency of your brand with multiple photos. So if I take a batch of like, we're talking about doordash. If I'm a restaurant, I take a batch of dishes with like different, you know, different aspect ratio, different positioning and everything. I can center everything and make it like, like super professional in a minute, you know, when you are a chef, you don't want to spend your minutes like, you know, taking photos. You rather be cooking.

It's a seemingly trivial example, but we have all had this issue where you want to normalize a group of photos. The photographers, you know, might have taken them from different angles, etc. And you want all the salads to have, you know, the same size and the same distance around them, the same cropping, etc. And wow, that's just done perfectly instantly. And if you think if you're scrolling around a modern app like a Doordash or an Uber Eats you, you wonder how they make it look so good. It's these kind of tools. So is this tool going to be built into the. And are you going to like sort of API it into the DoorDash app? Or does DoorDash just do this themselves for the restaurants? How. How do you see that going down?

Yeah, you know, it's. So we partner with some of the biggest marketplaces and it depends on the approach. Some will put that in their backend and they already have a photo workflow that is working. Some like Depop, so it's part of Etsy. They put that in the user interface. If you want to resell and sell some of fashion items, you're just flipping, then you just go in the app and you can find some of the photo technology in there. So it depends on the approach and how much they want to have control still on the images, but both flows are possible and for the small businesses, they use the app.

All right, so this tool can be used by a large number of different people, but you have a business to run. What was the go to market strategy here? I'm curious how, how you got started. And then getting to 50 million in annualized revenue is a big number. It's not easy getting to a couple of million in the AI space. I don't want to say it's easy, but it's obtainable. But the jump between, in my experience, people can get to 1 to 5 million who can then get to 20 or 50. These are different type of jumps. So walk us through as a founder, how you got there.

Yeah, so, I mean it was quite fast, but it wasn't like instant. So we started mid 2019 and really like what we did a lot in the early days, and I think that was quite important is we would, well, build and talk to users. So we had an office, we would build the day and then go down to the McDonald's that was downstairs and talk to our users. We would go in the line, ask people if they were okay to like let us pay for their. Their meal and in exchange for answering and testing this little app and telling us how they use photo apps and.

Now that for weeks. Yeah, well, this is, this technique is the listening lab technique I think some people call it. Is that where you found out about it?

How did you become aware of was from Zenia. Like the co founders that we knew that told us but they probably learned and there is like a mom test book that is amazing for like how to ask questions actually for listeners.

Yeah, it's super important that you don't lead the witness as it. As it were. You have to ask a generic question like, hey, your friend. I think the best one I saw was from Creative Good, an agency in New York where they just said, hey, your friend sent you this app and they said try it out, it's cool. What do you do next? And then you watch them, they say, okay, what do you want me to do? And they're like, you're supposed to just not speak and say your friend sent you the app.

Seems easy.

It's difficult, it's difficult to not say. And then they do like two steps and you're like, like I don't know what to do next. And it says submit where it says process image. And you're just like, want to just either absolutely punch this person in the face or you want to punch your designer in the face because it's so frustrating. And you're like, generate result. What else can we put on this button?

No, I mean it's exactly that. Like you go, you build this amazing feature and user would never get to the feature that is down the floor. They would always like stop before and so you always discover something. But it is so frustrating and you really want to grab the phone from the hands and like push this button, please. You'll get there. And that's absolutely not what you want to do. It's very. You're ashamed of your product if you're doing that. Well, like really, I was ashamed of the result with my co founder earlier. Like we were. But that's the only way and that's. That's how we got the best feedback. Actually we did it so much that I was banned from the this McDonald's. The security guy like didn't want me, didn't want to see me anymore.

So yeah, that would be in France. Like you accosting the people and tell them your pay for the meal is a little bit weird. Do it, do it. In a Starbucks? Yeah, I don't think they. Do they even have Starbucks in Paris yet?

They do. We did try. I mean, we did Starbucks. Starbucks, but it's not exactly the same audience, so. And people, how do the French look at Starbucks now?

Have they given in to it?

Yeah, yeah, yeah. It's quite successful in Paris is also a lot of tourists, but yeah, Starbucks is quite successful.

I think that was like the last city they think they could ever land in. North Korea and Paris were the two. So let's talk a little bit about how you got into the bigger companies and then this customization. Because how do you charge for products in the AI, you know, era of AI because people don't have a lot of seats. So do you do it on consumption based pricing? So there's really two questions there. How do you get into those big customers? And then pricing?

Yeah, of course. So. Well, it grew up quite well. Like we started to have market fit with ebay sellers. Gary Vee talked about us in 2020 and we grew to like the 10 million in revenue. And then did you reach out to.

Him or did he just stumble upon it?

I need to refine the tweet, but I asked him like this retreat for an entrepreneur and he was like, really kind. I got like, that was my first 100k views.

You asked Gary Vee for a retweet and he did it.

Yeah.

Look at you. So bold. Wow. And shout out to my friend and friends with Gary V. Amazing 20 years. And that's. That is pure Gary Vee. There is just like he, he appreciates the hustle. Very nicely done.

Yeah.

So you, you get this, you get the sellers and at what point does ebay in the marketplace find out about you or do you just tell them it's twofold?

So first, like we, I remember out of Hackathon, we built an API because people were asking and we put it on the side, but we didn't have the like team to focus on it. And one day there's someone in customer support that tells us like, please could I use your API? And are you sure you can handle the load? And well, I think we can do for two weeks. No one's talking and we don't hear from them anymore. And one day, boom, we get millions of requests on one day. And it was like the Barbie movie that was using our API, promote the movie. And there were like 20 million requests in a few weeks. We had Beyonce, the Kardashians were using our product, our technology inside, like make a Barbie poster and that was the beginning of our second revenue line which was the API. And then the API was like well a mix of credibility with the Barbie, the logo and at the same time the best depop door as they're seeing, they were seeing the app and they say can you do that at scale? That's exactly what I need. I just need it at scale. And so we started working with some of the early adopters of AI and this led to well getting some of the like smaller like international part of DoorDash using US and they had like the best photo of the company. So then DoorDash is asking how do you make these photos to this subsidiary? And, and they say like photo. So it's really like product that grow at scale. And same with G Pop. G Pop was kind of using you know, the sellers they knew they were using photo. So if my best sellers, if my best restaurants, if my best like you know resellers, real estate owners are using a tech then why not using them? It seems to be working. And so that's really like pure product led growth for all these marketplaces.

Amazing. Let's talk about how you're running the company. How many team members now and are you experiencing the phenomenon of your team members as the founder, the ones who do embrace AI And I suspect if you're an AI company, the team members embrace it. In running the company, how much more efficient is each team member becoming every quarter, every year in your estimation given all the other tools out there that help people run companies?

I think we're doing things that we couldn't do. So I would say in double digit percentage every year I think it accelerated more. Most recently of course we use our own technology for image editing but we also use like you know, audio, we use 11 labs, we use lovable like we build like special apps on our website. So we got, we got, I mean the way you do growth is totally different from two years ago. We are building apps all the time. Like you know as TikTok was like everyone is making video now everyone is building apps and it's easy for non developers to build that and the developer they use like either cursor or cloud code. Same double digit, more efficient. Yeah, we really try to lead by example. Like you know that's the age like if you're an AI company your biggest asset is speed. You're faster because you use good tools and because your team is relatively small, we are 100 or edge is to move faster than the, than the incubants. So I'm really trying to demo. We try to build in public inside the company to see how we can use. We use AI and that kind of pushes everyone, every function. We even have an internal podcast that is built on AI telling us it's user interview. It's not Book LLM about what we learn from the interviews. And so the way you teach your users is also about your employees, but also AI.

Yeah, I had the founder of Notebook LLM on the podcast, Stephen Berlin Johnson, and We are using NotebookLM internally a lot. Because when we do research for this podcast or we're researching or we're doing a deal, like we're closing a deal at the venture firm, you might put all the documents or on the editorial side, a bunch of videos or PDFs, and then you can summarize it, ask your questions, or even make a mini podcast out of it. It's really phenomenal. If you haven't played with NotebookLM, that is a crazy unlock.

Yeah, it's so engaging. Like, if you want. I mean, yeah, if you want to teach someone to the end, like say something to the full company. Well, people, they don't. I mean, you need to repeat all the time. So you do your all hands, you do your podcast, you do your slack, but that's like a phenomenal. And if you do like tens of podcasts, tons of user interview. Well, we do a Notebook LLM to like share the result. At the same time, we have agents that would, like, find the 5 seconds of snippet of this video that you want to really have the exact word as the pm, for instance.

That is genius. So somehow somebody just uploads to NotebookLM all of those interviews, and then you can summarize them, ask questions of them, or even make a podcast out of them. That is so genius. Now talk to me about the models. Obviously, you're not building a foundational model. You're leveraging other people's models. How is that going? Open source, closed source. Just the velocity at which these models are being released is crazy. Is crazy. And they're getting really good. Although there is a theory that maybe it's slowing down a little bit, or maybe parity, where they all seem to beat each other by 5%. What are your thoughts on those issues?

Yeah, so actually we did train a foundation model for imagery, but we specialize. We trained one that has specialized for all features, that was creating the background, our philosophy. And I think, well, if, like for any founder, what I would say is, well, first you need to be good at evaluating models because there is so much off the shelf for open source. Try to learn and eval what model is the best for your use case. And we really take what's off the shelf most of the time when we want to build a new feature. And our goal is really to push it so much that then we reach kind of a limit for our audience, which is commerce. And what do we do then that is better for them? So background removal. Well, initially we took open source and then we trained our own model. Now we have the best model in the world for the segmentation, which is a key feature for us. So it's core. So we would train it internally. Now, the text to image model, just for anything that is not core to us or the idea of generating a video or any LLM you use in the app here, we would try to evaluate what's the best for our user and, and really compare and to the point where we kind of feel, okay, this model is generalist. It's never going to be good enough for us because too smart. So it takes time, all the time to think we need to train our specialized model on top of that. Well, just at this moment we tried training internally. So we do a mix because our goal is really to get the best. We're an app player company. Like, we give the best to our users. So whatever is better at one point, we give it to our users. It can be internally trained, it can be someone else.

How do you red team these and sort of build trust? You know, these things can make mistakes. We saw, you know, all kinds of weird behaviors recently. Mecca, Hitler, whatever. It was some crazy, you know, thing that happened at a grok where, you know, people were tricking it or something. So you can have weird things happen. The red teams are in place for a reason. So. And you have a higher duty, which is, hey, you know, doordash or whoever, you know, or ebay, they need it to be perfect. You know, hallucinations or mistakes or, you know, just a brand safety issue. So have you had brand safety issues? How do you think about them? How do you avoid them? Or is this kind of going away now and not as big of an issue?

No, it's still a problem. You still have like in people or people. When you have 300 million downloads, like, people will generate plenty of things. It's as wide as the Internet. So what we do. Well, first we're okay in the sense that we keep the subject. So most of the use cases are like this. So it's the background. It's not as big as a full text to image, like a wide usage. So the usage is a bit limited in this case. The other. We have some safety checks in generation at all level, at the image level, at the prompt level. And what we do, I think for the big brands, and that's one of our value, is we have the B2C side of it. We have the app. So we have like, you know, 30.

Million activities, another data stream that you could watch to know people's behavior. Because you could do something completely innocuous like put a bride and groom in the background, you know, having their, you know, you may kiss the bride moment at their wedding. Totally charming. And it's for, you know, somebody who's in the audience in the foreground or something. Or it could be, hey, let's have these two celebrities kiss each other. And now you got a lawsuit on your hands because you decided it would be funny to do something like that. It is interesting to see young founders and people. I just had somebody on the podcast who made a deep fake of me in their advertisement for their company. And I said, yeah, don't do that. You're confusing people. It's too good. So I asked them to remove me and then they put in a. A Muppet version of me and then put not Jason Calacanis on it. And I was like, so what did you do? I had them on the pod just to goof about it and talk about it. I said, there's a thing called the right to privacy in America, and it's even stronger in France, by the way, in the eu. But celebrities have the right to their likeness. You can't use celebrities in any kind of advertisement without explicit permission. And you can't use a celebrity voice, even like a voice actor doing, I don't know, Robert De Niro's voice. You're not allowed to do that. So do you have people trying to do those? And have you run into that problem yet of just people don't understand how advertising is supposed to work and do you protect against that?

Weirdly enough, we didn't have too many problems. We tried to respect as much as possible. I think being focused, really focusing on commerce kind of give us some leeway here. Everyone is going to go like, we were starting to experiment with video. I know it's going to be a bit trickier here. Being able to. A B test here is key. And then you really need to be able to filter at the prompt and the image level what you can do. Yeah, it is definitely a challenge. It is moving super fast. So we try to Do. I mean our philosophy is to try to do best and really every time someone is bypassing really filter for that. Yeah, correct anything really have these signals where you can reflect some images.

You're the tool. So I don't think people expect the tool to have the responsibility for the output but you know, you should have some level of awareness. I saw a chatgpt now if you try to get it to make the Star wars characters or any of the Disney characters because I think they've had some legal issues with companies like that. It just won't do it anymore. So you can't go there and say make me into a Jedi knight. It just won't work. Or make me a birthday card with C3PO on it or a Disney princess. They'll just be like, yeah, no, we can't do that.

And we do buy some of like we do buy content from artists, from photographers, from some of the big.

Oh really?

So it is like we're trying to do good on that side.

Oh, so you will use it for your training data or for galleries or something like that. I think this is very wise. There are all these stock libraries out there. I'm sure that if they could get some reoccurring or you know, decent usage, they would be delighted to have an extra revenue source from different tools. And then that gives you a differentiator as somebody who's doing it correctly. And I think that's the piece I'm hoping the industry comes to which is thoughtful people like yourself saying I'm going to proactively respect content. How do you look at. I saw this week, it was just in the past week, I guess there was Vogue was. It might have been Vogue or Vanity Fair. They got into a little bit of hot water because they did their first like AI photo shoot. You heard about this. Was it Vogue?

Yeah, I saw that. I mean we just released like we were one of the partner for GPT Image in MAID. So there are like five, 10 partners. We're the first on the photo side and one of the feature we launched was well vistral trial. So it's one of the really key features now in photom. You can as I showed you, you can like yeah, try that with your shoes but also a T shirt and Well, I think it's. I mean it's quite engaging. So some like we'll see more and more on that and like for small businesses, like our idea is that we are leveling the playing field. So if you're a small business, like you can really create Your photo shoot, you can create more. I think we'll see like half a billion of entrepreneurs in the future and this is really a tool for them. Now. Vogue is. Well, I think creative will create. So Vogue is like there are plenty of creative there and they want just to try the new tools. Like cinema was a new technology. Disney, it was a new technology. Also Pixar, it's the same like technology is the best like creativity space for.

For us there was actually a backlash with like some of those tools for making DH celebrities, et cetera. And you're like, oh my God, would this be fair or not? But you, you know, you can't make certain movies without using cgi. It's just not possible, really. And this could open up a whole new venue. You could have models who say, I want to make my catalog of images of me available. And anybody who wants to use it can pay me $100 flat fee and use it. And that model could get a thousand people a year to pay 100 bucks for their summerwear beach photo shoot.

Or you rent the likeness of.

You rent the likeness. Yeah, I mean, so why is that unfair? I mean the only people who would complain about that are the people on the top who have made it already, you know, like when these tools come out. So there could be all kinds of new, all kinds of new revenue streams coming in from this. What about a B testing stuff dynamically? I know people like dynamic testing and I wonder if you could do human out of the loop dynamic testing with your tool and have people ask for that.

Yeah, actually, so our API is quite powerful for that. You can just write in description of the image in one line and have like five of them in your website and just ab test all of this. So yeah, I think we see the future is that everyone gets custom ads. You have infinite ads, infinite CRM imagery and. Well, yeah, just in your. I don't know what kind of furniture style you like, but I can put a chair you're seeing in the style of apartments that you like so you can be more engaged and we can abstase that first based on segments and then really have a unique image for everyone, like white glove marketing for you.

And if you knew I was of a certain age and I had kids and I lived in Austin, Texas, you could show a car that would fit a family of, you know, whatever driving down the road in Austin, Texas.

Exactly.

It'd be super compelling to see the car you're thinking about getting or a car you weren't thinking about getting. Even better with Five people in it or six people in it driving down a road near your home and you can see yourself in it.

Right.

It's kind of brave new world.

The same way when you like, you know, you write CRM messages like you say, like I name. Hi, Jason. Like you've buy this product, you've looked at this product in our site. Well, the imagery can be customized the same way. Like you can really recursive, tag your name there, put the product, put the other product that we think people will buy together and send that as an image. That's our things that people are starting to do with thoughts room with our API.

Imagine if Zuckerberg said on Instagram, allow me to be put in ads, sent to me like a little checkbox, put me in my ads. And then, you know, you're swiping up and you see yourself in this incredible Tom Ford suit with this incredible watch. And it has links. Here's the suit, here's the watch, here's the glasses.

Exactly.

I mean, that would be so effective and so engaging. You're putting me in the advertisement.

Yeah. The question now is very interesting is like, do you want to see yourself like scrolling on Instagram for thousands of imagery or do you want to like, sometimes you might want to see yourself. You might want to see people that are same like shame. Some might be aspirational, depending on the brand, but give so many opportunities for marketers to actually find the way they want to like show that, you know. So, yeah, this is definitely the kind of anybody.

Have you ever heard anybody as the first. I just had this idea though of put me in my ads. Have you ever heard anybody talk about that in the industry? It must be. It's such an obvious idea. Has.

Yeah, I think anybody done it. There is an app that you can put like, you can browse the Internet with your, like with, with your likeness. Yeah. The learning, I think product wise is you want to see yourself all the time or do you want depends on the people. But it's starting to appear.

It would be so cool. I mean, can you imagine if you just got rid of the need for a tool to do that and it just happened on Instagram, you know, you're swiping up and you just see yourself at vacation destinations or driving a certain car. Amazing. Matthew continues success with the company. If people want to find out more or try the app, where do they go?

So we forum.com, we're on the App Store, Play Store, and for the biggest company, we have the API. Matt Chew with so on Twitter, LinkedIn, Instagram.

Amazing. Thank you so much for doing this and sharing so much knowledge with our founders. All right, thanks again to Matt for joining us for AI Basics. If you want to check out Google's cloud report, Future of AI Perspectives for Startups, all you have to do is go to Goo gle Future of AI. You're going to find predictions, real world examples and practical advice like you got today just for founders. And don't forget all episodes. Legal, AI, Marketing, accounting. It's all there for you waiting this week in startups.com basics. Late night weekends. You want to get Sharper, go to thisweekinstartups.com basics. Thanks again and we'll see you the next time on this Week in Startups.