Matthieu Rouif: How Photoroom Creates a Standout AI Product

0:05Hey everyone and welcome to Generative Now. I am Michael Magnano. I am a partner at Lightspeed. Now, there are great generative AI models, and then there are great products built with these models that ultimately deliver value to customers and users. And this week, I talked to the founder and CEO of a company that's delivering a ton of value. It's Matthew Reif, co-founder and CEO of PhotoRoom, a photo editor that uses generative AI technology that can alter everything from an image's background to a subject's shadows, making this the perfect tool for many small business owners and solopreneurs who are selling their products online.

0:41Matt and I talk about all the lessons he's learned as the founder of multiple startups, finding product market fit early, and the unconventional way in which he's discovered what's really valuable to customers. So enjoy this conversation I had with Matthew Ruif, co-founder and CEO of PhotoRoom. Hey, Matt. Hey, Michael. How you doing? Good to see you. Yeah. Good to see you again. Thanks for doing this. My pleasure. Thanks for inviting me. Of course, of course. So there's so much I want to get into. PhotoRoom has been one of these amazing AI success stories, right? There's so much talk about AI, about generative AI across enterprise, across consumer, large language models, photo models, video models, music models.

1:32but like there haven't been that many success stories yet in terms of actual products like for as much hype as the category has gotten there's probably only a handful of actual products that have really broken out and perform really really well over a sustained period of time and photo room is one of them like no doubt hands down you've had a ton of success with photo room and i want to get into it there's so much to discuss with it but i think we should start with your background because it's really interesting you know uh you've been building in the app store and for the google play store for for kind of a long time right um yeah so i'm i'm an engineer by training uh i actually moved to the u.s uh to to do my master at stanford and that was like 15 15 years ago and basically i mean i had good um science math training which matches with AI.

2:28But I wanted to, at that point already, I wanted to start a photo business. So that's a long, long story. You know, it started early. Why photos? I like the technical part. I like arts and visual arts. So I've done a lot of drawing. I do that on my side, on the side. I spend a lot of time on photography and I like memories. I was building a ton of photo memories with books before, I think. And I wanted to print, actually, yeah, I wanted to print, we wanted to print like postcards and photos from travel back then with my back then co-founder. And so I came to Stanford with this idea of photo business.

3:10We're already having, I don't know, budget for printers, screens, digital screens and all that. And well, it was 28 and there was the first iPhone class. and for a graduate student it was so much an engineer on top of that it was so much more interesting to get into like okay i'm going to program a little app that does photo editing then negotiating contracts to print and install in some places in the world printers and that's i got into well photo first and mobile apps in the very early days i actually remember a friend of mine like back then 28 saying like oh yeah the early days of apps that apps are all over everything has been done it was six months in so to tell you how like it was not obvious at that time that uh there were a lot of things coming with the app but it was just fun for me to build software this way and basically that's how i got into uh mobile photo and yeah connecting i think already arts and science that's the thing that i love and really how do you mean very creative craft i think and how do you connect them to create something that's useful i i totally know what you mean about building mobile apps versus other types of software so i i was also uh i was a computer science major and yeah first few years out of school it was i i was totally bored you know i'm like programming these systems and these applications for things that i didn't even know what they did.

4:38And then, like you said, when the App Store comes out, all of a sudden, it's just a whole new lens. Before PhotoRoom, you worked on a bunch of different apps. Yeah. Talk to us about some of your early experiences. Yeah, I got into the App Store and iPhone development back, I mean, on campus at Stanford. I started my first company around, I wanted to do photo, but we pivoted and we did a bunch of ski resort apps. So we had a lot of apps for ski resort that will sell to ski resorts all over the world. And yeah, that was quite successful for 22 years old. After one year, I sold the company. So, wow, great experience.

5:17And yeah, a great way to get started into entrepreneurship and mobile apps. And then I started another company around social survey on mobile apps, A Crowd, that you would vote on questions. It was quite, we did that for three years. but yeah, I was missing a bit, I think, the visual part, so I came back to, I joined another startup in 2014 that was editing the Replay app. It's a video editor doing, yeah, like your photo, video memories from traveling, so back to my first love on photography, you know, traveling around the world, and they had this amazing tech and just pushing it after web to mobile, and so I joined the team, and it was quite successful.

6:03where we were on stage for the keynote app of the year also app of the year with Google yeah we were the first to do use AI to kind of make this video memories, stories of you automatically you're in the director's seat and not the editor's seat already 10 years ago and yeah it was a really amazing time using the beginning of AI at least on device and yeah and so it was very successful and gopro bought us in 2016 and i became in charge of the photo video apps at gopro for for two years that that's incredible i mean early in your career you've now already sold or been a part of two teams that have have gotten acquired i mean what's what's that experience like yeah i think it teaches you a lot i mean it first financially like it changed your approach to a lot of things to risk i mean you do believe in the story and that is possible um i i also think it makes you value a lot like what what was special in a culture because culture of a startup is not like a it it's not always concrete it's difficult to like uh materialize it you know it exists when you join that's why i mean that's So it feels special, but it's not always clear.

7:30And then when you join another company with the acquisition, it's another culture. So it's interesting to see how the two collide, if it's working well. So that's been a very interesting experience. GoPro is an amazing brand and tons of huge fan base. So very interesting, but it's a very different culture with a strong brand and marketing and doing hardware. it's not the same way as you do with for software where you ship fast and you try things and it makes you realize what's what well things that work and things that don't that you were not really aware of i think that's a big learning for me what do you mean by that things that work and things that don't uh at like the bigger company what do you mean by that no i mean i think what we did is like we in software you often like you can ship a lot of stuff and sometimes it works sometimes it doesn't and it's not a big deal and when you are a gopro you can't ship things that don't work or uh and same with a brand and so you realize that it's a strength of a small startup to be able to try things and have the have some forgiveness uh no one you're not good losing a lot when you you you're shipping something new that might not work and so it it's one of the asset of small startups that we tend to underestimate and that's that's a big learning totally yeah you can take more risk right like maybe it's because like you said maybe there's more forgiveness from the users or to the general public because they know you're a startup they know it's early or yeah you can definitely try things at a smaller company that you can't really at a big company because you could like you could really disrupt the financials of the business if You ship something that's broken or, or you, you know, you damage the brand or something.

9:18So it's so true, like being able to take those types of risks, both, I think per your point earlier, if you've already maybe had an acquisition can take more personal risk and being at a startup where it's young, take some risk on the business side. It's like a really, really powerful combination for doing special things. What was it like integrating at a much bigger company? It's also, you know, wasn't your, your company that was being acquired. What was that integration? Like, is it painful? Was it easy? Yeah, a bunch of people didn't stay on the long term. So I don't think it was, I mean, the app is great, but it's sad to see that a few things, like because there are bigger projects kind of collapse and people leave because they were in love with the small team and not in the big team.

10:01So you realize that sometimes it doesn't take a lot to have a bunch of people leaving, I think, here. Yeah, per your point earlier about the culture of a startup, that experience the acquisition the integration like it really highlights just how fragile and how special that is and how like quickly it can go away right if the culture changes yeah well that acquisition i believe turned out to be really important for your career because i think photo room ultimately came from that experience so tell us tell us a little bit about the journey through gopro to to starting photo room yeah i mean the exception for me of the idea of Photorum comes from, we were still like very independent and I was in charge of the apps.

10:44And I think as a PM, you always need to tell the story of the product internally, externally. And really, I became frustrated. I remember this day of trying to build an asset and my designer wasn't here. And so the US are sleeping and you need to promote that for, well, for tonight or the next day morning. And yeah, I just picked Photoshop and I had this exact story I wanted to build in mind. And I mean, I've been using Photoshop for 15 years. And I'm reasonably good with computers. I'm a developer. I build designs. I'm a builder, I think. And I became very frustrated by this experience of spending the afternoon and especially spending time on non-creative tasks that you shouldn't be spending time on.

11:31And at the same time, like we were we had the good ai lab at gopro in paris and i could see the paper they were working on it was 2016 and you you're like you start to have like good detection paper good segmentation paper and and you can only use it as a scientist where you go on this like this special uh code and you need to run the code by itself and it doesn't make any sense to have like this non-generative that has uh for like the mass audience and all the only the scientists that can use all this amazing AI back then. So that's the inception where I say, OK, we need to make editing photo more accessible.

12:09Programming got easier with no code. Editing photo should be simpler now because we can understand what's in an image. And that's how the idea of PhotoRoom started. There are obviously, I hope you don't mind me saying this, there's so many photo editing apps out there. You and I have talked about this before. I used to work on a photo editing app a long time ago, Aviary. even back then so many photo editing apps how do you stand out like when you decide hey we're going to start a new a new photo editing app like how i guess it's this feature it's the ai uh it's the machine learning you were talking about really that were that was able to to to make photoroom stand out from all the others we we came with kind of this unique approach of you should be able to i mean one of the non-creative tasks is like removing background And there were good papers back then, and the state of the art started to be very good, but on the web.

13:03And I think most people tend to underestimate, especially developers, that most people are editing photos on their smartphone. And let's do the best photo editor and the best background remover as a starting point on mobile. And that's where we got started. Yeah, and so very good quality on this feature, background remover. And then how do you, so that's like your first approach and be the best at it, be fast, be very good at the algorithm, understand there is demand for that. So I knew there were a lot of people searching for that on the app store from being in the app ecosystem for 10 years.

13:41And so I think you can crack it and be innovative when you come with this unique feature that's important to people. I mean, from that, you can build like a sizable business. And I think the extra part that brings us after that to what's next and why Photorum is very, very big now and very successful is that fundamentally, and in the eye space, people knew it, like removing background is a task that's called segmentation. And segmentation is basically parsing photos. So what you need to be the best at this task is understanding what every pixel of photo is and understanding what's in this photo, what's silent, what people are focusing on.

14:20And at the same time, you're building something that's very valuable as a feature-only company, but you also have many adjacent possibilities because you're understanding better than anyone on market, any other photo editor what's in the photo. And you can build so much from that. And that was kind of a bet, but we were confident it would make sense in the long run. So my takeaway from that is so many people on mobile at the time were just focused on like really basic stuff, right? Applying a filter, maybe, you know, messing with the saturation, messing with the content, like very, very basic sort of single double tap sort of things.

15:00Whereas on the desktop with programs like Photoshop and Lightroom, people are getting into more professional style work. But nobody was really doing the stuff that was really hard to do on mobile, but maybe easier to do on desktop on the mobile phone. But as you said, it was pretty clear to you that there was demand. So it was like, hey, let's just nail this feature, segmentation, something you can't really do on a phone. And we're going to base the product on that, at least initially. Yeah. Yeah, it's really smart. And you do have like tons of like, the question for a product manager is like, do you have depth on that?

15:37So are they like, are they tons of creatives that like by removing background, like segmenting that you can build that are unique and more accessible that all the other photo editor out there. And you go in the street, like you look at the ads, you look at the marketing assets, you look at many things. Like it's so common to be able to segment an object because you tell a story of many objects, and it's everywhere. So if that's at the top layer of your product, then there's going to be a lot of depth on the templates, the features that you're going to put forth to all the users. And this is going to be so much easier in PhotoRoom than in any other photo editor, because you start by that.

16:19In PhotoRoom, you import a photo, we segment it, so we remove the background, and you start from the object and not from the square of pixels. and that well maybe we're not going to be the best at like 100 of what you do but for 20 10 of all the possibilities we make you one step closer to make something amazing and that's good enough on a vertical photo editing that's like concerning five billion people yeah so what were some of those first use cases that you saw that maybe proved that you really did have depth to this feature right like i'm sure like you said there's you see you see segmentation everywhere but in photo room like what were the ways people were using it to start where you're like oh wow we you know we really have something here maybe we found product market fit yeah so we we we something we did very well that comes from the zenly teams uh team in paris is we would go to mcdonald and interview so we'd go in the line we would pay for a meal in exchange they were ordering in exchange for answering five, six questions on what their photo editing habits and testing the app.

17:26So it helped us remove a lot of friction in the outboarding and create an amazing outboarding experience. Make sure people get to understand why Photorum is special. But it also got us tons of ideas of what we can do that's special. A few things that resonated with the user in these interviews were, for instance, ID photo, passport, YouTube cover, YouTube thumbnail, But the biggest one by far was reselling on marketplaces. So the Poshmark, Zpop, eBay of the world, where some of them actually require a wide background that you do by background removal. And that's where we found product market fit.

18:08In early 2020, just before COVID, we started to have Gary Vee and other eBay influencers that talked about us, did YouTube videos about us, because, well, not of eBay sellers were hustling. They were doing like photos on the spot and they needed an app to go faster here because for them, time is money. And we were saving so much time for them. The quality of the algorithm was much better than what was out there on competition. And so that's where product market started. And did you learn that in McDonald's? Like you learned about the reselling need just by talking to users in McDonald's or did you already have that as a hypothesis?

18:45us? It's a good question. I think it's a, I mean, like product market fit is a bunch of things. But I definitely remember this user telling us like, my girlfriend would kill for an app like this to sell on Debo. And one thing we did very early, like we shipped the first version of the app after two weeks of programming, like and we put payment right there in the app subscription, right in the beginning. And the we listened a lot more to people paying than people not paying. And we realized a lot of people were coming back to us paying and asking about commerce, reselling features, eBay integration.

19:23So this part helped us a lot. And so this is like, I think it's a mix of things. We also did the product market fit survey where it was clear that some people would really miss the app. They would be very frustrated. So a bunch of things, but I definitely remember this deep up interview in McDonald's. So you're at product market fit. You find all these use cases and this depth of product segmentation. Where do you go from there? Right, so that's 2020, February. Basically, one month later, COVID hit. And that's, I think, I mean, startup, unsuccessful startup in particular, you have like also timing and luck is a big part of it.

20:01So while COVID was very tough personally and with the kids and all of that, for PhotoRoom, it meant like everyone needs to sell online because you can't, well, first, you can't take photos with photographer because you can't get contact. And then you need to go online because no one is going in physical store anymore. So that's like massive growth. We did YC remotely two months later. And really we got to 1 million in AR very fast like in the first year of 4 Term Story. So that amazing growth with YC, COVID and eating product market fit in one year. And talk a little bit about the algorithm for doing the segmentation.

20:44What were you doing back then, technically speaking, and maybe compare that to what's happening today on the imaging side? Obviously, there's been so much innovation in terms of innovation models just over the past year or two. So maybe explain to us what you were doing back then compared to today and maybe how have you changed as a result? yeah so i think our approach is really talking to user as i said and i remember this like uh i mean the way we pitched but photoreum is uh special because we do useful machine learning and that was our our pitch before gnai or anything but i think it's still very true today like we we do useful machine learning in the sense that we talk to user and try to understand how the ai is helping them more than starting from what's possible from the tech.

21:33We just try to be the best and use AI to fix the problem as best as possible. How do you do useful machine learning? We start from open source because we want to understand what's the pain from user. Well, the design school I did at Stanford is like show, don't tell. You want to show the AI to a user and understand what sticks with them, How did they project themselves? What's good? What's not good enough? Do you want to make it faster or more accurate? And you can't invent that. And certainly if you're in a lab and trying to do the next general purpose AI, you're not optimizing for that. So our strategy at Photoream is we start from open source.

22:15We put that in the ends of user as fast as possible. Then we understand what are the pains and the blockers of this special architecture. and then we train our own model that's best on market for this specific use case. So today, for the background remover, our tech is best on market. You can use it in the app, in the API, but any benchmark will tell you what's best on market. And that's because over time, we built data from our user participating and training a special architecture, using open source to improve it, participating in open source sometimes, to really adapt to our user use case, which is, well, photo editing.

22:54So it's not like, you know, self-driving cars. Like if you take some general purpose AI, like SAM from like segment anything from some model from Meta, it's like, it's a segmentation model that's made for everything. It's for robotic, it's for self-driving cars, it's for software, it's for many different things. And it's not optimized for what the user wants. And that's, I think, the value and why, I mean, the user, in our case, using software, what Photorum does is so valuable. And so that's the background remover. Then we moved on to other AI. So obviously, today, Photorum is much more than the background remover.

23:36It's a full photo editor with a lot of Gen AI features. And, well, that's the second part of the story where we, after being really good at a wide background, we started thinking of what's next as a photo editor. And that's just what is now called Gen.AI, but wasn't back then. We had the first DALI first version. We started playing and we realized, OK, sometimes what background or segmentation is not the best for users. They want a visual story. So they want for jewelry, for cars, for furniture. They want something that tells the story of the subject. And that's where we started investing in Gen.AI.

24:15So that's how we started investing in our Gen.AI models. So when you say tell the story of an object, are you referring specifically to the object in maybe the use case of reselling? Of, hey, these sneakers, this purse, this bag, let's give people Gen.I.I. tools to help tell the story of this item, this product that somebody is going to sell? Is that specifically like the use case and service area you think of or you think of other types of objects and use cases as well? No, I think, I mean, our mission is to power commerce photography. So really we're thinking of, okay, you come with an object and as a business, like you're showcasing, you're telling a story, you're selling, you're telling the story of your product.

25:02And you tell the story with multiple things. Well, first, even how you look at the seller, how is your profile pic, but also how do you put your object in situation? So sometimes you're showcasing in your shop, like in the physical store, you showcase in the window. And that's like, you really want to have like a, you do lifestyle, you have story that people can project themselves. And there's the same on photos on the internet. And then there's, when you go in the shop, you have like, okay, you have the full catalog. And then, so here it's about like seeing the different features, zooming on the material, if you take the fashion, zooming on the product itself.

25:38So that's more like catalog listing. You want a wide background, just want to focus on the product itself. So you really have very different kind of photography you want to do in the marketing life, in the marketing materials. And they're all different, but your scene that GNI enables is telling this story. And that's what we started doing with GNI. The white background is great for a catalog and product listing, but then all the lifestyle photos are supposed to tell the story of what you're doing. If you're going for sneakers, for instance, or you want to see them in the outdoor, so you want to see them in the mountain, how do you do that?

26:20Are you taking a full crew in the Alps to shoot that and you pay like 100K? Or are you doing that with Gen.AI? And Gen.AI enables so many things for that and to try things, to start A-B testing the different stories you want to do. what are some of the Gen.ai kind of technologies or model approaches you need to build to unlock some of these features? I mean, are we talking things like in-painting, out-painting? Like, how do you accomplish some of these things? Yeah, so today we have a full set of Gen.ai features. They come from our photo room instant diffusion model. So we train from scratch. And then from this model that's trained to do any photography, We do multiple things, AI backgrounds.

27:07So you start from an object and you generate a realistic scene that adapts to the light and the shadows of the product. We have a retouch feature where you use an AI to either erase an object from the photo or actually redraw, fill with a new object that you can specify. We have an AI resize feature where you can use any material. So you have a square image and you want to create a story. We just expand that to a new format. That's always a problem with stock photography or photos you have. They don't fully fill the space you have. Or maybe you want to use it in multiple places and you don't have all the pixels you would need.

27:56And so we also have an AI shadow model that's a specific model that's trying to create realistic shadow. And the value really gives volume to your product. But you can control the color of the background if you have a brand or you have a specific backdrop that you want to use for this one. So at the end of the day, it's like five different AI features that are used on top of our photorealismant diffusion model. What you talked about earlier with starting with open source is really, really smart because my interpretation of what you were saying was it basically allows you to test and validate hypotheses really, really quickly before training your own models or investing in your own technology.

28:36Will you even, like with some of the features that you just mentioned, will you even do that in the actual production product to test things before investing in building your own tech? Yeah, we do test off-the-shelf. If they are interesting, we have many different AI features that we put in either in beta internally to test them or try with user and do user testing to understand what they would value in that. So yeah, we do test many different models to see what the user value. At the end of the day, the goal, I mean, it's a bit provocative, but I don't really, I mean, I think there is value in owning the vertical stack.

29:22So because you can reinforce and use the user feedback to improve your model, But if the goal is to learn and improve your model, then you should ship as fast as possible with what's off-the-shelf or open source to understand the user need and really build from that. And if it's not our tech, if it's valuable for the user, I mean, the number one priority for us is our users. So we build our own tech because we think it's what's going to bring the most value to our user. But if it's someone else tech, we're fine using it. I think that's a really, really smart lesson that other teams should follow for sure.

29:58The other thing I'm thinking of as you're telling me all these features is, you know, by verticalizing all these features into a single app, PhotoRoom, there's probably just a tremendous amount of value you can unlock for your company and your product. But these features probably would also be really valuable in other products and other experiences as well. I believe you have an API. How did you land on the API? Like, talk a little bit about that. Yeah, it started as an hackathon, like many features, and we built like a, you get tons of ideas from that. It's kind of hackathon.photomall. We call them photo hack, some of the best part of the year for us.

30:41But basically, we did a quick prototype of the, yeah, for background remover page, and it used our API. and it started like this and we realized well back to our like to our being doing useful ai we understand that some people they told us like i don't want to download an app for that like for white background for instance especially as we were talking to like uh marketplaces where for them the flow is super important and we were super proud of our tech and the result but they told us like activation flow for a seller is just a single most important metric in the company and I'm not going to block the user by telling them you need to download an app now.