Latest / The Interline Podcast / How Far Can Generative Tooling For Fashion Go?

Transcript

How Far Can Generative Tooling For Fashion Go? · The Interline Podcast

• speaker-0: Hey, welcome back to the Interline Podcast. In case it's escaped your notice, there's a lot of generative AI in fashion right now. Just a couple of weeks ago, we published our annual report about it, as a matter of fact. So if you haven't read that, I'd encourage you to grab a copy. I think it's gonna sit well alongside today's show. And I think there'll be some things in both that you might feel like cross-referencing. I also recently released a standalone essay about the top findings from the survey portion of that report. You can go read that as a human written. Article called Arm's Length AI, or you can use the accompanying toolkit for AI to work with some of the most important findings in your agent of choice. That agent toolkit is a new thing for us, ⁓ and it's something I think we'll only do a couple of times a year at most for big projects, depending on how it's received. It came about because after we released the report, I was inundated with LinkedIn carousels and newsletter emails from companies and people who were taking the survey results by passing the full 230 page PDF and interpreting those results in some let's just say like interesting ways and not always with the right amount of nuance or care. And well there's nothing wrong with sharing your thoughts on free to read content. We think there's a better way for readers and listeners to interact with the big things we research and write. So if you want to try that agent toolkit out, I'm curious keen to hear what you think. It's something we'll iterate on over time for the big essays that warrant it, and I think it should be shaped by reader interest or disinterest. The real reason I bring this all up though is that one of the pillars of that survey data set is the finding that people in fashion on aggregate think AI is the most mature in the image heavy early and late stages of the product journey. So think Creative experimentation and ideation, and then think all the way downstream in content creation, marketing, e-commerce, and communications. The opposite's true in the middle. People see technical design, pattern making, sourcing, production, and so on as the places that AI is the least mature and the least advanced. Or if I wanted to be a bit cheeky about it, making pictures of fashion is approaching being kind of a solved problem, turning those pictures into Real garments is a whole other problem. And one that I think a lot of the companies that make generative image first tools for our industry are wrestling with. If you're making one of those tools, these are the kinds of questions you'd be asking yourself. Like, how far do we go? Are we a design tool, a development tool, a marketing platform? Do we just build our own pixel and vector editing functionality into these things? Are we a PLM, a dam? Are we all of that at once? Does being AI native and generative at the core support those kinds of ambitions, or does it work against them? This is all deeply interesting to me. Now, like a lot of people, I think recently I just sort of mentally filed image generation away in that weird kind of bracket where you appreciate that something is technically amazing and that it's advanced in a remarkable way in a very short span of time, but it's also kind of become background noise. It's a bit like drones, which Unless you live in a war zone, are a prime example of how quickly a cutting-edge, sci-fi-y sort of technology can disappear into the texture of everyday life. Talking to today's guest, though, reminded me that what we're dealing with really is a much bigger rebuilding of the creative process, and that image generation and editing models aren't always interesting in their own right, but that the platforms people build on top of them just might be. That guest is Weber Wong. He's the CEO of Flora. Which is a generative creative platform used by more than a million people, according to the company's own metrics, including teams at some of the world's biggest brands. Flora raised forty-two million dollars at the start of this year, and one of the biggest outcomes of that investment has been a strategy of industry verticalization that started just a few days before we recorded this conversation with fashion. You'll hear this as we get into it, but Weber's plan is to basically create a new version of the Adobe Creative Cloud for the generative paradigm. And to do that in a way that's driven by different industry needs. But as everybody listening to this knows, fashion doesn't get built in Adobe. It gets conceived and sketched and pixel edited and color-graded and motion graphics and everything else in that suite. But the product data lives in PLM and in other places, and the decisions are made there. There's a lot of other places in a typical enterprise. Flora's Fashion Studio, which is their kind of packaged Version of the product for fashion professionals started unsurprisingly with a lot of image-focused tools. But the company's goal is to go from, and I quote, sketch to finish campaign in one place without skipping over the middle, which means engaging with all of those other places we just talked about. And probably the most telling part of that strategy is the fact that the roadmap includes tech pack creation. Managing trims, importing from Clow Browseware, and a few other capabilities that vault it firmly out of generative image tool and into platform aimed at disrupting how fashion gets designed and developed. And that's all just a fundamentally interesting proposition for what is, at least based on outward appearances, until pretty recently. It's a node-based canvas for interacting with a combination of generative image, video, and text models, and it has some traditional deterministic tools like layering and color grading built in. So I think the watchword for this episode is ambition. And whether all of that ambition ends up being realized in Flora or not, I think the scope that Weber and I are going to talk about should be a reminder to everyone that we're really not done with generative tools changing the definition of design, communication, and maybe development as well. As a final disclosure. The Interline uses Flora in our generative image workflows. We don't ask for or accept free licenses from any technology company, so we've been paying users for the last, I think, seven or eight months or so. We also use plenty of other generative tools across text, image, research, etc. So us being customers doesn't influence how I talk to CEOs like Weber. Except where it gives me some extra insight into pricing or into features that I wouldn't have as an outside observer. This does, though, I think, represent the first time that I've interviewed the CEO of a company that we are also paying customers of. So I wanted to draw attention to it. For now, though, let's talk creative platforms for the generative era with Weber Wong, CEO of Flora. Okay, Weber Wong, welcome to the Interline Podcast. No at all. I'm excited about this conversation as I will extoll. speaker-1: It's great to be here. speaker-0: Several times in the course of the questions that I want to ask you today. Before I get too far ahead, though, we start every one of these shows with the same two things. We ⁓ try to build a snapshot of what the guests' day-to-day work looks like. And we ask them to define something that we hope elicits a slightly unexpected answer. for the day-to-day, I think the question to you is as much about how you arrived where you are as it is about what you do when you sit down and you open your laptop in the morning. ⁓ You've had what I'll call like a long a no fairly nonlinear route to being the CEO of a 30 person strong AI native creative platform that's raised more than fifty million dollars in funding. ⁓ your bio, if I'm correct, is you were in investment banking, then venture capital, then you did an art program in New York. ⁓ and now you lead Flora as a CEO. You also seem to spear ahead a lot the philosophy and the thinking behind it and do a bit of essay writing and that side of things. How did you get here? And how does that trajectory influence what your day to day looks like? speaker-1: Yeah. I think ⁓ yeah probably a nonlinear path, but my first love was actually poetry and I really loved art. ⁓ I did realize at some point that, you know, I needed to make money for a living, so that's why I wanted to invest in banking. I thought, hey, if I'm gonna sell out, might as well go all the way out, you know. And ⁓ it also, you know, I also do enjoy some aspects of it and learning about, you know, how businesses run ⁓ and things there and I especially drew adventure. ⁓ my thinking there was I work at a top venture capital firm. Invest in the best startups and then learn how to do my own. So I worked at this firm called Menlo Ventures. I was the youngest ever investor hire, looked at everything from anthropic to runway ML. ⁓ and you know, I was mainly there to learn how to do my own thing. But I realized after some time there, you know, a couple things. One is that none of the companies I had seen, I would have wanted to start myself. You know, they could be making a lot of money, but they just weren't that creatively interesting to me personally. I wouldn't start that company. I also realized that, you know, I wouldn't back myself. At the time I didn't feel I was a one of one founder and I have very high standards for founders. I didn't have any unique insight or worldview. And the third thing I realized is that I really missed art. You know, I spent three years in finance, grinding away, doing a lot of great work, and it turned out that I was pretty good at it, but I just wanted to make things. So as part of my quarter life crisis, I quit my job and moved to New York to work at a coffee shop and explore the art world. And it was through that that I found out about NYUITP, the art graduate program that I went to. Where they use technology to make art. Because I think during all of my time in finance and in venture, you know, I saw all these people trying to use technology to make money. I didn't know you could use technology to make art. It just blew my mind that you could do that. And if it turned out to be exactly what I was interested in. So I started doing a lot of that. And I talked more into that program. They gave me a full scholarship. And yeah, I just started making art. It was actually the course of doing these kind of art installations. I was experimenting with like, Real-time AI installation, interactive projects, and a bunch of other things, that I started building a creative tool for myself. And that turned out to be the first version of Floor. a lot of people started using it, and then eventually, you know, my venture brain kicked in and I realized like, hey, there's a big opportunity here. And at the time, the opportunity was quite obvious. No one was building professional creative tools. It was all just consumer AI stuff or model companies to like quickly prompt things. But as someone that had learned 10 to 15 different creative tools of that program, had built my own tech sacks to do my own art projects, I I we knew that there was a huge gap in the market and that creative professionals were not being served. So I decided to build, you know, Fora to gonna go do that. So at my heart, I'm a I'm a creative tool builder. ⁓ it's been two and a half years since then and, you know, we've been growing quickly since have been used by folks like Pentagram, audio, Nike, a bunch of folks that we're proud to call customers. But yeah, we're just getting started. speaker-0: Okay, excellent. And as somebody who studies classical literature at university, let me tell you it's not exactly a straightforward route to making money out of that either. speaker-1: Yeah. ⁓ but it's fun. It's great fun and that's what's important. speaker-0: So for the definition, I'm gonna cheat slightly and I'm gonna put two things side by side. And I'm gonna ask you to tell me how they fit together in the creative process. So one is generative models, and the second is what you would call creative systems. So things that are made up of reusable tools and components and repeatable workflows. Now, the creative process until pretty recently has had one creative entity in it, the human designer, creative designer, tech designer, who to be a bit reductive, brought the ideas and then one toolkit, which is made up out of well has been made up out of clearly scoped, deterministic, analog and digital tools that the creative then used to execute on those ideas. Things look pretty different today. You now have two entities potentially doing the ideas or more, humans and generative models and ancients. And then you have a toolkit that's also not exclusively deterministic anymore. It incorporates more and more of a blend of traditional software and new, more kind of probabilistic and AI native steps. Now, your bet with Flora is clearly that you can build repeatable components and systems and workflows out of AI foundations. But that feels like an easy thing for us to just sit here and say, and a much harder one, I think, for the people listening to this, particularly designers, to map to the way that they've worked for their entire careers. So walk me through how you define like an AI native creative process and what it means to then stack that up to make a creative system. speaker-1: Yeah. So at Flora, fundamentally what we do is we look at what are the creative technologies available. Then we look at what is the creative process. And then we try to build creative tools that best reflect the creative process using the best technologies available. So I think a great way to help understand, you know, how the creative process is different now is that, you know, it all there's two layers here. There's a creative technology of a given era. And there's creative interface for a given era. ⁓ Before AI came out, you know, it was basically just normal computing. So Adobe was founded in the 1980s to build creative interfaces for the personal computing paradigm. That was about controlling what was called the graphical user interface, which is a concept of being able to control every single pixel on the screen. So the screens you look at, you know, being able to change the pixel color of one pixel, ⁓ that's personal computing. It was about making one piece of media at a time. And that creation paradigm has been the same for the last like 30, 40 years. ⁓ Illustrator came out in 1987. And all image editing is about altering and reordering layers. Adobe Premiere came out in 1991. And since then, all video editing is about altering and reordering time. It's ⁓ like crafting. It's very bottoms up, almost kind of programmatic. You slowly build up, you know, what you're trying to do. Then, you know, a new kind of computing paradigm came out. Generative computing. It's very different. It's non-deterministic. It's kind of random. You ask for what you want. You get it really quick, and that's great, but it's not exactly what you want. So the the benefit there is speed. You ask for what you want and you get it with a snap of your fingers. The downside is control. You don't control every pixel that comes out, right? So the the benefit of that technology is that you can explore extremely quickly now. You can explore hundreds of logo variations. dozens of different designs and just see all those things in front of you and pick one and then refine from there. It doesn't give you the control of exactly making the garment where you want, or having the the pattern have the curvature of the exact angle you want. But it gives you that starting base really, really well. And for what we've done is we've kind of combined the best of both worlds. We have this node-based tool originally that helps you kind of generate a lot of things at once. But then can also use an image editor or a video editor directly in Flora that looks like Premiere Illustrator to then change the thing specifically. Great. So you know that's kind of like the difference in process. And of course, one benefit of the fact that you can make a single piece of media with the snap of your fingers is you can connect them together. And now you can come in and kind of almost map out a creative process. So, you know. What used to be like making one piece of media in Adobe Illustrator taking two hours now takes two seconds. So now you can zoom out one layer of abstraction. And that's really, really powerful. So I would say that's kind of like one of the biggest changes here. And I think it mostly benefits the first half of the creative process when you're in that kind of divergent exploration phase of many different ideas. It's not great yet at, you know, giving you exact control. But I think that's where the role of, you know, more traditional creative tooling paradigms, image editing, video editing kind of come into play. So that's kind of how we phrase it. speaker-0: Yeah, and I I do have some formal questions, a couple of those. So just just thinking about what you've just described though. So how often do you think the typical creative, the typical Flora user takes what is their final output, let's say, from Flora, and it then becomes an input to those traditional tool links? So just just you know, just for argument's sake, I am if I'm going into it and I'm trying I'm trying to create a piece of social ad creative or something along those lines. And I get to where I'm I'm happy enough with it. Am I then pretty universally taking it into Photoshop and ⁓ and elsewhere to prep it for final use? speaker-1: I mean it depends. I think like, you know One way you can think about it is that when the models were really bad, only like five percent of things that someone would want to do could be finished completely in a generative tool. As the models get better, maybe there's a broader amount. But there's a lot of nuance there, right? Like the biggest use case for, you know, AI right now that can be done in one generation is stuff like, you know, those really flashy social media image like videos that you see, right? But you can't make a brand book, you know, with one generation. I generally don't think you can make a great fashion item with one generation. You can make a cool concept, but if you want to have deep control over every aspect, you still need to go in and and have a deeper amount of control, which we have more of that, and we're adding even more over time, especially for fashion specifically. ⁓ but yeah, it really depends on kind of what you're doing. speaker-0: Okay. All right. And I do want to draw deeper on some of the things that you just talked about. Now, I only got around to reaching out to you pretty recently because we finished our ⁓ AI report twenty twenty six. And as as part of that, I was like, ⁓ damn, I've been to been meaning reach out to that guy for a while. And the reason I've been meaning to reach out to you was ⁓ I read an essay you published around the time of your series A. So I think that was last year. Correct me if I'm wrong. ⁓ and I think in in that You drew a distinction between traditional creative workflows and tools and AI native ones. And it's one that stuck with me since you alluded to a minute ago. Now I'm gonna paraphrase you slightly for the speed and expediency, but your argument was that most traditional digital creative tools work bottom up. You said that before. People use them to assemble the final product piece by piece. AI is more top-down. ⁓ you described it as like a version of sculpting with a big shortcut at the start where a journeyman sculptor has done eighty percent of the work and then you, the master, You come in and chisel away at the rest. I spent a ton of time thinking about writing about speaking about AI in general, and I haven't come up with a better analogy than that. I really do like it. I think it comes loaded with some positive and negative connotations. Chief among them is the fact that when you're building something up layer by layer, you can take the last layer away if you mess it up. It's reducible to components. And this is a bit of a torturous analogy, but if you're sculpting with marble, for example, do you you're thinking when a mistake gets big enough to see it, you kinda have to start all all over again. You might be able to get another piece of marble from the same quarry, it's not gonna look exactly the same and so on. That that really does mirror my experience of generative image workflows. You know, yes, you can create elements and techniques and and loops and reusable things, or can go back to earlier nodes or what have you, or you can use generative in painting to address something that went wrong, but you it still feels to me at least when I've played around with these things that we're dealing with something brittle and idiosyncratic in the same way that you are with Stone, which is why I liked your analogy so much. Now, not to bog down in philosophy too early, but tell me a bit more about what prompted you to write that essay and if you feel like the state of the art has moved on since. speaker-1: Yeah. Well, what prompted me to write the essay was I explained that have probably like a thousand, ten thousand times already. Cause I've, you know, went to a lot of different creative firms and talked to a lot of people individually because, you know, there's a lot of skepticism around this and for good reason. ⁓ but the reason why I'm building this tool is I I genuinely believe that it can help your creative process, like it helped mine. ⁓ and I I don't want creative professionals being left behind here. ⁓ I think the mindset for a lot of creative professionals is that this is going to take my job. This is a bad time to be a creative. ⁓ when actually I think this is the best time to be a creative. You can see your idea in like five minutes at greater depth than you ever could before. And we're doing everything we can to make it so you can get to your the final result of your idea. So I kind of wanted to explain why we believe that. And when I explain it individually to people, it works, but I want to try to get it out to more people. ⁓ that's an interesting point on like destructiveness. You know, in traditional Tooling, there's Command Z, which is, you know, fantastic, of course. For generative tooling, ⁓ in this era, you know, we do have an image editor. And even if you're not using floors, you still can generate something and then like segment out stuff. Like we have this in Fashion Studio where you can segment out a given fashion item, like on the person, and then edit that specifically. So there are versions of that where you can kind of do it with traditional means. But also for generations, you know, let's say you have an image reference of a clothing item and you want to kind of do a different concept of it. Let's you generate out four versions, you don't like any of them, you can go back to the starting reference and then kind of do it again. So it's non-destructive in in that sense. So it's ⁓ it's a little bit different, but there is, especially in Flora, you know, that kind of view of different steps. And even in Fashion Studio, we've kind of kept that as well. Fashion Studio kind of looks more like Adobe Illustrator, although although a lot simpler, where you have the asset in the center, like let's say you have a fashion item, and you can use any of these fashion specific tools. Like, you know. making a flat lay a ghost form, or vice versa. And when you do that, you still see the initial ghost form that you can go back to and then try something else. So it's non-destructive in that sense. speaker-0: Okay. I think I think I think that's right. And I think maybe the mindset shift for for the team here and the and that we that we've seen as well is you have to think about it differently. You have to think about it as multiple generations, as as you described earlier. It's not, it's not a single shot, but neither is it the traditional paradigm of I will steadily layer things up and then subtract them through command Z or control Z when the when I need to. ⁓ i it it's a slightly different mindset, I think. speaker-1: Yeah, and one thing I will say, and part of the reason why we built Fashion Studio is that we wanted to deliver the most powerful generative tools in a creative tooling interface and context that more people are used to. A big blocker is just the fact that Flora is node-based. It's the most powerful way to use Flora, but you have to learn which model to use, although we have auto mode to help pick the best one. You have to learn how to prompt, although we try to help you with that on the back end. And you have to learn how to think about workflows, which, you know, it's a new Way for a lot of people to use a creative tool. In Fashion Studio, we've basically taken the generative workflows that we see most common among all the fashion teams we've worked with, like the 10 to 15 ones. We've thought about what is the end-to-end fashion process and then what tools correspond to that in sequence. And we've kind of delivered that in, you know, concepting, you know, refining the garment, and then like showcasing and having a set of tools that map to that. And then over time adding things like vector editing, tech packs, which is probably next week. ⁓ to be able to get even more and more control there, including some of the traditional stuff. So as a result, we're framing it less around the technology and more around your creative process. That way you can think about it less. And my hope with you know, studios is that ⁓ you know, these generative, these really powerful generative workflows that typically are like a couple steps to really turn a great sketch into a render, ⁓ is just a click of a button. And it just feels no different than like last sewing at Adobe Illustrator. ⁓ that's my hope. Because then, you know, we really can give professionals the power they want without them even needing to think about a prompt or think about AI at all. Because I don't I don't think they should have to. ⁓ it just gets in the way a little bit. speaker-0: No, I think I think that's right. And I think like I've I'm somebody who's who's used these tools. and disclosure at the Interline ⁓ does does use Flora, although we use a bunch of ⁓ other AI tools. ⁓ I've always used them in the node based interface 'cause it's one that actually does make sense to me, I think. ⁓ but I I get that it's not intuitive and instinctive for everybody. speaker-1: Like I, you know, I I believe ⁓ both that node base is platonically correct in that it gives you most direct control over this creative technology. But I also know now from, you know, you know, we have over a million users, but ⁓ Adobe has 40 million. ⁓ there's a lot more people used to different interfaces, and I do think there are ways to deliver the same power that you would get from a node based tool ⁓ in a much simpler entry point. So really focused on doing that to help a broader class of creative professionals. speaker-0: And we're going to talk a little bit more about Fashion Studio as we go. ⁓ just as a ⁓ heads up for ⁓ for the audience, these ⁓ these conversations sometimes get recorded a couple weeks before they go live. So ⁓ if it there's a good chance that some of the features that Weber's describing will be live by the time ⁓ that you that you listen to this. The perils of interviewing ⁓ fast-growing companies who ship a lot is ⁓ there's a there's a delta between the the conversation you have and ⁓ and the conversation that people hear. Now I I do want to take some of the theory that we just talked about there and I think look at how it manifests itself in in the capabilities of this tooling as as you evolve it, right? ⁓ because one of the qualitative responses we got to the survey we ran in the AI report this year was generative design tools are, and I'm gonna quote here, good at creating very rough ideas, but terrible at refining those ideas. ⁓ you've written and said a bunch about that, ⁓ when you've talked about getting speed at the beginning and then losing control and precision when it matters. ⁓ people largely in fashion seem to agree with the idea that these tools are most mature and kind of most effective at the beginning and the end of the kind of product journey. So when you're dealing with initial design direction for a collection, or you're taking existing products and bringing them to life in marketing content and so on. the real value though does lie in the precision and the control. ⁓ and like a lot of people, like I I spend more time than I care to on LinkedIn. ⁓ and there's there's a huge cohort of people on there who seem variously proud at different different stages of the kind of Coherence and fidelity and consistency that they've been able to get from generative workflows. You'll routinely see people being like, Hey, look, here's six images that I've generated from this completely synthetic campaign of a synthetic model wearing various pieces and so on. and the issue is not anymore at least, do those images look realistic? One by one, they do. You know, like ⁓ I'm reasonably well trained, not ⁓ not a massive expert, and it's becoming harder and harder for me to distinguish between ⁓ a generated image and a photo and a photo. Where they fall down is in the consistency across all five or six of those generations. So you would have one where a model is wearing glasses, for instance, and the glasses look real in every individual shot, but the arms of the glasses are different in each of them, despite beginning presumably with the same prompt and so on. How far do you think you can push that precision and coherence and consistency between generations? speaker-1: Yeah. I think it's good enough to look at, but not good enough to put on a billboard, is where we're currently at, if that makes sense. It looks good on LinkedIn and it's useful for concepting and stuff, but I wouldn't necessarily recommend that, you know, you do that for like a massive billboard necessarily. Maybe for social media. But you know, I I think is really useful for the concepting stage for sure. ⁓ and just to kind of touch on the like literal kind of model limitations there. ⁓ there's some stuff that is in the purview of the model companies to improve it better. We actually, you know, have an applied AI team ⁓ that actually comes from Scale AI, which helped train a lot of the models initially. ⁓ that we work with the the Frontier Labs directly and kind of, you know, tell them our learnings from working directly with fashion professionals because they care about improving this for, you know, real world use cases as well. And we'll tell them, like, you know, sketch to render, not so great because of X, Y, and Z. ⁓ you know, garment swap or consistency across sunglass on different models, you know, not great. So we'll inform them there too. So part of it's just like the model. The other part is like how you use, you know, these different models together. How I would do that specific use case is have one image reference for the the sunglass item and then kind of ⁓ you know, connect that to different shots and try to use the same model there. It is somewhat varying. Right. I think a lot of them work quite well, but it's I guess like that side, there's not much more to do there. So I would say like it's mainly in the model. And I think the models will continue to get better, ⁓ and you can't already do a lot with it today. speaker-0: I I am to be clear, that I'm I'm picking at the details here. I am consistently impressed at how far image generation as as a class of technology has come in a very, very short span of time. so yeah, I I I do I do want to reiterate that I think the eighty percent now that you got to before, like people the the amount that you can get done quickly ⁓ is is seriously impressive and it does change the way that we think about these workflows. I just think it eventually does hit this precision. barrier, I don't know how to call it that, like roadblock, whatever, whatever you want to call it. ⁓ but I I agree with you. I think it's on the model providers to ⁓ to improve some of that. where it's not on the model providers, I think, is how you distinguish Flora at the kind of industry verticalization ⁓ specific tooling ⁓ level. You've mentioned Fashion Studio, ⁓ which we'll I will I'll quiz you on a little bit more as you go, but ⁓ you know, Flora began life as a cross industry tool. ⁓ you know, artistic and creative for a range of different purposes. there's a difference between saying, okay, I'm gonna take this cross-industry tool and target fashion with it, and then actually building verticalized best practices, tools, integrations, capabilities and so on. It sounds like you've done a good amount of that ⁓ already. There are multiple fashion focused AI creative and content workspaces out there already. You have ⁓ off the top of my head, non selective Raspberry, Fermat, Vizcom, Mursa, which I know used to be called Kala, on the creative end and then on the visualization end, Chimera, Botica, there's a there's a whole range of those ⁓ kind of campaign and ⁓ marketing and content focused ones. Where do you think Flora sits in that landscape? speaker-1: Yeah. So the context for why we're focusing more vertical specific is ⁓ again, our product philosophy is study the creative process, study the best technologies, and then build the best creative tool that reflects the creative process, helps creators go from the idea to end result with as much control and speed as possible. How do we help them make great work? we started by thinking that there's a universal creative process because every team operates differently. So let's build a general tool ⁓ that you can shape to map to any creative process. Which is a node-based tool. It was almost more built around the technology. We obviously got a lot of users in in fashion. and they were using node-based canvas both for concepting, ⁓ you know, garment explorations, and also in brand and marketing for exploring there, and also for, you know, batch generation and bulk generation at scale, which the workflow tool is uniquely good at and that no one else in the industry had. ⁓ increasingly, though, a lot of fashion designers at, you know, a lot of these firms were like, Hey, this is really powerful. This is more powerful than any of these. industry specific tools that are, you know, they're more like AI wrapper types. It's a bunch of different small tools. It doesn't feel like a creative tool. It's not a medium for the fashion professional. It's a way to access models if in some ways, right? So they wanted something for the broader team. And we realized that, you know, a lot of these fashion professionals in the workflows, they're doing the same 10 to 15 things. So why don't we just take those generative workflows, optimize them really, really well, like a sketch-runner workflow, for instance. Just optimize the exact right model, the exact right prompt, ⁓ so they don't have to think about it. Let's collect all these tools together into one studio where you can focus on the asset. And now you can just go through the full creative process with the 10 to 15 tools that best reflect what you actually need that generative tools can provide. And let's also build in, like for what it's worth, you know, traditional image editing there too. So like cropping, color grading, you know, we're adding Pantone color select selection this week. ⁓ we're gonna add vector editing too. It turns out it's quite easy to build Adobe Illustrator. ⁓ so you know, we're gonna add that in too. So now you can just have it all in one place, a proper creative tool. Not like a place where you're going just for like an you know, sketch or render thing here or like another thing there, like a full-on creative tool, specifically for fashion professionals. We're building the new Adobe Suite, right? And it's because like there was so much value there that people just couldn't get to. ⁓ people that got pilled on the canvas really loved it, but It was just taking too long and there were so many people that could see that value but just couldn't get to it. So by focusing more vertical specific, we can design the creative tool not around how the technology works, but around how you work. ⁓ we can study that process end to end and do exactly what we need there. Right now, I think what we're missing is deeper control in certain areas, vector editing, tech packs, storing the metadata of the swatch so it can go into the tech pack. So when you place that order, all the metadata is there perfectly. We we want to try to help you with the full fashion process. From studying it over the past like two months of talking to like 200 different fashion professionals, I don't think the current way it's being done is very effective. It's switching between all these different tools, Illustrator, Excel, like mood boarding and Figma. Like it it can be done a lot better, I think. And we're not just an AI company, like we're a creative tooling company. And we're gonna try to figure out the best way to do it there. So think of it as like what the first kind of tool in the new Adobe Suite. But instead of group being grouped around functionality. like image manipulation for Photoshop, or vector editing for Illustrator, or video editing for Premiere, is grouped around one creative process, fashion studio, then film studio, then brand studio. So we can really tailor it around you and make it bespoke. speaker-0: Okay. So you mentioned tech packs a couple of times and you just you just said something else there that I want to pick up on. So being the new illustrator or being i is is a is an interesting goal for fashion for the reasons that you've just described. So illustrator is where 2D design work gets done. Creative design, technical design, and so on. It is not where tech's tech packs technical specifications get done. get made and builds bills and materials and points of measure and so on. It is part of an integrated suite, ideally. There was a whole I was around for it, a whole messy push about five to ten years ago to integrate Illustrator into the dominant product lifecycle management PLM platforms of the day, because what you needed was a handoff from the sketch to the platform that housed the product data that was then required to manufacture it. ⁓ And I think people see this now when they look at generative workspaces in particular. ⁓ and I I'm gonna quote somebody else from our survey again, talking about generative images as saying it it doesn't exist, it's just pixels. Somebody else needs to go and make the pattern, somebody needs to sew it, somebody needs to fix it, and so on. Or to put it another way, that you can sell tools that make pictures of fashion or you can sell tools that make fashion. I'm that sounds like I'm picking on being derogatory. I don't mean it that way. There's a lot of value in visuals. speaker-1: Yeah. speaker-0: But there's arguably a lot more value in doing the whole thing. So th tell me how you see mm because you're not just trying to build Illustrator if you're trying to do the whole fashion workflow. You're trying to build Illustrator with a PLM attached to it. speaker-1: Yeah, we've looked at the PLM. ⁓ I've talked to a couple of damn librarians. Yeah, there's there's a way to do this end to end where it's not just the pretty pictures, but all the details you need to actually make the thing. ⁓ it is deep though. And, you know, we're gonna start with the parts that are, you know, lower hanging fruit, I suppose. And in some ways this problem technically could have been solved before generative AI. ⁓ In some ways. ⁓ but it just wasn't, I guess. and if we think about trying to help with the full process, what you're talking about is a natural end conclusion. I do think one additional reason why it's easier for us to go after this over time is that the speed of building software is a lot faster now. And you know, for us, we're a very high-growth venture-backed startup with a very technical team. We are pretty good at building creative tools well for professionals. And fairly quickly. So, you know, it is more likely that we can go after this. For now, you know, we're gonna start with the more straightforward stuff, the more digital side of the process, but yeah, we obviously see the the need here and we'll move more into that over time. speaker-0: Okay, 'cause the ⁓ 'cause the other element as well is using real inputs to the generative steps. ⁓ so if I if I s so I've I've just focused on you have a sketch and then you want to turn it into a real garment, the other way is you have a material library and you have a color p library and you have pre or pro seasonal colours, you have ⁓ materials n i properly The kind of way you've used them across the there's a lot of different ways to then take that into a workspace and instead of going, let me manually draw a new pattern or a new block and let me put some swatches next to it and maybe I will experiment in 3D or what have you. If it's if you can conceivably do it with the technology is up to it at the model level and the platform level, you can drop all of that into either the kinds of kind of more intuitive tooling you're talking about, or the node-based canvas. That's a very different way of creating then. That that I s I see that very strongly. speaker-1: ⁓ yeah, definitely, Ben. I mean, like we have a recolor tool and there is an input for a swatch and you can upload a library of your swatches. ⁓ you know, you need to do that with your colors as well. ⁓ right now you can already do that, but we want to make it ⁓ like across the workspace. So like for your team, everyone has access to exactly the right swatches and whatnot. And if you use that swatch, we wanna store that metadata. So when you turn that into a tech pack, you know, it ex exactly says what it needs to say. So yeah, that stuff like we we want to go in that direction as well. Honestly, candidly, like I this, this I actually don't know if we'll go in this direction. Cause there is like the sum, you know, ⁓ I have to kind of separate what I think is creatively interesting and, you know, what is a good scope for us for the business. But, you know, the most extreme version of this is like any single fashion designer can go into Flora, make their item, end up with a tech pack, and we actually just go like help them manufacture it because we have relationships with like the mills or whatnot. Right. And we have all that stored. That's the most vertically integrated version of it. ⁓ yeah it's a lot of work. But that would be really cool too at some point. speaker-0: It it would be it would it would turn you into a fashion company exclusively, I suspect, depending on how depending on how much you want to stuff up. Yeah. Okay. Now just very quickly the final thing on studio. So you I think we've I think we've done all of the functionality and all of the ambition stuff ⁓ to death now. But who do you think is the audience for Fashion Studio right now? So I know you've done some you've got some brand testimonials and things out there you might wanna reference. Who who is your ICP within fashion for this? And where who do you think becomes your target customer as you ⁓ extend the footprint out either earlier into the design phase or later into the ⁓ design ⁓ the development and production side of things? speaker-1: Primarily fashion, ⁓ like professional fashion designers at at larger firms and the brand and marketing teams there. ⁓ an interesting thing Flora is it doesn't serve just like the fashion design process, but also the brand and marketing side. And one studio project is collaborative. So you can go in there, work together, kind of use that as a system of record for one project. ⁓ also getting a lot of traction from you know, freelance designers and also indie brand owners that are kind of experimenting on their own. ⁓ And this is sort of a new way for them to explore different ideas and whatnot. ⁓ I'm gonna launch a big kind of Shopify integration soon to make it really easy to go from here onto Shopify. ⁓ but I would say that's kind of our main focus. And yeah, some of the customers we're working with, Jordan brand, Prada, Sketches we've been working with for a while, ⁓ and a bunch more that we're also iterating with very closely and have been instrumental in helping us develop this. speaker-0: All right, cool. That's helpful. So economics of image generation are really interesting to me. ⁓ because if I just put my cold commercial hat on, it's incredibly compelling to look at the unit cost of a traditional photograph. I I I don't mean like on film, I mean a digital photograph here, and the unit cost of an image generation. ⁓ it's it's a slum dunk. You I can generate a one K, two K, four K image, pull in el elements and inspiration from wherever I want, and it'll cost me less than a dollar. I can, you know, I can do a whole campaign for the price of a stop at the coffee shop on the way to what would have been the studio or the airport to the location. I'd probably have some challenges if I did that around the provenance of the inputs and stuff. And there's ⁓ a bunch of legal conversations I've had on the show recently. But if I wanted to do fashion photography faster and cheaper, this would be an absolute slam dunk for me. I find the pricing bizarre. ⁓ like in general workspaces. And to be fair to you, that's not It's not a problem that's unique to image generation. I think anyone listening to this who's looked at a clawed usage page recently will be like, How the hell did we get here? Like three, four different kinds of limits, usage credits, promotions, boosts, all that sort of stuff. You've been through a bit of a pricing shift for floor as well. now I think you used to have token, like pool token billing. and then everyone on the team would share credits. You didn't have a need to pay for seats, which makes sense on that unit economic level, but It also puts a weird layer of abstraction between it where you're pricing image generation in credits instead of cents and dollars. And you have to do quick maths to figure that out. Now you've got clearer pricing in that every generation has a cent and dollar cost, which I appreciate, but you've also moved to a seat model combined with token allowances, which means customers need to be more selective about who should be a user. The short version of all this is I I don't understand what the final form of pricing for images is going to be. whether it's subsidized or not, that you have people who are doing guaranteed outputs and you don't pay for things that that that d don't be a quality bar if you're an enterprise customer and so on. Cause what's your take on this? Because like on the one hand, I can understand people looking at it and saying, this is so much more effective. It's faster, it's cheaper, it's better than traditional photography. And then on the other hand, going, actually how we pay for this and how it's priced and how much it's gonna what the costs are going to be incurred in the long run is Arcane to figure out. speaker-1: Yeah. Well, I'll start by saying the two main assumptions here that determine pricing are, you know, how much do these models cost per inference and what is easy for customers to buy. So for the most part, both for self serve, ⁓ like people that just come in and buy, and also for enterprise, there seems to be a preference for just buying a seat and just getting started because that's what people are used to. In terms of the cost of things, ⁓ yeah, as mentioned, like we switched away from this abstract unit of credits to just telling you exactly how much you have, which is typically more than the actual price of a subscription. Because it's just easier to track, right? ⁓ and in some ways by buying seats and pairing that with a certain amount of usage, it just helps them determine how much to buy. Otherwise, it's a little bit confusing. In terms of where this goes, I think, you know, those two factors may shift. Maybe people become, you know, more inclined with paying for usage. But some of our enterprise customers that are quite used to it now, ⁓ that's kind of more of the arrangement we have. They roughly know what they spend, so they just kind of buy that up front and we price more on that than seats. ⁓ so I typically find that the more used to it people are, the more they want to kind of buy based on usage because it's, you know, they have a sense for it, if that makes sense. Of like how much that gets them. I think the other thing is like, the price of it. So one thing about media models is they are relatively expensive compared to LLMs. So some LLM products like you know a basic chat GPT subscription, it's just $20 and it's kind of like unlimited usage. You don't think about it. ⁓ but an LLM generation is like the minimus. I don't even know how much it costs. It's very little. A video generation can cost like $2, right for a really good one. So you definitely can't give that unlimited. And you do have kind of have to limit that a bit. Having credits or usage limits you know, is typically helpful for that. If it got really, really cheap, I think that pricing ends up being seat-based. Because you're kind of just buying software again. And it's a de minimis cost in the same way that, you know, using Adobe Illustrator and the electricity it takes to run it is also de minimis and you're basically paying for software. So it kind of depends there. In terms of like models in our space, the models have actually slowed down an improvement. I don't know if you agree, but it feels like it to me. I'm relaxed speaker-0: Yeah, no, I I would I would I would agree with that. I feel like it I don't know if plateauing is necessarily the right word. ⁓ but I think I the I mentioned earlier that I am continually impressed by the rapid progress that was made in image generation. I think that was all shoved into like a six-month window. ⁓ and then since then I don't think I've seen anything that is massively impressive. And to be honest with you, I ⁓ I find the newest GPT image model, GPT image two or whatever it is. ⁓ I find that one very weird. Like I see generations from it and they have this like textured, granular, like overly detailed look to them that I think is worse than what went before. ⁓ and and is incredibly obvious. So yeah, I'm I'm with you. I think I think plateauing, but also in some areas I think they're developing like an idiosyncratic kind of look that is off putting in a way that the previous ones weren't. speaker-1: Yeah. And ⁓ we we're pretty well connected with the Frontier Labs and a fairly technical team. And I think my answer here is that they don't have good data to train on. They've been kind of training off the aggregate of the internet, but they're not getting the feedback of creative professionals. So there is a gap that, you know, can be bridged there that can help improve the models even better for creative professionals. And there's stuff that, you know, we'll work on there as well. Part of which is defining what the actual professional use cases are, which if you look at our studios, the tools that we've chosen there basically reflect our opinion on on what are like based on our research, the most valuable use cases for fashion professionals. So I think once that sort of data pipeline, which exists for many other things in like LLM world, for instance, get built out, we should start seeing more of an improvement. In ⁓ the media models. speaker-0: Okay. at the risk of oversimplifying things, a lot of what you describe both in Fashion Studio and in the node-based Canvas is an interface. You sell an interface. You're not a lab in the sense that you don't have a front you don't have a frontier model of your own. And that describes a lot of companies in ⁓ in in AI. unlike kind of like a CRM or a PLM or an ERP, though, which is again it's based on commodity ⁓ LLMs underneath. ⁓ People the goal for you is for people to like sitting down and working with Flora and to build those new creative workflows and so on, which means that the the web interface, the canvas is the product. Or that's that's the way that I would think about it. So help me understand the MCP API CLI play here. speaker-1: I actually think it's not just the interface. Like we have an applied AI team that helps deliver better generations in Florida than other places. So there's the models, but there's also the the workflow behind Sketch to Render, which is proprietary to us. We've optimized the prompt, the workflow, the agent behind it essentially. and we will hill climb on that and improve that in particular, deliver better generations in the rest of the industry. That is probably something that we do that other folks don't. Another example here is we have this thing called auto mode. We have text image video node. Most places you have to go choose the right model and figure it out yourself. We'll pick the best model for you based on a combination of do you want it fast or do you want it good? So we have a slider. So we have, we also pride ourselves on delivering better generations. And I think that shows through in terms of our usage and also why people like it. I will say, yeah, most of it has historically been interface. In terms of the API MCP. I view that as a way to operate flora from the outside. So the most interesting thing is ⁓ I'm seeing some people use MCP and Claude to operate flora for them almost like autonomously. And this is pretty crazy because like in Claude, ⁓ people will, in their instructions, say, you know, I'm a you know generative workflow builder. Here's how I approach prompting. You know, here's things I do and I don't do. And then they'll open up a project and put in the brief, chat about it. And then just like ask it to generate entire workflows in Flora. You have Flora open on the side, you just see it spawning like 50 images at once, completely on brand. And then you can take a screenshot, put it back in Claude and say, actually, you know, make the background for all these black or make it more like this or that. Claude will think about it with the best in class text model and then just run it in Flora again. And at the end of it, that'll kind of take that chat, ask it, hey, summarize what worked and what didn't, feed it back to my instructions so I get better at prompting. And then all of a sudden, like the way they are using that, they just look superhuman comparatively to even the average Flora user. And the average person that uses that generates about a hundred X more ⁓ than you know the average user in Flora. And of course, the average user in Flora maybe generates 10 or 50 X more than the user of a very simple tool where you're just hitting one generate button at once. So there there are like levels. I think the hard part for me is, you know. Even I'm not that deep into the MCP because I I need to like run the business of stuff these days. Like I haven't been to tool as much as I'd like to. But even like really deep creative professionals sometimes aren't oriented into using like Claude or like know what an MCP is. So candidly, one thing I'm trying to figure out is how to bring that power to everyone else. ⁓ still trying to figure that out, but there is some crazy stuff kind of going on there. speaker-0: Yeah, I will I will say to anyone who thinks that generative workflows, generative image models and so on are kind of just picking away at ⁓ a traditional surface or like a bit of a shortcut to something, describing those kinds of workflows, ⁓ people doing that kind of orchestration and stuff, it it underlines to me that this is a fundamentally different way of working. speaker-1: I think like one parallel I would say is like, ⁓ unfortunately I used to be in finance, I used to use like Microsoft Excel a lot. It's kind of like watching someone use Microsoft Excel instead of watching someone use a calculator. Like they're just calculating a bunch of stuff at once and like it's it's pretty crazy to see. speaker-0: Well, there are World Excel championships for a reason as well. I have a couple of friends who work in finance in London and ⁓ they are they're in they're in awe of some of those guys. two very quick final questions. So everyone's obsessed with the idea of taste right now, whether we're talking about text, images, general decision making. The big conversation swirling around AI, the big debate is that execution is cheap. And I think you just described that, right? If if you're saying, you know, generating a hundred X, what somebody else would be like doing the doing. Is relatively cheap, the discernment is where the value sits. And I buy that idea, I think, as as somebody who likes to think I have taste that I earned through my classical literature education we talked about. I'm no artist, but ⁓ I do question how that develops over time. Do you think this taste thing is a bit of a cat and mouse game over time? Because it feels like when you the models get better, the reusable, repeatable techniques and things get better. feels like we're kind of progressively encoding taste. And then over time the taste becomes less important. speaker-1: Interesting. well, you know, one of our first customers was a Pentagram, and I remember them telling me how about in ⁓ like the nineteen nineties when Adobe Illustrator came out and they started switching from hand drawing fonts to using fonts in Adobe Illustrator, it was a very similar mood to the initial view on AI of like there's a lot of taste in I didn't use the word taste, a lot of craft in, you know, doing it by hand versus doing Adobe Illustrator. Now Adobe Illustrator holds the role of what Doing it by hand used to be. They view Adobe Illustrator as doing it by hand. Now we've worked with Pentagram a lot, and I think probably over half of the folks there are using Floor a decent amount now. ⁓ but yeah, I I think there's some parallel there to what you're talking about. But I feel like there is some aspect of taste or like decision making there that ⁓ probably will never be touched. I don't know. The the problem is like AI literally doesn't inherently care. Like it's literally just like I have an app I have a poster up at the office that I'm looking at right now and I think it's a nice poster. I would it can look at that and replicate it, but it won't actually be like I like that, if that makes sense. It's just kind of putting it back out. So at some point, someone has to decide that this is good and that is not. ⁓ sure, you can train a model and like on aggregate to do that, but that's not going to be fit for your exact creative problem you're trying to solve, or even just what you want. So at some point there is some core seed of agency that. just is fundamentally human and ⁓ is a prerequisite to starting a creative project almost. ⁓ speaker-0: Agree with that. Yeah. I think I think that's right. I think if you sit down and you want to do something creative, that is a fundamentally human act. ⁓ and the the tools that you use for that, ⁓ that that's where the frontier moves. Very final question is about forward-deployed creatives, which is a term of art of yours and is a play on the forward-deployed engineer concept that Palantir pioneered. how far do you see forward-deployed creatives mirroring what forward-deployed engineers did? Because I think people get confused that. ⁓ FDEs were s basically just consultants and advisors and they were there to help people get maximum value out of things that already existed. that's a consultant's job, an engineer's job is to build. ⁓ so what does a forward deployed creative actually do at Flora? Where are they embedded right now? And if they do their kind of field work properly, what do you expect them to bring back and build that improves the viability of Flora for industry wide fashion workflows in the near future? speaker-1: Yeah. So The reason why we kind of created that rule back in December of last year was ⁓ we were going to these firms and I was presenting Flora. And ⁓ as soon as I showed them the value of what you could do there and framed it into their creative context, they saw a lot of value there. But before that, it requires like a different way of thinking. ⁓ so, you know, we started bringing in our first FTC was Kat from Pentagram, actually. She had been using the tool for a year at that point. And she was just far superior to me, of course, at explaining that creative process and how to work with it and how to think about it than than even I was. And every time that she would go and do a demo at like Red Antler or like a different agency in New York, everyone there would just be blown away and just have like their minds, you know, blown away about like what they can do and immediately get it and get to a lot of value there. ⁓ so it was just very obvious that we need to show it to them. Typically what they show is like they maybe first talk a little bit about the high level concepts of it. Tops down, bottoms up thing at time. ⁓ it's different for different industries. So for instance, Kate, you know, she blows up a lot on LinkedIn. She's our four deployed creator for fashion. We have two actually. And yeah, she'll kind of just go into a firm, ⁓ show them like a bunch of examples of how they can use it, how to think about it, how it fits into their use case and helps kind of train them up a little bit. Like any powerful creative tool, you know, it takes a little bit of effort to learn. Honestly, I think it's easier than Figma even. Once you learn the core basics, it's just a node based thing sometimes confuses people. So they kind of help bridge the gap there. And then they work with those teams over time, build relationships, and also do a lot of kind of content education ⁓ to kind of get this way of building and the power of these tools out to more people. So that's kind of roughly what they do. Perfect. speaker-0: So I think over time we'll see that loop develop between what they bring back from those customers and ⁓ and how the platform speaker-1: Like they're in every product meeting. Like I'll just bring them in and be like, what do you think about this or that? 'Cause like it's just literally having the customer in the room. They represent like fifty customers. They've talked to fifty customers. They know exactly what's going on and they're in many ways the most valuable f you know, input as we think about product. Where, you know, I still talk to a bunch of customers, but they talk to even more. And they've been doing it for a living for like ten years. So they just know it like the back of their hands even better than I do. speaker-0: All right, perfect. ⁓ Weber, thank you so much for your time today. I really enjoyed this conversation. I think ⁓ I think we got through a lot. ⁓ I'll be keeping tabs on how Fashion Studio and Flora in general evolves from here. love to have you back at some point in a year or so, see how things have evolved. But for now, thanks for taking the time to chat to me. speaker-1: Yeah, definitely. And as a floor user, definitely let me know if there's any way I can improve the product for you. speaker-0: All right, Profit. Thank you. And that's the end of my conversation with Weber. I enjoyed this one a lot, and I hope you've also come away from it with some fresh ideas about how far generative tools might end up being pushed, even if the underlying models don't make any more big leaps in the near future. I think if you put this episode in your head next to the one I did with Gloria and Luke from Fabro, you'll be in the right mindset for understanding the potential shakeup that's coming for design and development as technology categories and design and development as disciplines. It's an interesting time, and we haven't properly touched on how AI might be putting long entrenched categories like PLM back up for reinterrogation and rebuilding either. And that's something I want to come back to pretty soon. Something different on this interview show next week though. ⁓ and make sure you come back on Tuesday for the next edition of the edit, where Grace and I spend twenty-five minutes or so running through the top headlines from the last seven days. It's a breezier, more conversational, more commute friendly show than this one, but I think you'll like it if you've never listened to it before. But now though, ⁓ thanks for listening. I'll speak to you again really soon.