{
  "video_id": "Uv7IfH8JD9w",
  "channel_slug": "dataindependent",
  "channel_handle": "dataindependent",
  "title": "I interview the man behind AI Virtual Try-On",
  "duration_seconds": 1031.0,
  "url": "https://www.youtube.com/watch?v=Uv7IfH8JD9w",
  "upload_date": "",
  "transcript": "we're getting you know inbound emails\nfrom from brands that you know I looked\nup to and and always thought were like\nyou know these are this is the Pinnacle\nof fashion that's Calvin Chen co-founder\nof copia the company trying to solve AI\nvirtual Tryon image generation tools are\ngreat at producing realistic images\nthere's just one problem it's really\nhard to see what we look like in new\nclothes you have you right your face\nyour body you have the actual item and\nso that's three inputs already Calvin is\ngiving top brands a way for their\ncustomers to virtually try on new\nclothes leading to more conversion and\nless returns if you were to describe the\ntype of demand that you're getting the\ninbound demand how how would you\ndescribe it Brands want to understand AI\nthey want to know what they can use it\nfor yeah we're finding some really\ninteresting data on on increase of\nconversions but from a couple brands\nthat we're working with there's a range\nbut it definitely is you know double\ndigit increase in conversion which is\npretty interesting in this interview\nwith Calvin you're going to learn the\nstate of the AI virtual Tryon industry\nhow recent diffusion model updates have\nchanged the game why product photography\nisn't actually The Right Use case and\nwhy virtual Tryon is just the beginning\nfor kobio hello good\n[Music]\npeople we all know about virtual Tron\nbut what's the niche or what's the\ncorner that you're really trying to\ntackle yeah I I think so we really are\nfocusing on kind of apparel virtual\nTryon really started when my co-found\nand I realized like you know ultimately\nwhen you shop online you're looking at\nother people right models wearing these\nitems and it's hard many times it's hard\nfor people to visualize what they look\nlike it's also hard to get you know an\nidea of what size I should get and\nthat's really why we started working on\nthis I have a background in kind of\ne-commerce and apparel starting a few\nstartups in high school and in school so\nit seemed obvious to me that this was\nkind of a a huge problem and probably\nthe hardest problem to solve just the\nquestion of is technology at a place\nwhere it actually can be solved and we\nfelt like with with AI and diffusion it\nwas well so let's talk about the user\nexperience with one of your customers\nright now so pretend I'm a user and I go\non to one of your customers websites\nwhat would that look like like yeah so\nit would really be as you're shopping\nlet's say you're interested in a jacket\nyou'd click on the product Details page\nyou'd see a couple photos some text and\nto cart sizing whatever there would be a\nbutton that just kind of says try this\non with AI you'd click it you'd go\nthrough a simple kind of minute long\nonboarding phase where we ask you\nquestions like your height your weight\nwhat size you normally wear how you like\nthings to fit just a couple questions\nabout your body so uh we don't ask for\nany photos of your body but we do ask\nyou a couple questions like are your\nshoulders narrow are they Broad you how\nbig is your stomach do you have you know\nkind of are you fit or not and we can\nactually assign kind of a a body type to\nyou and then we ask you for one uh\nselfie or photo or headshot of your face\nand that's it and that profile or\ndigital Avatar as you like to call it is\nstored and you can kind of travel with\nthat across all the brands that we work\nwith that's cool well I tell you what\nbefore we get into the tech and like\nwhat's the hard part I'd love to hear\nabout your state of the union of the\nvirtual Tron space is there anybody\ndoing this well right now or is it it\nstill we're all waiting for the tech to\nget there I think it's one of those\nideas where it's been tried maybe a\nhundred different times over the last\ndecade um you know we've seen it tried\nwith different Technologies such as like\n3D modeling 3D simulation uh AR VR um\nand I think we came to the opinion that\nthere was a reason why people weren't\nusing kind of a standard Tryon solution\nand uh you know there was a reason why\npeople didn't really use any of these uh\nit was frankly just because um it was is\neither hard to use it wasn't accurate or\nit wasn't very realistic and consumers\njust didn't love using it first thing we\nwanted to test was does diffusion models\nand does AI change that does it make it\neasier to use does is is it realistic um\nand we kind of found very quickly the\nanswer was yes and I think the state of\nthe union or like where we are in kind\nof the market today is I feel like it's\nstill early and since that diffusion is\nstill very new and people are trying to\nfigure it out um but I feel like it's\nquite obvious to me that it will exist\nin the next 2 to three years um I'm\nalmost 100% confident of that it's just\nthe question of like who is able to\nbuild it the best um and so far you know\nI think we have the best model so pretty\nexcited if you were to describe the type\nof demand that you're getting the\ninbound demand how how would you\ndescribe it I would you say crazy we're\ngetting you know inbound emails from\nfrom brands that you know I looked up to\nand and always thought we're like you\nknow these are this is the Pinnacle of\nfashion you know these Brands move very\nslow so it's not like hey you know we\nwant to give you our money right now but\nI think that the the environment of AI\nbeing so hot you know Brands want to\nunderstand AI they want to know what\nthey can use it for they're looking for\nSolutions I think that environment you\nknow is very conducive and helpful for\nus and then the other piece of we think\nwell AI actually can solve these\nproblems we believe and so these two\ncombinations make a perfect kind of\ntheory for why this is the perfect time\nto kind of start this kind of company um\nand why it can be ex extremely massive\nwho do people try to compare you to what\nother players in the market do people\nbring you up in conversation with yeah\nit's really interesting actually because\nmaybe we just don't know about uh the\nother players in the market but you know\nspecifically related to like building a\ndiffusion model we haven't really seen a\nlot of other companies try it from this\nperspective you know we've seen some\npapers from Google research Alibaba\nresearch Amazon research we've actually\nbuilt those repos a lot of them are\nopen- sourced and we just found the you\nknow no offense to them but the results\nweren't that great and so from a startup\nperspective you know we haven't really\nseen much and I think maybe partially\nit's because it's such a niche view of\ndiffusion I think a lot of the really\ntalented folks in diffusion are building\nfoundational level models right so one\nthat comes to mind is like Pika uh\nthey're building uh like AI video uh\nthat team is incredibly talented and so\na lot of The Talented folks want to work\non foundational stuff but I think it's\nit's it's quite obvious to me well what\nif you know in 5 years or 3 years when\nthese foundational level models are\namazing these results are incredible\nwhat use cases could you use that for\nand I think you know e-commerce virtual\nTryon is quite clearly one of them that\nhas just an immense\nopportunity so as you approach these\nlarger customers what's the value pitch\nthat you're pitching on them as to why\nyour product will help them out yeah I\nthink a lot of these brands are actually\ninbounding to us which is interesting uh\ngiven that we're still very new uh I\nthink it just um you know they've seen\nthe results they've seen the images and\nare really excited by it I think the\nthing that they're excited about is they\nwant to help the customers have a better\ncustomer Journey also what they want to\nsee on the back end is that leads to\nmetrics right so uh that leads to\nincrease in conversion or reduce in\nreturn rates um and those are the two\nthings that we're really uh kind of\nlooking to improve from a data point of\nview cool well I would love to talk\nabout um customers that you've chosen to\nwork with and so who are some of the um\ncustomers that you on board the platform\nright now yeah so we just launched two\nweeks ago so it's still all a bit um\nunder wraps because we're doing a lot of\npiloting and doing a lot of testing on\nthe back end so nothing I can publicly\nsay in terms of names but we're working\nwith a few large kind of name brand\ndesigner Brands a couple name brand\nretailers and a couple of different use\ncases with like uh different technology\ncompanies which is pretty interesting\nand is there any early results or\nfindings you can share about either\ndecreased returns or increased in\nconversions yeah we're finding some\nreally interesting data on on increase\nof conversions so early so nothing like\nstatistically significant but from a\ncouple brands that we're working with\nthere's a range but it definitely is you\nknow double digit increasing conversion\nwhich is pretty\ninteresting so if you were to sum up\nwhat is it about diffusion models that\nnow make this Tech possible what is it\nanybody has ever played with kind of mid\nJourney or dolly or any of those kind of\nin image generation tools what you'll\nsee is that they're able to generate\nincredibly realistic images\nand that's just because diffusion models\ndo a really good job of understanding\nthe images that they're trained on and\nreplicating them in different ways and\nso the fundamental aspect of them being\nyou know incredibly realistic is kind of\none thing that really adds as a benefit\nto Virtual Tron I would say the one\nthing that is a drawback of diffusion is\nthat it can be inaccurate or it can\nhallucinate like llms right and that\nreally comes down to how you train the\nmodel how you kind of um tackle those\nhallucination problems but they can be\nsolved and that's kind of the bet that\nwe're making is that they can be\nextremely accurate over time what are\nthe hard parts that you're building\nright now that is really your advantage\nfor your product yeah I would say you\nknow the problem itself is pretty\ndifficult so if you think about\ndiffusion models from a technology\nstandpoint most of them are text to\nimage or image to image right so um you\nknow one input of a prompt or you know\nyou give it a reference photo and it\ntries to kind of adhere to that what\nwe're doing is because virtual why\nvirtual TR is so hard is because you\nhave multiple inputs right you have you\nright your your face your body you have\nthe actual item and so that's three\ninputs already and so you know the model\nthat we've created is is not very\nstandard for kind of a a standard Fusion\nmodel it's built on top of kind of the\nopen source stuff but it's definitely a\nlittle bit more custom and built custom\narchitecture so it's pretty difficult I\nsay that to kind of say I think\ntechnology is somewhat emote now but I\nthink technology always becomes\ncommoditized so ultimately you do need\nkind of a a business or network effect\nmode to build a kind of generational\ntype company in a prior conversation you\nmentioned how mix and match is really\ntough can you describe what mix and\nmatch is and why it's so hard yeah so\nmix and match is kind of our our term\nfor really you want to try on let's say\nthis black T-shirt but you also want to\ntry on a jacket over it or let's say you\nwant to try on those two things and a\npair of pants so that's kind of mixed\nand matched that's what we're calling it\nit's extremely difficult cuz you're\nyou're trying to essentially whether\nit's generating multiple kind of layers\nor generating multiple pieces on the\nsame image but holding congruency you\nknow it gets very difficult so you know\nthere's a couple ways to approach it you\ncan do it with like different steps\nright so you could almost see yourself\nlike okay let me generate myself in this\nt-shirt first and then the jacket on top\nof that image but you run into\ncongruency problems much like a lot of\nthe video kind of problems that people\nhave been facing in Ai and so you know\nthese are problems that I think no no\ndoubt in my mind will be solved in the\nnext even 3 to 6 months it's just a\nquestion of do we have the resources to\nsolve it\nreally from your perspective as a\nFounder I like to hear about the risks\nthat you anticipate so it seems like\nthere's a little bit of a tech risk here\nbecause in order to get to that 100%\nthere will be further developments\nneeded what other sorts of risks do you\nimagine for your for your business yeah\nthat's a great question um I think\ntechnical risk is one that we feel\nactually quite confident in you know but\nthere always is a risk for sure if the\ntechnology will allow this to be built\nbut as I said before I'm almost 100%\nconfident so maybe I'm just delusional\num I think another risk is you know I\nthink with this specific IDE idea one of\nthe things that people have really\nstruggled to achieve is consumer\nadoption right do consumers love using\nthis and I think one of the risks that\nwe face and one of the challenges really\nis can we make an experience that\nconsumers love using and almost never\nwant to shop again without and that's\nnot maybe a a risk per se it's I just\nforesee that as kind of our main\nobjective and kind of our main Challenge\nand if we can solve that I think a lot\nof the other things become easier for us\nvirtual Tryon seems like it would be a\ngood place to start the the beach Head\nMarket and then go expand elsewhere so\nas you're thinking about the future how\nmuch of it is we're going to nail\nvirtual Tryon versus we're going to go\nto ancillary markets right next to it\nyeah uh this is where I get excited uh I\nI think you know virtual try is just\nplatform shift right it's just kind of\nthe beginning of making a much much\nbetter user experience for consumers and\npeople I think once we achieve kind of a\nvirtual Shrine that people love using\nand and our you know trust is really\nimportant important as well uh I think\nthere's so many things that we can do in\nterms of personalization in terms of\nunderstanding you know Greg I can\nunderstand exactly what you like to wear\nI can understand exactly what might look\ngood on you and actually just show you\nimages of stuff that you might like in\ndifferent backgrounds or with your\nfriends and that's the way and I think\nthat's the future of how images will be\ndisplayed to people and actually unlock\nthe the the ideas of people trying to\ntackle kind of an aggregated marketplace\nright you know maybe you could come on\nto kia.com and Shop from all the brands\nthat we work with where we show you\nimages of you in different outfits\ndifferent uh pieces from different\nbrands and we know exactly you know that\nyou're going to love them um and we know\nexactly what size you're going to get\nand maybe you can even share that with\nyour friends I think the possibilities\nthat virtual Tryon unlocks really are\nreally interesting in a prior\nconversation you and I had actually\ntalked about product photography versus\nvirtual Tron and how they they might\nseem similar but they're actually quite\ndifferent could you explain the\ndifference between the two and why you\nfavor one over the other yeah product\nphotography is something that almost\nevery brand talks to us about and it\nmakes a lot of sense in my mind so\nthere's two things as to why we do it\nbut not as our main focus and our main\nvision and Mission so the main thing for\nus is is one in my opinion it's kind of\nputting a Band-Aid on the problem right\nso as a consumer the real problem is I\ncan't see what this looks like on me and\nI I'm not sure what size to get and so\nit's really do you do you tackle it from\na consumer problem or do you tackle it\nfrom a business problem and The Business\nbusiness problem is well I'm paying so\nmuch money for these models and these\nphoto shoots and I'm spending a ton of\ntime you know could we use AI to\ngenerate this stuff so uh we don't have\nto spend that much time on it and so it\nmakes sense from that perspective but I\nthink we hold to our hearts like the\nconsumer is really who we resonate with\nthe most and that's who we want to solve\nfor uh secondly I think diffusion\nultimately has a very hard time\ngenerating 100% accurate images you know\nI think we've been able to achieve you\nknow anywhere in the bounds of like 98%\nI'd say um but 100% means uh like that\nblue shirt you're wearing those stripes\nor those uh you know those little\naccents are exactly where it'll be at\nevery image that you generate um and I\nthink when you want that as a brand\nbecause you're putting it as your image\non your product right people are going\nto be looking at this for um you know\nmaybe 30 45 seconds plus zooming in\nreally trying to understand what it is\nand I think that if it's even 1% or 2%\noff you're going to see higher return\nrates and Brands aren't going to be\nhappy with that so I think ultimately\nvirtual TR is is the thing that I think\nreplaces product photography and is also\nthe thing where I think diffusion model\nprobably has the best um advantages for\nand I would imagine there's a fair\namount of user psychology that goes into\nwhen you present a photo of somebody in\none of these items where if you showed\nit just vanilla pure the pure data it\nmay appear one way but it may be\nadvantageous to the brand to make the\nmodel look actually better in these\nphotos is is there any of that that goes\non behind the scenes yeah I think behind\nthe scenes it's been brought up to us\nnot by Brands actually um just kind of\npeople have floated this idea I think me\npersonally my view is you know it's not\nmy job to play God right it's not my job\nto make people look better than they are\nI think our job is to make people look\nyou know as they would in a real fitting\nroom if they were looking in the mirror\nand you know it's not my job to play\nwith your psychology to make you a\nlittle bit skinnier or a little bit\nfitter it's just the kind of increased\nconversion\non that track um you said you're going\nto be raising around um before the end\nof YC or at the end of your YC tenure\nhere how much are you raising and what\nare you going to be doing without cash\nyeah so uh you know you caught us in the\nright time so we're actually starting\nfundraising next week kind of gearing up\nfor YC demo day on April 3rd we're\nraising our seed around so we're we're\nlooking to raise a couple million you\nknow we're really excited to kind of\njust find the right partner hopefully\nget this you know get the money that we\nneed uh you know find the right people\nthat we should partner with\num and do it very quickly so we can get\nback to building uh you know we're get\nwe're seeing a ton of demand and and\nhonestly that we can't keep up so we're\nexcited to get some money hopefully you\nknow make some key hires kind of get\nalong the road yeah I love that so um\nwhat are the type of key hires that\nyou're looking forward to making just\nmore uh you know if anyone out there is\nreally interested in diffusion has\ndiffusion experience or really at the\nminimum kind of machine learning or AI\nbackground uh please reach out we we're\nalways looking to kind of bring new\npeople onto the team to help us with our\nmodel if I'm an e-commerce store owner\nand I want to get involved with copia\nwhat do I do yeah you can go to our\nwebsite so triopia docomo kopia um or\nyou can email me at Calvin\ntropia um and uh you know we can set up\na time chat I read every email that\ncomes through my inbox and uh pride\nmyself on responding pretty quickly um\nso you know feel free to reach out tell\nme a little about yourself uh tell me\nwhat you've done before um and we can\nset up a time\nchat",
  "transcript_chars": 19432,
  "ingested_at": "2026-05-15T04:40:50.991354+00:00",
  "source": "channel",
  "yt_meta": {
    "view_count": 2395,
    "like_count": 66,
    "channel_id": "UCyR2Ct3pDOeZSRyZH5hPO-Q",
    "categories": [
      "Science & Technology"
    ],
    "tags": []
  }
}