Real Fit Solutions with AI and Physics | 3DRC, University of British Columbia | PI Apparel
In this presentation, the speaker (a university researcher and industry lead) introduces VitalFit, a physics‑based soft‑avatar platform that predicts garment fit with high accuracy. The talk outlines the challenges of realistic fit prediction, emphasizing the need for detailed body tissue measurements and sophisticated garment material models. The speaker describes the end‑to‑end workflow: starting from a high‑quality graded digital pattern, generating thousands of fit simulations across a diverse set of soft avatars, extracting fit attributes (compression, pressure, ease, placement), and compressing this rich data into a compact neural network for fast e‑commerce recommendations. Key innovations include goal‑directed pattern editing via a paint‑to‑adjust interface, personalization through a privacy‑preserving Vital Body Code derived from body scans or measurements, and virtual product launch tools that validate size tables across regions using regional body scan datasets. The platform integrates with Browzwear via a VitalFit Bridge and uses a proprietary VMG garment file format that can be generated directly or via plugins. Dynamic fit testing for athletic apparel and multi‑layer garment simulations are also supported. The speaker concludes with a vision of made‑to‑measure customization and encourages the industry to adopt this intentional AI‑driven fit solution.
Chapters
Thank you so much. The 3DRC has been like an amazing sponsor and driver of 3D adoption in this industry. And it's really an honor to receive this award. And it's actually kind of my second such award. This one is for the university research, which we did with one of my former students, Megha Shastri, previously Recently in 2021, we are honored to also get one for this, for the industry category for the soft avatars that we introduced. So thank you so much for the support. Okay, so I know it's been a very difficult year, both in terms of what's happening in the apparel industry and also the huge changes that are happening in AI that everybody is trying to get their hands around. I'd like to share something about our recent experience about how you can use AI carefully and get real fit solutions. This is my cartoon sketch of how product creation works. My apologies if this is too much of a simplification. But the idea is that you start with a business goal and some design concepts. But at the end of the day, you want to sell a physical product that is actually going to be worn by real people. So the first challenge is, of course, to make a physical prototype. And so that is the problem we initially started addressing, which is, how do you create realistic physical virtual prototypes of your physical garment and the fit? And we introduced soft avatars about 3 to 4 years ago, 4 years ago now, I think, for virtual fit testing. I'll say a little bit more about that later. And it's pretty mature. And in fact, thanks to Browzwear's open platform, we also have a plugin for Browzwear that lets you move garments that have been created in Browzwear into VitalFit. So this is great. But I've always felt that the industry as a whole is optimizing or over-optimizing for this one-size garment, the base size for a perfect one-fit model and iterating on that, trying to get that right, when the real problem that you want to solve is what happens at retail, which is How do you fit to, you know, tens of sizes to hundreds of thousands of customers? And so that's where I think the really big problems arise. So today I wanna talk about how once, you know, you have a good 3D pattern, you can then do something like a virtual product launch so that you can anticipate fit problems that would, that would occur maybe well in advance as soon as you created your initial patterns. And once you have done that, you can also use VitalFit with AI to do much better recommendations which are based not just on body measurements, which is what I think the typical fit recommendations are, but on how that particular garment would fit on a particular body. I know that you've heard this promises of fit technologies before. Lots, I'm sure every year there is some such solution. I would like to slightly open it up, open up the box and tell you a little bit about how we do this in 4 easy steps. Maybe not easy, but 4 steps. The first step is to start with quality of fit prediction. Quality and not speed. So people often ask us, VitalFit would take about 5 minutes to do a particular fit test. And they say, but here's this other piece of software that can do it in 5 seconds. Why are you so slow? And the reason is that our focus has always been on trying to get the most accurate fit prediction. Because we knew that the speed issues can be solved later. But what that means is we first of all want to acknowledge that the body, the human body is a very, very complex thing. And there is a huge variation of body types. You can just look around the room here. So it means that you need to acknowledge that it's highly complex. You have to actually pay attention to data. So in fact, in my research lab at the university, we developed new instrumentation to measure human body soft tissue properties. We've measured more than 200 at this point. So we built the SoftAvatar technology based on top of that. The other thing you have to do is to respect the garment. And this is actually, in some sense, less appreciated too, because especially by the AI community, that real garments are incredibly complex. Which means that you can't just use a simple cloth model. You have to use sophisticated, anisotropic, nonlinear, spatially varying material properties for your fabrics. You have to take into account things like seams and elastics and underwire and molding and all of these types of things that go into real garments. And it also means understanding how to get the garment to sit properly in the first place on soft tissues. Because of friction, the garment could sit in some random place. So you need to be able to provide things in your soft avatars so that you can, for example, get the underwire into the inframammary fold so that it sits properly where it's supposed to. So paying attention to these kind of details means that our customers can now do complex garments like this one. This is from Featuring, which is doing, for example, this is a pretty complex bra with dozens of pieces in it. It including has underwire inside a wire sheathing. It has lace. It's on a larger avatar with quite a bit of soft tissue movement. So you can get realistic fit predictions for these types of difficult cases. By the way, it isn't just for soft bodies which require a lot of compression that you can use this for. You can use it for shirts and other things as well. But this is the— I guess, in some sense, this is the most difficult cases I'm showing you. So that's also possible. So the next step then, once you have the foundation of a good fit simulation, the next step is to generate good data. And if there is like one take-home message you should have about AI, you hear so much about AI, and that is that good data requires, I mean, good AI requires good data. Because you can't just talk about AI without knowing the provenance of the data that it was trained on. But if you have a good high-quality graded pattern, so we can take that and now we can run the full-size run on hundreds of soft avatars. You don't have to just think about fitting on this perfect, single fit model, you can run it on hundreds of soft avatars in parallel. It takes the same amount of time and get thousands of fits in your style. And so from these thousands of fits specifically on your style, we can collect fit attribute data such as compression, pressure, how much stretch there is in the garment, where is the— what's the ease, what's the placement of the garment on the body. You collect this huge amount of data and then you can take this data and try to distill the results into the thing that you wanted, which is something that is really fast and small. We take this fit data and also very important, the fit intent of the garment because there is no such thing as something that just fits, without knowing what it was supposed to fit for, or even personal preference, for example. So you'd be taking that data, and then we run our machine learning algorithms, combine it with the previous data that we have, and we compress it down to a compact and blazingly fast small neural network that you can deploy as you wish. So here's that, like I'm just showing you an example in e-commerce-like application, but you could do all kinds of other things. And maybe one of the best examples is this R&D project that we won the 3DRC Grand Challenge for. So this was addressing a different and sort of a future-oriented idea of automatically doing pattern adjustments based on the goals. So what you're seeing there is like a typical problem situation. Say you're trying something on and you find in your fit simulation that there's too much pressure on the left shoulder. How do you get rid of that pressure? Well, you could try to adjust the pattern. You could bring all your experience. Maybe there's one person in your organization who has 35 years of experience who knows, OK, if I make this change, it will happen. But wouldn't it be much nicer if you can just paint the desired pressure on the right spot with a— as an interface and say, I want the pressure here to be reduced by this much. And then the AI can figure out what should be the pattern adjustment to reduce the pressure, which is your goal. That's what we call goal-directed pattern editing. The short answer is that yes, we can do that. And how it works is, again, the same recipe I just showed you in these first 3 steps, which is we utilize the power of digital and 3D, which is that once you have this pattern, then we can generate a huge number of pattern variations. And we can run all the different fit tests with VitalFit. And we train a generative AI model to be able to make the pattern adjustments at runtime very efficiently. So here are a couple of other examples. There's leggings here. There was too much ease or wrinkling around the knees. That's what you see on the left side. You can paint what you would like. So far, it's just changing the color, saying that, OK, I want the— the ease to be less by that much, and then it generates the pattern adjustment. Then on the right, you see the final results re-simulated with a pattern adjustment. It's not perfect, but within the constraints of the fact that you have to have a pattern to do that, it does as well as it could. And then in the bottom, you see a simple tank top for which we also even made a physical prototype just to compare the prediction with the physical prototype. So the final step is then how do you personalize the fit to a particular shopper? And for that, you need to get some information about the shopper, your customer. And this is, of course, a bit of a tricky balance. How do you reduce the friction in getting that? One way you can do it is by entering measurements. That's, I think, pretty obvious. But we actually train our AI on full-body fits. So we can also take body scans that might be acquired. We don't work on body scan acquisition, but we are happy to work with third parties who— there's a lot of great options out there on the shop— so on the on the exhibition floor that you can check out. But if you have a body shape data, so we can encode that into a privacy-preserving and compact code that we call a Vital Body Code. So that then can be used for testing the fit. Okay, so that's— those are the basic, you know, those are the basic steps. Of course, you can also combine, combine the two because you can start with, okay, this is what your BodyCode says you are, but I put on a little bit of weight here, or the scan wasn't right. I measured myself, it's slightly different. You can adjust it in real time. Okay, so I hope I've given you a little bit of a look under the hood about how these fit predictions can work and how you can use AI in a intentional way, I want to just sketch a few scenarios in which about how you can use it and what you can do with this VitalFit AI. Let me start with this virtual product launch that I said. As soon as you have your full graded size, you want to maybe test whether the size table makes sense for your particular garment. So a typical size table— here's again a cartoon— will divide up the space of bodies into boxes. And the key point is that these are based on just body measurements. You say that, OK, your underbust measurement is this and this, and your full bust measurement is this and this, so you are a medium. But is your garment that you made actually going to fit that medium? You're advertising this as a medium to your customer, your customer expects that that particular garment is a medium, but is it actually a medium? So you can validate that. What you can do is you can use a sampling of of bodies that I've kind of sketched there in that space to cover all the boxes that you expect your customers to be in. And we can run these thousands of fit tests on all of these bodies in parallel. And so then you can validate and say, these are my actual fits that particular— again, I want to emphasize, this is not just based on body measurements. It's actually based on the garment that you are advertising as medium. And is that particular one actually going to fit— I guess I need the pointer here. Is it actually going to fit that box? Or the true shape might be some curvy line like that, and you might have left out a large portion of the people who are not actually being fit. By what you thought was your large. You can do similar things in various other things. Suppose you're introducing a new slightly different product. You want this person who bought medium in the other garment to also buy a medium here, perhaps, but it might fit very differently. So once again, you have a digital pattern. You can then run all of these fit tests and you can validate the size table. And even more interesting and exciting thing is to fit to a different demographic. So suppose you are huge— you have a huge bestseller in the southern United States. And now you want to be huge in Japan. But the body shapes are going to be different there. But if you have some idea of the demographic of your target market, then you can do the similar thing of running lots of fit tests on it. So this is just a cartoon of it. The best real such data that I know of is from Humanetics, who has measured with their Size America, Size Europe, et cetera, They have specific data for different regions that you can access through their i-Size portal. But what's interesting is that in addition to these little dots that show you the distribution of measurements, they actually have 3D body scans and high-quality 3D body scans that are available specifically sampled from a region. So we can convert those scans into soft avatars. We can actually anonymize and remove the clothing that is used in the scanner. And then we can run a fit test with virtual people from that region and, again, collect all these fit metrics. For people from that region. And then you can get much more accurate, I think, regional sizing questions. So we can turn to a little bit to the— then suppose you now launch the product. It's in the stores. And now how do you use this to— guide your customers and provide a better e-commerce experience. To recall, at this point, the AI has already been trained and you've run all the fits, etc. You can then use this to create a shopping experience. For example, I can show you a very quick demo if this works with this thing. I just have to— Figure out how to move my mouse to the right screen. So this is just a mockup of— this is not a real website— of a typical e-commerce website. You can see maybe different pictures of the bra. But the main point is you can, in real time, using a very compact neural network, you can say, if this is my size, I would be a medium by default. That's what it says. But maybe my full bust is actually slightly bigger. And then I have a slight higher probability of being a large. So again, we provide you just the neural network. How exactly you present it to your customer is really up to you. Maybe I'll go back here. And finally, if you actually have access to body scans of your guests, so then you can convert that into— at one time, it can be done during guest registration or something like that. It's not like every time you have to do this. It's part of your loyalty program or something. You have a Vital Body code. Then that can be used to really create new and exciting shopping experiences, including, for example, you could filter what you show people based on what you know their body is. You're not showing stuff that they'll never fit or something like that. It reduces the friction of figuring out what is the actual correct size. I think. So this is my last slide. And so I think we covered a lot of material. So I just want to very briefly summarize what we talked about. So the basic takeaway is that you've got to go beyond 3D. And by combining realistic VitalFit physics with soft bodies and good— garment physics. With AI, you can produce practical fit solutions. And the best part is that all you need is a good graded digital pattern. And once you have that, you can, with very little work, generate a virtual product launch to see how this would fit. I think on the real goal, which is your customers, as well as create new shopping experiences. So I think— and in the future, of course, I hope this will happen where we can actually make made-to-measure. Then you can actually customize and say, I want pants where my left leg is bigger than my right leg. I want my left leg to be slightly looser. Can you make your pattern like that? And sure, why not? And I think that can happen. There is, frankly, the technology is already there. To make all of that happen. I think we just need to get into that mindset of really doing this personalization. So it's a very exciting time. Thank you for listening. And thank you once again for the recognition. Thanks. So I believe we have a couple of few minutes for questions. Either you can use Slido or this is a small room, please raise your hand and happy to take questions. Yes. Briefly, you mentioned having hundreds of soft avatars. Yeah. In one place. Thank you. Yeah. But I was also curious how you make sure that the fit is correct on all of those 100 avatars with very different bodies. So we are actually not making sure that the fit is correct. We are trying to find fit problems on all of those things. So what we do is we take the avatars that you think you're nominally trying to target with your different sizes, maybe a few more. So we are randomly, in some sense, sampling the population of bodies and then running fit tests as if somebody there bought yours, you're a size 42 or something and tried it on. And you collect these statistics about what is the probability that it would fit them. Yeah.
Wonderful presentation.
Thank you.
Thank you. You mentioned that you were using Braswork Beast Teacher as part of the 3D simulation program. Can you use any other 3D 3D software to fit into this product.
Yeah, sorry for the confusion. We don't use Browzwear for the simulation. The simulation engine is our own and it kind of runs on our servers. We have our own garment file format called Vital Mechanics Garment, VMG. And what we do is just to make it easy for Browzwear customers to access VitalFit, we have a a plugin called the VitalFit Bridge for Browzwear, so that if you have a garment that's already draped and you've done all your processes in Browzwear, it's very frictionless. You press two buttons and it sends it to VitalFit. It converts it to VMG. But we also have customers who can directly generate the VMG file, and you can run a fit test directly from that. Yeah. Yes. Great. I can hear you. You can speak.
So you briefly showed something where you could paint the fit, and then it would automatically adjust the patterns. What is it that you trained on?
It's a good question. So what we train is we, in that particular example, We train on— we generate a lot of pattern variations ahead of time. So once you have your base pattern, we generate based on standard measures. So you increase the rise or something like that. So we generate— actually, in that particular case, I think we generated thousands of— maybe 10,000 or something pattern variations. And for every one of them, we run a fit test on the body. So that way, what we are learning is how do you go from a change that you want to make in a body attribute— remember, this is like the goal. We don't want people to say change the pattern. We want people to say reduce the pressure, reduce the ease. So we learn to map. That very complex map from change in the pressure to the change in the garment.
Quick follow-up.
Yeah.
How does that work with more complex garments? I assume you would have to try to—
That is a very good question. That's why I mentioned this as R&D and future. We can do simple garments. But as I mentioned in one of my slides, I have the deepest appreciation for the complexities of real garments. So for a lot of the things, you don't have to wait for the future. If you have a good graded garment that you've— you might have been producing it for 5 years already. But you can then directly do a fit test at scale. And use that data. Thank you. Yeah, Natalia.
I'm just wondering, could you apply this technology for solving layering problems? For— For solving layering problems. So if you had a couple of garments and you wanted to have a base layer and then another layer and figure out how the upper layer or layers might fit.
Yeah, so actually, all our simulations deal with multi-layered garments. So for example, even that bra that I showed you has 3 or 4 layers. And so we deal with collision and contact between the different layers.
So I'm talking less about layers within one garment and more like if you have a garment system, like if you—
Yeah. So we would treat it in exactly the same way. We would treat it as an outfit is a— is a garment with pieces not connected to each other. Yeah. Please.
Is there any accounting for the amount of softness in different parts of the body or from one individual to the next that might work out more?
That's a very good question. When for VitalFit, where we actually build We've worked with different brands to build custom avatars based on— some brands deal with 50-plus people, and some people deal with children or something. So we can do that at that level. For the fit testing at scale, at the moment, we don't directly take that into account. But it is not that hard for us to then also sample from different age groups and other body types and potentially like morphotypes and even, as I said, regional variations like I mentioned, if you have that kind of data. Yeah, so that's a very good point. And it's, again, trying to model real bodies is, I think, is an ultimate grand challenge. But you should not be discouraged by saying it's a hard problem, because even if you get it 80% right, you are doing a lot better than getting it— having no knowledge at all. So that's about how the body is. So 80% or 90% is better. The perfect is the enemy of the good. Yeah, question.
So you're talking about pressure on that bra. Have you been able to determine anything with support, like with the bra bounce test or anything?
Yeah, that's a very good question. We do do dynamic fit tests as well. And particularly, we get that request from athletic apparel companies. So we can do dynamic fit tests, which take a lot longer to run. But we can get things like, what's the amount of excursion it is producing in the bus point and things like that. So we can collect those type of statistics. But it just takes longer, a lot more compute to do that. Oh, sorry, I missed that. So a couple of questions from— I think I can probably read it from here as well. So what file types work with VitalFit, and are you planning to integrate with other 3D garment simulation programs? So in principle, you know, the VMG file format is platform-neutral. We are willing and happy to integrate with all 3D garment platforms that have our open open enough. Browzwear was open and has a relatively easy way to integrate with their— we are, we are very happy to work with others. You know, please ask your favorite 3D garment simulation platform to say that you want to use VitalFit, and we can make that happen. Yeah. The second question is, is the technology currently being used and verified by a brand or manufacturer. It is being used mostly in the innovation departments of many— I can't mention them, but big athletic apparel companies. We also have other things like Featuring that I mentioned, who's been a customer for a couple of years doing this type of testing. Okay, I think time's— okay, great. Thank you very much.
