Hacker News new | past | comments | ask | show | jobs | submit login
Launch HN: Revery.AI (YC S21) Scalable deep learning-based virtual dressing room
103 points by mchong6 5 months ago | hide | past | favorite | 62 comments
Hi HN! We are Kedan, Jeff, and Min Jin and we are co-founders of Revery AI. We've built a virtual dressing room for online retailers that allows customers to visualize any combination of garments on any model.

The rise of online shopping has posed significant challenges for fashion retailers. The lack of ability to try on and visualize outfits has made shopping less interactive, contributing to low conversion rates and high return rates compared to brick-and-mortar shopping. Virtual dressing rooms can recreate the lost experience of trying on clothes in person. There are other companies working on a virtual dressing room. However, the reason why this is not taking off is scalability. Fashion ecommerce platforms have thousands, if not millions of SKUs. Current approaches generally require custom Photoshop work or expensive 3D models which are difficult to scale. In contrast, our solution leverages our machine learning research to automate the entire process, resulting in the first scalable virtual dressing room that can be easily integrated with any large e-commerce platform with millions of SKUs.

Rather than time-consuming 3d modeling, our system works with basic images. The goal, of course, is to produce accurate and realistic visualizations of outfits on people. A naive solution would be to simply copy-paste the garment onto the model. This presents two problems. 1) If the poses of the model/garment are mismatched, copy-paste does not work. 2) Even with ideal poses, copy-paste does not take into account garment-garment, garment-model interactions and also ignores lighting, shadows, etc. We use deep learning to overcome this problem. For problem 1) we use a series of image warpers to warp the garment onto an approximate body location in the appropriate pose. This differs from current approaches that typically use only a single warp which is extremely limited. For 2) we train an image generator that takes in relevant inputs (includes the model image, garment image, pose, etc) and produces a realistic image of the model wearing the garment. Our system produces significant improvements in size, fit, and drape compared to prior art, allowing us to create realistic images of any model wearing any combination of garments. If anyone is interested in additional details, we published an earlier version of our system here https://arxiv.org/abs/2003.10817. We also have another paper that will appear in the CVPR2021 conference soon.

This approach makes integration with retailers far easier because it requires only a single garment image on a uniform background per SKU. Upon receiving their catalog, our team processes them at a rate of 1 million images per week. We then work with the retailer to create a widget that can be easily injected into their website. The simplicity of this solution means that clients can have a virtual dressing room live in as quick as a few days. A live demo can be viewed here:https://revery.ai/demo.html.

We’ve successfully integrated with several fashion e-commerce retailers. Through working with our clients, we’ve shown that our dressing room improves the average engagement of users by 6x and, more importantly, the conversion rate by 6x. Additionally, we’ve seen increases in average order value (AOV) and decreases in return rates. Our solution also presents several use cases beyond the virtual dressing room. Because image generation is at the heart of our business, clients have also expressed interest in using our services to generate photoshoot images to forgo expensive studio photography.

Funnily enough, we have never envisioned ourselves doing a start-up in the fashion space as our backgrounds are all in computer science and research. We are all computer vision Ph.D. students at the University of Illinois at Urbana-Champaign and virtual try-on was initially just an academic pursuit. Kedan was researching fashion AI applications such as product recommendation while Min Jin was working on image generation and manipulation. Jeff was working on applied machine learning to practical problems like medicine and image search. We quickly realized that our individual expertise was compatible in tackling this difficult yet exciting problem. While image-based virtual try-on is an active research field in academia, no one has yet been able to productionize this technology. The transition from research to product is non-trivial - published research often operates on a largely simplified version of the problem. Generating realistic and accurate high-fidelity images of people and clothing is harder than it sounds. Inaccuracies are simply unacceptable for customers. People will not be happy if their miniskirt turns out to be a long skirt! It took us a year to get satisfactory results and at that point, we realized that this academic exercise can actually be a tool that real users want to use. That’s when we decided to launch Revery AI to bring a virtual dressing room shopping experience to all shoppers and retailers.

We would love to hear any feedback or answer any questions!

Interesting and congratulations for launching! :)

ON a side note the bags look ugly, I was expecting them to be resized and appearing as held by hands. Instead I got this: https://imgur.com/a/xsJAKR6

Is this something you're aware of? PS: I really like the wrinkling on clothes.

Ah, the bag feature is very new, and sth we're still working on to improve. Sorry for the disappointment, but thx for the feedback.

Why release something this awful? There is zero contextual awareness of bag type/shape and model hand positioning

Hi, as someone interested in both ML and fashion, this is pretty cool!

How are you addressing the idiosyncrasies of how clothing can fit differently? Personally, I like wearing a variety of different fits (slim/boxy/etc).

From your demo, it seems like the model is learning a single type of fit (the average of your training data?) and then mapping a texture + details onto this single fit.

In physical reality, the fit of a garment is obviously affected by both body type and garment measurements (as alluded to in other comments). Data quality (e.g., consistent scale and angle) also would be important. I imagine that because of these factors, the fit estimation problem is very challenging.

Best of luck with these interesting challenges!

You right, we don't handle fitting at the moment. It's difficult bc most e-commerce don't have diverse body shape model and many other reasons. We are working on it, will take some time for sure.

Based on your demo, I would suggest going this route "clients have also expressed interest in using our services to generate photoshoot images to forgo expensive studio photography."

You can make a ton of money doing that. I personally don't find a ton of value in this as a consumer. I want to see how the clothes look on me, not on a model.

I can see people wanting to throw together pretend outfits on a model, but I'm not sure how you monetize that.

Yeah, we are gonna do it for a few clients and see how well that works. Could be a quicker way to market than the dressing room.

"I personally don't find a ton of value in this as a consumer. I want to see how the clothes look on me, not on a model. I can see people wanting to throw together pretend outfits on a model, but I'm not sure how you monetize that."

Agree, I bet everyone wants to see outfit on themself or a person that resembles themselves. It's just HARD to build such product, period. I think at least for some people if not for everyone, styling is a important part. Also, style color mismatch accounts for about 18-20% of the returns. So far, we see a substantial conversion rate increase from the dressing room on our clients' website. Trying to get more validations to figure out the actual value of the dressing room.

It’s hard to build a product that shows you how clothes fit on someone like you as a B2B service. Retailers don’t want to showcase their clothes on anyone who isn’t anatomically perfect. Plus, if you try to source a more diverse set of “models” from real people wearing clothes, you run into the problem that most people are uncomfortable sharing photos of themselves “modeling” clothes publicly.

You’re also right that fit is only a part of the picture, and even the terms fit and style don’t quite capture what’s really going on you really want to see what clothes are going to look like on someone who looks like you and dresses like you (same preferences for fit, style, etc). Again, hard as B2B for sure.

I’ve been working on a B2C solution in this space for a while (fitfirst.app)…all too familiar with the nuances and intricacies in this space.

Fun fact and totally tangential: if you have two clothes with the exact same measurements and material that are dyed different colors, the darker dyed version tends to feel tighter than the lighter dyed one. Has to do with how the dye feels on the skin. It’s a nuance you can’t get from a photo or rendering.

haha yeah, there's a lot of nuance about a fitting room that's hard for 1 product to solve. Our current product focus more on the styling / outfits / engagement, not claiming on the exact fit. Hopefully it brings positive value to conversion and AOV, which would be enough to justify a B2B case. We've also build an app (Style Space), but are not experts in running it.

I definitely can see this becoming a standard thing. It's not an earth shattering improvement but enough of one for everyone to want it.

Makes a lot of sense. Thanks for the reply!

I spent significant time working on a virtual dressing room system using 3D avatars created from user photos, and clothing designs from fashion design firms, their brands and manufacturers. With a 3D game developer and film VFX background, I got an MBA to help with my pursuit of the idea. However, once in full time business development with a working system, I found the secrecy and competitive suspicion within the Fashion Industry to be a progress stopper. The clothing designers are so secretive, the clothing manufactures are not given the designs until the last possible moment - to prevent copying. This situation creates a very difficult time frame whereby the clothing designs need to be photographed for marketing literally 2-3 days before they hit the stores. In order to have a virtual clothing try-on system ready when the clothing becomes available, the company producing the virtual clothing system needs to be inside the secrecy veil, with every client, and uber security between clients. If you'd like to discuss what I learned, I'd be happy to share: bsenftner at earthlink dot net. I globally patented the process, but lost the patent in bankruptcy; this is a difficult problem to crack.

I'm definitely aware the kind of thing you're talking. We're lucky that with our solution, we deal more with retailers than designer brands. So far it's been not that bad for us.

Curious what's the name of company you are at?

I created and ran a startup called the 3D Avatar Store between '08 and '15. All that's left online is twitter: https://twitter.com/3DAvatarStore/media . Since that time, been working in FR. But am looking for something new atm.

Awesome product. As others have said, this has huge potential in eliminating costly photoshoots. If you can find a few big brands to partner with, you can completely disrupt the modeling industry.

Haha yeah, photoshoots are definitely not the first thing in our mind, but seems like there's real need for it.

I'm going to hazard a guess that it's because the tech isn't quite there yet? People pay expensive money for photoshoots because of the quality but as far as I'm aware generative technology isn't quite there yet? Or who knows, you guys could be the breakthrough!

Yeah, it's slowly getting there but hard to reach 100% so will need QA to touch up for a long time. But the hybrid approach gets better rendering than photoshop vendor image (at least shopper have difficult spotting).

Congrats on launching. Very interesting to see as someone who's into the intersection of fashion/design x tech.

This is based more on availability from brand than retailer but would make a huge difference in conversion: Images with more isometric/perspective views of clothed models convey more information (replacing need for up to 3 separate pictures [front+side+back] and/or rotation in many cases), thus more valuable to the consumer. However, this is less of an issue with standardized "flat" items and related categories like athleisure (t-shirts, sweatshirts, joggers, etc).

Good move with the skin shade feature - visualizing how an outfit contrasts/complements one's complexion is as important as fit. Needs some fine tuning though; there's some mismatch between models and listed shades.

Per people wanting to see models of similar build, try a regional or brand base approach: Generate a list of model types to use based on researching a) most common body types for each sex in a given geography (extra layer of detailing would be sub-categorization to account for differences in physiology across various ethnicities where applicable) or b) most common body types for each sex in a given brand's clientele or across multiple brands in given fashion category.

Can you perform the same model/garment fit with an uploaded model image? If so this would be revolutionary in reducing returns of online garment shopping. To see how a garment stretches when put on consumer's body type would be very beneficial. Ideally this would be done in real time but even a minute's wait to see how the item will look on you might be acceptable for consumers if it means less disappointment when you try on the item physically at home and elimination of a return.

This is our ultimate goal! It will take some time :)

It's very hard to ensure good quality rendering on user uploaded image (a lot of out of distribution). We've seen others who try to do that, but quality not yet great.

hey, congrats on launching- the demo is impressive :)

at unspun.io our customers create a body scan and we use that to make completely custom garments.

have you all considered supporting a use case similar to ours: virtual try-on with a 3d model as the input instead of user uploaded image?

We're considered the 3D option, it's difficult bc most retailers don't have 3D garments and it take very long time to create them. Maybe when that content becomes more available, we'll eventually replace our current system with 3D.

Interested to learn more about the type of 3D model you make. Drop me an email kedan@revery.ai if you're interested to chat :)

Awesome product.

Two things:

1) would love to rotate the model to see how it looks from all angles

2) being able to customize the model to look similar to me (so I can get a sense for how the clothes would fit on my body) would be awesome.


1) would love to rotate the model to see how it looks from all angles.

This our current product can do, if we were are able to get the side and back garment image from retailer (surprisingly a lot of them don't carry these images).

2) being able to customize the model to look similar to me (so I can get a sense for how the clothes would fit on my body) would be awesome.

This one is very hard, despite a lot of ask on it. We're working on it!

From a purely layman's perspective, why couldn't you get a head to toe photo of a subject, then have the subject enter in their height, then use that measurement to get the relative measurements of other body parts? You'd need a way to quickly identify the points of the body, but i can't imagine that is the difficult part here? The head to toe photo of the subject would need to be wearing skin tight apparel, such as underwear to ensure clothing doesn't throw the relative measurements off.

Interested to know where my assumptions here don't line up with reality.

Haha, cool idea! I thought we might had this discussion sometime internally. The truth is we haven't got time to explore it yet.

One possible difficulty I see is still how do we collect the face data / body data. Image generation has a strong bias toward domain -- meaning if we use one kind of faces during training, we will need that same kind of face during inference (with the same angle, lighting, etc..). It's possible but need more thoughts on how to ask users for good image.

I think that's one reason we avoid user uploaded images so far, bc it's hard for users to understand exactly what kind of image we need, and why their images doesn't work well sometimes. There's a lot to explore on this front before we can get a market ready product.

Regardless, cool idea, appreciated:)

This looks great! The demo currently shows front poses only, can it also do other poses? Instead of a virtual fitting room, could you use this to generate product images with models? Would the pricing be different since you're delivering a set number of images (generated once) instead of a widget that generates different combinations?

Also, what's the smallest store you're willing to work with (min # of SKU)?

"product images with models" Some of our customer has requested this service and we are doing it for them. There is difference: we use a large deep learning model that has higher latency and we do more QA on these images.

The pricing will be obviously more expensive. We charge slightly cheaper than market rate with more realistic image quality.

"Also, what's the smallest store you're willing to work with (min # of SKU)?"

There's not strict line. a store with a few hundred SKUs would work well. It depends on how shoppers's shopping habits: we noticed for smaller shop, shoppers more often come looking for specific items than browsing. Our product is more valuable for shopper with exploratory mindset.

Congrats on the launch! Can't wait to see how online fashion retail will evolve in the near future. I was just wondering how you guys are dealing with image copyrights, did you simply scrape public images and train your model on those? And which shop owns the copyright of the product photos used for your demo? Again, congrats! Impressive product!

Could you train a model on the reverse and make a background less image out of a normal picture?

I'm think if you created an app where people can let friends or others try on cloths from their closet.

Also, Wondering about a shopify plug in?

Also, can you just take clothing images from amazon other merchants use an affiliate program to sell from a virtual try on app.

What's the smallest merchant you work with?

Do you mean background matting that removes the background of an image? Perhaps something like this https://github.com/senguptaumd/Background-Matting

Yes! We definitely want to allow people to try on clothes from their closet. This will involve digitizing your own closet which is something we are working on to make it easy for everybody.

For shopify, we have plans on releasing that in the near future. We do have an app on IOS https://apps.apple.com/us/app/style-space/id1535818149 that pulls garments from different websites for you to visualize different looks.

Right now we are working with platforms with larger SKUs. We will support smaller merchants with our shopify plugin.

I also meant the clothes aren't flat and wrapped around the person. My thought was they could be unwrapped to show a flat version.

I'm saying that you currently take a flat without background picture of a piece of clothing add a model and get a piece of clothing that is made to fit the person.

What if you swapped your output image as your input image and trained a model on the reverse of your output and input.

new model current output image -> current input image

now instead of needed a perfectly flat no background image as input you can take a messier image and convert it into a flat no background image to be used by your current algo.

Having worked for a big luxury ecommerce brand, I can say this has huge potential -> "clients have also expressed interest in using our services to generate photoshoot images to forgo expensive studio photography."

Studio costs can be high, not to mention the delay in getting new product images uploaded.

True. We actually didn't thought about this, but some of our client we try to sell the dressing room ask us to make photoshoots since it's more urgent. So now we are doing that too haha. How large do you think the total addressable market is for this?

Super interesting product. Would it be possible for me to use this with images I find online?

It's possible, and I imagine would be a very cool feature for a shopping app!

We went to the B2B path after launching Style Space (https://apps.apple.com/us/app/style-space/id1535818149) and realize we don't know how to run e-commerce haha. Might revisit the shopping app / marketplace idea once we have more validation and tractions.

Apologies for the late reply. I downloaded Style Space yesterday after you mentioned here, and the main problem really just was not being able to import my own photos.

A consumer play might be super interesting here; let users import their own photos, make their own outfits, share it with their friends, and add referral/affiliate links on the way to checkout. I work on consumer tech, so email's in my profile if you want to chat more.

Curation is becoming a big part of the next wave of online influence and you have a very unique opportunity to capitalize on that.

Frankly, I think it's just very hard to get the kind of effect you're imagining. https://www.formatech.com I know this company that does what you want, but I don't think their rendering is close to what you're imagining. We recognize it's difficulty, so want to start with a more constrained case, maybe will eventually get there :)

Ah yea, I didn't mean importing yourself as a model. Just a digital wardrobe of sorts with the model being a default like you have in Style Space.


I remember seven or eight years ago thinking about how this would be a great use case for VR and would transform the way we shop for clothes. Cool to see you starting to make this a reality. Good luck!

I love the idea of this!

One issue I noticed is that adding a jacket does not seem to work. It shows me the loading circle, then goes back to the shirt / pants / no jacket image.

Thank you and thanks for trying out the demo! The jackets should work, can you point to which jacket(s) don't work? It is also possible you caught us updating something in the backend while trying.

It appears to be working now. Great launch demo!

Do you think there's an opportunity in shoes too? Probably less variance, but I wonder if you can deliver still positive impacts.

Feet are incredibly diverse, so shoe sizing is not trivial.

This would require either multiple measurements of the feet or a 3D scan, to be able to get a good fit on a shoe.

Yeah, we are working on supporting shoes!

This is such a great idea. I can see it being really useful. Can't wait to see it evolve!

How will you compete with Facebook, which is already doing this and giving this functionality and more away for free, and also owns the only channels that matter for fashion digital ads?

They wouldn't be our competitor on B2B, more like a user of this tech. They might have it at some point (either through us or in house), but only available for merchants on their platform.

Wasn't there a famous startup doing this in 2000/2001 ?


Thank you.

Yea I believe they raised some absurd amount of money ($100 million+) and blew it mostly on marketing. We definitely won't make that mistake haha :)

Congratulations on the launch, guys! Super cool application for deep learning.

I appreciate that your offering diverse skin tone gradient, but the first four skin tones are way too similarly pale at the moment. It’s not a gradient yet, it’s 4 light and 2 dark.

Looks like a super useful product. congrats

Gotya, we'll check that. The selections are some time pretty random, since we have a large model pool.

Love the product and really impressive results! Congrats on your work!! :)

As someone who loves fashion, only shops online and hates the virtual dressing room experience, I am extremely excited with what you guys are building. Congrats!

wow, that's awesome, thanks! Any suggestions / improvements? Most retailers are asking for diverse body shape supports -- super hard, we're working on it.

Looks super cool! The demo is awesome :) How are you guys able to support retailers so quickly? That's insane!

Haha, because we build the first virtual try-on system that's fully automated! We've seen 50+ other companies doing virtual try-on, but they need a lot of manual work to process each garment, thus have difficulty scaleing.

Guidelines | FAQ | Lists | API | Security | Legal | Apply to YC | Contact