AI Virtual Fitting Function Cures Impulsive Consumption

AI also judges by appearance.

Every time you shop for clothes online, it's a chance to redefine your self - perception. With the recent 618 Shopping Festival, it once again proved the old saying - you see the buyer's show and think you'll look like a martial arts master, but when you put the clothes on, you look like a down - and - out soul. Let's not even mention whether the size fits. You can't imagine how much courage it takes for me to tell the customer service my height and weight, but I never get accurate advice. "Dear, you should choose a slightly larger size." Forget it. I'll just buy an XL... What's most annoying is returning clothes. From another perspective, the shop owners must be exhausted too. Thanks to the seven - day no - questions - asked return policy, many buyers take advantage of it and treat Taobao stores like fitting rooms, returning the clothes after taking pictures. Meanwhile, those of us who can never find the right clothes online are afraid of being blocked by the shop owners. As AI is becoming more and more capable, I've tried a few "AI clothing swap" tools on GitHub before. I uploaded a photo of myself and a picture of a yellow dress. After waiting a few seconds, I excitedly got a picture that made me look like a "big yellow pineapple"... Recently, at Google I/O, I learned that Google has launched an "Try On" AI feature for e - commerce shopping. The results finally show some signs of AI tailoring. < h2 > A Boon for Shopaholics < / h2 > The clothing style is like a recipe, and your body is like the ingredients. The model photos look great, with smooth lines and good drape, because they've got the right ingredients according to the recipe. However, ordinary people come in all shapes and sizes. Whether you have broad shoulders or sloping shoulders, a wide or narrow chest and back, or a high or low waistline - these details all affect how the clothes look on you. The same set of clothes can look refreshing on a thin person and bulky on a muscular person. What impressed me most is that this feature can show whether the clothing style suits you. Guess which of the following pictures is a real photo and which is AI - generated? < img data - img - size - val = "" src = "https://img.36krcdn.com/hsossms/20250617/v2_5f8309e55d4f4f208648cd275cc0fa97@000000_oswg86768oswg1080oswg784_img_000?x - oss - process = image/format,jpg/interlace,1" > Answer: They are all AI - generated. Dark colors make you look thinner, white makes you look fatter. The bulge of a beer belly and the folds at the bottom of the clothes when you put your hands in your pockets all look natural. To put it this way, I feel like I could use the photos it generates to become a Uniqlo fashion blogger on Xiaohongshu . < img data - img - size - val = "" src = "https://img.36krcdn.com/hsossms/20250617/v2_828bc6a8e4cc47fb98edcbd118357f7c@000000_img_000?x - oss - process = image/format,jpg/interlace,1" > A reporter from How To Geek was shocked after trying the Try On feature. He said, "I uploaded a photo of myself wearing long pants, but when I tried on shorts, it actually showed my calves. I swear, they look exactly like my own calves. To be honest, it gave me the creeps. " < img data - img - size - val = "" src = "https://img.36krcdn.com/hsossms/20250617/v2_fdc5e912ec6a4bd3a4aee78bf8a4d44c@000000_oswg79539oswg1080oswg743_img_000?x - oss - process = image/format,jpg/interlace,1" > Left: After trying on Right: Original photo | Image source: How To Geek " This is the first time I've thought that AI clothing try - on isn't just a gimmick ." - From tech blogger @MKBHD. It's a pity that the Try On feature is only in the internal testing phase. You need a US IP address and a US Google account to apply for SearchLabs and unlock it. The actual results it generates are different from the previous "clothing swap" tools. The reason is simple. Photo editing is about stacking existing layers, while Try On creates a brand - new layer from scratch . < h2 > AI Custom Tailor < / h2 > I dug into the technical details behind Try On. Step 1: Portrait Recognition First, it relies on a high - precision segmentation model to create an accurate "portrait contour model". Its multi - scale semantic segmentation technology is used to identify hair, ears, gaps between fingers, the direction of legs, and areas covered by clothing... So even if you have long, loose hair, wear earrings, or hold a phone, the AI will calculate the boundaries and preserve them in the subsequent generation. Step 2: Generative AI The simulation images generated by Try On are not only realistic but also three - dimensional compared to other similar tools. You can see not only the front view but also the side view and the view at a slightly turned angle. < img data - img - size - val = "" src = "https://img.36krcdn.com/hsossms/20250617/v2_f4ed007c973d4de6a2601a310092374b@000000_img_000?x - oss - process = image/format,jpg/interlace,1" > This shows that simply "cutting out" the person from the base image is not enough. The AI also has to analyze your shoulder width, chest circumference, waist circumference, body posture, standing position, and the angle of light in the photo... It does body modeling based on Google's self - developed Vision AI model and pose estimation model. < img data - img - size - val = "" src = "https://img.36krcdn.com/hsossms/20250617/v2_9ea039e561fc43d9bf7c39e69a462807@000000_oswg66412oswg900oswg505_img_000?x - oss - process = image/format,jpg/interlace,1" > Motion capture works on the same principle | Image source: Google This is why even if your calves aren't visible, the AI can calculate what they should look like based on the body model. At the same time, the model receives information from the clothing image, such as the texture of the fabric, the number of layers, and whether the style is form - fitting or oversized... Through the cross - attention mechanism, it fuses the information from the two images and uses a diffusion model to generate a new image of you wearing the new clothes. We know that the knowledge graph is the "brain" of traditional Google Search, storing all the information used to answer users' queries. The Shopping Graph is a branch of it that provides information about shopping intentions and products. Based on the vast database in the Shopping Graph, the model is trained to recognize how clothes look on different body types, skin tones, and poses, which improves its generalization ability. < img data - img - size - val = "" src = "https://img.36krcdn.com/hsossms/20250617/v2_b7562d5c3a3843a39f893f56c2709908@000000_oswg515002oswg1080oswg603_img_000?x - oss - process = image/format,jpg/interlace,1" > Try - on photos of different body types as basic data | Image source: Google But at this stage, the try - on results are still a bit "rough", like when you pick up clothes in a store and hold them in front of a mirror. Will they really fit when you put them on? Step 3: Image Synthesis Google uses image alignment and pose - aware warping networks to align the clothing with the body posture. For example, when you put your hands in your pockets, bend over, or raise your hands, the AI can make the clothes deform naturally. There might be a bulge at the knees, the corners of the clothes might flutter, and there might be folds on the shoulders when you raise your hands. < img data - img - size - val = "" src = "https://img.36krcdn.com/hsossms/20250617/v2_c1cd8cf451c242a5a57f4d43b7f02db4@000000_oswg41529oswg1080oswg478_img_000?x - oss - process = image/format,jpg/interlace,1" > Model the original clothes first, then "change the skin" through 3D rendering | Image source: Google Some "difficult areas" are given special treatment. The cross - attention mechanism will determine how long, flowing hair and clothes should connect naturally, like "which should be in front, the hair or the collar". For example, if the person in the base image is wearing a dark inner layer and wants to try on a semi - transparent sun protection jacket, the "dynamic layered mapping" technology will automatically adjust the visual hierarchy of the clothes to ensure that the sun protection jacket layer correctly covers the shirt layer visually, making sure the light, texture, and transparency are all in harmony. There's also light matching to make the lighting of the new clothes match the light source in the base image. However, this technology is still quite "picky" about clothes and is far from being a universal wardrobe. Currently, it's best at handling form - fitting, well - defined upper - body clothing, such as T - shirts, shirts, and sweaters. It's also better if the person in the base image stands straight facing the camera, keeps their hands still, and the lighting is stable. Only then can the AI build the model without any problems. If the clothes are too loose and complex or there are multiple layers of clothing, the AI will be confused. It also doesn't support accessories like hats, glasses, and shoes. This is understandable because it's complicated to handle the occlusion of these accessories. Dynamic clothing try - on is also not supported - you can't spin around, take a few steps, or try a sitting pose like in the dressing room of The Sims. < img data - img - size - val = "" src = "https://img.36krcdn.com/hsossms/20250617/v2_0751cb4dd40d40e698138e84bf74cf6c@000000_img_000?x - oss - process = image/format,jpg/interlace,1" > Did this technology appear in a 1995 movie? | Image source: Clueless < h2 > More Realistic Than Online Shopping, More Comfortable Than In - Store Shopping < / h2 > More importantly, the Try On feature may give rise to a whole "clothing experience chain". For example, when you try on a Cuban collar shirt, it might suggest a thin gold necklace, a pair of white sneakers, and a leather watch. That's how you get a "tropical OOTD". Looking further ahead, static clothing try - on could develop into dynamic try - on. You could put on a headset and enter a virtual dressing room, completely revolutionizing the online shopping experience. This is also good news for merchants. The return rate has always been a headache for clothing e - commerce. From the time a piece of clothing leaves the warehouse to the time it's returned, there are costs for logistics, packaging, and labor, and it also affects the turnover efficiency of the goods. Google also demonstrated a scenario. It shows that AI's transformation of e - commerce won't be limited to just clothing try - on. < img data - img - size - val = "" src = "https://img.36krcdn.com/hsossms/20250617/v2_012a8ff9d288457fab952c86e7abfcf0@000000_oswg458025oswg1080oswg641_img_000?x - oss - process = image/format,jpg/interlace,1" > Currently, when we shop online, we still have to type keywords like "cashmere carpet" or "minimalist living room carpet" in the search bar. In the future, you could write something like "I have a light gray sofa and want to buy a matching carpet to make the room look brighter. I have four kids, and we like to play around the sofa. But I have a cat at home and don't want it to scratch the carpet." With such specific requirements, the AI can find the most suitable products from thousands of items. < img data - img - size - val = "" src = "https://img.36krcdn.com/hsossms/20250617/v2_38b083be79e842dd9f621b159b3505bb@000000_oswg563567oswg1080oswg744_img_000?x - oss - process = image/format,jpg/interlace,1" > Left: Specific requirements from the user Right: Recommended products | Image source: Google The logic of the recommendation algorithm is to "guess what you want", while the ultimate goal of AI is to precisely filter and save people time and improve efficiency. After all the brainstorming, we have to face reality. There are already signs of misuse of this new feature. Some people have used the feature to put women's clothing on US Vice President J.D. Vance, turning a "redneck" into a "drag queen". Others have used photos of minors to put on revealing clothing. < img data - img - size - val = "" src = "https://img.36krcdn.com/hsossms/20250617/v2_9d1d4dab10ab402388e8dc74ff58acb7@000000_oswg460961oswg1080oswg666_img_000?x - oss - process = image/format,jpg/interlace,1" > This is what it looks like | The Atlantic

该文观点仅代表作者本人，36氪平台仅提供信息存储空间服务。

This AI virtual fitting function has cured me of impulsive consumption.