Human AI Interaction for Better Fashion and Textiles | AiDLab | PI Apparel

Hello, I'm Calvin. As the piece entitled said, what I would like to share with you today is our recently two developed and commercialized technologies and case studies, which is about the AI inspection on textile material and also how we can use AI to facilitate fashion designers creation process. And the key point is that how we can make use of our human creativity and knowledge to interact with AI, you know, to make AI which can really be practical, you know, adopted by the industry. Okay, before I share with you about, you know, the details, maybe allow me to spend around a couple of minutes to introduce, you know, what is AID Lab. AID Lab is jointly established by two universities, Hong Kong Polytechnic University and also the Royal College of Art London, and which is fully funded by a flagship research program of Hong Kong SAR Government, InnoHK Research Castles. There's a lot of sentences, you know, on this slide. In a word, the mission and vision of AID Lab is to conduct, you know, research projects integrating AI with related design sectors. For example, fashion and textiles. And one more important thing is that we really want to transform our research outcomes into practical use adopted by the industry. I would like to spend a minute, you know, to show you a short video about What is AIDLab? AIDLab is the first research platform to integrate AI with design. Working closely with our collaborators, we bring interdisciplinary innovations to shape a future where AI and design combine creatively to offer real-world applications for commercial and societal impact. The strategic partnership is between the Royal College of Art and Hong Kong Polytechnic University, building on many years of partnership collaboration, but taking this to a new formal level. To explore the ways in which artificial intelligence and design come together to solve some of the problems and challenges to improve human lives.

AID Lab conducts research focused on three main themes: ergonomic and inclusive design, innovation in product and service design, and intelligent fashion design and quality control.

We are situated at Hong Kong Science Park. Right now we have over 100 research associates and assistants. We also have around 40 PhD students to carry out, you know, over 20 research projects under these 3 thematic research focuses which you have come across in the video. This is our lab environment. It's very spectacular sea view because we are on the top floor of 19W building. We have various laboratories including Fashion AI and Intelligent Textiles in Fashion Lab, 4D Body Scanning Lab, 4D Foot Scanning Lab, Human Centered AI Design Lab, and also our supercomputer. GPU for training our AI model. Okay, having said that, there are two technologies which have been recently commercialized by AID Lab. The first one is AIDA, which is an AI-based interactive design assistant for fashion, and the second one is WiseEye, which is an AI-based interactive textile material inspection. Okay, let's deep dive the First one, which is WiseEye. We have just spun off a startup company called WiseEye Automation Limited to market this AI-based textile material inspection system. I think all of you must understand, you know, the ping pong of our fashion and textile industry, particularly about, you know, the material, fabric, accessories, garment inspection. Because I can't say 100%, but at least 99-point-something percent of, you know, the textile mills and apparel manufacturers are still relying on human visual inspection method, right? Of course, it is very unreliable. The norm is around 70% of inspection accuracy. It means that 30% of the defects, you know, will go to the downstream operators of garment manufacturing, which costs a lot of money, resources, we work even related to sustainability issue. And the efficiency reaches around, say, 12 to 15 meters per minute inspection speed. So that's why we rely on a huge amount of manpower to do inspection. Even for the textile mill and also the apparel factories, I mean, before delivering the fabric rolls to the apparel manufacturers, the textile mill must do the final inspection on the fabric rolls. And then after the apparel manufacturers receive the batch of fabric rolls delivered by the textile mill, they still do the same thing again. It's ridiculous, right? So how can we streamline this process? Whether we need to take out some inspection tasks. I'm not sure whether you have come across, you know, some AI textile inspection systems. If you visit, you know, the ITMA Asia, you know, TraceO, you can come across, you know, right, some, you know, AI inspection systems launched by the market already. I think even in China, I think around 10, this kind of companies. In Europe, we have 2 to 3. How about the performance? Right now we don't know. You know when my visit, you know these booths? You know all these companies told me, ah, I have collaborated with these brands, these factories. They have already adopted my systems. But at the end of the day, these companies come to my booth and told me the truth. Actually, you know, which is, I can't challenge or criticize their limitations, their weakness, but the point is we are now just at the very, very initial stage. I would say that AI inspection is a new industry, a new sector under our fashion and textile industries because we are still in the testing process. Why? What we have observed is that all these AI textile inspection companies or providers, you know, I can't 100% confirm, but most probably, you know, I predicted that they are relying on one AI model. What does it mean? AI model to inspect all different types of fabric structures, all types of defects. Even, you know, all types of inspection standards. You know, for different fashion brands, different textile mills, different, you know, apparel manufacturers, they use different, you know, definition. I mean, to inspect the defects. For the same defect on the same fabric, for fashion brand A, they will consider this is a major defect. For the others, they consider, ah, this is not a defect, right? How can we use just AI model, you know, to inspect everything? It's impossible. It doesn't work. So that's why right now this kind of inspection system is not popular in the market. And then you ask, so what's the difference between our technology with the others? So I would like to share with you a video about our recently developed AI inspection technology, which is called WiseEye.

In view of the numerous fabric structures that give great variations in fabric texture and defect types, automatic fabric defect detection has been a challenging and unaccomplished mission in the past two decades. Wiseye is a standalone intelligent system embedded with advanced AI technologies that carries out automatic inspection on textile materials. The intelligent inspection system revolutionizes conventional manual inspection in the textile and apparel industries. Its inspection accuracy is greater than 90%. It is efficient and provides high-speed inspection at up to 60 meters per minute. In addition, Wiseye helps to alleviate the shortage of highly skilled quality inspectors and minimize downstream wastage. Wiseye can inspect both finished woven and knitted fabric defects on real-time basis, including plain, twill, dobby, single jersey, double jersey structures, and more in different types of solid color and pattern, such as striped and checked, etc. With the integration of machine vision algorithms and deep learning techniques, AI technologies can also be used for other types of material, such as leather. Wiseye Inspection covers over 40 common fabric defects with a greater than 90% detection accuracy. The inspection results are more consistent and reliable. In addition to automatic and efficient inspection, Wiseye can also help in auto-labeling the defects. When the inspection is completed, it will provide an inspection report with a detailed defect location map. Wiseye is not only a technological breakthrough that meets industry needs, it also marks a significant milestone in the quality control automation for the traditional textile and apparel industry.

What you have seen, which is the fourth prototype, and this prototype I have just scrapped. Actually, research is an evolutionary process. We need to do some improvements. Anyway, when you come across the appearance of this system, this machine, actually it's quite similar with those you have visited in the booth in ITMA. I mean, the other competitors. But the key point is about the AI model. Okay, I would like to highlight some uniqueness and also the advantages of device-eye technology. The first one is the specialized AI inspection model. Actually, my team have started this project, I think, 10 years before. So in the past 10 years, on day 1, when we kick off this project, we develop the AI model based on textile material. Unlike the other companies, because for the other companies, I realize that they borrow the AI model from the other industries, like electronic printed circuit boards, something like that, which is very different. So we have a tailor-made model for textile materials. And we have a very large-scale real-world trained fabric defect dataset. Right now, we have already collected over 100,000 images for model training. All those are embedded with the fabric defects, with different fabric structures, whatever, the woven fabric or the knitted fabric. The second one is the most important part, which is the human-AI interaction. So this is a key topic I would like to tell you. Because just relying on human-interaction, we can train up different, you know, models which can cater for different types of fabric, for different types of inspection standards, and also for different, you know, inspection requirements from different brands and clients, which is a key point. So we are not talking about just relying on AI, just one AI model to detect everything is not feasible. Okay, we need to train up with different, you know, models to cater for different needs. So this is a key point. I would like to elaborate more about this in the next slide. And other one is the non-native learning capability. So through, you know, the human-AI interaction, you know, process, we can continuously improve the inspection performance by, you know, by improving, you know, the robustness of the AI model. And the last one is the advanced inspection system architecture, because which can, you know, harness the full potential of our larger AI inspection model in the future, which can also accommodate, you know, the huge number of parameter settings and also the image resolution. So this is the key thing, you know, our key technology. Certainly, I will not deep dive, you know the details, but when we kick off to inspect the fabric, what we need to do is that we must have a specialized base AI textile material model. So if we can recognize this type of fabric which we have already inspect, we can trigger a specific expert model to detect the defects on this particular type of fabric. So if we don't have this kind of model, what we can do is human-computer or human-machine interaction. You know, every time, once you know our model can detect the defects, you know our system or our machine can stop, and then through the verification of the human, they can decide whether this is a defect. Defect, or not a defect. And even they can also adjust the defect types. For all this information, all this input which can feedback to the AI model, it means that we can create another expert model for this particular type of fabric. So through this non-stopping process, we can train up different type of expert systems, expert model. To cater for different types of fabric until we can achieve the fully automatic inspection process. So through the human-computer or human-machine interaction process, the frequency of this interaction will be gradually reduced until we can achieve full automation of inspection. So this is the process. This kind of inspection architecture we have already built in, in our system. So because when we discuss with quite a lot of apparel or the textile manufacturers, they presume that, OK, once I buy your machine, AI inspection machine, just plug in, switch on the engine or the system or the machine, we can do everything. And then once they find that, okay, for this type of fabric or this type of defect, the system or machine can't find it and they fail. This machine is not good. That's it. It's not the case. We should not treat the AI inspection system as this kind of implementation. Okay. I will say that no AI inspection system can fulfill all types of requirements. And even for those manufacturers, if they are interested in implementing this kind of system, which is a kind of learning process, evolutionary process, it takes time. So the Wiseye solution, which is ease of use, online learning, and also customized inspection capability, which I have mentioned before. And I would like to share with you some, you know, detection results which are the results for the knitted fabric. And even we can also classify the types of fabrics, not only to tell you the location of the defects, we can also classify the types of defects, which is important, you know, particularly for those textile mills because which can facilitate, you know, the investigation on the root causes of the problem. Quality problem. And then they can maybe investigate about the yarn quality, or even the machine setting of their weaving machine or the knitting machine, which is important for them to improve the fabric quality. Roven fabric, same as knitted fabric. So the features of Wiseye technology include we can automatically detect the defects. We can classify the types of defects. We can grade the defects based on whatever the 4-point or the 10-point systems. We can generate the inspection report. And also, we can record all those defect locations for further analysis. And recently, we are now developing another hardware part based on WiSeye, which is called WiSpar, which is an add-on inspection module, which is actually a retrofit, which can be attached on the existing traditional fabric inspection machine, circular knitting machine, and also the warp knitting machine. Because believe we know for those textile mills, they consider that once we can identify the defects during the fabric manufacturing process, they can stop the production to avoid the production of substandard fabric. Okay, this is what I would like to share with you about the first case, which is Wiseye, and the second one is AIDA, which is AI-based interactive design assistant for fashion system. And AID Lab has already spun off another company which is called Code Grey to market Ada. And today we also have our CEO Kim Wong, Miss Kim Wong, to join this seminar. Okay, I think in the past 1 to 2 years, I think most of us have already tried ChatGPT, Midjourney, using a lot of even experience, different types of the generative AI fashion apps or system. Am I right? Can you sense their strength and also the limitations? So all of us are talking about generative AI, but at the same time, the core asset of fashion designers is human creativity. How we can combine them together? So that's why we are talking about human AI. Interaction again. So on this side, can you see this very stunning outfit? Very stunning design, which actually is the design of Jae Lim, which is a creative director of Korean fashion designer NAEBOL. And Jae, you know, has perfectly— adopt Ada to create this outfit. Okay, before that, before I introduce, you know, the technology of Ada, so on this slide I have a question. On this slide there are 6 outfits, right? Can you identify the difference or similarity? The same thing is that, you know, all these are the One-piece dress in black color, right? What are the differences? Take the second one, about the collar and also the cuff, which are in white color, right? So I would like to pinpoint that when we are talking about AI model for fashion, we must have a model, AI model, which can differentiate, I mean, these 6 alphabets. So this is the core thing, the fundamental thing. So that's why it's necessary to train up an AI model which can understand faction. So that's why our research team, you know, has spent a lot of years, you know, to develop a faction AI dataset including menswear, womenswear, and we are now expanding the dataset to children's wear. And so in this dataset, we consider about, you know, over 250 fashion attributes across, you know, all the womenswear and menswear main product category, and also the AI key points. AI key points is very important because We need to train up an AI model who can locate the position of the component of a fashion image. Like, you know, the necktie, shoulder points, the hems, right? So we need to teach the AI model to understand the fashion attributes, which means design elements, and also the position of the component of an apparel. So we have we have collected 150,000 fashion images for annotation labeling. So for the women's wear, which consists of bras, outerwear, skirts, trousers, dresses, jumpsuits, for the men's wear, we cover, you know, the outerwear including light jacket, blazer, pants, something like that. So on the basis of this, you know, faction AI dataset, we have trained up, you know, an AI model which can understand faction. It's very important because when we rely on AI, it means that we must trust the AI model, the quality of AI model. The AI model must understand faction. Such that designers can work with AI, interact with AI to create fashion. So this is the flow of human-AI interaction of our Ada system. On the left-hand side, you can see the mood board, pins board, the sketch board, right? All these come from the fashion designers. Having said that, the core asset of fashion designers is human creativity. So that's why when the designers adopt our system, They still need to do market survey to prepare their mood board, to forecast the fashion trend, and even to consider about the DNA of their brands, the design elements. They still need to do this. There's no difference between using AI and the conventional design process. Instead of just test the prompt and then I want something like this, and then the Midjourney or some app generate something for you, the print or the new fashion for you, I'm sure that it may not really meet the actual requirements of fashion designers. And even in a large extent, it seems like in the field, you know, the conventional design process. So that's why for the input, it's still the designer creativity. And then all these images of Mood Board will input to our ADA system. So that's why for our ADA system, which is composed of different types of AI models including style recognition, pinch generation, attribute recognition, and the backbone of course including Stable Diffusion, you know, UNet, ResNet, different types of AI models we have adopted. But this model, we do different types of things. The first thing is to extract the key features, the design elements of the mood board. Another thing is based on these new features the model can understand, we can recommend— not we, the model, the system— can recommend similar things for the designers for further inspiration and brainstorming. So this is the whole design process in the middle. Before lunch, you know, some speakers said that, you know, it's quite difficult, you know, to manage the AI model to do some editing, right? But, you know, in our Ada system, designers, you know, can manage, can control what they want. Even, you know, The Ada system can generate the designs for the fashion designers. They can still use their human creativity to edit or modify some certain parts of the designs. So this is the key difference between the Ada and the others generative AI systems and apps. And certainly the right-hand side is our output. So I would like to show you a demo video which is around 5 minutes, right? You can sense about, you know, the key features and also the operation of the Ada system. And actually, you know, we have already added more new features, but this one is not the most updated one. Okay, anyway, you know, you can sense, you know, the operation of Ada.

Ada is the world's first designer-led AI system. To log in, type your username and password. After that, type your registered email address. You will then receive a one-time verification code from that email. Type the code to log in. You can kickstart by uploading the mood board images from your computer or select from your library. Alternatively, you can create new mood board images by entering the descriptive prompts and click Generate button. We will generate 4 images for you to choose from. Zoom in to examine the image details. Click the heart icon to save it to your personal library. You can design the mood board freely. We offer a variety of layer options for you to choose from. Each layer can be modified independently. You can stretch, rotate, and apply bevels to it. You can also drag them to anywhere. The image selected with the mouse will be placed on the top layer Once you're happy with your mood board, click Submit to save it. You can upload print from your device. Ada also enables you generate new print. By default, image-only mode creates prints based on your uploads. There are 3 styles of generation to choose from. 4 new prints are generated for you to choose from. Select the print and add it to your print board. For a text-only generation. Simply enter your desired print features in the prompt box. The third approach combines image with text to generate prints. Let's try out the illustration style this time. Pin prints will be prioritized in your designs. Then you can choose your preferred colors. Main colors can be extracted from an uploaded image for you to choose from. You can also pick colors from the color palette or key in the TCX color code. Next step is to choose the sketches. You can upload or select from your sketch library. You can also generate new sketches. Type in the desired prompt and click Generate. 4 new sketches will be generated. Select the appropriate clothing category for these sketches. After you press the Design button, a collection of 8 outfits will be generated in around 8 seconds. Click little red heart to save your favorite design. Click on the result to edit its details. You can change its sketch, print, and color. To add a print to your design, choose your preferred print and click the layout. The overall mode overlay print across entire piece, and you can adjust its scale. Single mode allowing placing individual prints at desired locations. You can adjust the position, size, and rotation of prints as desired. Click the cross sign to add garments. Upload or select sketch from your library. Select the appropriate category for the uploaded sketch. Its print and color are also editable. Click here to enter the canvas page. Here you can freely adjust the size and placement of each item to meet your design specifications. Click this button to compare changes with initial result. You can also undo or redo any action. Click the Redesign if you would like to look for more design options. Ada will then generate another 8 new designs for your collection. After finalizing your design, use the Export to download collection. Here shows the exported collection. If you wish to continue with your previous work, you can find your saved collection in the History. You can Rename the collection to any name you prefer. Simply click Retrieve to resume the last working state. All selections remain editable. Meet your personal fashion assistant powered by large language model. You can ask it questions for expert advice and insights. Looking for sketches? Just ask, and it will retrieve them from the database for you. Enhance your search accuracy by adding more attributes. Light up the little red heart to save your preferred sketches to the library. Our assistant also can access the internet for gathering information as well. New to our platform? No problem! Simply ask the assistant for a tutorial and you'll be guided through a comprehensive introduction.

So when we compare, you know, Ada system with the other existing, you know, 3D fashion pattern making or the, you know, the animation softwares, we just put a focus on inspiration part. Okay, we are not the expert in 3D rendering. I'm sorry, they are doing They are doing very well, right? So what we've put the focus is just on the inspiration and creativity process. The main purpose is just one, you know, to spark designers' inspiration and creativity. And we do hope that designers, after they use our system, they can think more about their design, and then they can fine-tune it and also confirm it before prototyping. So this is our main purpose. We have already launched this system, which is a cloud-based system. It means that designers over the world just need to pay the subscription fee, which is very, very reasonable. I can't say cheap, but very reasonable and affordable. Even we have a trial a few days far, 5 days far, you can test it. If you are satisfied with the Ada features, you can subscribe. Okay, time is a bit running out. Anyway, so this 4 outfits, having said that, Jay, Korean designer label, has perfectly used our system. He co-created with Ada to develop this outfit. These are the outcomes of the outfits, use our Ada system. It has been showcased in the New York Fashion Week. And just a very short comment from Jay about Ada. Those keywords that come to my mind when describing the system are safe, fast, friendly, smart. And futuristic. That's how I felt when I was using them. I want to stress on the word futuristic and it's hassle-free because like if you think of future, like things should get really, really easy to put your hands on. And with this system, like things were really, really straightforward and really easy to use. And that is the main reason why I use the word futuristic. Jai mentioned that it's very user-friendly, it's easy to use. I would say that I think less than half an hour you can make use of Ada, just less than half an hour. Okay, these are the niche advantages, just take a quick look. User control, you can sense that, you know, designer can control all the process when using Ada. Integrating with, you know, the existing workflows. We don't want, you know, to interfere, you know, the conventional design process. Reduce time to market. For those designers, when they adopt Ada system, they commented that over 70% of their design and development time can be saved. 70%. Originality and brand identity, because, you know, the designers still need to upload their mood board, you know, to our system, and then we can generate something for them. So that's why the originality and brand identity can also be kept. By the fashion brand and fashion designers. Cohesive collection creation, which is another, you know, niche different, you know, from the other generative AI system. For the other AI system, you know, each time, you know, they can only generate an individual design. But for us, you can see that we can generate a collection for the designers. And the last one is sustainability. So designers, you know, can create, edit, and confirm the best design before prototyping. So if you are interested in trying Ada, so you can go to our website, CodeCreate.com, and for 5 days trial, right? So yes, if you would like to know more about AID Lab, you want to please scan our QR code such that you can understand more and keep update with all our recent events and also the technologies. So this is what I would like to share with you today. Thank you so much. Time is running out. So any questions about my talk? Is output from Ada ready for production? Sorry, that is just for inspiration, for creativity. Because every time, particularly for those fashion brands, they ask the same question. Whether our system can be input to Brownswear, Karl Lagerfeld 3D, Style3D. I would say no. Strictly speaking, no, because we just put a focus on the inspirational part. If we can integrate them together, it will be ideal. And one more thing is that we are also developing to transform our 2D images to 3D animations. We are now doing on this. I do hope that in a later version, you can see the 3D appearance. Of the outfit. Okay, what's the IP for Ada? We are not managing the IP issues because we are just a platform. The IP things should be burdened by the designers ourselves because we can't control what they have uploaded, right? But for those generative schedules and prints which are owned by Ada system. Can you also create store track reports? This is a very good question because time is limited because originally I plan to introduce another technology which is called MISI. Okay, which is another system which we can do the auto tagging on those fashion images or the products uploaded by the fashion brands and retailers. And because we have also developed a system which can manage the stock and also even provide mix and match recommendations automatically to the sales assistant. And then they can present to the customers. So this is another technology we have already developed. Yes, it's a very good question. I think it is unavoidable to stop AI, right? Nobody can stop it. But how we can embrace AI technologies into the conventional process, even fashion design process. I think how you can treat AI in your process, in your operation. So this is the reason why ADASystem, the full name is AI-based Interactive Design Assistant for Fashion, which is an assistant to help you, to support you in the design process. Okay, so you treat it as a tool. Okay, I think we can embrace AI technologies in the design process. Yes, any more questions? Okay, thank you very much.