Generative AI 는 수많은 양의 데이터로 부터 패턴을 학습하고, 그것을 바탕으로 한번도 본적이 없는 새로운 데이터를 생성해내는 AI 기술을 말합니다. 이 세상에는 존재하지 않는 사람의 얼굴을 생성할 수도 있고, 특정 작가의 스타일을 가지는 기사를 작문할 수도 있고 , 또는 어떤 배경이나 상황을 설명하는 새로운 이미지를 생성할 수도 있습니다. 최근 ChatGPT , Stable Diffusion , Imagen 등을 필두로 이러한 생성 AI가 큰 화제가 되고 있는데요. 과연 우리가 지금 보고 있는 ChatGPT와 같은 생성 AI를 패션 산업에 어떻게 활용할 수 있을까요? 이 질문에 대한 답을 찾아보기 위해 ChatGPT에게 질문을 해보았습니다.

처음 던진 질문은 아래와 같습니다.

💡
Q: How can we utilize ChatGPT in fashion industry?

답변

  • Generating product descriptions: ChatGPT can be trained on a dataset of fashion product descriptions, and then used to generate new, unique descriptions for new products.
  • Personal styling: ChatGPT can be used to provide personalized styling recommendations to customers based on their preferences and body type.
  • Customer service: ChatGPT can be integrated into a company's customer service chatbot to provide assistance to customers with their orders, returns, and other inquiries.
  • Social media content creation: ChatGPT can be used to generate engaging and relevant content for a fashion brand's social media accounts.
  • Trend forecasting: ChatGPT can be used to analyze large amounts of data on fashion trends and predict future trends.
  • Virtual try-on: ChatGPT can be used to generate realistic virtual images of clothes on avatars of different body types, so that customers can see how a piece of clothing would look on them before buying it.
  • Personalized marketing: ChatGPT can be used to generate highly personalized marketing messages for customers based on their browsing and purchase history.

ChatGPT의 본질은 대화형 시스템 기반의 텍스트 생성 모델이라는 점에서 상품 설명 생성 (product description), 고객 서비스를 위한 챗봇, SNS 컨텐츠 생성 등은 자연스러운 방향으로 보여지는데요. 이를 검증하기 위해 상품 설명을 생성하는 task를 시켜보았습니다. 이미지 속의 패션 속성을 자동으로 태깅해주는 옴니어스 태거 (reference: https://omnicommerce.ai/) 를 활용하여 샘플 이미지를 태깅한 결과를 ChatGPT의 입력으로 넣고, 이에 대한 상품 설명을 생성해보았습니다.

OMNIOUS TAGGER 결과 예시
💡
Q: Here is the fashion attribute detail of a coat, half length, long sleeve, tailored collar, normal fit, grey color, check print, wool / cashmere texture, buttoned detail. Can you generate product description based on this input?

답변

Sure, here's a possible product description based on the input you provided:

"Stay warm and stylish with our new half-length coat. Tailored to perfection, it features a tailored collar, long sleeves and a normal fit, designed to flatter your figure. The grey color and check print add a touch of sophistication to this wool and cashmere blend coat. The buttoned detail completes the look, making it perfect for both casual and formal occasions. This coat is perfect for the colder months and will keep you looking stylish and feeling warm all season long."

Please note that this is just one example of a product description that could be generated based on the input provided. Depending on the specific model and dataset used, there could be different variations of the product description. Additionally, the generated text should be proofread and edited by human editors before they are published, as language models are not able to understand the context and the meaning of the text they generate.

신기한점은, 위 이미지에 대한 스타일 정보를 제외하고 입력으로 넣었지만  making it perfect for both casual and formal occasions 와 같이 스타일 정보를 유추해 냈다는 점입니다. 이렇듯 ChatGPT를 활용하면 우리가 가지고 있는 데이터 또는 AI 모델을 활용하여 더 자연스럽고 더 풍부한 정보를 생성하는 것이 가능합니다.

하지만, ChatGPT는 텍스트 생성 모델이기 때문에, virtual try-on에 어떻게 적용이 가능할지는 자연스럽게 그려지지 않습니다. 그래서 질문을 좀 더 이어 나가보았습니다.

💡
Q: How can we apply ChatGPT in virtual try-on? Isn't ChatGPT is language model, which generates text given text, but how can we integrate it with virtual try-on where we should generate images?

답변

ChatGPT is primarily a language generation model, so its direct application to generating images for virtual try-on may be limited. However, it can still be used in conjunction with other technologies to create a virtual try-on experience.

One way to use ChatGPT for virtual try-on is to generate text-based descriptions of how an item of clothing or accessory might look on a person. For example, ChatGPT could be trained on images and descriptions of clothing and accessories, and then be used to generate text such as "this dress would look great on someone with an hourglass figure," or "these earrings would add a pop of color to any outfit." This text-based description can be used to give customers an idea of how an item might look on them, without having to generate an image.

Another way to use ChatGPT in virtual try-on is to use it as a part of a chatbot that can assist customers in finding the right clothing and accessories for them, by answering their questions and making suggestions based on their preferences and body type.

Additionally, GPT-3 can be integrated with computer vision models to generate images. This can be done by fine-tuning GPT-3 on the text-image pairs and with the help of GPT-3, a computer vision model could be able to generate images of the similar style or pattern on a given text input.

It's worth noting that, although ChatGPT can be integrated with other technologies for virtual try-on, it would likely be just one part of a larger system that would also involve computer vision and other technologies to generate images.

위 답변으로 부터 볼 수 있는 ChatGPT의 힘은 단순히 주어진 문장 또는 질문에 대한 그럴듯한 답변을 생성해내는 것에 그치지 않고, 다양한 다른 분야의 AI 기술에 쉽게 접목 가능하다는 것입니다. ChatGPT는 언어라는 공통의 해석 가능한 커뮤니케이션 도구를 통해 다양한 분야로의 컨텍스트 변환 (context switching)이 가능하고, 이를 통해 우리는 지금까지 활발히 연구가 되어온 다른 도메인의 AI 기술과 융합할 수 있습니다. 그것이 위의 예시에서 보여준 Computer vision 과 Graphics 기술이 용합된 Virtual try-on 분야일 수도, 개인화 추천 시스템일수도 있고, 그 가능성은 무궁무진 합니다. 우리가 가지고있는 컨텍스트를 어떻게 ChatGPT에게 적절한 prompt로 전달하느냐에 따라서 우리는 ChatGPT를 다양한 분야로 확장하는 도구로 활용할 수 있습니다.

물론, 아직은 ChatGPT와 같은 생성 AI가 널리 활용되는 것에 회의적인 시선도 많은데요. 실제로 존재하지 않는 논문을 마치 어떤 학회에 발표된 것처럼 참조 한다거나, 어떤 것이 진실인지에 대한 증거 없이 그럴듯한 답변을 생성해내는 것은 사회적, 도덕적 문제를 야기할 수도 있습니다. 최근 Yann Lecun은 LLM (Large Language Model)이 writing assistant로는 훌륭하지만, 검색 엔진 등으로 확장하기에는 무리가 있다며 RLHF (Reinforcement Learning with Human Feedbacks) 로 학습한 ChatGPT의 근본적인 한계에 대한 우려를 보이기도 했습니다. 또한, AI 분야의 저명한 학회인 International Conference on Machine Learning (ICML) 에서는 ChatGPT와 같은 LLM으로 작성된 text가 실험 결과가 아닌 다른 부분에 포함되는 것을 금지한다는 발표를 하기도 했습니다.

오늘 글에서는 최근 화제가 되고 있는 generative AI 인 ChatGPT가 어떻게 패션 산업에 활용될 수 있을까 라는 주제를 통해 그에 대한 시사점에 대해 알아 보았는데요. 최근 ChatGPT 를 접목한 어플리케이션들이 이미 많이 나오고 있고 (vscode, chatGPT-for-google, youtube video summary 등) , Microsoft에서는 LLM을 더 잘 활용하기 위한 LMOps라는 개념에 대해 발표하기도 했습니다. Microsoft Azure에서도 ChatGPT를 활용한 서비스 도입에 대한 발표와 더불어 최근 검색 엔진인 Bing에서도 공식적으로 ChatGPT가 Integration되었다는 소식입니다. 하루가 다르게 변화하는 지금, 우리는 ChatGPT를 어떻게 받아들이는 것이 좋을까요? 많은 우려들이 있지만 분명한 것은, 우리가 어떻게 활용하느냐에 따라 그 가능성은 열려 있다는 점입니다.

, Tech