StyleMe: Towards Intelligent Fashion Generation with Designer Style

Title of the Paper

StyleMe: Towards Intelligent Fashion Generation with Designer Style

Paper Information

• Subject Area: Fashion Design and Artificial Intelligence
• Keywords: Fashion Design, Hand-drawn Sketch Generation, Sketch Coloring, Generative Adversarial Networks, Generative AI Tools

Research Background and Issues

• Research Background: Fashion design is a form of expressing individuality and style. Traditional fashion design relies heavily on designers' creativity, which is subjective and uncertain. With the development of big data and artificial intelligence, it has become possible to use deep learning and AI tools to assist in design.
• Problems and Challenges:
  • Drawing hand-drawn sketches and coloring them are among the most tedious yet essential steps in fashion design.
  • Existing automated sketch generation methods primarily extract image outlines, failing to capture the details and stylistic variations required for hand-drawn sketches.
  • After generating sketches, quickly applying harmonious and aesthetically pleasing colors remains a significant challenge.
• Importance: This research aims to enhance the efficiency of fashion design through technological means and assist designers in creating personalized designs, reducing repetitive tasks, and better integrating designers' styles.
• Research Motivation: There is a lack of tools capable of simultaneously generating personalized sketches and transforming sketches into stylized fashion images.

Solution

• Method or Solution:
  • Proposed an intelligent design system called StyleMe, which can generate fashion sketches with designer styles and achieve style transfer from sketches to fashion images.
• Innovations:
  • Introduced two Generative Adversarial Networks (GANs) for sketch generation and style transfer.
  • Incorporated Channel Attention Module (CAM) and Adaptive Layer Instance Normalization Module (AdaLIN) to improve the quality of generated images.
  • Integrated a user interface allowing users to browse, select, and edit generated sketches, enhancing interactivity.
• Implementation Steps and Key Technologies:
  1. Sketch Generation Model:
     • Utilized a unidirectional GAN model to extract content and style features from fashion images via an encoder.
     • Applied the Channel Attention Module to refine sketch details, ensuring consistency with the designer's hand-drawn style.
     • Optimized sketch quality through loss functions, making them closer to the designer's style.
  2. Style Transfer Model:
     • Used a dual-encoder architecture to separately extract content features from input sketches and style features from reference images.
     • Employed a decoder to generate new fashion images with specified styles while retaining the content features of the sketches.
     • Combined adaptive normalization layers to ensure color harmony and style consistency in the generated images.
  3. User Interface:
     • Provided flexible online functionalities such as generating coarse/fine line sketches, sketch editing, style specification, and downloading.

Research Outcomes

• Specific Outcomes:
  • Achieved automatic generation of personalized fashion sketches, with sketch quality comparable to those drawn by human designers.
  • Generated diverse, stylistically rich, and harmoniously colored fashion images based on reference styles.
• Comparative Advantages:
  • Outperformed other methods in quantitative metrics such as Frechet Inception Distance (FID) and Learned Perceptual Image Patch Similarity (LPIPS).
  • In user studies, StyleMe's generated designs were widely recognized as indistinguishable from human designers' works.
• Experimental Results:
  • Improved design efficiency in terms of time and output quality.
  • Multiple user tests demonstrated that StyleMe helps designers quickly complete design tasks while providing diverse design options.
• Limitations and Future Directions:
  • Currently requires a substantial amount of designer sketch data for model training; future optimization could focus on few-shot learning.
  • May fail to generate images for fashion items with complex backgrounds or highly detailed designs.
  • Future improvements include enhancing support for local details, expanding the library of diverse design elements, and enabling cross-device collaborative design functionalities.
  • Optimizing evaluation systems by introducing objective quantitative standards to assess the aesthetic quality and professionalism of generated results.