Would Your Clothes Look Good on Me? Towards Transferring Clothing Styles with Adaptive Instance Normalization

Area of Science:

• Computer Science
• Artificial Intelligence
• Fashion Technology

Background:

• Deep learning applications in fashion, including image synthesis, are gaining traction.
• Generating realistic clothing images is a complex, unsolved challenge with significant creative potential for designers.

Purpose of the Study:

• To address the challenge of style transfer for clothing images between different individuals.
• To adapt and enhance the StarGANv2 architecture for fashion-specific style transfer applications.

Main Methods:

• Modified the StarGANv2 architecture to eliminate the need for distinct domain separation.
• Incorporated a perceptual loss function to improve the accuracy of style transfer between target and source clothing.
• Edited the style encoder to better capture and represent the stylistic features of target clothing.

Main Results:

• Demonstrated the efficacy of the proposed method through qualitative and quantitative experiments on the DeepFashion2 dataset.
• Validated the novelty of the adapted architecture for fashion image style transfer.
• Achieved successful style transfer of clothing between different individuals, enhancing creative possibilities.

Conclusions:

• The adapted StarGANv2 architecture offers a promising solution for clothing style transfer in the fashion industry.
• The method enhances designer creativity by enabling novel fashion image generation.
• Further research can explore advanced style manipulation and personalization in fashion AI.