Mannequin2Real: A Two-Stage Generation Framework for Transforming Mannequin Images Into Photorealistic Model Images for Clothing Display

The rapid development of e-commerce has significantly influenced consumer behavior, and online clothing purchases have been increasing. To effectively showcase clothing items to consumers, merchants often require high-quality fashion display images, which can be acquired by hiring human models for photography for a high cost. Leveraging the power of generative models, this study develops an automated generation framework called Mannequin2Real to translate mannequin images into photorealistic model images for fashion display purposes. The designed framework comprises two stages: model head generation and skin generation. In the head generation stage, the relevant features of the model head regions are first extracted and used as inputs to the head generation network, which is responsible for synthesizing a photorealistic head image. Subsequently, in the skin generation stage, the skin mask and pose features of a model body image are extracted and fed into the skin generation network, accomplishing the generation of photorealistic skin. Finally, the synthesized head region and skin region are combined to produce a photorealistic model image. To examine the effectiveness of our developed Mannequin2Real model, we first evaluated it on a high-resolution virtual try-on dataset. In addition, we constructed a dataset of images of mannequins captured in real-world scenarios. The experimental results demonstrate the effectiveness of our approach compared to other image generation algorithms.