Reproduce “FitGAN: Fit- and Shape-Realistic Generative Adversarial Networks for Fashion” on OpenResearch

About this paper

Sonia Pecenakova, Nour Karessli, Reza Shirvany

Zalando SE developed FitGAN, a generative adversarial network that synthesizes photo-realistic fashion images with explicit control over garment fit and shape, moving beyond general visual appeal to ensure physical realism. This framework enables scalable generation of diverse fit visualizations, validated by domain experts and automated classifiers.

Key insights

• The model implicitly learns data entanglements present in the training data (e.g., oversized garments associated with lighter colors), necessitating robust conditioning methods.
• The semi-supervised conditioning approach effectively addresses these entanglements, providing finer control over implicit attributes and improving disentanglement in generated outputs.
• The learned latent space implicitly encodes the ordinality of fit classes (e.g., "Bodycon" is semantically closer to "Skinny" than to "Oversized"), demonstrating a meaningful organization of attributes.

Summary by alphaXiv.