BC-GAN: A Generative Adversarial Network for Synthesizing a Batch of Collocated Clothing

TL;DR

BC-GAN tackles batch collocated clothing synthesis by leveraging a pre-trained StyleGAN and two discriminators: a style-embedding discriminator for latent-space realism and a contrastive compatibility discriminator for fashion compatibility. It formalizes two-way translation between upper and lower garments, encodes inputs into the latent W-space, and optimizes with adversarial, diversity, and compatibility losses, validated on the DiverseOutfits dataset with improvements in LPIPS, FID, and F^2BT over baselines. The method enables designers and fashion platforms to generate multiple diverse, compatible outfits from a single item, with potential extensions to diffusion models and multi-category fashion synthesis. This work advances practical fashion intelligence by combining batch generation, latent-space manipulation, and contrastive supervision to achieve high diversity, realism, and compatibility in generated outfits.

Abstract

Collocated clothing synthesis using generative networks has become an emerging topic in the field of fashion intelligence, as it has significant potential economic value to increase revenue in the fashion industry. In previous studies, several works have attempted to synthesize visually-collocated clothing based on a given clothing item using generative adversarial networks (GANs) with promising results. These works, however, can only accomplish the synthesis of one collocated clothing item each time. Nevertheless, users may require different clothing items to meet their multiple choices due to their personal tastes and different dressing scenarios. To address this limitation, we introduce a novel batch clothing generation framework, named BC-GAN, which is able to synthesize multiple visually-collocated clothing images simultaneously. In particular, to further improve the fashion compatibility of synthetic results, BC-GAN proposes a new fashion compatibility discriminator in a contrastive learning perspective by fully exploiting the collocation relationship among all clothing items. Our model was examined in a large-scale dataset with compatible outfits constructed by ourselves. Extensive experiment results confirmed the effectiveness of our proposed BC-GAN in comparison to state-of-the-art methods in terms of diversity, visual authenticity, and fashion compatibility.

Figures (8)

• Figure 1: Multiple visually-collocated fashion items synthesized by our BC-GAN.
• Figure 2: Comparable frameworks of the StyleGAN-based models: (a) StyleGAN karras2019style, (b) GAN inversion for edit framework zhu2020domain, and (c) BC-GAN (ours). Here, the pre-trained model contains mapping network $f$ and synthesis network $g$.
• Figure 3: Key components of our proposed BC-GAN: (a) The generator $G$ which translates an input image $\mathbf{x}$ and a randomly sampled latent code $\mathbf{z}_i$ into an output image $\mathbf{y}_i$, (b) the style embedding discriminator $D_{se}$ which distinguishes between real and fake style embeddings, and (c) the compatibility discriminator $D_{cmp}$ which provides fashion compatibility supervision for synthesized clothing images in a contrastive learning perspective.
• Figure 4: Statistics of our DiverseOutfits dataset: (a) an illustration of multiple possible matching outfits, (b) statistics of matching pairs (for each distinct item of upper clothing), and (c) statistics of matching pairs (for each distinct item of lower clothing).
• Figure 5: Comparisons between our BC-GAN and other multimodal I2I translation methods, which are MUNIT huang2018multimodal, DRIT lee2018diverse, DRIT++ lee2020drit++, StarGAN-v2 choi2020starganv2, and SAVI2I mao2022continuous, in terms of (a) translation on upper $\rightarrow$ lower setting, and (b) translation on lower $\rightarrow$ upper setting.