The rapid evolution of generative adversarial networks (GANs) has significantly impacted various domains, including fashion, where design typicality is a critical factor in consumer preference. This thesis presents a novel framework for fine-grained, disentangled manipulation of visual attributes in fashion images using generative adversarial networks. It leverages a StyleGAN2-Ada model trained on a custom dataset of over 14,000 dresses from Zalando—with transfer learning from FFHQ—to create high-fidelity images even from limited data. The work systematically evaluates GAN inversion techniques, including encoder-based (e4e) and hybrid methods (PTI, ReStyle, HyperStyle), to accurately reconstruct real images into a manipulable latent space. Two complementary typicality measures are developed: one using pre-trained DINOv2 embeddings and another employing a disentangled embedding model that isolates key aesthetic attributes such as color, fabric, and fit via distance correlation loss. By applying the InterFaceGAN method, the thesis achieves controlled latent space manipulations that subtly adjust design typicality while preserving overall image identity. Comprehensive evaluation using metrics like FID, IS, and LPIPS demonstrates the potential of this approach to enhance fashion design by aligning products more closely with consumer preferences.
To showcase the fine-tuned generator model, a toy example website was built in hommage to the website based on the original StyleGAN2-Ada implementation. The website can be accessed here.
Curated from Zalando's extensive dress collection, this dataset comprises over 14,000 high-quality images of women's
dresses paired with rich article metadata. Only standardized, centered packshot images were retained and processed to a
uniform 1024×1024 resolution. The metadata focuses on key attributes such as brand, price, category, fabric, fit,
neckline, pattern, collar, length, shape, and sleeve length, ensuring reliable and clean labels.
In addition, color information was automatically extracted using a pre-trained CLIP model based on a fixed palette of 14
colors. Representing 643 unique brands, this dataset provides a diverse yet focused foundation for training advanced
generative models and performing fine-grained, disentangled visual attribute manipulation.
The method presents a modular pipeline that integrates high-fidelity image generation, robust inversion of real images into the latent space, quantitative assessment of design typicality, and targeted latent space manipulation. This workflow enables precise, disentangled editing of fashion images while preserving the core identity of the design.
In the generation stage, the approach leverages the StyleGAN2-Ada architecture to synthesize realistic dress images. A transfer learning strategy from the FFHQ dataset, combined with data augmentations such as x-flip, ensures that the model captures the nuanced aesthetics of the fashion domain. Evaluation metrics like FID, IS, and LPIPS guide the selection of the optimal model checkpoint, balancing image quality and diversity.
The inversion phase maps real images into the latent space using state-of-the-art methods. An encoder-based model (e4e)
serves as the baseline, while hybrid approaches such as PTI, ReStyle, and HyperStyle are also explored. This step is
essential to accurately reconstruct images and maintain latent code editability for subsequent manipulation.
Typicality measures are computed using two complementary strategies. One approach employs pre-trained DINOv2 embeddings
to compare individual dress features to a composite morph, capturing an overall sense of typicality. The other approach
refines this assessment by learning disentangled sub-embeddings for key attributes—such as color, fabric, and fit—using
distance correlation losses, which together provide a robust and nuanced quantification of design typicality.
For latent space manipulation, the method utilizes InterFaceGAN to identify linear semantic boundaries corresponding to design attributes. By traversing these boundaries, targeted adjustments can be made that subtly modify the typicality of a design without disrupting its overall structure. This controlled editing mechanism allows for fine-grained modifications that align the aesthetic attributes of fashion items with consumer preferences.
The results demonstrate that the proposed framework successfully integrates high-quality image generation, robust inversion, and precise attribute manipulation into a unified pipeline. The experiments confirm that the approach achieves both visually realistic outputs and controlled, fine-grained edits aligned with design typicality.
In generator training, the StyleGAN2-Ada model, fine-tuned on the curated Zalando dataset, produces high-fidelity dress images that closely resemble real-world samples. Quantitative metrics—such as FID, IS, and LPIPS—validate the quality and diversity of the generated images, while visual samples further underscore the model's ability to capture nuanced fashion details.
Inversion experiments reveal that the selected methods, including the encoder-based e4e and hybrid approaches like PTI, ReStyle, and HyperStyle, reliably map real images into a latent space suitable for subsequent manipulation. The results highlight a balance between reconstruction accuracy and editability, ensuring that the latent codes maintain semantic consistency while remaining amenable to targeted edits.
The typicality measures, computed using both pre-trained DINOv2 embeddings and a custom disentangled embedding model, effectively capture the aesthetic attributes of each design. These measures provide a robust quantification of design typicality, correlating well with perceptual evaluations and offering a solid foundation for supervised manipulation.
Finally, manipulation experiments using InterFaceGAN demonstrate that controlled latent space navigation can subtly adjust the typicality of fashion images. By identifying and traversing semantic boundaries, the method achieves targeted modifications that enhance or diminish design typicality while preserving the overall identity of the garment.