WardrobeWizard 🧙‍♂️👔

Implementation of cycle GAN-based model capable transforming a person’s clothing style based on user-specified attribute Paper: https://arxiv.org/pdf/1710.07346

Project Overview

WardrobeWizard is an AI-powered fashion manipulation system that uses advanced deep learning techniques to modify clothing styles in images based on text descriptions. The project implements a two-stage GAN (Generative Adversarial Network) architecture for precise clothing manipulation while preserving the person's identity.

Key Features

• Text-Guided Image Manipulation: Transform clothing styles based on natural language descriptions
• Two-Stage Generation:
  • Shape Generation: Creates segmentation masks for desired clothing
  • Image Generation: Generates realistic clothing while preserving identity
• Identity Preservation: Maintains facial features and hair during clothing transformation
• Conditional Generation: Supports multiple clothing types and styles

Technical Implementation

1. Reimplement G_shape GAN
   1. Create Data loader for training loop
   2. Implement GAN structure (generator / discriminator)
   3. Train
   4. Test
2. Reimplement G_image

Architecture

• Shape Generator: Creates semantic segmentation masks from text descriptions
• Image Generator: Transforms segmentation masks into realistic clothing
• Discriminators: Ensure realistic and accurate generation for both stages
• Conditional Inputs:
  • Text embeddings for style description
  • Segmentation masks for clothing regions
  • Noise vectors for style variation

Model Components

• G_shape.py: Shape generation pipeline
• G_image.py: Image generation pipeline
• G_shape_model.py: Shape generator architecture
• G_image_model.py: Image generator architecture
• dataloader.py: Data preprocessing and loading utilities
• segmentation.py: Image segmentation utilities

Results and Visualization

• plot_G_shape_results.py: Visualize shape generation results
• plot_G_image_results.py: Visualize final image generation results
• net_graph_sr1.py: Network architecture visualization

Training and Evaluation

• Custom loss functions for realistic generation
• Identity preservation metrics
• Style transfer accuracy evaluation
• Segmentation accuracy metrics

Technologies Used

• PyTorch
• CUDA for GPU acceleration
• NumPy
• Matplotlib
• H5py for data management

Future Improvements

• Enhanced text understanding capabilities
• More diverse clothing style options
• Improved resolution and detail generation
• Real-time processing capabilities
• Mobile deployment support

Installation and Usage

1. Clone the repository
2. Install dependencies:

pip install torch numpy h5py matplotlib

3. Prepare your data following the format in dataloader.py
4. Run the main script:

python main.py

Author

• jak-weston
• Sean-Fuhrman

License

This project is open source and available under the MIT License.