| --- |
| license: apache-2.0 |
| tags: |
| - image-to-image |
| - gan |
| - facial-restoration |
| - pytorch |
| datasets: |
| - CelebA |
| - FFHQ |
| metrics: |
| - psnr |
| - ssim |
| - lpips |
| pipeline_tag: image-to-image |
| library_name: pytorch |
| --- |
| |
| # SocialFace-Restore-GAN |
|
|
| A Component-Based Hybrid GAN architecture designed to restore high-fidelity facial features (eyes, nose, mouth) in images degraded by social media compression, motion blur, and sensor noise. |
|
|
| ## 🎯 Goal |
|
|
| The primary objective is to achieve perceptual realism while strictly preserving the subject's unique identity. This model leverages specialized discriminators for local facial components and an identity-consistency module. |
|
|
| ## 🧵 Dataset |
|
|
| The model was trained using paired training data generated via synthetic distortion from: |
| - **CelebA**: Large-scale face attributes dataset. |
| - **FFHQ**: Flickr-Faces-HQ dataset. |
|
|
| ## 🧾 Description |
|
|
| Images on social media often suffer from aggressive lossy compression. This project implements a **Hybrid GAN** featuring: |
| - **Generator (U-Net)**: A symmetric encoder-decoder with residual learning. |
| - **Global Discriminator (PatchGAN)**: For full-image texture realism. |
| - **Local Feature Discriminator**: Specialized networks focusing on eyes, nose, and mouth. |
| - **Identity Preserver**: VGG-based feature extraction for identity-consistency loss. |
|
|
| ## 📊 Performance Metrics |
|
|
| The model achieves the following target benchmarks: |
| - **PSNR**: > 28dB |
| - **SSIM**: > 0.85 |
| - **LPIPS**: < 0.15 (indicating high human-perceived realism) |
|
|
| ## 🚀 Usage |
|
|
| You can load the model in PyTorch using: |
|
|
| ```python |
| import torch |
| # Assuming your model architecture class 'Generator' is defined |
| model = Generator() |
| model.load_state_dict(torch.load("socialface_restore_gan.pth", map_location="cpu")) |
| model.eval() |
