Restoration of Distorted Social Media Images

[dheeraj00001]

Pull Request: Restoration of Distorted Social Media Images

Issue: #639
Contributor: Gurijala Dheeraj Reddy (dheeraj00001)
Program: GSSoC'26

Closes #639

Summary

This PR adds a facial restoration project that addresses image degradation from social media compression, motion blur, and sensor noise using a Component-Based Hybrid GAN architecture.

Changes

• Added project directory Restoration_of_Distorted_Social_Media_Images/ following repository structure conventions
• Implemented Hybrid GAN in PyTorch with:
  • U-Net generator with residual learning
  • Global PatchGAN discriminator for full-image realism
  • Local feature discriminators for eyes, nose, and mouth regions
  • Identity-preserving loss using VGG-16 feature extraction
• Included documentation: technical README, architecture overview, and metric analysis (PSNR, SSIM, LPIPS)
• Hosted model weights (694MB) on Hugging Face; download instructions in Model/README.md

Testing

• Quantitative evaluation: PSNR >28dB, SSIM >0.85, LPIPS <0.15 on held-out test set
• Qualitative validation: side-by-side comparison grids of distorted inputs, restored outputs, and ground truth
• Inference pipeline verified in local environment with Python 3.11 and PyTorch 2.0

Type of Change

• New feature (non-breaking change which adds functionality)
• Documentation update

Checklist

• Code follows project style guidelines
• Self-review of implementation completed
• Code commented where logic is non-obvious
• Documentation updated to reflect changes
• No new warnings introduced
• Feature validated with test cases and metric thresholds
• No unmerged dependencies

Notes

• Model weights exceed GitHub LFS limits; hosted externally with verified download link
• Training scripts and synthetic distortion pipeline included for reproducibility
• Compatible with CUDA 11.7+; CPU inference supported with reduced throughput

[github-actions]

Our team will soon review your PR. Thanks @dheeraj00001 :)

[abhisheks008]

Looks good to me. But the rendering of the notebook in Github web is taking time, hence I am not able to see the code you shared. Let the glitch fixed, will update you on the merger of the ppull request.

@dheeraj00001

[dheeraj00001]

Thanks for the review. If the notebook rendering issue persists on GitHub, the code can still be reviewed directly from my commit history and commit details.
Link: 9d36e0d
Thank You.

[abhisheks008]

Hi @dheeraj00001 can you please upload the complete .ipynb file, in which all the EDA, data processing and all the models are implemented and showcased for this project. I can see only one model, but you need to showcase all the models you have built for this project.

[dheeraj00001]

Yeah, I've updated everything you've mentioned. Please merge the pr and close the issue