zsxkib / bsrgan

Upscale videos + images with BSRGAN

  • Public
  • 184 runs
  • GitHub
  • Weights
  • Paper
  • License
Run with an API
Playground API Examples README Versions

BSRGAN: Practical Image and Video Restoration 🖼️🎥

A deep learning model designed to improve low-quality images and videos by addressing real-world degradations like blur, noise, and compression artifacts.

What It Does

BSRGAN helps restore visual content by: - Enhancing resolution (2x or 4x upscaling) - Reducing blur and noise in old photos - Improving compressed/low-quality video frames - Handling unknown degradation types automatically

For videos: Processes each frame individually then reconstructs them into a coherent sequence

How It Works

Trained using a realistic degradation model that simulates: - Multiple blur types (motion, lens, sensor) - Mixed noise patterns (Gaussian + camera-specific) - Various downsampling methods - JPEG compression artifacts

Unlike ESRGAN and other GAN-based methods that assume ideal degradation scenarios, BSRGAN specifically addresses the unpredictable quality loss found in real-world images and footage.

Key Features

  • Blind Restoration: Works without prior knowledge of image degradation
  • Frame Consistency: Maintains temporal coherence when processing video
  • Flexible Input: Accepts images (JPG/PNG) and videos (MP4/MOV)
  • Quality Control: Adjustable parameters for noise reduction and sharpness

Common Use Cases

  1. Restoring historical photos/family albums
  2. Upscaling low-resolution surveillance footage
  3. Improving compressed social media content
  4. Pre-processing for archival digitization projects
  5. Enhancing video conference quality

Limitations

  • Maximum resolution: 1536x1536 for images
  • Recommended video length: <2 minutes
  • Processing time varies by hardware (~10 sec/image on GPUs)
  • May struggle with extreme degradation cases

Credits

Developed by researchers at ETH Zurich: 

@inproceedings{zhang2021designing,
  title={Designing a Practical Degradation Model for Deep Blind Image Super-Resolution},
  author={Zhang, Kai and Liang, Jingyun and Van Gool, Luc and Timofte, Radu},
  booktitle={ICCV},
  year={2021}
}

Original implementation: cszn/BSRGAN on GitHub
Maintained by @zsxkib for Replicate integration

Note: Video processing works by applying the image model frame-by-frame. For optimal results, combine with temporal stabilization tools in post-processing.