Neural networks that add, or remove, censorship squares

What it does DeepMosaics is a PyTorch-based tool that detects mosaic regions in images and video, then either removes them by inpainting the underlying content or adds fresh ones. It handles both still frames and video sequences, and ships with pre-trained models tuned for different effects. The project also exposes a web demo and a Windows GUI for users who prefer not to touch source code.

The interesting bit The pipeline borrows from CycleGAN, pix2pixHD, and BiSeNet, treating mosaic manipulation as a semantic-segmentation-plus-translation problem rather than simple filtering. That same backbone apparently moonlights in style transfer, as the examples include Van Gogh and winter conversions.

Key highlights

• Supports both mosaic addition and removal on images and video.
• Provides a Windows GUI binary and a web demo (face-only removal).
• Requires specific pre-trained model files, including mosaic_position.pth, for core operations.
• Can be retrained on custom datasets using the provided training docs.
• GPU acceleration via CUDA is supported but not mandatory.

Caveats

• The public web demo only handles face mosaic removal; full functionality requires local execution.
• The pre-built Windows GUI lags behind the source code in updates.
• Output videos may need a player like PotPlayer to render correctly.

Verdict Worth a look for computer-vision researchers or content-restoration hobbyists working with censored media. If you are not comfortable managing PyTorch model files and ffmpeg versions, the pre-built GUI is a better entry point than source.