Restoring images with neural nets that never learned

What it does This repository provides Jupyter Notebooks to reproduce every figure from the CVPR 2018 paper Deep Image Prior. It demonstrates image restoration using neural networks that require no pre-training or external datasets—what the authors describe as restoration “without learning.”

The interesting bit The premise is that a deep network can act as its own image prior. The code restores corrupted images by optimizing network weights without any training data, using only the model’s structure to guide the result.

Key highlights

• Reproduces all paper figures via ready-to-run notebooks
• Requires no training data or pre-trained weights
• Authors candidly note that most hyperparameters are “far from optimal”
• Includes Docker and Google Colab configurations
• Targets PyTorch 0.4 and Python 3.6

Caveats

• Optimization may fail to converge on Tesla V100 and P40 GPUs; the authors suggest double precision or disabling cuDNN
• Dependency stack is several years old, which may complicate execution on current systems

Verdict Useful if you want to reproduce the paper or study why network architecture alone imposes strong image priors; otherwise, it is a dated research reference rather than a practical tool.