The Lua Script That Taught Neural Networks to Paint

What it does This is the reference Torch/Lua implementation of Gatys et al.’s “A Neural Algorithm of Artistic Style.” Feed it a photograph and a painting; it uses a pretrained VGG-19 network to synthesize a new image that keeps the photo’s structure while adopting the painting’s textures, colors, and brushwork. You can dial the content-style tradeoff, blend multiple artworks, preserve original colors, or adjust the -style_scale to control which visual features transfer.

The interesting bit Released in 2015, this repo became the proof-of-concept that launched a thousand art filters. It treats style as a statistical property of convolutional feature correlations—an idea that now feels obvious but was genuinely surprising at the time. At 18k stars, it is a time capsule from the era when generating an image meant running iterative optimization for 1000 iterations instead of a single forward pass.

Key highlights

• Implements the exact algorithm from the original Gatys, Ecker, and Bethge paper
• Supports multiple style images with weighted blending and interpolation
• Adjustable -style_scale to control granularity of transferred artistic features
• -original_colors flag preserves the content photo’s palette while transferring only texture
• Backends for CPU, CUDA, cuDNN, and OpenCL (via the old Torch ecosystem)

Caveats

• Built on Torch7 and Lua, a stack that has been largely abandoned in favor of PyTorch/Python
• Memory-hungry: the default VGG-19 model needs a decent GPU; the README suggests falling back to a smaller NIN model for “smaller memory” cards
• L-BFGS optimizer gives better results but uses more memory than ADAM

Verdict A must-read for ML archaeology students and anyone tracing the lineage of generative image models. If you need a modern, real-time style filter, this is a museum piece.