What it does This repo implements two ways to make your photos look like Van Gogh got to them. MSG-Net is the fast, feed-forward version that can stylize images in one shot — there’s even a camera demo. The second mode implements the original Gatys et al. optimization-based approach, which is slower but included for completeness.

The interesting bit The real value is the multi-style angle: one trained model handles 21 different artistic styles without retraining for each. The README is admirably direct about what needs what — GPU optional, CPU fallback with a flag flip.

Key highlights

• Pre-trained 21-style model available via shell script
• Real-time camera demo (camera_demo.py) for live stylization
• Training pipeline included: 4 epochs default, COCO dataset auto-downloaded
• Also implements slow neural style transfer (Gatys CVPR 2016) for comparison
• Cross-framework: sister repos exist in Torch and MXNet/Gluon

Caveats

• The README has a “Tabe of content” typo and HTML-table formatting that feels circa-2017
• No mention of model size, memory requirements, or speed benchmarks beyond “real-time”
• Training details are sparse — hyperparameters, loss curves, or convergence behavior aren’t discussed

Verdict Grab this if you need a working, pre-trained style transfer baseline in PyTorch with minimal fuss. Skip if you want state-of-the-art quality or modern diffusion-based methods; this is solid 2017-era work kept functional.