What it does Deep Image Analogy finds semantically-meaningful dense correspondences between two images by adapting classical image analogy to deep CNN features. It can transfer a painting’s style onto a photo, swap styles between artworks, convert sketches to photorealistic images, or perform color transfer between photos.

The interesting bit The trick is running PatchMatch in feature space rather than raw pixels. The algorithm matches deep VGG-19 features across layers, then backpropagates correspondences through the network hierarchy. This preserves structure where it matters and transfers appearance where it doesn’t.

Key highlights

• Official C++/CUDA implementation of the SIGGRAPH 2017 paper
• Built on Caffe with VGG-19; requires separate model download
• Supports four transfer modes: photo→style, style→style, style→photo, photo→photo
• Includes optional WLS filter to preserve photo structure during style transfer
• Provides pre-built Windows executable for command-line use
• Blend weight and resize ratio parameters control fidelity vs. stylization

Caveats

• Windows-only officially (Linux exists on a branch, macOS unsupported)
• CUDA 7.5/8 only; tested on specific Nvidia GPUs from Titan X down to GTX 770
• Input size effectively capped around 700×500 at full resolution; larger images need ratio downsampling
• Requires Visual Studio 2013 and manual Caffe build

Verdict Worth a look if you’re researching neural style transfer or need a reference implementation of deep PatchMatch. Skip it if you want modern PyTorch convenience or need to run on current CUDA versions without friction.