What it does

You pick a few frames of a video, paint or stylize them by hand, and this tool learns the style from 32×32 pixel patches. It then redraws every remaining frame to match. There’s also a webcam mode for live stylization, though it crops to square and resizes for speed.

The interesting bit

Temporal consistency is deliberately not baked into the network. That sounds like a bug, but it’s a feature: training stays fast and parallelizable. The tradeoff is flickering, which the authors tackle with a separate post-processing pipeline of optical flow and bilateral filtering. They also use auxiliary “gaussian mixture” images to disambiguate similar-looking patches—like two chunks of sky that should map to differently stylized outputs.

Key highlights

• Trains on patch-level correspondences, not full frames, from as few as 1–4 stylized keyframes
• Optional temporal consistency tools: disflow for optical flow, bilateralAdv for noise filtering, gauss for patch disambiguation
• Pre-trained models and test data provided via Google Drive links
• Webcam demo included (generate_webcam.py)
• PyTorch Lightning reimplementation available from community contributor

Caveats

• Windows-first tooling: build scripts and prebuilt binaries are for Windows; Linux/MacOS users get “get inspired by the build script” guidance
• TensorFlow 1.15.3 is listed as a dependency but only used in logger.py; author notes it will be removed
• Webcam mode always crops to square and resizes, so aspect ratio and resolution take a hit

Verdict

Worth a look if you’re doing research in few-shot video stylization or need a baseline for comparison. Probably not a drop-in production tool: the temporal consistency workflow is involved, and the Windows-centric tooling will slow down non-Windows shops.