Cole Feuer’s Unpaired Style Transfer

About the Project

This work from Cole Feuer presents a two-stage, unpaired approach: first a perceptual pretraining phase on anime datasets; then a CycleGAN-based domain adaptation using 8-channel inputs (RGB, depth, edges, blurred prior frames) to enforce visual and temporal fidelity without paired data. Ideal for studios seeking automated stylization at scale.

Key Technical Features

• 8-Channel Encoding: Combines scene structure (depth, edges) with motion context (prior frame) for stable outputs.
• GAN Adaptation: Dual discriminators with identity, adversarial, and cycle-consistency losses on unpaired data.
• Temporal Coherence: Minimizes frame-to-frame flicker via blurred prior-frame input and style loss.
• Performance: Real-time inference demo with TensorRT & ONNX build.

Code & Documentation

View the complete implementation, training scripts, and setup instructions on GitHub:

github.com/CDFire/UnpairedStyleTransfer

Future Directions

Planned enhancements include explicit optical-flow losses, multi-style conditional transfer, and user-driven style parameter controls. Feedback and collaborations welcomed.

Contact

Questions or collaborations? Reach out to coledfeuer@gmail.com or connect on LinkedIn.