Reproduce “Modernizing Old Photos Using Multiple References via Photorealistic Style Transfer” on OpenResearch

About this paper

Agus Gunawan, Soo Ye Kim, Hyeonjun Sim, Jae-Ho Lee, Munchurl Kim

This work introduces Multi-Reference-based Old Photo Modernization (MROPM), a framework that transforms aged photographs to appear as if captured with contemporary cameras. It leverages multiple modern images to guide stylization and utilizes a novel synthetic data generation scheme, outperforming traditional restoration and single-reference methods on quantitative metrics and user studies while also contributing the Cultural Heritage Dataset (CHD).

Key insights

• Different regions of an old photograph often require distinct modern stylistic influences, necessitating a multi-reference approach for effective modernization.
• Photorealistic local style transfer can be achieved without explicit semantic segmentation masks by strategically aligning local style codes from reference images to content image regions.
• Synthetic data generation, involving semantic region-specific style-variant transformations, enables training deep learning models for old photo modernization without requiring impossible paired old-modern datasets.

Summary by alphaXiv.