SA-LUT: spatial adaptive 4D look-up table for photorealistic style transfer

Abstract

With the rapid advancement of technology, photorealistic style transfer (PST) has become increasingly significant in film post-production and professional photography, owing to its capability to adapt the color characteristics of a reference image while preserving the structural content. However, current methods exhibit notable limitations for practical implementation. Generation-based techniques often prioritize stylistic fidelity at the expense of content integrity and computational efficiency, whereas global color transformation methods, such as those based on look-up tables (LUTs), maintain structural consistency but lack spatial adaptability. To address these limitations, this thesis presents the Spatial Adaptive 4D Look-Up Table (SA-LUT), a novel framework that integrates the efficiency of LUTs with the adaptability of neural networks. The proposed approach comprises two key compo- nents: a Style-guided 4D LUT Generator, which extracts multi-scale features from the style image to predict a 4D LUT; and a Context Generator, which utilizes content-style cross-attention to produce a context map for spatially adaptive adjustment. This context map enables precise, location-specific color transformations while preserving structural integrity. Furthermore, this thesis introduces PST50, the first benchmark dataset specifically designed for the evaluation of photorealistic style transfer performance. Experimental results demonstrate that the proposed SA-LUT framework significantly outperforms existing state-of-the-art methods, achieving a 66.7% reduction in perceptual distance relative to 3D LUT-based approaches, while maintaining real-time video stylization performance at 16 frames per second.