Multi-scale light field super-resolution via learnable codebook priors

Light field (LF) imaging is an effective technique for acquiring 3D information, yet it remains constrained by the inherent trade-off between spatial and angular resolution. Super-resolution (SR) methods offer a promising solution by enhancing spatial resolution while preserving angular consistency. However, existing LF SR approaches face limitations in large-scale SR tasks, commonly resulting in the loss of high-frequency details and overly smooth textures due to insufficient information for accurate reconstruction. Moreover, these methods typically lack scalability across different SR factors, often requiring separate training for each target scale, which is computationally inefficient and impractical for real-world deployment. To overcome these challenges, we propose a novel Light Field Vector Quantization Super-Resolution Network (LF-VQSR) that integrates high-resolution priors into a learning-based framework. Our model learns a quantized feature codebook that captures high-resolution structural priors, enabling super-resolution across multiple scale factors (from 2× to 8×) within a single unified network. To the best of our knowledge, this is the first deep learning-based method capable of multi-scale LF image super-resolution. Experimental results on public benchmarks demonstrate that LF-VQSR outperforms existing methods in terms of both SSIM and LPIPS scores, delivering more realistic details and maintaining strong angular consistency across a wide range of magnification scales.