BSSR: Binarization and Sparsity For Image Super-Resolution

ICLR 2026 Conference Submission16117 Authors

Keywords: Image super resolution, model pruning, binary quantization

Abstract: Lighter models and faster inference remain the focus in the field of image super-resolution. Quantization and pruning are both effective methods for compressing deep models. Unfortunately, existing approaches often optimize quantization and pruning independently: standalone binarization reduces storage but underutilizes sparsity, while N:M sparsity on weights accelerates inference but leaves high-bit storage overhead. Notably, no prior work has explored N:M sparse binary SR networks. In this paper, we combine quantization and sparsity to propose an extreme compression method for super-resolution tasks, namely BSSR. Within this framework, we introduce two key components: Binarized N:M Sparse Quantizer (BSQ) and Binarized Sparse Gradient Adjuster (BSGA). Firstly, BSQ is a sparse binarization operation across dimensions, simultaneously performing activation and weight binarization while imposing N:M sparsity on weights, significantly reducing storage and computational resource requirements. Secondly, BSGA employs a learnable clipping interval and distinct gradient scaling factors for preserved and masked elements to overcome the non-differentiability of binarization and sparse masking, thereby enabling stable gradient propagation and improving training convergence in sparse binary networks. Extensive experiments on SR benchmarks demonstrate that BSSR achieves state-of-the-art performance with significant improvements in PSNR and SSIM over compression methods.

Primary Area: applications to computer vision, audio, language, and other modalities

Submission Number: 16117