<p>JPEG remains one of the most widely used image formats due to its compatibility and low computational cost. However, its limited compression efficiency has become a growing burden for large-scale cloud storage platforms, where massive numbers of JPEG images must be archived and strict bit-level integrity requires fully lossless recompression. This paper presents a bit-consistent JPEG recompression framework based on joint spatial–transform domain prediction, which explicitly exploits structural redundancies that are not utilized by the original JPEG encoder. The method reconstructs the quantized DCT coefficients from the JPEG bitstream, generates prediction blocks using a set of intra prediction modes, and transforms them into the DCT domain to form accurate coefficient predictors. A mask-based refinement is applied to suppress unreliable high-frequency components, and the resulting residuals are encoded using a hybrid entropy coding scheme that combines lightweight CABAC for dense coefficients with Brotli for sparse data. Experiments on the Kodak and DIV2K datasets demonstrate that the proposed approach achieves average bit savings of 22.97% and 22.88%, respectively, providing an effective and storage-efficient solution for cloud platforms that require strict preservation of original JPEG bitstreams.</p>