NOBUGS 2026

High Energy Photon Source (HEPS) experiments generate large volumes of image sequences, whose wide dynamic range, non-negligible noise, and complex temporal redundancy make it difficult for existing lossless compression methods to reduce data volume effectively. We propose a two-stage lossless compression method based on the Mamba state space model. The original images are first transformed into token sequences by inter-frame differencing and Flag quantization. A stage-one Mamba model predicts tokens position by position, and prediction residuals are computed in the physical-value domain. Since these residuals have a more concentrated distribution, a clipping-and-escape mechanism is used to preserve large residuals exactly, while a stage-two Mamba model performs entropy modeling and arithmetic coding on residual tokens. During decompression, residual decoding, stage-one re-inference, and token recovery reconstruct the original images exactly, with pixel-level verification ensuring full losslessness. Tests on real datasets show that our method achieves an average compressed-to-original size ratio of 46.52%, corresponding to a compression ratio of 2.1496x, which is 29%-81% higher than those of three other methods, JPEG-LS, JPEG-XR, and gzip level 9, whose compression ratios on the same dataset are 1.6563x, 1.6241x, and 1.1855x, respectively. These results demonstrate that two-stage residual modeling can effectively improve the lossless compression performance of synchrotron image sequences.