Learned Lossless Image Compression with Interleaved Parallel Inference and Irregular Causal Reasoning

Lossless image compression typically leverages already decoded pixels to reconstruct the undecoded ones. Existing learned lossless image compression methods typically adopt either pixel-by-pixel or block-based decoding strategies. Pixel-by-pixel decoding fully exploits spatial priors but introduces high computational overhead. In contrast, block-based decoding improves parallelism but degrades performance due to missing spatial priors at the first decoding step. To balance computational overhead and missing spatial priors, we propose Interleaved Parallel Inference (IPI), which enables interleaved decoding across block groups to reduce the number of pixels decoded without spatial priors while keeping the overhead low. Furthermore, we introduce the Irregular Causal Reasoning Module (ICRM), which aligns model inference with irregularly available spatial priors induced by interleaved parallel decoding, enabling adaptive context-aware prediction and improved coding efficiency in learned lossless image compression. Meanwhile, we design three ICRM variants to address the three scenarios that collectively cover most real-world cases. Experimental results on 9 benchmark datasets and three scenarios demonstrate that IPI and the ICRM achieve state-of-the-art compression performance. Compared with representative baselines from prior work, our method achieves up to 20.61% bpd savings.