Low-Complexity Parallel Decoding for Non-Autoregressive Learned Image Compression

Autoregressive probability models improve the rate–distortion performance of learned image compression, but their spatial conditioning introduces serial neural inference and high decoding latency. This paper presents a latency-oriented learned image compression framework that jointly addresses sequential entropy estimation, decoder computation, and dense intermediate activation responses. First, a dependency-decoupled entropy model estimates all latent distribution parameters in parallel from hyperprior statistics, eliminating the neural autoregressive loop. Second, a lightweight synthesis transform uses compact convolutional mappings, reduced channel widths, and a short progressive reconstruction path. Third, a differentiable activation-magnitude regularizer penalizes the normalized L1 magnitude of intermediate decoder activations and is introduced gradually during rate–distortion training to encourage sparse feature responses. The activation regularizer is treated as a representation-level training constraint rather than as a direct estimator of dense-convolution arithmetic. Experiments on Kodak and BSD100 show that the proposed model consistently improves on the hyperprior baseline over the evaluated bitrate range. On Kodak, it achieves BD-rate reductions of 7.9% and 9.7% relative to hyperprior in terms of PSNR and MS-SSIM, respectively. At approximately 0.30 bpp, the proposed model achieves 32.94 dB PSNR and 0.9745 MS-SSIM with 8.10 M parameters, 165.3 GMACs, and a decoding time of 35.6±1.0 ms. Compared with the serial autoregressive joint baseline at a comparable operating point, this corresponds to a 95.0% reduction in decoding latency.