Adapting Diffusion Language Models for Lossless Pixel-Level Image Transmission

Researchers propose DDM-SSCC, a discrete diffusion model framework that improves lossless image transmission over noisy channels by combining pixel-level restoration with arithmetic coding. The approach outperforms existing lossless and semantic communication baselines on standard datasets, offering practical improvements for exact-recovery image transmission scenarios.

This research addresses a specialized but important problem in information theory: transmitting images perfectly over noisy communication channels without any data loss. Traditional approaches rely on raster-order autoregressive coding, which processes pixels sequentially and limits context availability. The DDM-SSCC framework diverges by applying diffusion language models—typically used for generation tasks—to the inverse problem of reconstructing masked pixels with high probability accuracy.

The innovation stems from recognizing that diffusion models naturally excel at modeling complex probability distributions over pixels, a requirement for lossless transmission. By synchronizing the denoising process with arithmetic coding and processing multiple masked tokens per step, the authors achieve computational efficiency gains. The bidirectional attention mechanism allows newly restored pixels to inform future restoration decisions, creating richer context than unidirectional methods provide.

The technical contributions—Halton-guided denoising order, mask-ratio-aware scheduling, and temperature calibration—are deliberate engineering choices addressing the gap between generation-oriented diffusion models and strict lossless coding requirements. Experiments across CIFAR10, DIV2K, and Kodak datasets under realistic channel conditions (AWGN and Rayleigh fading) demonstrate consistent performance improvements.

While this research operates in an academic domain rather than commercial cryptocurrency or finance, it represents meaningful progress in channel coding—foundational infrastructure for reliable communication systems. Such advances eventually propagate into practical applications: satellite imagery transmission, medical imaging networks, and blockchain-based distributed storage systems. The work demonstrates how modern deep learning architectures can solve classical information theory problems more elegantly than legacy approaches.

• →Discrete diffusion models can outperform traditional autoregressive methods for lossless image transmission by processing multiple pixels simultaneously with bidirectional context
• →The framework achieves exact-recovery performance improvements over semantic and lossless communication baselines on standard benchmarks
• →Temperature calibration and adaptive denoising schedules bridge the gap between generative modeling and arithmetic coding requirements
• →Bidirectional attention in the restoration process creates favorable source representations that improve robustness over noisy channels
• →Results demonstrate consistent gains across multiple datasets and channel models including AWGN and Rayleigh fading scenarios