Skip to main content
Back

DDM-SSCC Turns Diffusion AI Into a Lossless Wireless Image Transmitter

A new paper proposes DDM-SSCC, using discrete diffusion models to achieve lossless pixel-perfect image transmission over noisy wireless channels — outperforming both classical and semantic communication baselines.

Paper, Full Access
  • New framework: DDM-SSCC adapts discrete diffusion language models for lossless, pixel-perfect image transmission over noisy wireless channels.
  • Core idea: Bidirectional attention in diffusion models provides richer context than autoregressive models, enabling better probability estimates for arithmetic coding.
  • Three key modules: Halton-guided denoising order, mask-ratio-aware cosine schedule, and temperature calibration bridge the gap between generation and lossless compression.
  • Results: Outperforms lossless, autoregressive, and semantic communication baselines on CIFAR10, DIV2K-LR-X4, and Kodak across AWGN and Rayleigh fading channels.
  • Best use case: Scenarios requiring exact pixel recovery — medical imaging, satellite data, forensic or archival systems.
  • No code released: The paper is available on arXiv but no public implementation has been shared alongside it.

Most AI research on wireless image transmission has chased a single goal: send a compressed, good-enough version of an image that looks plausible at the receiver. Semantic communication systems do this well, but they fundamentally accept some information loss. A new paper posted to arXiv takes a different stance , what if you needed every single pixel to arrive perfectly intact, even over a noisy channel?

The work introduces DDM-SSCC (Discrete Diffusion Model-based Separate Source-Channel Coding), a framework that adapts discrete diffusion language models to the problem of lossless image transmission. The core insight is that the iterative, bidirectional nature of diffusion-based token restoration is actually a better fit for lossless compression than the autoregressive models that dominate the field today.

Why lossless transmission is a harder problem than it looks

The dominant paradigm in learned wireless image transmission is semantic communication , systems that compress an image into a compact latent representation, transmit it, and reconstruct something visually plausible on the other end. Semantic communication aims to improve efficiency by transmitting only the semantic meaning of the source data rather than the complete set of bits. This works beautifully for human viewing, but it is fundamentally lossy.

There are real-world scenarios where pixel-perfect recovery is non-negotiable: medical imaging, satellite data relay, forensic evidence, and archival systems. Lossless pixel-level image transmission is a fundamental regime beyond semantic communications, because exact recovery requires both accurate symbol probability modeling and reliable delivery over noisy channels. Both conditions must hold simultaneously , and that is the crux of the difficulty.

The standard approach for lossless source coding is arithmetic coding, a technique that encodes data as a single number by recursively narrowing a probability interval based on each symbol's predicted likelihood. The better your probability model, the shorter the resulting bitstream. State-of-the-art systems pair arithmetic coding with autoregressive language models, which predict each pixel token one at a time in a fixed left-to-right order. The problem: separate source-channel coding's effectiveness in noisy transmission is fundamentally constrained by the fragility of autoregressive source decoding , in low-SNR regimes, even a small number of residual bit errors after channel decoding may derail the subsequent lossless reconstruction process.