OSD-JSCC: Neural Joint Source-Channel Coding with One-Step Diffusion

arXiv:2611.06437 · eess.SP, cs.IT, math.IT · Submitted 2026-11-04 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Electrical Engineering. The day in electrical engineering: signals, systems and the hardware that carries them.

Rosa: We're looking at the paper: "OSD-JSCC: Neural Joint Source-Channel Coding with One-Step Diffusion".

Dev: Here's the gist of the paper. Neural joint source–channel coding (JSCC) enables efficient image transmission over noisy channels, yet distortion-optimized approaches often produce blurry reconstructions The proposed OSD-JSCC achieves superior perceptual quality and significantly lower decoding latency compared to existing diffusion-based JSCC methods at the same bandwidth compression ratio Pruning redundant reasoning tokens The paper introduces OSD-JSCC, an SNR-adaptive JSCC framework that integrates a pretrained diffusion prior for high-fidelity image transmission with single-step decoding This framework addresses the limitation of iterative sampling in diffusion priors by proposing a single diffusion step The core idea is to transmit pretrained variational autoencoder (VAE) latents through an SNR-conditioned JSCC codec How the training signal is built The system couples an SNR-conditioned latent JSCC (L-JSCC) codec with a pretrained one-step diffusion prior The L-JSCC decoder predicts a Gaussian state and extracts image-specific semantic tokens to guide the reconstruction process The timestep is adjusted according to the channel SNR using a mechanism that combines feature modulation and diffusion timestep selection Where the gains come from The authors introduce two key designs for OSD-JSCC 1.

Rosa: A stochastic diffusion-state interface that enables the JSCC decoder to predict a Gaussian diffusion state rather than a clean latent, while extracting imagespecific semantic tokens 2.

Dev: SNRconditioned feature modulation and adaptive timestep selection to jointly adapt latent transmission and diffusion refinement to varying channel conditions The training strategy involves a three-stage process with low-rank adaptation (LoRA) to efficiently adapt the pretrained diffusion prior to the JSCC task Performance and Ablation Insights Experiments on Kodak and CLIC2020 demonstrate that OSD-JSCC achieves superior perceptual quality and significantly lower decoding latency than existing diffusion-based JSCC methods Specifically, OSD-JSCC reduces LPIPS by approximately 43% on Kodak and 53% on CLIC2020, while achieving up to a 50.6× reduction in decoding latency Ablation studies show that removing SNR modulation increases the required SNR by an average of 0.66 dB and 0.58 dB at matched LPIPS and DISTS, respectively Furthermore, routing all SNRs to t = 249 leaves results above 10 dB almost unchanged but raises LPIPS at 0 dB from 0.169 to 0.

Paper summary: Rosa: To wrap up this look at OSD-JSCC: Neural Joint Source–Channel Coding with One-Step Diffusion, the authors are essentially proposing a way to use a single diffusion step within an SNR adaptive JSCC framework to handle noisy channels while keeping things sharp and fast.

Dev: The title itself suggests the core innovation is merging the source coding aspect with diffusion in a way that isn't reliant on slow, iterative sampling. It tackles the latency problem directly by proposing this one-step approach.

Taro: So, what does this mean for us when we think about future autonomy? It means we can expect systems that handle channel degradation much more smoothly because they have this built-in mechanism to adjust their refinement based on real-time signal conditions.

Rosa: The work demonstrates that integrating the stochastic diffusion state interface with semantic tokens is a powerful way to guide the reconstruction process, providing better spatial structure preservation than just predicting a clean latent.

Dev: And from an engineering standpoint, the SNR conditioned feature modulation and adaptive timestep selection are key mechanisms that allow this framework to actually perform well across different channel conditions without needing massive retraining.

Taro: It shows that we don't always need a huge, complex model running for every scenario; instead, a smart interface between the codec and the prior is enough to keep things robust.

Rosa: So, in short, OSD-JSCC isn't just about achieving better numbers on Kodak or CLIC2020; it’s about creating a more practical transmission system that balances perceptual quality with decoding speed under real-world channel noise.

Conclusion: Rosa: So we're wrapping up on OSD-JSCC, which stands for Neural Joint Source–Channel Coding with One-Step Diffusion, and we need to look at what that title actually means for us.

Dev: It’s about combining source coding and channel coding with a single diffusion step instead of the usual long iterative sampling. That’s the core idea here.

Taro: So, if you think about it, they are taking a pretrained diffusion model and using it to help encode images over noisy channels, but they skip all those slow refinement steps.

Rosa: Exactly. It's not just about getting a better image; it's about how fast we can get that image decoded when the channel is bad.

Dev: And the authors are really focused on making sure this works under real-world conditions, specifically through SNR adaptation and feature modulation to handle different noise levels.

Taro: I see them using a stochastic diffusion state interface that lets the coder predict a Gaussian state instead of just a clean latent image. That sounds like it gives the reconstruction more control.

Rosa: It does. They are extracting image-specific tokens along with that state, which guides the process better than just sending raw data through the channel and hoping for the best.

Dev: And they're using this to achieve really low latency, up to fifty point six times faster than some other diffusion based methods at similar compression ratios. That’s a big win for real-time systems.

Taro: What they show in their experiments on Kodak and CLIC2020 is that the perceptual quality actually improves significantly, with things like the LPIPS score dropping by about forty three percent on Kodak.

Rosa: It's not just a small tweak to the image; it’s a substantial jump in how visually good what you get when you receive it.

Dev: But there are trade-offs, because they did show that if you remove the SNR modulation, your required signal-to-noise ratio goes up by about zero point six six decibels.

Taro: That tells us the SNR adaptation is pretty crucial for making this system robust across different channel conditions. If you don't adapt, it struggles when the noise changes unexpectedly.

Rosa: It’s a balance, right? They’ve found a way to get high fidelity without needing an impossibly high signal quality all the time.

Dev: And they also showed that even if you just send everything through one specific step in the diffusion process, say at timestep two hundred ninety-nine, things stay pretty good above ten decibels.

Taro: That suggests the single-step approach is surprisingly resilient when things are not too bad initially. But they did flag that routing all SNRs to that one specific time step hurts quality if the channel is actually weak.

Rosa: So, OSD-JSCC seems to be a clever way to marry high-quality image synthesis with the speed needed for practical communication under noisy conditions.

Dev: It’s a solid piece of work because it tackles both the quality and the latency issues head-on with this specific architectural choice.

Taro: Now we need to see if this kind of one-step refinement can scale up to handle much more complex or unpredictable failures in a real autonomous system.

Shunpu Tang, Qianqian Yang

Zhejiang University

eess.SP, cs.IT, math.IT

Submitted: 2026-11-04

Updated: 2026-11-04

Comments: 6 pages, 5 figures. Code: https://github.com/tangshunpu/OSD-JSCC

Code: https://github.com/tangshunpu/OSD-JSCC

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

Importance score: 94/100

The gist: Neural joint source–channel coding (JSCC) enables efficient image transmission over noisy channels, yet distortion-optimized approaches often produce blurry reconstructions The proposed OSD-JSCC…

Key concepts

Neural Joint Source–Channel Coding (JSCC)
This is a method for efficiently sending images over noisy channels. It combines compressing the image source with coding it for the channel simultaneously, aiming to minimize distortion while adapting to channel noise.
One-Step Diffusion Prior
This is a pre-trained model used to refine images during transmission. Instead of many iterative steps, this framework uses a single diffusion step guided by an SNR-conditioned JSCC codec to achieve high fidelity reconstruction.
SNR-Conditioned Feature Modulation
This technique adjusts how features are transmitted based on the channel's Signal-to-Noise Ratio (SNR). It dynamically modifies the latent transmission and diffusion refinement process to optimally handle varying levels of channel noise.
Adaptive Timestep Selection
The system intelligently chooses the appropriate diffusion timestep based on the channel SNR. This ensures that image refinement is tailored precisely to the current transmission conditions, improving quality and efficiency.

Terminology

Summary

Neural joint source–channel coding (JSCC) enables efficient image transmission over noisy channels, yet distortion-optimized approaches often produce blurry reconstructions The proposed OSD-JSCC achieves superior perceptual quality and significantly lower decoding latency compared to existing diffusion-based JSCC methods at the same bandwidth compression ratio

Pruning redundant reasoning tokens

The paper introduces OSD-JSCC, an SNR-adaptive JSCC framework that integrates a pretrained diffusion prior for high-fidelity image transmission with single-step decoding This framework addresses the limitation of iterative sampling in diffusion priors by proposing a single diffusion step The core idea is to transmit pretrained variational autoencoder (VAE) latents through an SNR-conditioned JSCC codec

How the training signal is built

The system couples an SNR-conditioned latent JSCC (L-JSCC) codec with a pretrained one-step diffusion prior The L-JSCC decoder predicts a Gaussian state and extracts image-specific semantic tokens to guide the reconstruction process The timestep is adjusted according to the channel SNR using a mechanism that combines feature modulation and diffusion timestep selection

Where the gains come from

The authors introduce two key designs for OSD-JSCC 1. A stochastic diffusion-state interface that enables the JSCC decoder to predict a Gaussian diffusion state rather than a clean latent, while extracting imagespecific semantic tokens 2. p. 0 SNRconditioned feature modulation and adaptive timestep selection to jointly adapt latent transmission and diffusion refinement to varying channel conditions The training strategy involves a three-stage process with low-rank adaptation (LoRA) to efficiently adapt the pretrained diffusion prior to the JSCC task

Performance and Ablation Insights

Experiments on Kodak and CLIC2020 demonstrate that OSD-JSCC achieves superior perceptual quality and significantly lower decoding latency than existing diffusion-based JSCC methods Specifically, OSD-JSCC reduces LPIPS by approximately 43% on Kodak and 53% on CLIC2020, while achieving up to a 50.6× reduction in decoding latency Ablation studies show that removing SNR modulation increases the required SNR by an average of 0. p. 466 dB and 0.58 dB at matched LPIPS and DISTS, respectively Furthermore, routing all SNRs to t = 249 leaves results above 10 dB almost unchanged but raises LPIPS at 0 dB from 0. p. 4169 to 0.

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. SNR-adaptive latent JSCC framework allows for joint source-channel coding in VAE latent space while integrating a pretrained diffusion prior for high-fidelity image transmission with single-step decoding. This enables superior perceptual quality while significantly reducing decoding latency, achieving up to a 50.6 times reduction compared to existing diffusion-based JSCC baselines.

  2. Introduction of a stochastic diffusion-state interface allows the JSCC decoder to predict a Gaussian diffusion state rather than a clean latent, which preserves spatial structure and provides an initialization for one-step refinement while extracting image specific semantic tokens. These tokens, along with an SNR token, guide the reconstruction process through cross attention.

  3. Development of an SNR adaptive mechanism combining feature modulation and diffusion timestep selection jointly adapts latent transmission and diffusion refinement to varying channel conditions. This mechanism utilizes SNR embedding to adjust group normalization for both the JSCC encoder and decoder, allowing the codec to adapt its features while sharing network weights across SNRs.

  4. The system can perform high-fidelity image transmission over noisy channels by mapping an image directly into a latent space, converting it into channel symbols via an SNR conditioned JSCC codec, and then refining this latent using a single step guided by the diffusion model. This process maintains perceptual quality even under severe channel noise where traditional iterative sampling methods fail due to latency.

  5. The improved AI system can achieve superior reconstruction fidelity compared to baselines on Kodak and CLIC2020, demonstrating better perceptual quality at all SNRs with much lower LPIPS, DISTS, and FID scores. This performance is achieved while maintaining PSNR and MS-SSIM scores comparable to or better than generative JSCC methods.

  6. The system can drastically reduce decoding latency for image transmission by employing a single diffusion step for reconstruction instead of multiple iterative denoising steps required by typical diffusion-based methods. This makes the proposed OSD-JSCC suitable for real-time image transmission applications, as shown by its 50.6x reduction in decoding latency on CLIC2020.

Sources

Related papers