Fading Channels as Diffusion Steps: One-Step Generative Decoding for End-to-End Wireless Image Transmission

A diffusion model is an image generator trained to undo noise step by step. A wireless channel is a device that adds noise to whatever you send. Over the textbook additive white Gaussian noise (AWGN) channel, these two descriptions match exactly. If the transmitter sends the latent code of a pretrained diffusion model, the receiver gets that code plus Gaussian noise of known power. That is precisely an intermediate state of the diffusion model's forward noising process, at a timestep fixed by the signal-to-noise ratio (SNR). A single call to the denoiser can then estimate the clean code. Because decoding is a single differentiable step, the transmitter can be trained end-to-end through the generative receiver. Current diffusion-aided transmission systems cannot do this, because they backpropagate through dozens of sampling steps.

The catch is that the match holds only for the simplified channel. Over a real frequency-selective fading channel, the equalized signal is still roughly "code plus noise," but each transmitted symbol sees a different noise level, set by how deep its subcarrier fades. The noise is also inflated by channel-estimation error and channel aging in ways that are not Gaussian. This proposal makes that correspondence explicit and uses it to build a one-step generative receiver for fading channels. The central measurement is how far the AWGN-exact construction survives as the channel moves from AWGN to Rayleigh to 3GPP tapped-delay-line fading with mobility.

Core Claim

We claim three things. First, after equalization, an OFDM fading channel can be treated as a diffusion forward process with a known per-symbol noise level instead of a single timestep. Second, a one-step denoiser conditioned on that noise map can decode a diffusion-latent transmission at roughly the cost of one network evaluation. Third, this single step makes it affordable to train the transmitter's encoder end-to-end through the generative receiver, and end-to-end training gives measurable perceptual gains over decoupled training on fading channels.

Concretely, the transmitter sends a power-normalized latent $z_0$. After per-subcarrier zero-forcing equalization, symbol $k$ is received as

$$\tilde y_k = z_{0,k} + \frac{n_k}{h_k},$$

where $h_k$ is the channel gain on that symbol's subcarrier and $n_k$ is receiver noise. This is a forward diffusion state with per-symbol noise standard deviation $\sigma_k = \sigma_n/|h_k|$, where $\sigma_n$ is the receiver noise standard deviation. Under AWGN, $|h_k| = 1$ and every $\sigma_k$ collapses to one scalar. The relation is exact only when $h_k$ is known perfectly. With an estimated $\hat h_k$, a residual term $z_{0,k}(h_k - \hat h_k)/\hat h_k$ appears. It depends on the signal and is not Gaussian, and it is the precise sense in which the simplified model breaks.

Research Program

The work has three parts.

The first part is the receiver. We take a latent diffusion model and condition its denoiser on a spatial noise-level map $\{\sigma_k\}$ in place of a scalar timestep. We then distill it into a one-step decoder. Two one-step routes are compared. The first is the Tweedie estimate, a closed-form posterior-mean reconstruction $\hat z_0 = (\tilde y - \sigma \hat\epsilon)/\alpha$ from a single noise prediction. It is the proxy that ControlNet++ and ReNeg use to get cheap training gradients, and it is known to blur at high noise. The second is an average-velocity one-step map, the technique One-Step Generative Channel Estimation via Average Velocity Field uses for channel estimation. That paper learns the whole trajectory from a noisy observation to a clean sample in one evaluation, which is meant to avoid the posterior-mean blur. Both papers supply method here, not motivation.

The second part is the transmitter. The encoder maps an image into the diffusion latent space. It is trained end-to-end by sampling a channel realization, forming $\tilde y$, running the one-step decoder, decoding to pixels, and backpropagating a combined distortion-plus-perceptual loss. ControlNet++ reports that the one-step proxy cut training memory from roughly 340 GB to 6.8 GB relative to unrolling 50 steps. That figure is what makes this part feasible on ordinary hardware.

The third part is the stress test of the simplified model. We train under AWGN, Rayleigh block fading, and 3GPP tapped-delay-line (TDL-A/B/C) OFDM channels, and test across all of them. We then add imperfect channel estimates and Doppler-induced aging at speeds from 3 to 120 km/h. These test conditions follow the mobility setup of Distribution-Free Adaptive Modulation and Coding via Online Conformal Inference under Non-Stationary Fading. For each condition we locate the SNR, or per-symbol noise level, below which the one-step reconstruction stops carrying useful gradient. This is the wireless counterpart of the noise threshold beyond which ControlNet++ reports that its proxy degrades.

This positions the project against the closest work. DiffJSCC attaches a multi-step Stable Diffusion post-processor to a pretrained deep joint source-channel coding (JSCC) autoencoder. JSCC here means a neural encoder that maps the image directly to channel symbols, with no separate compression and error-correcting code. DiffJSCC is trained in decoupled stages and evaluated almost only on AWGN. We make decoding one step, train through it, and move to fading. DMCE uses a multi-step diffusion model to clean the channel estimate rather than the image, again with staged training. That motivates our claim that staging is forced by sampling cost, not chosen. Robust video JSCC over multipath fading handles OFDM fading but uses a small deterministic denoiser, which contributes only 0.57 dB of its 5.13 dB gain. We replace that piece with a generative one. The SNR-matched initialization in Diffusion-based Generative Prior for Low-Complexity MIMO Channel Estimation and Conditional Prior-based Non-stationary Channel Estimation is the scalar special case of our noise map. Both papers apply it to channel matrices, not to transmitted content.

Baselines

The naive reference is the same encoder and decoder with no diffusion prior. It is a standard deep JSCC autoencoder in the style of DeepJSCC-f (without feedback), trained end-to-end over each channel with the identical distortion-plus-perceptual loss and the same channel-symbol budget. DeepJSCC-f reports up to 3 dB PSNR over separation schemes on Kodak and CIFAR. Any generative receiver must beat this autoencoder on perceptual metrics without losing more than a small, pre-declared PSNR margin.

The strongest existing method is DiffJSCC. It is the only system in the file that puts a pretrained diffusion prior at the image receiver, and it targets exactly our low-SNR, low-bandwidth regime. The file reports that it reconstructs 768×512 Kodak images from 3072 symbols (under 0.008 symbols per pixel) at 1 dB SNR over AWGN. No reported LPIPS or FID values for DiffJSCC were found on Lacuna, so we will rerun it. It gets the same backbone, the same training images (OpenImages), the same symbol budget, and its native multi-step sampler at 20 and 50 steps. On fading channels it receives the same equalized input and SNR conditioning as our method. Beating it means matching its LPIPS and FID within the run-to-run spread at the same budget, while decoding in one evaluation and staying ahead on TDL fading.

Two ablations isolate the key ingredients. The first is a decoupled one-step system: an encoder trained without the prior, followed by a frozen one-step decoder. It tests whether end-to-end training through the one-step step matters. The second is scalar-SNR conditioning, which uses the average SNR as a single timestep in place of the per-symbol noise map. It tests whether the departure from the simplified channel model has to be modeled at all.

Why not just…?

Why not just train a plain end-to-end deep JSCC autoencoder with a perceptual or adversarial loss? It is cheaper and already end-to-end. The evidence against it is indirect. The small learned denoiser in the robust video JSCC work adds only 0.57 dB under multipath fading. DiffJSCC was motivated by the blur and cliff effect of such decoders at low SNR. Neither result settles the comparison under fading at matched symbol budgets, so the naive baseline above is the direct test. If it ties our method on LPIPS and FID over TDL channels, the generative prior is not earning its cost, and we would report that.

Why not just keep DiffJSCC's decoupled training and distill its sampler into one step at inference? This is the most serious threat. It would deliver the speed without any end-to-end machinery. The decoupled one-step ablation is exactly this system. Our claim survives only if the end-to-end version beats it, and we expect any gap to be largest on fading channels, where the encoder can learn to spread important content away from symbols likely to land in deep fades. If the gap is negligible even under TDL fading, the contribution reduces to the noise-map receiver, and the proposal will say so.

Why not just equalize and feed the denoiser the average SNR, treating fading as AWGN? This is the simplified channel model applied naively, and it is the scalar-conditioning ablation. Within one OFDM symbol, per-subcarrier noise can vary by tens of dB. A single timestep therefore under-denoises the faded symbols and over-smooths the good ones. We expect the cost to grow with delay spread, from TDL-A to TDL-C. If it does not, the simplified model is good enough here, which would itself be a useful result.

Proposed Experiments

Training uses OpenImages. Evaluation uses Kodak and CelebA-HQ at bandwidth ratios that include DiffJSCC's 3072-symbol setting. Channels are AWGN, Rayleigh block fading, and 3GPP TDL-A/B/C with OFDM. Test SNRs run from −5 to 20 dB, with perfect and estimated channel knowledge and mobility from 3 to 120 km/h. Metrics are PSNR, MS-SSIM, LPIPS, FID, decoder wall-clock time, and training memory.

Three checks decide the proposal. The first is the train–test channel grid: a model trained on AWGN and tested on TDL-C is compared against one trained on TDL-C. The size of this gap is the direct measure of how much the simplified model misleads a one-step generative receiver. The second is the ranking of end-to-end one-step, decoupled one-step, and multi-step DiffJSCC on LPIPS under TDL fading. The third is a sweep of channel-estimation error and Doppler. It locates the noise level at which the residual non-Gaussian term breaks the diffusion correspondence. We compare Tweedie and average-velocity decoders at that point to see whether the one-step map pushes the breakdown later than the posterior mean, as the AVF paper's low-SNR results suggest.