Conditional Diffusion Decoding Module

• CDDM is a neural module that implements conditional score-based denoising for reconstructing signals and images from noisy observations.
• It integrates structured side information such as channel state, semantic latents, and syndromes to steer its reverse diffusion process across varied applications.
• Empirical results demonstrate significant improvements in MSE, PSNR, and NMSE in wireless communications, semantic transmission, and error correction tasks.

A Conditional Diffusion Decoding Module (CDDM) is a neural module that implements a denoising diffusion probabilistic model whose reverse process is conditional on structured side information—such as channel state, semantic latents, quantized content, syndrome, or physical-layer observations. Originating in both information-theoretic image compression and wireless physical layer inference, CDDMs have emerged as highly flexible decoders that use iterative score-based denoising, steered by auxiliary or context variables, to approach statistically optimal signal recovery under realistic, non-Gaussian, and often highly structured uncertainty (Wu et al., 2023).

1. Core Mathematical Framework

A CDDM leverages the Markovian forward–reverse diffusion paradigm, adapting it to conditional (contextual) inference. The diffusion process consists of:

Forward (noising) process: Given a target $x_0$ (transmitted symbol, image, codeword, or latent), the process iteratively corrupts $x_0$ over $T$ steps: $$x_t = \sqrt{\alpha_t} x_{t-1} + \sqrt{1-\alpha_t}\, W_n\, \epsilon,$$ where $W_n$ is typically a context-dependent scaling (e.g., channel-dependent) and $\epsilon \sim \mathcal N(0, I)$. The closed-form marginal is

$$q(x_t|x_0, h_r) = \mathcal N\left(x_t; \sqrt{\bar{\alpha}_t} x_0, (1-\bar{\alpha}_t) W_n^2\right),$$

with $\bar{\alpha}_t = \prod_{i=1}^t \alpha_i$ (Wu et al., 2023).

Reverse (denoising) process: The conditional denoiser parameterizes

$$p_\theta(x_{t-1}|x_t, c) = \mathcal N\left(x_{t-1}; \mu_\theta(x_t, c, t), \sigma_t^2 I\right),$$

where $c$ is the conditioning variable (e.g., fading vector, JSCC or semantic latents), and

$$\mu_\theta(x_t, c, t) = \frac{1}{\sqrt{\alpha_t}}\left(x_t - \sqrt{1-\alpha_t} W_n \epsilon_\theta(x_t, c, t)\right).$$

The $\epsilon$-prediction network is trained to minimize MSE between the true and predicted noise or a denoising target, using a loss of the form

$$L_\text{CDDM}(\theta) = \mathbb E_{x_0, \epsilon \sim \mathcal N(0, I), t} \Big\| \epsilon - \epsilon_\theta(\sqrt{\bar{\alpha}_t} x_0 + \sqrt{1-\bar{\alpha}_t} W_n \epsilon, c, t) \Big\|^2,$$

with context injection throughout the network (Wu et al., 2023).

2. Conditioning Mechanisms

CDDMs condition the reverse process by various mechanisms, depending on application domain:

• Channel-aware conditioning: In wireless applications, $W_n$ and other parameters depend on the channel's fading vector $h_r$. CDDMs inject $h_r$ either through explicit input concatenation or normalized functions within the U-Net (Wu et al., 2023, Wu et al., 2023).
• Semantic/content guidance: In semantic communication and image compression, low-rate VAE or transform-coded latents $z_c$ are broadcast, concatenated, or injected via additional control modules into every diffusion denoiser resolution. Some implementations use additive skip (“zero-conv”) conditioning (Li et al., 2024).
• Multimodal fusion: For cell-free ISAC, sensing-derived embeddings and UE locations are fused via a multimodal transformer and provided as conditioning context for the MLP denoiser (Farzanullah et al., 7 Jun 2025).
• Syndrome or parity-based guidance: In error correction, syndrome weight or parity error count is mapped into embeddings and injected via FiLM-style modulation into the denoising backbone (Choukroun et al., 2022).

Typical conditioning approaches include:

• Concatenation at channel/spatial level in the U-Net.
• Affine or FiLM bias/gain modulation in normalization layers.
• Additive injection after spatial broadcast and MLP expansion.
• “Plug-in” control networks with reduced channel width projecting into the main denoiser blocks as in ControlNet-style architectures.

3. Architectural and Algorithmic Implementations

The CDDM backbone is typically a U-Net or an MLP, with stepwise processing as follows:

Component	Role	Conditioning
U-Net backbone	Main noise denoiser	Channel features, semantic
Control/MLP	Modality fusion	Multimodal, content latents
Time embedding	Stepwise control	Sinusoidal/MLP at each $t$
Input reshaping	Domain adaptivity	Preprocessing for structure

For sampling (inference), the mapping proceeds backward from an observed (possibly noisy) target $y_r$ or initial noise $x_T$, iteratively applying the learned reverse steps: $$x_{t-1} = \frac{1}{\sqrt{\alpha_t}}\left(x_t - \sqrt{1-\alpha_t} W_n \epsilon_\theta(x_t, c, t)\right)$$ with context $c$ injected as described above.

CDDMs allow acceleration via DDIM-style step reduction or “short-chain” initialization: for MMSE-equalized channels, the process initializes at a step $m$ corresponding to the post-equalizer variance and only executes $m$ reverse steps (Wu et al., 2023).

4. Application Domains

• Wireless Communications: CDDM post-processing following MMSE equalization mitigates residual channel noise, yielding up to +3.55 dB MSE improvement at low SNR in AWGN channels and PSNR gains in semantic image transmission beyond state-of-the-art codecs and JSCC systems (Wu et al., 2023, Wu et al., 2023).
• Semantic/JSCC Transmission: Integrated into end-to-end semantic communication systems, CDDMs enhance symbolic-to-image translation and improve perceptual metrics such as SSIM and LPIPS compared to standard autoencoders and VAEs (Letafati et al., 26 Sep 2025).
• Extreme Image Compression: In transform coding, CDDMs function as learned non-Gaussian decoders reconstructing high-frequency “texture” from low-rate content latents, achieving large BD-rate and perceptual gains versus deterministic decoders (Li et al., 2024, Yang et al., 2022).
• Error Correction: For BPSK linear codes, the forward diffusion models channel corruption, while the reverse CDDM iteratively reduces syndrome error, outperforming belief propagation and prior neural decoders both in BER and latency (Choukroun et al., 2022).
• ISAC Channel Estimation: Multimodal CDDMs fuse radar-based sensing and UE locations to jointly denoise LS estimates, achieving 8–9 dB NMSE improvements over LS/MMSE estimators and 27.8% over non-conditional DDMs (Farzanullah et al., 7 Jun 2025).
• General PHY Layer Tasks: CDDMs in frameworks like CoDiPhy generalize to detection, estimation, and predistortion, with U-Net denoisers guided by conditional encoders over side information, pilots, or physical observations, attaining near-LMMSE performance for OFDM and 6 dB gains for phase-noise estimation (Neshaastegaran et al., 13 Mar 2025).

5. Theoretical Guarantees and Consistency

CDDMs grounding in conditional score-matching and variational inference enables entropy reduction guarantees and estimator consistency under mild conditions. Theoretical analysis shows that for bounded MSE predictors, each sampling step reduces the conditional entropy of $x_{t-1}$ given $(x_0, h)$, up to a critical index (Wu et al., 2023). In semantic communication, M-estimation theory confirms that network minimizers converge in probability to the true minimizer as sample size increases (Letafati et al., 26 Sep 2025).

6. Quantitative Performance and Complexity

Empirical results across applications consistently demonstrate statistically meaningful gains. Key findings include:

• Wireless CDDM: +0.49 dB (SNR=20 dB) to +3.55 dB (SNR=5 dB) MSE improvements after MMSE equalization; up to 1.06 dB PSNR improvement over JSCC systems (Wu et al., 2023).
• ISAC Channel Estimation: 8–9 dB NMSE gain over LS/MMSE; 27.8% improvement over non-conditional DDMs (Farzanullah et al., 7 Jun 2025).
• Compression: Up to +35.77% BD-rate reduction and large perceptual improvements with ControlNet-style injection into frozen diffusion backbones (Li et al., 2024).
• Error Correction: 1–4 nats negative log BER improvement (Polar, LDPC, BCH); convergence in 1–3 reverse steps, matching optimal syndrome in ECC tasks (Choukroun et al., 2022).

Complexity is governed by the number of reverse steps $m$ (typically $\leq$100) and the per-sample network cost (single U-Net or MLP pass). Typically, sub-second inference per instance on modern hardware is reported (Wu et al., 2023, Wu et al., 2023).

7. Integration and Design Considerations

CDDMs are modular and act as “plug-in” denoising/decoding blocks. Key design choices:

• Conditioning type: Direct broadcast or control module, attention-based fusion, or FiLM modulation.
• Scheduled reverse-step initialization: Adaptive to observation-specific noise or channel variance.
• Staging in system pipelines: Multi-phase training (e.g., train encoder/decoder, then CDDM, then fine-tune decoder) is typical in JSCC systems (Wu et al., 2023, Wu et al., 2023).
• Adaptivity and robustness: CDDMs can be trained to handle variable channel conditions, bandwidth regimes, or interference scenarios by including relevant context or training adaptively (Letafati et al., 26 Sep 2025).

A plausible implication is that CDDMs, due to their statistical adaptivity and plug-and-play nature, are increasingly replacing classical Gaussian decoders in structured communication and perception tasks, providing a generic methodology for learning conditional posteriors under complex or multimodal uncertainty (Wu et al., 2023, Farzanullah et al., 7 Jun 2025, Li et al., 2024).