Joint Source-Channel-Generation Coding

• JSCGC is a novel communication framework that shifts from deterministic reconstruction to probabilistic semantic generation using conditional generative models.
• It integrates advanced diffusion and GAN-based architectures to optimize mutual information and ensure perceptually realistic outputs under bandwidth and SNR constraints.
• Experimental results show that JSCGC improves perceptual metrics (LPIPS, DINO) and maintains semantic fidelity compared to traditional JSCC approaches.

Joint Source-Channel-Generation Coding (JSCGC) is a recently introduced framework for communication systems that reframes the traditional source-channel coding paradigm by shifting the focus from deterministic reconstruction under distortion-centric metrics to probabilistic, information-guided semantic generation. JSCGC leverages advanced generative models at the receiver, maximizing the transmission of semantic information, and enables the channel output to parameterize controlled, perceptually realistic sampling from the natural data manifold. This approach generalizes and surpasses distortion-oriented joint source-channel coding (JSCC) approaches, especially in scenarios with severe bandwidth and signal-to-noise ratio (SNR) constraints, and delivers superior perceptual and semantic fidelity compared to conventional methods (Wu et al., 19 Jan 2026, Erdemir et al., 2022).

1. Paradigmatic Shift: From Deterministic Reconstruction to Semantic Generation

Traditional communication systems, including both source–channel separation and neural JSCC, are founded on deterministic reconstruction: compressing a source $x$ to bits or symbols, protecting them across the channel, and minimizing distortion (e.g., MSE) in the point estimate $\hat{x}$ at the receiver. Deep JSCC merges these stages via a neural architecture but remains distortion-oriented, typically under pixel-wise losses (MSE, PSNR), and produces deterministic reconstructions that often lack perceptual realism under tight rate or low SNR regimes (Erdemir et al., 2022).

JSCGC introduces a receiver-side conditional generative model in place of the classical decoder. Rather than yielding a unique $\hat{x}$, the receiver draws samples $\hat{x} \sim q_\theta(x|y)$ from a learned conditional distribution parameterized by the channel output $y$. The transmission goal shifts: maximize the mutual information $I(X; Y)$ under the constraints of the physical channel, with the receiver generating semantically consistent and perceptually realistic outputs by sampling on the learned data manifold, thus allowing for a controlled, probabilistically constrained trade-off between detail-preserving reconstruction and overall semantic fidelity (Wu et al., 19 Jan 2026).

2. Mathematical Foundations and Objective

The JSCGC problem is formally posed as follows: Let $x \in \mathbb{R}^m$ be the source, $f_\theta: \mathbb{R}^m \rightarrow \mathbb{C}^k$ the encoder, and $W$ the noisy channel. The receiver employs a conditional generative model $q_\theta(x|y)$, often instantiated as a diffusion model $G_\theta$.

The objective is

$$\max_{\phi,\,\theta} \; I(X; Y) \quad \text{subject to}\quad d_p\left(p(X),\,p(G_\theta(Y))\right) \leq \zeta$$

where $d_p$ is a perceptual divergence (e.g., LPIPS or adversarial loss), and $G_\theta$ is trained so its marginals match the target data distribution. This design ensures that the communication link is optimized for semantic content transmission and that the generated output remains highly perceptually faithful (Wu et al., 19 Jan 2026).

Within the generative JSCC context, an alternative objective combines MSE and LPIPS into a weighted loss: $$\mathcal{L}(x, \hat{x}) = \lambda \, \mathrm{MSE}(x, \hat{x}) + (1-\lambda)\, \mathrm{LPIPS}(x, \hat{x})$$ where $0 \leq \lambda \leq 1$ trades off pixel-wise fidelity for perceptual similarity (Erdemir et al., 2022).

3. Generative Architectures and Implementation Methodologies

JSCGC systems operationalize these principles by integrating state-of-the-art generative models, particularly pre-trained diffusion or GAN architectures. Two main instantiations are prominent:

• InverseJSCC: Starts from a pre-trained distortion-oriented DeepJSCC pair, followed by an unsupervised denoising process using a pre-trained StyleGAN-2 generator. The process projects noisy DeepJSCC outputs onto the natural data manifold by solving an inverse optimization problem in StyleGAN's latent space with regularization terms to encourage naturalness and semantic consistency (Erdemir et al., 2022).
• GenerativeJSCC: Trains source encoder and receiver jointly with a frozen generative (e.g., StyleGAN-2) backbone. The receiver splits features to project onto latent vector and multi-scale noise spaces for the generator’s synthesis network, enabling faithful and variable reconstructions. Training employs staged optimization and a mixed MSE-LPIPS loss, ensuring outputs maintain both pixel-wise and perceptual alignment with originals (Erdemir et al., 2022).
• Diffusion-based JSCGC: The receiver-side generator $G_\theta$ is realized as a diffusion model, where channel symbols condition the generation process via communication-aware adapters. The forward and reverse diffusion process ensures generated images adhere to the data manifold. During inference, the reverse ODE is solved with the channel-conditioned velocity field, guaranteeing authenticity and controlled variability (Wu et al., 19 Jan 2026).

All schemes maintain the critical bandwidth ratio $\rho = k/m \ll 1$, and are robust across wide SNR ranges through noise-randomized training and adaptive feature modules.

4. Information-Theoretic and Semantic Consistency Limits

JSCGC introduces a new theoretical regime, establishing bounds relating mutual information and semantic fidelity. Under the manifold assumption, where $X$ lies on a $d$-dimensional manifold in $\mathbb{R}^m$, the maximal semantic inconsistency $\delta^*$—the minimal radius below which the probability of semantically inconsistent reconstruction vanishes—is lower-bounded as: $\delta^* \geq C^{1/d} 2^{-\frac{I(X; Y) + 1}{d}}$ This result, derived by combining Kolmogorov–Tikhomirov covering arguments with Fano's inequality, quantifies the trade-off between channel capacity (mutual information) and achievable semantic matching. The implication is that improving $I(X;Y)$ via better coding or higher SNR reduces semantic divergence, but the decay rate is dictated by the intrinsic data dimension $d$—a constraint specific to the high-level semantics, not present in traditional distortion-centric bounds (Wu et al., 19 Jan 2026).

5. Experimental Validation and Empirical Benchmarks

Comprehensive experiments demonstrate that JSCGC delivers significant improvements in perceptual quality (e.g., LPIPS), semantic fidelity (e.g., DINO score), and distributional authenticity (e.g., rFID) compared to both classical separation-based and neural JSCC baselines.

• InverseJSCC achieves LPIPS improvements of $\approx$0.05–0.12 over DeepJSCC across SNRs and maintains PSNR, while demonstrating transferability even when the forward operator is trained out of domain.
• GenerativeJSCC at low BCR ($\rho = 0.0013$) increases PSNR by up to 2 dB, MS-SSIM by 0.10, and reduces LPIPS by up to 0.15 over DeepJSCC.
• Diffusion-based JSCGC (e.g., with Z-Image and MambaJSCC encoders) lowers LPIPS by 30–50%, halves rFID, and increases semantic DINO scores by 10–15% at moderate SNR, with visual outputs retaining semantic object identity and detail under severe channel conditions (Erdemir et al., 2022, Wu et al., 19 Jan 2026).

The following table summarizes experimental results from the key JSCGC methods:

Scheme	PSNR Improvement	LPIPS Reduction	Semantic Metric Gain
InverseJSCC	–	0.05–0.12	Preserves identity/detail
GenerativeJSCC	Up to 2 dB	Up to 0.15	MS-SSIM +0.10
Diffusion JSCGC	Drops (expected)	30–50%	DINO +10–15%; rFID ×0.5

A key observation is that while pixel-level distortion metrics (PSNR) may decrease due to the generative approach, perceptual and semantic metrics improve significantly, validating the central design principle of JSCGC (Erdemir et al., 2022, Wu et al., 19 Jan 2026).

6. Architectural and Training Considerations

Typical JSCGC instantiations for images employ:

• Deep residual and attention-based encoder backbones with spatial downsampling and power normalization to guarantee transmit power constraints.
• Pre-trained GANs (e.g., StyleGAN-2) or diffusion transformers (e.g., Z-Image) at the receiver, with architecture adapters injecting channel outputs into intermediate layers for fine-grained conditioning.
• Training on large-scale datasets (CelebA-HQ, Open Images), evaluation on standard sets (Kodak), and randomized SNR during training to ensure SNR-agnosticity.
• Optimization with mixed pixel-perceptual losses, two-stage receiver training (latent first, then noise scaling), and loss of pixelwise perfection in favor of genuine semantic and perceptual alignment with the natural data distribution.

7. Context, Related Approaches, and Outlook

While JSCC-PAC codes (Zheng et al., 2023) demonstrate empirical approaches to finite-length benchmark attainment in lossless digital settings, JSCGC advances the paradigm for complex, perceptual- and semantic-rich analog sources such as images. Whereas classical and deep JSCC systems target zero-excess distortion or minimal block-error probability, JSCGC explicitly optimizes for semantic channel capacity under perceptual constraints, structurally aligning with modern diffusion and GAN-based learning frameworks.

Ongoing research in JSCGC centers on tightening information-theoretic limits for high-dimensional manifolds, improving generative model adaptability, and extending the approach to broader modalities and tasks that demand semantic-level communication reliability and authenticity. The interplay between channel coding, source generative modeling, and semantic information theory positions JSCGC as a foundational framework for the next generation of perceptual communication systems (Wu et al., 19 Jan 2026, Erdemir et al., 2022).