Semantic Generative Communication (SGC)

• Semantic Generative Communication (SGC) is a paradigm that transmits compact semantic representations (e.g., prompts, tokens) so that generative models synthesize task-appropriate outputs.
• SGC architectures integrate semantic encoders, compact payloads, and generative decoders to support diverse applications such as image delivery, TTS synthesis, and remote monitoring.
• Empirical evaluations demonstrate that SGC reduces energy consumption, latency, and data rate while maintaining task-relevant semantic fidelity and perceptual quality.

Semantic Generative Communication (SGC), also termed Generative Semantic Communication (GSC) or Gen SemCom in parts of the literature, is a communication paradigm in which the transmitter sends semantic-level representations—such as prompts, tokens, indices, masks, textual descriptions, codebook entries, or latent variables—and the receiver uses generative models, often together with shared semantic knowledge bases, to synthesize task-appropriate outputs rather than reconstruct the exact source signal. Across the recent literature, SGC has been instantiated for downlink image delivery via text-to-image generation, text-to-speech synthesis, joint image reconstruction and segmentation, remote visual monitoring, prompt-based image generation, AGI-driven human-facing interfaces, and receiver-centric video query answering (Lee et al., 2023, Zheng et al., 2024, Yuan et al., 2024, Yang et al., 2023, Grassucci et al., 2023, Yuan et al., 21 Apr 2025, Liu et al., 2024).

1. Conceptual foundations

SGC is defined against two contrasts. Relative to Shannon-style communication, it does not require exact recovery of the original transmitted sequence of bits or symbols; substantial pixel-level or waveform-level distortion can be tolerated provided that the destination data preserves task-relevant semantics and remains perceptually aligned with natural human experience (Yuan et al., 21 Apr 2025). Relative to traditional semantic communication, which primarily focuses on data reconstruction tasks, SGC is explicitly generative: the received signal functions as a prompt or conditional input for a generative model, not merely as a compressed copy of the original data (Zheng et al., 2024).

A formal framing introduced for AGI-driven GSC models semantic information as a semantic graph

$\mathcal{G} = (\mathcal{S}, \mathcal{R}),$

where semantic nodes span multiple abstraction levels and edges encode relations. On this basis, the literature distinguishes a task-relevant induced subgraph for a general-sense task,

$$\mathcal{G}_\mathrm{GSC} \triangleq (\mathcal{S}_\mathrm{GSC}, \mathcal{R}_\mathrm{GSC}),$$

from a human-perception related subgraph,

$$\mathcal{G}_\mathrm{GSC}^\star \triangleq (\mathcal{S}_\mathrm{GSC}^\star, \mathcal{R}_\mathrm{GSC}^\star).$$

The associated design objective is not merely to preserve task semantics with low distortion, but also to produce a human-friendly interface in forms such as images, video, or text (Yuan et al., 21 Apr 2025).

Another conceptual line emphasizes that SGC relies on a shared prior. In the knowledge-base perspective, source messages are represented in low-dimensional semantic subspaces or semantic manifolds, spanned by semantic metalets or codebooks, and the receiver generates the desired content from compact prompts such as indices and residuals (Ren et al., 2023). In the generative-model perspective, prompts may be layered or sequential, so that partial semantic information can already support coarse semantic reconstruction while additional prompts progressively reduce semantic distortion,

$D_l = D(x, \hat{x}(p_l)), \qquad D_{l-1} \ge D_l,$

with $p_l = \{p_1,\dots,p_l\}$ (Grassucci et al., 2024).

The conceptual shift is therefore twofold: semantics are treated as the primary object of transmission, and generation is treated as the primary mechanism of reconstruction. This suggests that SGC is less a single architecture than a family of AI-native communication systems organized around semantic extraction, compact semantic transport, and receiver-side generative realization.

2. Architectural patterns and shared components

Despite major differences in modality and task, the literature exhibits a recurring architecture: a semantic encoder or semantic knowledge base at the transmitter, a compact semantic payload over the channel, and a generative decoder or generative knowledge base at the receiver (Zheng et al., 2024, Yuan et al., 2024, Ren et al., 2023, Li et al., 2024).

Instantiation	Semantic payload	Generative module
Downlink image delivery	Text prompts, GSI	Text-to-image generator
TTS synthesis	Text tokens, RVQ code indices, residual vector	Diffusion model + SoundStream
Joint image tasks	Task instruction, hierarchical semantic features	Diffusion reconstruction decoder
Remote monitoring	Semantic map, static scene information	Conditional DDPM
Prompt-based inpainting	Semantic mask, textual description	Stable Diffusion inpainting

In downlink image SGC, the base station transmits semantic-level text prompts instead of raw image data to generative users equipped with local text-to-image generators, while non-generative users receive conventional image transmissions. The system therefore combines semantic communication, generative AI at the edge, and joint communication–computation optimization (Lee et al., 2023).

In TTS-oriented SGC, the transmitter sends text tokens, residual semantic vectors, and RVQ code indices derived from WavLM features, while the receiver reconstructs speech semantics through a receiver semantic knowledge base and then synthesizes speech with a diffusion model and SoundStream (Zheng et al., 2024). In multi-task image SGC, task requirements are mapped into task instructions by a Task KB, source KBs select or generate task-specific features, and task-specific JSCC decoders produce either reconstructed images or segmentation masks (Yuan et al., 2024).

A distinct architectural strand uses shared semantic knowledge bases. One formulation splits the KB into source, task, and channel sub-KBs: the source KB provides semantic metalets and a semantic manifold, the task KB parses and plans task requirements, and the channel KB predicts environmental channel knowledge from HD maps and locations (Ren et al., 2023). Another formulation uses a lightweight SKB containing class-level attribute vectors so that only class indices and, when resources allow, low-dimensional latent codes are transmitted (Li et al., 2024).

Receiver-centric designs alter the information flow. In receiver-centric generative semantic communications, each transmission is initialized by the receiver, which first sends its request for the desired semantic information; the transmitter then extracts the required semantic information accordingly, using GPT-4 and specialized tools for detection and estimation (Liu et al., 2024). This makes the semantic payload explicitly query-conditioned rather than transmitter-chosen.

Across these systems, semantic representations take multiple forms: text prompts, semantic masks, segmentation maps, textual descriptions, depth maps, codebook indices, residual vectors, task instructions, attribute prototypes, and latent variables. The common feature is that they are sufficiently compact to reduce communication load while remaining rich enough to condition a strong generative prior.

3. Semantic knowledge, coding, and objective functions

A central design problem in SGC is how to define a semantic representation that is simultaneously compact, task-relevant, and generatively sufficient. Several papers resolve this through shared knowledge structures. In knowledge-base enabled systems, semantics are represented by semantic metalets, codebooks, or attribute prototypes, and the transmitter sends only the indices and residual information needed to instantiate the same semantic manifold at the receiver (Ren et al., 2023, Li et al., 2024). In TTS SGC, the transmitter and receiver share identical RVQ codebooks $\mathbf{C}_i$, and the meaning of a transmitted code index is guaranteed by this shared KB alignment (Zheng et al., 2024).

The optimization criteria vary by application. An AGI-driven formulation introduces Rate–Distortion–Perception theory: $\min \quad R(D,P)$ subject to a semantic distortion constraint

$$d(\hat{\boldsymbol{g}_\mathrm{GSC}}, \boldsymbol{g}_\mathrm{GSC}) \le D$$

and a perceptual quality constraint

$p(\hat{\boldsymbol{s}}) \ge P.$

Here semantic-NMSE serves as the distortion measure for task-relevant semantics, while FID, PIQE, or Kullback–Leibler divergence can quantify perceptual quality (Yuan et al., 21 Apr 2025). This is one of the clearest formal distinctions between SGC and classical rate–distortion formulations, since distortion is defined on semantic space rather than raw signal space.

In energy-aware downlink SGC, the core problem is to decide which users should be generative and which should be non-generative. A binary variable $a_k \in \{0,1\}$ indicates user mode, and the base station minimizes the total energy

$$\mathcal{G}_\mathrm{GSC} \triangleq (\mathcal{S}_\mathrm{GSC}, \mathcal{R}_\mathrm{GSC}),$$0

subject to an overall transmission time constraint

$$\mathcal{G}_\mathrm{GSC} \triangleq (\mathcal{S}_\mathrm{GSC}, \mathcal{R}_\mathrm{GSC}),$$1

Under channel inversion, this becomes a binary integer linear programming problem over user selection (Lee et al., 2023).

In prompt-based edge-device collaborative Gen SemCom, the objective is latency per unit semantic quality. The paper defines

$$\mathcal{G}_\mathrm{GSC} \triangleq (\mathcal{S}_\mathrm{GSC}, \mathcal{R}_\mathrm{GSC}),$$2

and minimizes the maximum CCQ across users by jointly optimizing prompt generation offloading, communication resource allocation, and computation resource allocation (Ren et al., 2024). In diffusion-based inpainting SGC, the optimization target is PSNR under a bandwidth budget, but the key control variable is semantic timeliness. There the notion of a Semantic Deadline is introduced as the minimum time that conditioning data is required to be injected to meet a given performance threshold (Choi et al., 18 Aug 2025).

Another line adopts the information bottleneck. Semantic communications based on adaptive generative models and information bottleneck optimize a compressed semantic representation that preserves task-relevant information under power, rate, and delay constraints, and uses probabilistic generative models to regenerate transmitted images or run classification tasks at the receiver side (Barbarossa et al., 2023).

These objective functions reveal a recurring principle: SGC is rarely optimized for a single metric. The design space is inherently multi-objective, involving semantic fidelity, perceptual realism, communication rate, computation energy, latency, robustness, and, in some systems, user fairness.

4. Modality-specific realizations and empirical evidence

The literature now contains several concrete SGC instantiations with quantitative evaluations. In energy-efficient downlink SGC with text-to-image generators, simulation results corroborate that the proposed generative user selection algorithm reduces total energy by up to 54% compared to a baseline with all non-generative users (Lee et al., 2023). The same study reports that more stringent latency constraints favor fewer offloading/non-generative users, hence more generative users, and that SGC is most attractive when server-side semantic data is reasonably small relative to local data and output size.

In TTS-oriented SGC, the system uses WavLM, residual vector quantization, a transformer encoder, a diffusion model, and SoundStream. Under AWGN and Rayleigh fading, the proposed framework achieves the lowest WER and the highest SPK among all five schemes across SNR, and under a fixed communication budget of 160 Kbits it remains superior to PCM+LDPC+YourTTS, PCM+LDPC+NS2, JSCC+YourTTS, and JSCC+NS2 (Zheng et al., 2024). The same paper reports a computation-and-storage profile of 3878 FLOPs and 873 MB for the proposed SGC, compared with 15213 FLOPs and 1908 MB for PCM+LDPC+NS2, and 21 FLOPs and 406 MB for PCM+LDPC+YourTTS.

In multi-task image SGC, transmitter and receiver each maintain semantic KBs comprising a source KB and a task KB. A unified residual block-based JSCC encoder is combined with a diffusion reconstruction decoder and a ResNet-based segmentation decoder. Experimental results show that this multi-task generative semantic communication system outperforms previous single-task communication systems in terms of peak signal-to-noise ratio and segmentation accuracy (Yuan et al., 2024).

In remote monitoring, semantic change driven generative SemCom transmits semantic maps rather than full images, and a conditional DDPM reconstructs scenes using the semantic map and local static scene information. The paper reports compressed full images of 93 kb, 96 kb, 82 kb, and 128 kb under different weather filters, versus a semantic map of 5 kb; segmentation performance includes average row correct of [99.5%, 71.9%], IoU of [97.9%, 65.6%], and mean IoU of 81.7% (Yang et al., 2023). The reconstructed scenes preserve object positions while allowing colors and details to differ, which is consistent with the semantic-equivalence objective.

In diffusion-guided image SGC, the GESCO framework transmits compressed semantic layouts and reconstructs photorealistic images via a conditional DDPM with SPADE. On Cityscapes resized to $$\mathcal{G}_\mathrm{GSC} \triangleq (\mathcal{S}_\mathrm{GSC}, \mathcal{R}_\mathrm{GSC}),$$3, full image transmission requires $$\mathcal{G}_\mathrm{GSC} \triangleq (\mathcal{S}_\mathrm{GSC}, \mathcal{R}_\mathrm{GSC}),$$4 bits, whereas the proposed semantic-only transmission uses on average $$\mathcal{G}_\mathrm{GSC} \triangleq (\mathcal{S}_\mathrm{GSC}, \mathcal{R}_\mathrm{GSC}),$$5 bits, corresponding to about 92% bit reduction; at PSNR $$\mathcal{G}_\mathrm{GSC} \triangleq (\mathcal{S}_\mathrm{GSC}, \mathcal{R}_\mathrm{GSC}),$$6 dB, GESCO still achieves mIoU $$\mathcal{G}_\mathrm{GSC} \triangleq (\mathcal{S}_\mathrm{GSC}, \mathcal{R}_\mathrm{GSC}),$$7, while several baselines collapse much more sharply (Grassucci et al., 2023).

In SKB-enabled end-to-end SGC for image classification and image generation, the transmitter completes classification locally and communicates only a class index and, when available, a latent vector. With $$\mathcal{G}_\mathrm{GSC} \triangleq (\mathcal{S}_\mathrm{GSC}, \mathcal{R}_\mathrm{GSC}),$$8 classes and $$\mathcal{G}_\mathrm{GSC} \triangleq (\mathcal{S}_\mathrm{GSC}, \mathcal{R}_\mathrm{GSC}),$$9 attributes, evaluation shows classification accuracy around 0.77 and semantic accuracy around 0.84, while at very low compression ratio the system can still generate recognizable, semantically correct images where JPEG+LDPC and Vanilla SemCom fail (Li et al., 2024).

Taken together, these results indicate that SGC is not limited to a single modality or metric. The empirical gains appear in energy, WER/SPK, PSNR, IoU, mIoU, FID, semantic accuracy, transmitted bits, and prompt-conditioned generation quality, depending on the operational objective.

5. Receiver-centric, deadline-aware, and low-latency SGC

A prominent misconception is that semantics can be extracted once at the transmitter and remain meaningful for every receiver. Receiver-centric generative semantic communications explicitly rejects this. It argues that when semantic information is extracted based on criteria at the transmitter alone, critical information of primary concern to the receiver may be lost, making the semantic transmission meaningless to the receiver (Liu et al., 2024). The proposed remedy is a receiver-initiated workflow in which GPT-4 constructs a plan of Video Sampler, Tool Selection, and Analysis; when direct semantic fulfillment is impossible, the transmitter falls back to frame selection. On a dataset of 700 natural-language semantic requests over 100 traffic surveillance video clips, task reflection accuracy exceeds 90% at $$\mathcal{G}_\mathrm{GSC}^\star \triangleq (\mathcal{S}_\mathrm{GSC}^\star, \mathcal{R}_\mathrm{GSC}^\star).$$0, receiver success rate in obtaining semantic information reaches 83.90% overall, and the average reduction is 81.70% in frame count and 66.33% in data size relative to full-video transmission (Liu et al., 2024).

Another latency-focused development is Gen SemCom via textual prompts. Here the semantic content is a prompt produced by a pre-trained M/VLM, and the optimization balances semantic quality against end-to-end latency. With $$\mathcal{G}_\mathrm{GSC}^\star \triangleq (\mathcal{S}_\mathrm{GSC}^\star, \mathcal{R}_\mathrm{GSC}^\star).$$1 Gcycles/s and $$\mathcal{G}_\mathrm{GSC}^\star \triangleq (\mathcal{S}_\mathrm{GSC}^\star, \mathcal{R}_\mathrm{GSC}^\star).$$2 bits, the proposed framework achieves CIDEr $$\mathcal{G}_\mathrm{GSC}^\star \triangleq (\mathcal{S}_\mathrm{GSC}^\star, \mathcal{R}_\mathrm{GSC}^\star).$$3 and latency $$\mathcal{G}_\mathrm{GSC}^\star \triangleq (\mathcal{S}_\mathrm{GSC}^\star, \mathcal{R}_\mathrm{GSC}^\star).$$4 s, compared with CIDEr $$\mathcal{G}_\mathrm{GSC}^\star \triangleq (\mathcal{S}_\mathrm{GSC}^\star, \mathcal{R}_\mathrm{GSC}^\star).$$5 and latency $$\mathcal{G}_\mathrm{GSC}^\star \triangleq (\mathcal{S}_\mathrm{GSC}^\star, \mathcal{R}_\mathrm{GSC}^\star).$$6 s for fully offloaded prompt generation, and CIDEr $$\mathcal{G}_\mathrm{GSC}^\star \triangleq (\mathcal{S}_\mathrm{GSC}^\star, \mathcal{R}_\mathrm{GSC}^\star).$$7 and latency $$\mathcal{G}_\mathrm{GSC}^\star \triangleq (\mathcal{S}_\mathrm{GSC}^\star, \mathcal{R}_\mathrm{GSC}^\star).$$8 s for full on-device prompt generation (Ren et al., 2024). The result is not that offloading is always preferable, but that semantic-aware collaboration between edge and device yields a better latency–quality operating point.

FAST-GSC pushes latency optimization inside the generative pipeline. It parallelizes semantic extraction at the transmitter and inference at the receiver, then adds transmitter-side reinforcement learning to learn temporal dependencies among semantic units and receiver-side semantic difference calculation with sequential conditional denoising. Extensive experiments demonstrate that the proposed architecture achieves a performance score comparable to the conventional GSC architecture while realizing a 52% reduction in residual task latency that extends beyond the fixed inference duration (Wang et al., 2024). The learned temporal prompt engineering policy prioritizes NOUN units early, then VERB units, then STYLE and ADJECTIVE units, and can terminate extraction early when low-value units are discarded.

Deadline-aware diffusion-based SGC for image inpainting introduces a different latency notion. Because semantic mask and textual description traverse separate wireless channels and arrive asynchronously, generation quality depends on when each modality is injected into the diffusion process. The paper defines a Semantic Deadline curve from the set of latest acceptable arrival pairs $$\mathcal{G}_\mathrm{GSC}^\star \triangleq (\mathcal{S}_\mathrm{GSC}^\star, \mathcal{R}_\mathrm{GSC}^\star).$$9 for a target PSNR threshold and proposes a bandwidth allocation scheme that ensures each semantic information can be transmitted within the corresponding semantic deadline (Choi et al., 18 Aug 2025). Experimental results show higher generation performance in terms of PSNR for a given bandwidth than traditional schemes that do not account for semantic deadlines.

These studies together suggest that timeliness in SGC is not merely transport delay. It is often a model-internal timing problem: the semantic payload must arrive at moments when the generative model can still respond effectively.

6. Applications, limitations, and research directions

SGC is already positioned as relevant to 6G semantic and goal-oriented networks, AGI services, metaverse and XR applications, telepresence, online meetings, road monitoring, smart cities, sensor networks, and digital twins (Yuan et al., 21 Apr 2025, Xia et al., 2023). In online-meeting and road-monitoring case studies, AGI-driven GSC uses facial segments plus depth maps, or BLIP-2 video descriptions plus depth maps, together with Stable Diffusion with ControlNet or Control-A-Video to outperform traditional communication in semantic-NMSE and PIQE under equal compressed bit-rate (Yuan et al., 21 Apr 2025). This indicates that SGC can serve both machine-facing tasks and human-facing perceptual reconstruction.

A second misconception is that SGC is simply “compression plus a fancy decoder.” The literature consistently treats shared semantic knowledge, model alignment, and task selection as first-class system elements. GSI in downlink image SGC is uploaded to let the base station coordinate prompt design and user selection (Lee et al., 2023). In TTS SGC, codebooks $D_l = D(x, \hat{x}(p_l)), \qquad D_{l-1} \ge D_l,$0 are identical at transmitter and receiver (Zheng et al., 2024). In KB-enabled systems, KB consistency and update are explicit research problems rather than implementation details (Ren et al., 2023).

The main limitations are also consistent across the literature. Real systems must contend with reliability and fidelity of generative models, hallucination and distribution shift, complexity and storage overhead of diffusion or foundation models, edge deployment constraints, privacy and security, semantic misrepresentation, and the lack of fully mature evaluation metrics. Several papers stress that semantic-NMSE and PIQE are useful but incomplete, and that better semantic similarity metrics, multi-task metrics, and human-aligned perceptual metrics are still needed (Yuan et al., 21 Apr 2025). FAST-GSC shows that reducing latency can compromise task performance without careful sequential design (Wang et al., 2024). Receiver-centric SGC shows that transmitter-centric semantic extraction can omit exactly the information a receiver values most (Liu et al., 2024).

The open research agenda is therefore broad. It includes physical-layer integration with MIMO, beamforming, RIS, and NOMA; lightweight, distributed GSC for edge intelligence; efficient semantic graph compression; robust guarantees on semantic fidelity under generative uncertainty; scalable multi-user orchestration; privacy-preserving and integrity-preserving semantic pipelines; and richer multimodal KBs and generative models (Yuan et al., 21 Apr 2025, Xia et al., 2023). A plausible implication is that future SGC systems will not be evaluated by a single scalar such as BER or PSNR, but by structured operating regions defined by task success, perceptual quality, latency, energy, and knowledge alignment.

In this sense, SGC has emerged as a distinct research area rather than a minor extension of semantic communication. Its unifying principle is stable across formulations: transmit meaning in a compact, model-compatible form, and let a generative receiver realize the destination content under explicitly optimized semantic, perceptual, and system-level constraints.