Lightweight Diffusion Models for Resource-Constrained Semantic Communication

Giovanni Pignata, Eleonora Grassucci, Giordano Cicchetti, Danilo Comminiello

Abstract

Recently, generative semantic communication models have proliferated as they are revolutionizing semantic communication frameworks, improving their performance, and opening the way to novel applications. Despite their impressive ability to regenerate content from the compressed semantic information received, generative models pose crucial challenges for communication systems in terms of high memory footprints and heavy computational load. In this paper, we present a novel Quantized GEnerative Semantic COmmunication framework, Q-GESCO. The core method of Q-GESCO is a quantized semantic diffusion model capable of regenerating transmitted images from the received semantic maps while simultaneously reducing computational load and memory footprint thanks to the proposed post-training quantization technique. Q-GESCO is robust to different channel noises and obtains comparable performance to the full precision counterpart in different scenarios saving up to 75% memory and 79% floating point operations. This allows resource-constrained devices to exploit the generative capabilities of Q-GESCO, widening the range of applications and systems for generative semantic communication frameworks. The code is available at https://github.com/ispamm/Q-GESCO.

Introduction

Semantic communication represents a paradigm shift from traditional bit-level transmission to the transmission of "meaning." In this framework, a sender transmits a compressed semantic representation (e.g., a label map), and the receiver employs a generative model—often a diffusion model—to reconstruct the original image. While diffusion models offer high-fidelity reconstruction, their iterative nature and large parameter counts present significant barriers for deployment on resource-constrained edge devices like IoT sensors or mobile phones.

Q-GESCO (Quantized GEnerative Semantic COmmunication) addresses these challenges by applying post-training quantization (PTQ) to diffusion models. This framework slashes memory usage by 75% and computational load by 79% while maintaining image quality, even under noisy channel conditions.

Proposed Method: Q-GESCO

The Q-GESCO framework optimizes the generative process through a four-step pipeline designed for robustness and efficiency. Unlike standard training, PTQ converts model weights from 32-bit floating-point (FP32) to 8-bit integers without requiring expensive retraining.

The Q-GESCO pipeline, illustrating the transition from noise-aware training to quantized inference at the receiver.

The pipeline includes:

  1. Noise-aware Training: The full-precision model is trained to handle semantic map noise, simulating real-world channel interference.
  2. Quantization: Weights are mapped to a discrete 8-bit integer set.
  3. Timestep-aware Calibration: A critical innovation where a calibration dataset samples intermediate denoising steps. This accounts for the shifting activation distributions inherent in diffusion processes. Synthetic Additive White Gaussian Noise (AWGN) is injected during calibration to ensure robustness.
  4. Inference: The final 8-bit calibrated model is deployed for efficient image reconstruction.

Experimental Evaluation

The framework was evaluated using the Cityscapes dataset, which contains urban street scenes at 256 × 512 resolution across 35 semantic classes. This dataset is highly relevant for applications such as autonomous driving and smart city infrastructure.

Key Results

Q-GESCO demonstrates that significant compression does not necessitate a loss in perceptual quality.

  • Efficiency Gains: The model achieved a 75% reduction in memory footprint (from ~2.7 GB to 674 MB) and a 79% reduction in floating-point operations (FLOPs).
  • Perceptual Fidelity: Metrics such as LPIPS and FID remained comparable to the full-precision baseline. Notably, at low noise levels (PSNR 100), Q-GESCO achieved an FID of 15.82, slightly better than the baseline's 16.43.
  • Robustness: Performance remained stable even in highly degraded channels (PSNR 10).
Visual comparison between the full-precision GESCO and the quantized Q-GESCO. The reconstructions are virtually indistinguishable.

Discussion

The observation that the 8-bit model occasionally outperforms the 32-bit version suggests that quantization may act as a regularizer, helping the model focus on essential features. Furthermore, the success of timestep-aware calibration highlights the necessity of accounting for the dynamic nature of diffusion activations during the compression process.

Conclusion and Future Work

Q-GESCO enables high-quality generative semantic communication on edge devices by significantly reducing resource requirements. Future research will focus on reducing the 13-hour calibration time, exploring more complex channel models beyond AWGN (such as fading channels), and scaling the technique to higher-resolution imagery and diverse architectures.

Summary updated

Abstract

Recently, generative semantic communication models have proliferated as they are revolutionizing semantic communication frameworks, improving their performance, and opening the way to novel applications. Despite their impressive ability to regenerate content from the compressed semantic information received, generative models pose crucial challenges for communication systems in terms of high memory footprints and heavy computational load. In this paper, we present a novel Quantized GEnerative Semantic COmmunication framework, Q-GESCO. The core method of Q-GESCO is a quantized semantic diffusion model capable of regenerating transmitted images from the received semantic maps while simultaneously reducing computational load and memory footprint thanks to the proposed post-training quantization technique. Q-GESCO is robust to different channel noises and obtains comparable performance to the full precision counterpart in different scenarios saving up to 75% memory and 79% floating point operations. This allows resource-constrained devices to exploit the generative capabilities of Q-GESCO, widening the range of applications and systems for generative semantic communication frameworks. The code is available at https://github.com/ispamm/Q-GESCO.
BibTeX
@article{pignata2025lightweight,
  author = {Giovanni Pignata and Eleonora Grassucci and Giordano Cicchetti and Danilo Comminiello},
  title = {{Lightweight Diffusion Models for Resource-Constrained Semantic Communication}},
  journal = {IEEE Wireless Communications Letters},
  year = {2025},
  doi = {10.1109/lwc.2025.3578724},
}
Figure