Beyond Prompts: Preserving Semantics in Diffusion-based Communication

Semantic communication seeks to transmit meaning rather than raw data. Recently, diffusion-based semantic communication has received a huge attention. Yet, recent diffusion-based approaches depend on external prompt generators, which can not preserve fine-grained details or require sending heavy image latents, incurring a robustness–bandwidth trade-off. We propose Textual Inversion-based Semantic Communication (TISC), a novel diffusion-based framework for semantic communication that learns compact, text-aligned latent tokens inside the model via textual inversion. These tokens serve as semantic carriers that condition generation, reducing reliance on external prompting and improving preservation of intended semantics under channel noise. TISC demonstrates consistently superior performance in communicating semantic meaning compared with strong baselines across a wide range of SNRs, and even under severely limited bandwidth.