This continuation project investigates semantic communication for future 6G wireless systems, with a focus on efficient, robust, and learning-based transmission of task-relevant information over wireless channels. The project builds on the previous NAISS project “Semantic Communication in 6G”, which supported research on computation-resource-efficient task-oriented communications, multimodal task-oriented communications, and generative wireless image transmission. Semantic communication aims to transmit the meaning or task-relevant information contained in source data, rather than simply reconstructing bit-level information. This paradigm is particularly relevant for future 6G applications such as intelligent sensing, immersive communication, autonomous systems, Internet of Things, and remote monitoring, where the receiver often needs to perform a task based on heterogeneous data such as images, text, audio, video, or multimodal observations. During the previous project period, our work on computation-resource-efficient task-oriented communications was published in IEEE Transactions on Communications. In the continuation project, we will focus on two main research directions. First, we will further investigate multimodal semantic and task-oriented communication. The goal is to design semantic encoders and decoders that can extract compact, complementary, and task-relevant representations from multiple data modalities, while reducing redundant information across modalities. We will study information bottleneck methods, redundancy-aware representation learning, multimodal fusion, and robustness under noisy wireless channels. Second, we will investigate generative semantic communication and wireless image transmission. In particular, we will develop and evaluate generative decoders based on flow matching and related generative models, where the wireless channel is incorporated into the reconstruction process. The aim is to improve visual quality, robustness, and decoding efficiency under AWGN, fading, and MIMO channel conditions. The methodology relies on machine learning and deep learning techniques, including autoencoder-based communication systems, variational information bottleneck methods, multimodal representation learning, and generative models such as flow matching. GPU resources are essential for training the neural encoders and decoders, running repeated experiments under different channel conditions, performing ablation studies, and benchmarking against classical and deep-learning-based communication baselines. The expected outcomes include new algorithms, reproducible experimental results, and journal/conference publications on semantic communication, task-oriented communication, and generative wireless transmission for future 6G systems.