In this paper, we propose FreqTS, a novel Frequency-Aware Token Selection approach for accelerating diffusion models without requiring retraining. Diffusion models have gained significant attention in the field of image synthesis due to their impressive generative capabilities. However, these models often suffer from high computational costs, primarily due to the sequential denoising process and large model size. Additionally, diffusion models tend to prioritize low-frequency features, leading to sub-optimal quantitative results. To address these challenges, FreqTS introduces an amplitude-based sorting method that separates Token features in the frequency domain of diffusion models into high-frequency and low-frequency subsets. It then utilizes fast Token Selection to reduce the presence of low-frequency features, effectively reducing the computational overhead. Moreover, FreqTS incorporates a Bayesian hyper-parameter search to dynamically assign different selection strategies for various denoising processes. Extensive experiments conducted on Stable Diffusion series models, PixArt-Alpha, LCM, and other models demonstrate that FreqTS achieves a minimum acceleration of 2.3× without the need for retraining. Furthermore, FreqTS showcases its versatility by being applicable to different sampling techniques and compatible with other dimension-specific acceleration algorithms.

Large Multimodal Models (LMMs) are increasingly essential in various real-time applications, yet their substantial parameter counts and complex architectures pose significant challenges. Traditional global compression methods often rely on trial-and-error experimentation, leading to inefficiencies. In this paper, we introduce new GS-MM, an Efficient Global Sparsification technique tailored for resource-conscious multimodal models. GS-MM assigns global sparsification strategies by extracting primitive importance from the model’s components. We first derive the importance of elements from the mapping values of fully weighted activations, based on the weights of the elements. Subsequently, we compute the average multimodal importance to establish a global importance score. This score is then linearly mapped to determine the global allocation ratio, enabling the realization of global sparsity in LMMs. We demonstrate the effectiveness of our approach through extensive experiments on diverse benchmarks, including visual question-answering and reasoning tasks. Our pruned models consistently outperform conventional pruning methods, setting new standards for compressed model performance. Notably, our approach exhibits remarkable resilience to increasing sparsity ratios, preserving model quality even under extreme compression.

Recently, intelligent reflecting surface (IRS) emerges as an effective technique for saving power consumption by customizing the wireless propagation environment. On the other hand, the symbol level precoding (SLP) technique provides a clever solution to interference exploitation by converting the multiuser interference (MUI) into a beneficial part of the desired signal. In this paper, we propose to jointly exploit IRS and SLP to cope with the power control and interference management issues in a heterogeneous network (HetNet). Considering the possible coordination between the macro base station (MBS) and the pico base station (PBS), we propose two corresponding schemes to manage the inter-cell and intra-cell interference. For both proposed schemes, the power minimization problems are studied by jointly optimizing the precoding matrices at the MBS and PBS as well as reflecting coefficients at the IRS. Due to the non-convexity of these problems, the precoding matrices and reflecting coefficients are optimized alternately. We propose two Lagrangian based algorithms to obtain the optimal solutions of the precoding matrices, where the precoding matrix of the MBS always yields a closed-form. A multiple-gradient descent algorithm based on the Riemannian manifold (MGD-RM) is proposed as well to enhance the received signal quality of each MUE and PUE for the reflecting design. Simulation results manifest a significant performance gain achieved by our proposed HetNet over benchmarks.

The non-orthogonal multiple access (NOMA) technique can provide higher spectral efficiency and massive connectivities to wireless networks. The wireless power transfer (WPT) technique is a controllable and promising way to solve the energy scarcity problem of wireless devices. In this paper, we introduce the NOMA and WPT techniques into a multi-user wireless network, where multiple users need to transmit their information to an information receiver within a very limited spectrum and they suffer from the energy scarcity problem. We consider a harvest-then-transmit protocol by dividing each transmission block into two time-slots. In the first time-slot, a power station sends dedicated energy to the users via wireless energy beamforming. In the second time-slot, using their harvested energy in the previous time-slot, the users transmit their information to the information receiver in the NOMA manner. In this network, the throughput can be optimized by jointly optimizing the energy beamforming at the power station, the transmit powers of the users, as well as the time allocation between the two time-slots. We propose an algorithm to find the optimal solution of the throughput maximizing joint energy beamforming and resource allocation problem. Simulation results show that the proposed algorithm achieves higher throughput than the benchmark scheme.