Text-to-Image (T2I) diffusion models have made remarkable advancements in generative modeling; however, they face a trade-off between inference speed and image quality, posing challenges for efficient deployment. Existing distilled T2I models can generate high-fidelity images with fewer sampling steps, but often struggle with diversity and quality, especially in one-step models. From our analysis, we observe redundant computations in the UNet encoders. Our findings suggest that, for T2I diffusion models, decoders are more adept at capturing richer and more explicit semantic information, while encoders can be effectively shared across decoders from diverse time steps. Based on these observations, we introduce the first Time-independent Unified Encoder (TiUE) for the student model UNet architecture, which is a loop-free image generation approach for distilling T2I diffusion models. Using a one-pass scheme, TiUE shares encoder features across multiple decoder time steps, enabling parallel sampling and significantly reducing inference time complexity. In addition, we incorporate a KL divergence term to regularize noise prediction, which enhances the perceptual realism and diversity of the generated images. Experimental results demonstrate that TiUE outperforms state-of-the-art methods, including LCM, SD-Turbo, and SwiftBrushv2, producing more diverse and realistic results while maintaining the computational efficiency. https://github.com/sen-mao/Loopfree

The diffusion models, in early stages focus on constructing basic image structures, while the refined details, including local features and textures, are generated in later stages. Thus the same network layers are forced to learn both structural and textural information simultaneously, significantly differing from the traditional deep learning architectures (e.g., ResNet or GANs) which captures or generates the image semantic information at different layers. This difference inspires us to explore the time-wise diffusion models. We initially investigate the key contributions of the U-Net parameters to the denoising process and identify that properly zeroing out certain parameters (including large parameters) contributes to denoising, substantially improving the generation quality on the fly. Capitalizing on this discovery, we propose a simple yet effective method—termed "MaskUNet"— that enhances generation quality with negligible parameter numbers. Our method fully leverages timestep- and sample-dependent effective U-Net parameters. To optimize MaskUNet, we offer two fine-tuning strategies: a training-based approach and a training-free approach, including tailored networks and optimization functions. In zero-shot inference on the COCO dataset, MaskUNet achieves the best FID score and further demonstrates its effectiveness in downstream task evaluations. Project page: https://gudaochangsheng.github.io/MaskUnet-Page/

REPA and its variants effectively mitigate training challenges in diffusion models by incorporating external visual representations from pretrained models, through alignment between the noisy hidden projections of denoising networks and foundational clean image representations. We argue that the external alignment, which is absent during the entire denoising inference process, falls short of fully harnessing the potential of discriminative representations. In this work, we propose a straightforward method called $\textit{$\textbf{R}$epresentation $\textbf{E}$ntanglement for $\textbf{G}$eneration}$ ($\textbf{REG}$), which entangles low-level image latents with a single high-level class token from pretrained foundation models for denoising. REG acquires the capability to produce coherent image-class pairs directly from pure noise, substantially improving both generation quality and training efficiency. This is accomplished with negligible additional inference overhead, requiring only one single additional token for denoising (<0.5\% increase in FLOPs and latency). The inference process concurrently reconstructs both image latents and their corresponding global semantics, where the acquired semantic knowledge actively guides and enhances the image generation process. On ImageNet 256$\times$256, SiT-XL/2 + REG demonstrates remarkable convergence acceleration, achieving $\textbf{63}\times$ and $\textbf{23}\times$ faster training than SiT-XL/2 and SiT-XL/2 + REPA, respectively. More impressively, SiT-L/2 + REG trained for merely 400K iterations outperforms SiT-XL/2 + REPA trained for 4M iterations ($\textbf{10}\times$ longer). Code is available at: https://github.com/Martinser/REG.

In this paper, we explore optimal subsampling strategies for estimating the parametric regression coefficients in partially linear models with unknown nuisance functions involving high-dimensional and potentially endogenous covariates. To address model misspecifications and the curse of dimensionality, we leverage flexible machine learning (ML) techniques to estimate the unknown nuisance functions. By constructing an unbiased subsampling Neyman-orthogonal score function, we eliminate regularization bias. A two-step algorithm is then used to obtain appropriate ML estimators of the nuisance functions, mitigating the risk of over-fitting. Using martingale techniques, we establish the unconditional consistency and asymptotic normality of the subsample estimators. Furthermore, we derive optimal subsampling probabilities, including A-optimal and L-optimal probabilities as special cases. The proposed optimal subsampling approach is extended to partially linear instrumental variable models to account for potential endogeneity through instrumental variables. Simulation studies and an empirical analysis of the Physicochemical Properties of Protein Tertiary Structure dataset demonstrate the superior performance of our subsample estimators.

For a relatively small labeled dataset from high-dimensional generalized linear models with block-wise missing covariates and a large unlabeled dataset, we utilize a model-assisted approach in the labeled dataset to address the issue of block-wise missing covariates and then integrate the unlabeled data to construct estimation equations for the coefficients without any imputation. A lasso-penalized semi-supervised estimator is obtained, and then its debiased estimator is proposed to establish asymptotic normality/confidence intervals. When the labeled data are distributed in multiple machines independently and only some machines have unlabeled data, we further propose a distributed debiased semi-supervised estimator for estimation and inference. The finite sample performance of our proposed two estimators is studied through simulations and further illustrated with a breast cancer dataset.

Recently, intelligent fault diagnosis has become the focus of research. The powerful feature extraction capability enables it to show better results for large amounts of data. Because of this, the data requirements of fault diagnosis methods consume a lot of manpower and material resources, especially the data annotation process. Sufficient labeled data cannot be met in real industrial conditions. We propose a novel framework for contrastive learning and semi-supervised learning, where the pseudo labels from the SemiSL model assist in the clustering of the contrast space by additional guidance, thereby improving the feature extraction effect of contrastive learning. Meanwhile, we design a new class-aware loss that can promote the compactness of learning similar representations and the separability of different samples. Through experimental verification on classic bearing datasets, the reliability of the method proposed has been sianificantly improved.

The Proximity Radio Access Network(P-RAN) is effective in minimizing small coverage holes in addition to improving the system capacity in high frequency bands cellular networks of future generations. [1] Proximity Link (PL) is expected to operate at out-of-band high frequencies for simplicity. Before PL can be completely standardized in 3GPP,Wi-Fi Direct(WFD) technology can be used because it enables 5G devices to directly form a single-hop D2D communication network through service discovery, information acquisition, and independent decision-making. In this paper, we design a P-RAN based on WFD protocol for devices to form multi-hop D2D communication network. Each device within this network can offload data to the neighboring devices, and finally relay to the cellular network. In order to study the performance and device energy consumption of such multi-hop P-RAN, we first studied the energy consumption model of 5G devices, and verified the rationality of the model by comparing the device energy consumption output from the actual laboratory test and simulation platform. Then, the performance of the multi-hop P-RAN was comprehensively evaluated. The key performance indicators such as throughput rate, latency and device power consumption in typical application scenarios were obtained, which provided accurate data basis for GO selection, networking scheme and data offloading algorithm optimization involved in P-RAN performance improvement, and to facilitate progress in the proximity network.