Recent advancements in visual generation technologies have markedly increased the scale and availability of video datasets, which are crucial for training effective video generation models. However, a significant lack of high-quality, human-centric video datasets presents a challenge to progress in this field. To bridge this gap, we introduce OpenHumanVid, a large-scale and high-quality humancentric video dataset characterized by precise and detailed captions that encompass both human appearance and motion states, along with supplementary human motion conditions, including skeleton sequences and speech audio. To validate the efficacy of this dataset and the associated training strategies, we propose an extension of existing classical diffusion transformer architectures and conduct further pretraining of our models on the proposed dataset. Our findings yield two critical insights: First, the incorporation of a large-scale, high-quality dataset substantially enhances evaluation metrics for generated human videos while preserving performance in general video generation tasks. Second, the effective alignment of text with human appearance, human motion, and facial motion is essential for producing high-quality video outputs. Based on these insights and corresponding methodologies, the straightforward extended network trained on the proposed dataset demonstrates an obvious improvement in the generation of human-centric videos. The source code and the dataset are available at: https://fudan-generative-vision.github.io/OpenHumanVid.

With the increasing integration of renewable energy devices, modern power systems have become complex, exhibiting diverse circuit topologies. Existing FPGA-based accelerators are optimized for fast single-topology simulation but lack the flexibility to handle multiple topology simulations. This paper proposes a scalable heterogeneous CPU-FPGA system designed for efficient simulation of diverse power system topologies. The proposed system enables real-time simulation for variable-scale power systems, distinguishing itself from conventional simulators by leveraging flexible matrix decomposition and a topology-aware approach. Our design specifically achieves the minimal latency of 100ns for both the boost and three-phase voltage source converter (VSC) models, highlighting its superior performance across diverse topologies. Introduced by dynamic fixed-point quantization, our proposal reduces 30.8% LUTs, 26.8% FFs, 22.2% BRAMs and 43.8% DSPs versus a fixed-point implementation with the same computational accuracy.

Generating highly dynamic and photorealistic portrait animations driven by audio and skeletal motion remains challenging due to the need for precise lip synchronization, natural facial expressions, and high-fidelity body motion dynamics. We propose a human-preference-aligned diffusion framework that addresses these challenges through two key innovations. First, we introduce direct preference optimization tailored for human-centric animation, leveraging a curated dataset of human preferences to align generated outputs with perceptual metrics for portrait motion-video alignment and naturalness of expression. Second, the proposed temporal motion modulation resolves spatiotemporal resolution mismatches by reshaping motion conditions into dimensionally aligned latent features through temporal channel redistribution and proportional feature expansion, preserving the fidelity of high-frequency motion details in diffusion-based synthesis. The proposed mechanism is complementary to existing UNet and DiT-based portrait diffusion approaches, and experiments demonstrate obvious improvements in lip-audio synchronization, expression vividness, body motion coherence over baseline methods, alongside notable gains in human preference metrics. Code and data for this paper are at https://github.com/fudan-generative-vision/hallo4.

Breast cancer continues to be a pressing global health concern, emphasizing the essential need for effective diagnostic techniques. Automated Breast Ultrasound Systems (ABUS) provide a promising advance in breast tumor detection, yet they require significant expertise in interpreting 3D ABUS images, a task fraught with distinctive challenges. Although Vision Transformers (ViT) display remarkable potential for image processing, their low inductive bias and significant data requirements pose obstacles, particularly in the data-constrained medical field. To mitigate these issues, we introduce a Mask-Recover strategy for pretraining Transformer models on 3D ABUS images, enhancing model adaptability and reducing the data demands of the ViT model. Moreover, recognizing the risk that ViTs’ average pooling approach may unintentionally mask small but vital features, we propose Dual-CapsViT, an inventive model combining Transformers and Capsule Networks. This integration affords efficient token routing while preserving fine-grained details. To reconcile potential inconsistencies between capsules and tokens, we engineer a novel dual-channel routing algorithm, strengthening the decoder’s performance. We benchmarked our models against well-known standards such as ResNet and ViT for classifying breast tumors in ABUS images. Our models exhibited superior performance, as evidenced by improved accuracy, specificity, and Area Under the Receiver Operating Characteristic Curve (AUC) metrics, thereby affirming Dual-CapsViT’s potential to enhance breast cancer diagnostics.