Yuchen Deng

Yuchen Deng, Zidang Cai, Hai-Tao Zheng et al.

Omnimodal large language models (Omni-LLMs) show strong capability in audio-video understanding, but their practical deployment remains limited by high inference cost of long video streams and dense audio sequences. Despite recent progress, existing compression methods for Omni-LLMs typically rely on fixed or native compression units, which can disrupt cross-modal correspondence and the complementary information required for audio-video reasoning, making it difficult to improve inference efficiency while stably preserving performance. To address this, we propose OmniRefine, a training-free two-stage framework for efficient audio-visual token compression in Omni-LLMs. First, Correspondence-Preserving Chunk Refinement refines native chunk boundaries into cross-modally aligned compression units through frame-audio similarity and dynamic programming. Second, Modality-Aware Cooperative Compression jointly compresses video and audio tokens within each refined unit to reduce redundancy while preserving critical evidence. Extensive experiments show that OmniRefine achieves a better efficiency-performance trade-off than strong baselines and maintains stable performance under lower compression ratios. On WorldSense, it still reaches 46.7% accuracy at a 44% token retention ratio, nearly matching the full-token baseline. The code and interface will be released to facilitate further research.

Naibo Wang, Yuchen Deng, Wenjie Feng et al.

Traditional federated learning mainly focuses on parallel settings (PFL), which can suffer significant communication and computation costs. In contrast, one-shot and sequential federated learning (SFL) have emerged as innovative paradigms to alleviate these costs. However, the issue of non-IID (Independent and Identically Distributed) data persists as a significant challenge in one-shot and SFL settings, exacerbated by the restricted communication between clients. In this paper, we improve the one-shot sequential federated learning for non-IID data by proposing a local model diversity-enhancing strategy. Specifically, to leverage the potential of local model diversity for improving model performance, we introduce a local model pool for each client that comprises diverse models generated during local training, and propose two distance measurements to further enhance the model diversity and mitigate the effect of non-IID data. Consequently, our proposed framework can improve the global model performance while maintaining low communication costs. Extensive experiments demonstrate that our method exhibits superior performance to existing one-shot PFL methods and achieves better accuracy compared with state-of-the-art one-shot SFL methods on both label-skew and domain-shift tasks (e.g., 6%+ accuracy improvement on the CIFAR-10 dataset).

Yuchen Deng, Xiuyang Wu, Hai-Tao Zheng et al.

Handling fast, complex, and highly non-linear motion patterns has long posed challenges for video frame interpolation. Although recent diffusion-based approaches improve upon traditional optical-flow-based methods, they still struggle to cover diverse application scenarios and often fail to produce sharp, temporally consistent frames in fine-grained motion tasks such as audio-visual synchronized interpolation. To address these limitations, we introduce BBF (Beyond Boundary Frames), a context-aware video frame interpolation framework, which could be guided by audio/visual semantics. First, we enhance the input design of the interpolation model so that it can flexibly handle multiple conditional modalities, including text, audio, images, and video. Second, we propose a decoupled multimodal fusion mechanism that sequentially injects different conditional signals into a DiT backbone. Finally, to maintain the generation abilities of the foundation model, we adopt a progressive multi-stage training paradigm, where the start-end frame difference embedding is used to dynamically adjust both the data sampling and the loss weighting. Extensive experimental results demonstrate that BBF outperforms specialized state-of-the-art methods on both generic interpolation and audio-visual synchronized interpolation tasks, establishing a unified framework for video frame interpolation under coordinated multi-channel conditioning.

Naibo Wang, Yuchen Deng, Shichen Fan et al.

Federated learning (FL) has attracted considerable interest in the medical domain due to its capacity to facilitate collaborative model training while maintaining data privacy. However, conventional FL methods typically necessitate multiple communication rounds, leading to significant communication overhead and delays, especially in environments with limited bandwidth. One-shot federated learning addresses these issues by conducting model training and aggregation in a single communication round, thereby reducing communication costs while preserving privacy. Among these, one-shot federated ensemble learning combines independently trained client models using ensemble techniques such as voting, further boosting performance in non-IID data scenarios. On the other hand, existing machine learning methods in healthcare predominantly use unimodal data (e.g., medical images or textual reports), which restricts their diagnostic accuracy and comprehensiveness. Therefore, the integration of multi-modal data is proposed to address these shortcomings. In this paper, we introduce FedMME, an innovative one-shot multi-modal federated ensemble learning framework that utilizes multi-modal data for medical image analysis. Specifically, FedMME capitalizes on vision large language models to produce textual reports from medical images, employs a BERT model to extract textual features from these reports, and amalgamates these features with visual features to improve diagnostic accuracy. Experimental results show that our method demonstrated superior performance compared to existing one-shot federated learning methods in healthcare scenarios across four datasets with various data distributions. For instance, it surpasses existing one-shot federated learning approaches by more than 17.5% in accuracy on the RSNA dataset when applying a Dirichlet distribution with ($α$ = 0.3).

Naibo Wang, Yuchen Deng, Wenjie Feng et al.

Federated Class Incremental Learning (FCIL) is a critical yet largely underexplored issue that deals with the dynamic incorporation of new classes within federated learning (FL). Existing methods often employ generative adversarial networks (GANs) to produce synthetic images to address privacy concerns in FL. However, GANs exhibit inherent instability and high sensitivity, compromising the effectiveness of these methods. In this paper, we introduce a novel data-free federated class incremental learning framework with diffusion-based generative memory (DFedDGM) to mitigate catastrophic forgetting by generating stable, high-quality images through diffusion models. We design a new balanced sampler to help train the diffusion models to alleviate the common non-IID problem in FL, and introduce an entropy-based sample filtering technique from an information theory perspective to enhance the quality of generative samples. Finally, we integrate knowledge distillation with a feature-based regularization term for better knowledge transfer. Our framework does not incur additional communication costs compared to the baseline FedAvg method. Extensive experiments across multiple datasets demonstrate that our method significantly outperforms existing baselines, e.g., over a 4% improvement in average accuracy on the Tiny-ImageNet dataset.

Shunkai Zhou, Zike Yan, Fei Xue et al.

We present Online3R, a new sequential reconstruction framework that is capable of adapting to new scenes through online learning, effectively resolving inconsistency issues. Specifically, we introduce a set of learnable lightweight visual prompts into a pretrained, frozen geometry foundation model to capture the knowledge of new environments while preserving the fundamental capability of the foundation model for geometry prediction. To solve the problems of missing groundtruth and the requirement of high efficiency when updating these visual prompts at test time, we introduce a local-global self-supervised learning strategy by enforcing the local and global consistency constraints on predictions. The local consistency constraints are conducted on intermediate and previously local fused results, enabling the model to be trained with high-quality pseudo groundtruth signals; the global consistency constraints are operated on sparse keyframes spanning long distances rather than per frame, allowing the model to learn from a consistent prediction over a long trajectory in an efficient way. Our experiments demonstrate that Online3R outperforms previous state-of-the-art methods on various benchmarks. Project page: https://shunkaizhou.github.io/online3r-1.0/

Yuchen Deng, Haibin Ling, Bingyao Huang

We propose LAPIG, a language guided projector image generation method with surface adaptation and stylization. LAPIG consists of a projector-camera system and a target textured projection surface. LAPIG takes the user text prompt as input and aims to transform the surface style using the projector. LAPIG's key challenge is that due to the projector's physical brightness limitation and the surface texture, the viewer's perceived projection may suffer from color saturation and artifacts in both dark and bright regions, such that even with the state-of-the-art projector compensation techniques, the viewer may see clear surface texture-related artifacts. Therefore, how to generate a projector image that follows the user's instruction while also displaying minimum surface artifacts is an open problem. To address this issue, we propose projection surface adaptation (PSA) that can generate compensable surface stylization. We first train two networks to simulate the projector compensation and project-and-capture processes, this allows us to find a satisfactory projector image without real project-and-capture and utilize gradient descent for fast convergence. Then, we design content and saturation losses to guide the projector image generation, such that the generated image shows no clearly perceivable artifacts when projected. Finally, the generated image is projected for visually pleasing surface style morphing effects. The source code and video are available on the project page: https://Yu-chen-Deng.github.io/LAPIG/.

Zimo Cao, Yuchen Deng, Haibin Ling et al.

Spatial augmented reality (SAR) directly projects digital content onto physical scenes using projectors, creating immersive experience without head-mounted displays. However, for SAR to support intelligent interaction, such as reasoning about the scene or answering user queries, it must semantically distinguish between the physical scene and the projected content. Standard Vision Language Models (VLMs) struggle with this virtual-physical ambiguity, often confusing the two contexts. To address this issue, we introduce ProCap, a novel framework that explicitly decouples projected content from physical scenes. ProCap employs a two-stage pipeline: first it visually isolates virtual and physical layers via automated segmentation; then it uses region-aware retrieval to avoid ambiguous semantic context due to projection distortion. To support this, we present RGBP (RGB + Projections), the first large-scale SAR semantic benchmark dataset, featuring 65 diverse physical scenes and over 180,000 projections with dense, decoupled annotations. Finally, we establish a dual-captioning evaluation protocol using task-specific tokens to assess physical scene and projection descriptions independently. Our experiments show that ProCap provides a robust semantic foundation for future SAR research. The source code, pre-trained models and the RGBP dataset are available on the project page: https://ZimoCao.github.io/ProCap/.

Yuchen Deng, Xiuyang Wu, Hai-Tao Zheng et al.

Talking-head animation focuses on generating realistic facial videos from audio input. Following Generative Adversarial Networks (GANs), diffusion models have become the mainstream, owing to their robust generative capacities. However, inherent limitations of the diffusion process often lead to inter-frame flicker and slow inference, restricting their practical deployment. To address this, we introduce AvatarSync, an autoregressive framework on phoneme representations that generates realistic and controllable talking-head animations from a single reference image, driven directly by text or audio input. To mitigate flicker and ensure continuity, AvatarSync leverages an autoregressive pipeline that enhances temporal modeling. In addition, to ensure controllability, we introduce phonemes, which are the basic units of speech sounds, and construct a many-to-one mapping from text/audio to phonemes, enabling precise phoneme-to-visual alignment. Additionally, to further accelerate inference, we adopt a two-stage generation strategy that decouples semantic modeling from visual dynamics, and incorporate a customized Phoneme-Frame Causal Attention Mask to support multi-step parallel acceleration. Extensive experiments conducted on both Chinese (CMLR) and English (HDTF) datasets demonstrate that AvatarSync outperforms existing talking-head animation methods in visual fidelity, temporal consistency, and computational efficiency, providing a scalable and controllable solution.