Zhihang Liu

Yinglu Li, Zhiying Lu, Zhihang Liu et al.

Multi-modal Retrieval-Augmented Generation (RAG) has become a critical method for empowering LLMs by leveraging candidate visual documents. However, current methods consider the entire document as the basic retrieval unit, introducing substantial irrelevant visual content in two ways: 1) Relevant documents often contain large regions unrelated to the query, diluting the focus on salient information; 2) Retrieving multiple documents to increase recall further introduces redundant and irrelevant documents. These redundant contexts distract the model's attention and further degrade the performance. To address this challenge, we propose \modelname, a novel framework that shifts the retrieval paradigm from the document level to the region level. During training, we design a hybrid supervision strategy from both labeled data and unlabeled data to pinpoint relevant patches. During inference, we propose a dynamic pipeline that intelligently groups salient patches into complete semantic regions. By delegating the task of identifying relevant regions to the retriever, \modelname enables the generator to focus solely on concise visual content relevant to queries, improving both efficiency and accuracy. Experiments on six benchmarks demonstrate that RegionRAG achieves state-of-the-art performance. Improves retrieval accuracy by 10.02\% in R@1 on average and increases question answering accuracy by 3.56\% while using only 71.42\% visual tokens compared to prior methods. The code will be available at https://github.com/Aeryn666/RegionRAG.

Zhihang Liu, Jun Li, Hongtao Xie et al.

Video Moment Retrieval (VMR) aims to retrieve temporal segments in untrimmed videos corresponding to a given language query by constructing cross-modal alignment strategies. However, these existing strategies are often sub-optimal since they ignore the modality imbalance problem, \textit{i.e.}, the semantic richness inherent in videos far exceeds that of a given limited-length sentence. Therefore, in pursuit of better alignment, a natural idea is enhancing the video modality to filter out query-irrelevant semantics, and enhancing the text modality to capture more segment-relevant knowledge. In this paper, we introduce Modal-Enhanced Semantic Modeling (MESM), a novel framework for more balanced alignment through enhancing features at two levels. First, we enhance the video modality at the frame-word level through word reconstruction. This strategy emphasizes the portions associated with query words in frame-level features while suppressing irrelevant parts. Therefore, the enhanced video contains less redundant semantics and is more balanced with the textual modality. Second, we enhance the textual modality at the segment-sentence level by learning complementary knowledge from context sentences and ground-truth segments. With the knowledge added to the query, the textual modality thus maintains more meaningful semantics and is more balanced with the video modality. By implementing two levels of MESM, the semantic information from both modalities is more balanced to align, thereby bridging the modality gap. Experiments on three widely used benchmarks, including the out-of-distribution settings, show that the proposed framework achieves a new start-of-the-art performance with notable generalization ability (e.g., 4.42% and 7.69% average gains of R1@0.7 on Charades-STA and Charades-CG). The code will be available at https://github.com/lntzm/MESM.

Zhihang Liu, Chen-Wei Xie, Pandeng Li et al.

Recent Multi-modal Large Language Models (MLLMs) have been challenged by the computational overhead resulting from massive video frames, often alleviated through compression strategies. However, the visual content is not equally contributed to user instructions, existing strategies (\eg, average pool) inevitably lead to the loss of potentially useful information. To tackle this, we propose the Hybrid-level Instruction Injection Strategy for Conditional Token Compression in MLLMs (HICom), utilizing the instruction as a condition to guide the compression from both local and global levels. This encourages the compression to retain the maximum amount of user-focused information while reducing visual tokens to minimize computational burden. Specifically, the instruction condition is injected into the grouped visual tokens at the local level and the learnable tokens at the global level, and we conduct the attention mechanism to complete the conditional compression. From the hybrid-level compression, the instruction-relevant visual parts are highlighted while the temporal-spatial structure is also preserved for easier understanding of LLMs. To further unleash the potential of HICom, we introduce a new conditional pre-training stage with our proposed dataset HICom-248K. Experiments show that our HICom can obtain distinguished video understanding ability with fewer tokens, increasing the performance by 2.43\% average on three multiple-choice QA benchmarks and saving 78.8\% tokens compared with the SOTA method. The code is available at https://github.com/lntzm/HICom.

Nianzu Yang, Pandeng Li, Liming Zhao et al.

Existing video tokenizers typically use the traditional Variational Autoencoder (VAE) architecture for video compression and reconstruction. However, to achieve good performance, its training process often relies on complex multi-stage training tricks that go beyond basic reconstruction loss and KL regularization. Among these tricks, the most challenging is the precise tuning of adversarial training with additional Generative Adversarial Networks (GANs) in the final stage, which can hinder stable convergence. In contrast to GANs, diffusion models offer more stable training processes and can generate higher-quality results. Inspired by these advantages, we propose CDT, a novel Conditioned Diffusion-based video Tokenizer, that replaces the GAN-based decoder with a conditional causal diffusion model. The encoder compresses spatio-temporal information into compact latents, while the decoder reconstructs videos through a reverse diffusion process conditioned on these latents. During inference, we incorporate a feature cache mechanism to generate videos of arbitrary length while maintaining temporal continuity and adopt sampling acceleration technique to enhance efficiency. Trained using only a basic MSE diffusion loss for reconstruction, along with KL term and LPIPS perceptual loss from scratch, extensive experiments demonstrate that CDT achieves state-of-the-art performance in video reconstruction tasks with just a single-step sampling. Even a scaled-down version of CDT (3$\times$ inference speedup) still performs comparably with top baselines. Moreover, the latent video generation model trained with CDT also exhibits superior performance. The source code and pretrained weights are available at https://github.com/ali-vilab/CDT.