Zero-Shot Composed Image Retrieval (ZS-CIR) has garnered increasing interest in recent years, which aims to retrieve a target image based on a query composed of a reference image and a modification text without training samples. Specifically, the modification text describes the distinction between the two images. To conduct ZS-CIR, the prevailing methods employ pre-trained image-to-text models to transform the query image and text into a single text, which is then projected into the common feature space by CLIP to retrieve the target image. However, these methods neglect that ZS-CIR is a typicalfuzzy retrieval task, where the semantics of the target image are not strictly defined by the query image and text. To overcome this limitation, this paper proposes a training-free LLM-based Divergent Reasoning and Ensemble (LDRE) method for ZS-CIR to capture diverse possible semantics of the composed result. Firstly, we employ a pre-trained captioning model to generate dense captions for the reference image, focusing on different semantic perspectives of the reference image. Then, we prompt Large Language Models (LLMs) to conduct divergent compositional reasoning based on the dense captions and modification text, deriving divergent edited captions that cover the possible semantics of the composed target. Finally, we design a divergent caption ensemble to obtain the ensemble caption feature weighted by semantic relevance scores, which is subsequently utilized to retrieve the target image in the CLIP feature space. Extensive experiments on three public datasets demonstrate that our proposed LDRE achieves the new state-of-the-art performance.
Zero-Shot Composed Image Retrieval (ZS-CIR) has attracted more attention in recent years, focusing on retrieving a specific image based on a query composed of a reference image and a relative text without training samples. Specifically, the relative text describes the differences between the two images. Prevailing ZS-CIR methods employ image-to-text (I2T) models to convert the query image into a single caption, which is further merged with the relative text by text-fusion approaches to form a composed text for retrieval. However, these methods neglect the fact that ZS-CIR entails considering not only the final similarity between the composed text and retrieved images but also the semantic increment during the compositional editing process. To address this limitation, this paper proposes a training-free method called Semantic Editing Increment for ZS-CIR (SEIZE) to retrieve the target image based on the query image and text without training. Firstly, we employ a pre-trained captioning model to generate diverse captions for the reference image and prompt Large Language Models (LLMs) to perform breadth compositional reasoning based on these captions and relative text, thereby covering the potential semantics of the target image. Then, we design a semantic editing search to incorporate the semantic editing increment contributed by the relative text into the retrieval process. Concretely, we comprehensively consider relative semantic increment and absolute similarity as the final retrieval score, which is subsequently utilized to retrieve the target image in the CLIP feature space. Extensive experiments on three public datasets demonstrate that our proposed SEIZE achieves the new state-of-the-art performance. The code is publicly available at https://github.com/yzy-bupt/SEIZE.
Conventional Video Large Language Models (Video-LLMs) operate under an offline paradigm: they ingest a complete video before producing any response. This post-hoc design fundamentally limits their ability to serve as real-time interactors that perceive, reason, and communicate while the video is still unfolding. Streaming video understanding breaks this constraint by requiring models to process frames as they arrive and generate temporally grounded responses on the fly, thereby transforming passive narrators into active, always-on interactors. In this survey, we present a comprehensive and structured overview of recent advances in streaming video understanding, with a particular focus on Video-LLM-based approaches. We categorize existing methods into two complementary paradigms: proactive streaming models, which autonomously determine when to initiate responses in dynamic visual environments, and reactive streaming models, which focus on how to maintain long-term context and deliver efficient inference upon user queries over unbounded video streams. Together, these two paradigms characterize the core capabilities required for online video interactors: temporal awareness for timely engagement and scalable memory for sustained reasoning. We further provide a systematic review of emerging benchmarks and datasets designed to evaluate streaming-specific capabilities, including multi-turn dialogue, real-time captioning, and proactive response timing. Finally, we identify the key open challenges and outline promising future directions toward building truly interactive, efficient, and scalable streaming video systems. An accompanying resource repository is maintained at https://github.com/sotayang/Awesome-Streaming-Video-Understanding.
Despite significant progress in Video Large Language Models (Video-LLMs) for offline video understanding, existing online Video-LLMs typically struggle to simultaneously process continuous frame-by-frame inputs and determine optimal response timing, often compromising real-time responsiveness and narrative coherence. To address these limitations, we introduce LiveStar, a pioneering live streaming assistant that achieves always-on proactive responses through adaptive streaming decoding. Specifically, LiveStar incorporates: (1) a training strategy enabling incremental video-language alignment for variable-length video streams, preserving temporal consistency across dynamically evolving frame sequences; (2) a response-silence decoding framework that determines optimal proactive response timing via a single forward pass verification; (3) memory-aware acceleration via peak-end memory compression for online inference on 10+ minute videos, combined with streaming key-value cache to achieve 1.53x faster inference. We also construct an OmniStar dataset, a comprehensive dataset for training and benchmarking that encompasses 15 diverse real-world scenarios and 5 evaluation tasks for online video understanding. Extensive experiments across three benchmarks demonstrate LiveStar's state-of-the-art performance, achieving an average 19.5% improvement in semantic correctness with 18.1% reduced timing difference compared to existing online Video-LLMs, while improving FPS by 12.0% across all five OmniStar tasks. Our model and dataset can be accessed at https://github.com/yzy-bupt/LiveStar.