Generative Frame Sampler for Long Video Understanding

Despite recent advances in Video Large Language Models (VideoLLMs), effectively understanding long-form videos remains a significant challenge. Perceiving lengthy videos containing thousands of frames poses substantial computational burden. To mitigate this issue, this paper introduces Generative Frame Sampler (GenS), a plug-and-play module integrated with VideoLLMs to facilitate efficient lengthy video perception. Built upon a lightweight VideoLLM, GenS leverages its inherent vision-language capabilities to identify question-relevant frames. To facilitate effective retrieval, we construct GenS-Video-150K, a large-scale video instruction dataset with dense frame relevance annotations. Extensive experiments demonstrate that GenS consistently boosts the performance of various VideoLLMs, including open-source models (Qwen2-VL-7B, Aria-25B, VILA-40B, LLaVA-Video-7B/72B) and proprietary assistants (GPT-4o, Gemini). When equipped with GenS, open-source VideoLLMs achieve impressive state-of-the-art results on long-form video benchmarks: LLaVA-Video-72B reaches 66.8 (+4.3) on LongVideoBench and 77.0 (+2.7) on MLVU, while Aria obtains 39.2 on HourVideo surpassing the Gemini-1.5-pro by 1.9 points. We will release all datasets and models at https://generative-sampler.github.io.

Temporal Reasoning Transfer from Text to Video

Video Large Language Models (Video LLMs) have shown promising capabilities in video comprehension, yet they struggle with tracking temporal changes and reasoning about temporal relationships. While previous research attributed this limitation to the ineffective temporal encoding of visual inputs, our diagnostic study reveals that video representations contain sufficient information for even small probing classifiers to achieve perfect accuracy. Surprisingly, we find that the key bottleneck in Video LLMs' temporal reasoning capability stems from the underlying LLM's inherent difficulty with temporal concepts, as evidenced by poor performance on textual temporal question-answering tasks. Building on this discovery, we introduce the Textual Temporal reasoning Transfer (T3). T3 synthesizes diverse temporal reasoning tasks in pure text format from existing image-text datasets, addressing the scarcity of video samples with complex temporal scenarios. Remarkably, without using any video data, T3 enhances LongVA-7B's temporal understanding, yielding a 5.3 absolute accuracy improvement on the challenging TempCompass benchmark, which enables our model to outperform ShareGPT4Video-8B trained on 28,000 video samples. Additionally, the enhanced LongVA-7B model achieves competitive performance on comprehensive video benchmarks. For example, it achieves a 49.7 accuracy on the Temporal Reasoning task of Video-MME, surpassing powerful large-scale models such as InternVL-Chat-V1.5-20B and VILA1.5-40B. Further analysis reveals a strong correlation between textual and video temporal task performance, validating the efficacy of transferring temporal reasoning abilities from text to video domains.

UBiSS: A Unified Framework for Bimodal Semantic Summarization of Videos

With the surge in the amount of video data, video summarization techniques, including visual-modal(VM) and textual-modal(TM) summarization, are attracting more and more attention. However, unimodal summarization inevitably loses the rich semantics of the video. In this paper, we focus on a more comprehensive video summarization task named Bimodal Semantic Summarization of Videos (BiSSV). Specifically, we first construct a large-scale dataset, BIDS, in (video, VM-Summary, TM-Summary) triplet format. Unlike traditional processing methods, our construction procedure contains a VM-Summary extraction algorithm aiming to preserve the most salient content within long videos. Based on BIDS, we propose a Unified framework UBiSS for the BiSSV task, which models the saliency information in the video and generates a TM-summary and VM-summary simultaneously. We further optimize our model with a list-wise ranking-based objective to improve its capacity to capture highlights. Lastly, we propose a metric, $NDCG_{MS}$, to provide a joint evaluation of the bimodal summary. Experiments show that our unified framework achieves better performance than multi-stage summarization pipelines. Code and data are available at https://github.com/MeiYutingg/UBiSS.

DeCo: Decoupling Token Compression from Semantic Abstraction in Multimodal Large Language Models

The visual projector, which bridges the vision and language modalities and facilitates cross-modal alignment, serves as a crucial component in MLLMs. However, measuring the effectiveness of projectors in vision-language alignment remains under-explored, which currently can only be inferred from the performance of MLLMs on downstream tasks. Motivated by the problem, this study examines the projector module by interpreting the vision-language semantic flow within MLLMs. Specifically, we trace back the semantic relevance flow from generated language tokens to raw visual encoder patches and the intermediate outputs produced by projectors. Our findings reveal that compressive projectors (e.g., QFormer), abstract visual patches into a limited set of semantic concepts, such as objects or attributes, resulting in a 'double abstraction' phenomenon. This involves a first visual semantic abstraction by the projector referring to pre-defined query tokens, and a second extraction by the LLM based on text instructions. The double abstraction is inefficient in training and will result in cumulative vision semantics deficiency. To mitigate this issue, we propose the key insight of 'Decouple Compression from Abstraction (DeCo), that is compressing the visual token number at the patch level by projectors and allowing the LLM to handle visual semantic abstraction entirely. Consequently, we adopt a simple compressor, i.e., 2D Adaptive Pooling, to downsample visual patches in a parameter-free manner. Empirical evaluation demonstrates that DeCo surpasses traditional compressive projectors regarding both performance and efficiency. It achieves performance gains of 0.9%, 7.1%, and 2.9% across the MLLM Benchmarks, Visual Localization, and Open-ended VQA tasks with fewer trainable parameters and faster convergence speed.

LaDiC: Are Diffusion Models Really Inferior to Autoregressive Counterparts for Image-to-Text Generation?

Diffusion models have exhibited remarkable capabilities in text-to-image generation. However, their performance in image-to-text generation, specifically image captioning, has lagged behind Auto-Regressive (AR) models, casting doubt on their applicability for such tasks. In this work, we revisit diffusion models, highlighting their capacity for holistic context modeling and parallel decoding. With these benefits, diffusion models can alleviate the inherent limitations of AR methods, including their slow inference speed, error propagation, and unidirectional constraints. Furthermore, we identify the prior underperformance of diffusion models stemming from the absence of an effective latent space for image-text alignment, and the discrepancy between continuous diffusion processes and discrete textual data. In response, we introduce a novel architecture, LaDiC, which utilizes a split BERT to create a dedicated latent space for captions and integrates a regularization module to manage varying text lengths. Our framework also includes a diffuser for semantic image-to-text conversion and a Back&Refine technique to enhance token interactivity during inference. LaDiC achieves state-of-the-art performance for diffusion-based methods on the MS COCO dataset with 38.2 BLEU@4 and 126.2 CIDEr, demonstrating exceptional performance without pre-training or ancillary modules. This indicates strong competitiveness with AR models, revealing the previously untapped potential of diffusion models in image-to-text generation.

TimeChat: A Time-sensitive Multimodal Large Language Model for Long Video Understanding

This work proposes TimeChat, a time-sensitive multimodal large language model specifically designed for long video understanding. Our model incorporates two key architectural contributions: (1) a timestamp-aware frame encoder that binds visual content with the timestamp of each frame, and (2) a sliding video Q-Former that produces a video token sequence of varying lengths to accommodate videos of various durations. Additionally, we construct an instruction-tuning dataset, encompassing 6 tasks and a total of 125K instances, to further enhance TimeChat's instruction-following performance. Experiment results across various video understanding tasks, such as dense captioning, temporal grounding, and highlight detection, demonstrate TimeChat's strong zero-shot temporal localization and reasoning capabilities. For example, it achieves +9.2 F1 score and +2.8 CIDEr on YouCook2, +5.8 HIT@1 on QVHighlights, and +27.5 R@1 (IoU=0.5) on Charades-STA, compared to state-of-the-art video large language models, holding the potential to serve as a versatile video assistant for long-form video comprehension tasks and satisfy realistic user requirements.

Edit As You Wish: Video Description Editing with Multi-grained Commands

Automatically narrating a video with natural language can assist people in grasping and managing massive videos on the Internet. From the perspective of video uploaders, they may have varied preferences for writing the desired video description to attract more potential followers, e.g. catching customers' attention for product videos. The Controllable Video Captioning task is therefore proposed to generate a description conditioned on the user demand and video content. However, existing works suffer from two shortcomings: 1) the control signal is fixed and can only express single-grained control; 2) the video description can not be further edited to meet dynamic user demands. In this paper, we propose a novel Video Description Editing (VDEdit) task to automatically revise an existing video description guided by flexible user requests. Inspired by human writing-revision habits, we design the user command as a {operation, position, attribute} triplet to cover multi-grained use requirements, which can express coarse-grained control (e.g. expand the description) as well as fine-grained control (e.g. add specified details in specified position) in a unified format. To facilitate the VDEdit task, we first automatically construct a large-scale benchmark dataset namely VATEX-EDIT in the open domain describing diverse human activities. Considering the real-life application scenario, we further manually collect an e-commerce benchmark dataset called EMMAD-EDIT. We propose a unified framework to convert the {operation, position, attribute} triplet into a textual control sequence to handle multi-grained editing commands. For VDEdit evaluation, we adopt comprehensive metrics to measure three aspects of model performance, including caption quality, caption-command consistency, and caption-video alignment.

Rethinking Benchmarks for Cross-modal Image-text Retrieval

Image-text retrieval, as a fundamental and important branch of information retrieval, has attracted extensive research attentions. The main challenge of this task is cross-modal semantic understanding and matching. Some recent works focus more on fine-grained cross-modal semantic matching. With the prevalence of large scale multimodal pretraining models, several state-of-the-art models (e.g. X-VLM) have achieved near-perfect performance on widely-used image-text retrieval benchmarks, i.e. MSCOCO-Test-5K and Flickr30K-Test-1K. In this paper, we review the two common benchmarks and observe that they are insufficient to assess the true capability of models on fine-grained cross-modal semantic matching. The reason is that a large amount of images and texts in the benchmarks are coarse-grained. Based on the observation, we renovate the coarse-grained images and texts in the old benchmarks and establish the improved benchmarks called MSCOCO-FG and Flickr30K-FG. Specifically, on the image side, we enlarge the original image pool by adopting more similar images. On the text side, we propose a novel semi-automatic renovation approach to refine coarse-grained sentences into finer-grained ones with little human effort. Furthermore, we evaluate representative image-text retrieval models on our new benchmarks to demonstrate the effectiveness of our method. We also analyze the capability of models on fine-grained semantic comprehension through extensive experiments. The results show that even the state-of-the-art models have much room for improvement in fine-grained semantic understanding, especially in distinguishing attributes of close objects in images. Our code and improved benchmark datasets are publicly available at: https://github.com/cwj1412/MSCOCO-Flikcr30K_FG, which we hope will inspire further in-depth research on cross-modal retrieval.

CapEnrich: Enriching Caption Semantics for Web Images via Cross-modal Pre-trained Knowledge

Automatically generating textual descriptions for massive unlabeled images on the web can greatly benefit realistic web applications, e.g. multimodal retrieval and recommendation. However, existing models suffer from the problem of generating ``over-generic'' descriptions, such as their tendency to generate repetitive sentences with common concepts for different images. These generic descriptions fail to provide sufficient textual semantics for ever-changing web images. Inspired by the recent success of Vision-Language Pre-training (VLP) models that learn diverse image-text concept alignment during pretraining, we explore leveraging their cross-modal pre-trained knowledge to automatically enrich the textual semantics of image descriptions. With no need for additional human annotations, we propose a plug-and-play framework, i.e CapEnrich, to complement the generic image descriptions with more semantic details. Specifically, we first propose an automatic data-building strategy to get desired training sentences, based on which we then adopt prompting strategies, i.e. learnable and template prompts, to incentivize VLP models to generate more textual details. For learnable templates, we fix the whole VLP model and only tune the prompt vectors, which leads to two advantages: 1) the pre-training knowledge of VLP models can be reserved as much as possible to describe diverse visual concepts; 2) only lightweight trainable parameters are required, so it is friendly to low data resources. Extensive experiments show that our method significantly improves the descriptiveness and diversity of generated sentences for web images. Our code will be released.

Image Difference Captioning with Pre-training and Contrastive Learning

The Image Difference Captioning (IDC) task aims to describe the visual differences between two similar images with natural language. The major challenges of this task lie in two aspects: 1) fine-grained visual differences that require learning stronger vision and language association and 2) high-cost of manual annotations that leads to limited supervised data. To address these challenges, we propose a new modeling framework following the pre-training-finetuning paradigm. Specifically, we design three self-supervised tasks and contrastive learning strategies to align visual differences and text descriptions at a fine-grained level. Moreover, we propose a data expansion strategy to utilize extra cross-task supervision information, such as data for fine-grained image classification, to alleviate the limitation of available supervised IDC data. Extensive experiments on two IDC benchmark datasets, CLEVR-Change and Birds-to-Words, demonstrate the effectiveness of the proposed modeling framework. The codes and models will be released at https://github.com/yaolinli/IDC.