CVPR 2025 Papers

Visual language models (VLMs) have made significant advances in accuracy in recent years. However, their efficiency has received much less attention. This paper introduces NVILA, a family of open VLMs designed to optimize both efficiency and accuracy. Building on top of VILA, we improve its model architecture by first scaling up the spatial and temporal resolutions, and then compressing visual tokens. This "scale-then-compress" approach enables NVILA to efficiently process high-resolution images and long videos. We also conduct a systematic investigation to enhance the efficiency of NVILA throughout its entire lifecycle, from training to deployment. NVILA matches or surpasses the accuracy of many leading open and proprietary VLMs across a wide range of image and video benchmarks. At the same time, it reduces training costs by 1.9-5.1X, prefilling latency by 1.6-2.2X, and decoding latency by 1.2-2.8X.

Visual-Language Tracking (VLT) is emerging as a promising paradigm to bridge the human-machine performance gap. For single objects, VLT broadens the problem scope to text-driven video comprehension. Yet, this direction is still confined to 2D spatial extents, currently lacking the ability to deal with 3D tracking in the confines of monocular video. Unfortunately, advances in 3D tracking mainly rely on expensive sensor inputs, e.g., point clouds, depth measurements, radar. Absence of language counterpart for the outputs of these mildly democratized sensors in the literature also hinders VLT expansion to 3D tracking. Addressing that, we make the first attempt towards extending VLT to 3D tracking based on monocular video. We present a comprehensive framework, introducing (i) the Monocular-Video-based 3D Visual Language Tracking (Mono3DVLT) task, (ii) a large-scale dataset for the task, called Mono3DVLT-V2X, and (iii) a customized neural model for the task. Our dataset is carefully curated, leveraging a Large Langauge Model (LLM) followed by human verification, composing natural language descriptions for 79,158 video sequences aiming at single object tracking, providing 2D and 3D bounding box annotations. Our neural model, termed Mono3DVLT-MT, is the first targeted approach for the Mono3DVLT task. Comprising the pipeline of multi-modal feature extractor, visual-language encoder, tracking decoder and a tracking head, our model sets a strong baseline for the task on Mono3DVLT-V2X. Experimental results show that our method significantly outperforms existing techniques on the Mono3DVLT-V2X dataset. Our dataset and code are available in https://github.com/hongkai-wei/Mono3DVLT.

In this paper, we focus on monolithic Multimodal Large Language Models (MLLMs) that integrate visual encoding and language decoding into a single LLM. In particular, we identify that existing pre-training strategies for monolithic MLLMs often suffer from unstable optimization or catastrophic forgetting. To address this issue, our core idea is to embed a new visual parameter space into a pre-trained LLM, thereby stably learning visual knowledge from noisy data while freezing the LLM. Based on this principle, we present Mono-InternVL, a novel monolithic MLLM that seamlessly integrates a set of visual experts via a multimodal mixture-of-experts structure. Moreover, we propose an innovative pre-training strategy to maximize the visual capability of Mono-InternVL, namely Endogenous Visual Pre-training (EViP). In particular, EViP is designed as a progressive learning process for visual experts, which aims to fully exploit the visual knowledge from noisy data to high-quality data. To validate our approach, we conduct extensive experiments on 16 benchmarks. Experimental results confirm the superior performance of Mono-InternVL over existing monolithic MLLMs on 13 of 16 multimodal benchmarks, e.g., +80 points over Emu3 on OCRBench. Compared to the modular baseline, i.e., InternVL-1.5, Mono-InternVL still retains comparable multimodal performance while reducing up to 67% first token latency. Our project is available at https://internvl.github.io/blog/2024-10-10-Mono-InternVL/.

Previous research on lightweight models has primarily focused on CNNs and Transformer-based designs. CNNs, with their local receptive fields, struggle to capture long-range dependencies, while Transformers, despite their global modeling capabilities, are limited by quadratic computational complexity in high-resolution scenarios. Recently, state-space models have gained popularity in the visual domain due to their linear computational complexity. Despite their low FLOPs, current lightweight Mamba-based models exhibit suboptimal throughput. In this work, we propose the MobileMamba framework, which balances efficiency and performance. We design a three-stage network to enhance inference speed significantly. At a fine-grained level, we introduce the Multi-Receptive Field Feature Interaction (MRFFI) module, comprising the Long-Range Wavelet Transform-Enhanced Mamba (WTE-Mamba), Efficient Multi-Kernel Depthwise Deconvolution (MK-DeConv), and Eliminate Redundant Identity components. This module integrates multi-receptive field information and enhances high-frequency detail extraction. Additionally, we employ training and testing strategies to further improve performance and efficiency. MobileMamba achieves up to 83.6% on Top-1, surpassing existing state-of-the-art methods which is maximum x21| faster than LocalVim on GPU. Extensive experiments on high-resolution downstream tasks demonstrate that MobileMamba surpasses current efficient models, achieving an optimal balance between speed and accuracy.

Learning generalizable visual representations across different embodied environments is essential for effective robotic manipulation in real-world scenarios. However, the limited scale and diversity of robot demonstration data pose a significant challenge. Recent research has explored leveraging large-scale human activity data for pre-training, but the substantial morphological differences between humans and robots introduce a significant human-robot domain discrepancy, hindering the generalization of these models to downstream manipulation tasks. To overcome this, we propose a novel adaptation paradigm that leverages readily available paired human-robot video data to bridge the domain gap. Our method employs a human-robot contrastive alignment loss to align the semantics of human and robot videos, adapting pre-trained models to the robot domain in a parameter-efficient manner. Experiments on 20 simulated tasks across two different benchmarks and five real-world tasks demonstrate significant improvements. These results span both single-task and language-conditioned multi-task settings, evaluated using two different pre-trained models. Compared to existing pre-trained models, our adaptation method improves the average success rate by over 7% across multiple tasks on both simulated benchmarks and real-world evaluations.

The automated analysis of chemical literature holds promise to accelerate discovery in fields such as material science and drug development. In particular, search capabilities for chemical structures and Markush structures (chemical structure templates) within patent documents are valuable, e.g., for prior-art search. Advancements have been made in the automatic extraction of chemical structures from text and images, yet the Markush structures remain largely unexplored due to their complex multi-modal nature. In this work, we present MarkushGrapher, a multi-modal approach for recognizing Markush structures in documents. Our method jointly encodes text, image, and layout information through a Vision-Text-Layout encoder and an Optical Chemical Structure Recognition vision encoder. These representations are merged and used to auto-regressively generate a sequential graph representation of the Markush structure along with a table defining its variable groups. To overcome the lack of real-world training data, we propose a synthetic data generation pipeline that produces a wide range of realistic Markush structures. Additionally, we present M2S, the first annotated benchmark of real-world Markush structures, to advance research on this challenging task. Extensive experiments demonstrate that our approach outperforms state-of-the-art chemistry-specific and general-purpose vision-language models in most evaluation settings. Code, models, and datasets are available.

Recent State Space Models (SSM), especially Mamba, have demonstrated impressive performance in visual modeling and possess superior model efficiency. However, the application of Mamba to visual tasks suffers inferior performance due to three main constraints existing in the sequential model: 1) Casual computing is incapable of accessing global context; 2) Long-range forgetting occurs when computing the current hidden states; 3) Weak spatial structural modeling due to the transformed sequential input. To address these issues, we investigate a simple yet powerful vision task adapter for Mamba models, which consists of two functional modules: Adaptor-T and Adapator-S. When solving the hidden states for SSM, we apply a casual prediction module Adaptor-T to select a set of learnable locations as memory augmentation feature states to ease long-range forgetting issues. Moreover, we leverage Adapator-S, composed of multi-scale dilated convolutional kernels, to enhance the spatial modeling and introduce the image inductive bias into the feature output. Both modules can enlarge the context modeling in casual computing, as the output is enhanced by the inaccessible features. We explore three usages of Mamba-Adaptor: A general visual backbone for various vision tasks; A booster module to raise the performance of pretrained backbones; A highly efficient fine-tuning module that adapts the base model for transfer learning tasks. Extensive experiments verify the effectiveness of Mamba-Adapter in three settings. Notably, our Mamba-Adapter achieves state-of-the-art performance on the ImageNet and COCO benchmarks.

We propose to synthesize high-quality and synchronized audio, given video and optional text conditions, using a novel multimodal joint training framework (MMAudio). In contrast to single-modality training conditioned on (limited) video data only, MMAudio is jointly trained with larger-scale, readily available text-audio data to learn to generate semantically aligned high-quality audio samples. Additionally, we improve audio-visual synchrony with a conditional synchronization module that aligns video conditions with audio latents at the frame level. Trained with a flow matching objective, MMAudio achieves new video-to-audio state-of-the-art among public models in terms of audio quality, semantic alignment, and audio-visual synchronization, while having a low inference time (1.23s to generate an 8s clip) and just 157M parameters. MMAudio also achieves surprisingly competitive performance in text-to-audio generation, showing that joint training does not hinder single-modality performance. Code, models, and demo are available at: hkchengrex.github.io/MMAudio

Multi-view object tracking (MVOT) offers promising solutions to challenges such as occlusion and target loss, which are common in traditional single-view tracking. However, progress has been limited by the lack of comprehensive multi-view datasets and effective cross-view integration methods. To overcome these limitations, we compiled a Multi-View object Tracking (MVTrack) dataset of 234K high-quality annotated frames featuring 27 distinct objects across various scenes. In conjunction with this dataset, we introduce a novel MVOT method, Multi-View Integration Tracker (MITracker), to efficiently integrate multi-view object features and provide stable tracking outcomes. MITracker can track any object in video frames of arbitrary length from arbitrary viewpoints. The key advancements of our method over traditional single-view approaches come from two aspects: (1) MITracker transforms 2D image features into a 3D feature volume and compresses it into a bird's eye view (BEV) plane, facilitating inter-view information fusion; (2) we propose an attention mechanism that leverages geometric information from fused 3D feature volume to refine the tracking results at each view. MITracker outperforms existing methods on the MVTrack and GMTD datasets, achieving state-of-the-art performance.

Multimodal large language models (MLLMs) improve performance on vision-language tasks by integrating visual features from pre-trained vision encoders into large language models (LLMs). However, how MLLMs process and utilize visual information remains unclear. In this paper, a shift in the dominant flow of visual information is uncovered: (1) in shallow layers, strong interactions are observed between image tokens and instruction tokens, where most visual information is injected into instruction tokens to form cross-modal semantic representations; (2) in deeper layers, image tokens primarily interact with each other, aggregating the remaining visual information to optimize semantic representations within the visual modality. Based on these insights, we propose Hierarchical Modality-Aware Pruning (HiMAP), a plug-and-play inference acceleration method that dynamically prunes image tokens at specific layers, reducing computational costs by approximately 65% without sacrificing performance. Our findings offer a new understanding of visual information processing in MLLMs and provide a state-of-the-art solution for efficient inference.

Recent years have seen a significant increase in video content creation and consumption. Crafting engaging content requires the careful curation of both visual and audio elements. While visual cue curation, through techniques like optimal viewpoint selection or post-editing, has been central to media production, its natural counterpart, audio, has not undergone equivalent advancements. This often results in a disconnect between visual and acoustic saliency. To bridge this gap, we introduce a novel task: visually-guided acoustic highlighting, which aims to transform audio to deliver appropriate highlighting effects guided by the accompanying video, ultimately creating a more harmonious audio-visual experience. We propose a flexible, transformer-based multimodal framework to solve this task. To train our model, we also introduce a new dataset--the muddy mix dataset, leveraging the meticulous audio and video crafting found in movies, which provides a form of free supervision. We develop a pseudo-data generation process to simulate poorly mixed audio, mimicking real-world scenarios through a three-step process---separation, adjustment, and remixing. Our approach consistently outperforms several baselines in both quantitative and subjective evaluation. We also systematically study the impact of different types of contextual guidance and difficulty levels of the dataset.

Although text-to-image (T2I) models have recently thrived as visual generative priors, their reliance on high-quality text-image pairs makes scaling up expensive. We argue that grasping the cross-modality alignment is not a necessity for a sound visual generative prior, whose focus should be on texture modeling. Such a philosophy inspires us to study image-to-image (I2I) generation, where models can learn from in-the-wild images in a self-supervised manner. We first develop a pure vision-based training framework, Lumos, and confirm the feasibility and the scalability of learning I2I models. We then find that, as an upstream task of T2I, our I2I model serves as a more foundational visual prior and achieves on-par or better performance than existing T2I models using only 1/10 text-image pairs for fine-tuning. We further demonstrate the superiority of I2I priors over T2I priors on some text-irrelevant vision tasks, like image-to-3D and image-to-video.

Visual imagery does not consist of solitary objects, but in-stead reflects the composition of a multitude of fluid con-cepts. While there have been great advances in visual repre-sentation learning, such advances have focused on buildingbetter representations for a small number of discrete objectsbereft of an understanding of how these objects are inter-acting. One can observe this limitation in representationslearned through captions or contrastive learning - wherethe learned model treats an image essentially as a bag ofwords. Several works have attempted to address this lim-itation through the development of bespoke architectures.In this work, we focus on simple and scalable approaches.In particular, we demonstrate that by improving weakly la-beled data, i.e. captions, we can vastly improve the perfor-mance of standard contrastive learning approaches. Previ-ous CLIP models achieved near chance rate on challengingtasks probing compositional learning. However, our sim-ple approach boosts performance of CLIP substantially andachieves state of the art results on compositional bench-marks such as ARO and SugarCrepe. Furthermore, weshowcase our results on a relatively new captioning bench-mark derived from DOCCI. We demonstrate through a se-ries of ablations that a standard CLIP model trained withenhanced data may demonstrate impressive performance onimage retrieval tasks.

Video-text retrieval, the task of retrieving videos based on a textual query or vice versa, is of paramount importance for video understanding and multimodal information retrieval. Recent methods in this area rely primarily on visual and textual features and often ignore audio, although it helps enhance overall comprehension of video content.Moreover, traditional models that incorporate audio blindly utilize the audio input regardless of whether it is useful or not, resulting in suboptimal video representation. To address these limitations, we propose a novel video-text retrieval framework, Audio-guided VIdeo representation learning with GATEd attention (AVIGATE), that effectively leverages audio cues through a gated attention mechanism that selectively filters out uninformative audio signals.In addition, we propose an adaptive margin-based contrastive loss to deal with the inherently unclear positive-negative relationship between video and text, which facilitates learning better video-text alignment.Our extensive experiments demonstrate that AVIGATE achieves state-of-the-art performance on all the public benchmarks.

Long-term sports assessment is a challenging task in video understanding since it requires judging complex movement variations and action-music coordination. However, there is no direct correlation between the diverse background music and movements in sporting events. Previous works require a large number of model parameters to learn potential associations between actions and music. To address this issue, we propose a language-guided audio-visual learning (MLAVL) framework that models "audio-action-visual" correlations guided by low-cost language modality. In our framework, multidimensional domain-based actions form action knowledge graphs, motivating audio-visual modalities to focus on task-relevant actions. We further design a shared-specific context encoder to integrate deep multimodal semantics, and an audio-visual cross-modal fusion module to evaluate action-music consistency. To match the sport's rules, we then propose a dual-branch prompt-guided grading module to weigh both visual and audio-visual performance. Extensive experiments demonstrate that our approach achieves state-of-the-art on four public long-term sports benchmarks while maintaining low parameters. Our code is available at https://github.com/XuHuangbiao/MLAVL.

Existing state-of-the-art feature matchers capture long-range dependencies with Transformers but are hindered by high spatial complexity,leading to demanding training and high-latency inference.Striking a better balance between performance and efficiency remains a critical challenge in feature matching.Inspired by the linear complexity \mathcal O (N) of Mamba, we propose an ultra-lightweight Mamba-based matcher, named JamMa, which converges on a single GPU and achieves an impressive performance-efficiency balance in inference.To unlock the potential of Mamba for feature matching,we propose Joint Mamba with a scan-merge strategy named JEGO, which enables:(1) Joint scan of two images to achieve high-frequency mutual interaction, (2) Efficient scan with skip steps to reduce sequence length, (3) Global receptive field, and (4) Omnidirectional feature representation.With the above properties, the JEGO strategy significantly outperforms the scan-merge strategies proposed in VMamba and EVMamba in the feature matching task.Compared to attention-based sparse and semi-dense matchers, JamMa demonstrates a notably superior balance between performance and efficiency,delivering better performance with less than 50% of the parameters and FLOPs.

This paper aims to address universal segmentation for image and video perception with the strong reasoning ability empowered by Visual Large Language Models (VLLMs). Despite significant progress in current unified segmentation methods, limitations in adaptation to both image and video scenarios, as well as the complex reasoning segmentation, make it difficult for them to handle various challenging instructions and achieve an accurate understanding of fine-grained vision-language correlations. We propose HyperSeg, the first VLLM-based universal segmentation model for pixel-level image and video perception, encompassing generic segmentation tasks and more complex reasoning perception tasks requiring powerful reasoning abilities and world knowledge. Besides, to fully leverage the recognition capabilities of VLLMs and the fine-grained visual information, HyperSeg incorporates hybrid entity recognition and fine-grained visual perceiver modules for various segmentation tasks. Combined with the temporal adapter, HyperSeg achieves a comprehensive understanding of temporal information. Experimental results validate the effectiveness of our insights in resolving universal image and video segmentation tasks, including the more complex reasoning perception tasks. Our code is available at https://github.com/congvvc/HyperSeg.

State-space models (SSMs) have recently shown promise in capturing long-range dependencies with subquadratic computational complexity, making them attractive for various applications. However, purely SSM-based models face critical challenges related to stability and achieving state-of-the-art performance in computer vision tasks. Our paper addresses the challenges of scaling SSM-based models for computer vision, particularly the instability and inefficiency of large model sizes. We introduce a parameter-efficient modulated group mamba layer that divides the input channels into four groups and applies our proposed SSM-based efficient Visual Single Selective Scanning (VSSS) block independently to each group, with each VSSS block scanning in one of the four spatial directions. The Modulated Group Mamba layer also wraps the four VSSS blocks into a channel modulation operator to improve cross-channel communication. Furthermore, we introduce a distillation-based training objective to stabilize the training of large models, leading to consistent performance gains. Our comprehensive experiments demonstrate the merits of the proposed contributions, leading to superior performance over existing methods for image classification on ImageNet-1K, object detection, instance segmentation on MS-COCO, and semantic segmentation on ADE20K. Our tiny variant with 23M parameters achieves state-of-the-art performance with a classification top-1 accuracy of 83.3% on ImageNet-1K, while being 26% efficient in terms of parameters, compared to the best existing Mamba design of same model size. Code and models are available at: https://github.com/Amshaker/GroupMamba

Visual localization, the task of determining the position and orientation of a camera, typically involves three core components: offline construction of a keyframe database, efficient online keyframes retrieval, and robust local feature matching. However, significant challenges arise when there are large viewpoint disparities between the query view and the database, such as attempting localization in a corridor previously build from an opposing direction. Intuitively, this issue can be addressed by synthesizing a set of virtual keyframes that cover all viewpoints. However, existing methods for synthesizing novel views to assist localization often fail to ensure geometric accuracy under large viewpoint changes. In this paper, we introduce a confidence-aware geometric prior into 2D Gaussian splatting to ensure the geometric accuracy of the scene. Then we can render novel views through the mesh with clear structures and accurate geometry, even under significant viewpoint changes, enabling the synthesis of a comprehensive set of virtual keyframes. Incorporating this geometry-preserving virtual keyframe database into the localization pipeline significantly enhances the robustness of visual localization.