CVPR 2025 论文列表

Large multimodal models (LMMs) have achieved impressive progress in vision-language understanding, yet they face limitations in real-world applications requiring complex reasoning over a large number of images. Existing benchmarks for multi-image question-answering are limited in scope, each question is paired with only up to 30 images, which does not fully capture the demands of large-scale retrieval tasks encountered in the real-world usages. To reduce these gaps, we introduce two document haystack benchmarks, dubbed DocHaystack and InfoHaystack, designed to evaluate LMM performance on large-scale visual document retrieval and understanding. Additionally, we propose V-RAG, a novel, vision-centric retrieval-augmented generation (RAG) framework that leverages a suite of multimodal vision encoders, each optimized for specific strengths, and a dedicated question-document relevance module. V-RAG sets a new standard, with a 9% and 11% improvement in Recall@1 on the challenging DocHaystack-1000 and InfoHaystack-1000 benchmarks, respectively, compared to the previous best baseline models. Additionally, integrating V-RAG with LMMs enables them to efficiently operate across thousands of images, yielding significant improvements on our DocHaystack and InfoHaystack benchmarks. Our code and datasets will be made publicly available.

Despite significant progress in multimodal large language models (MLLMs), their performance on complex, multi-page document comprehension remains inadequate, largely due to the lack of high-quality, document-level datasets. While current retrieval-augmented generation (RAG) methods offer partial solutions, they suffer from issues, such as fragmented retrieval contexts, multi-stage error accumulation, and extra time costs of retrieval. In this work, we present a high-quality document-level dataset, Doc-750K, designed to support in-depth understanding of multimodal documents. This dataset includes diverse document structures, extensive cross-page dependencies, and real question-answer pairs derived from the original documents. Building on the dataset, we develop a native multimodal model--Docopilot, which can accurately handle document-level dependencies without relying on RAG. Experiments demonstrate that Docopilot achieves superior coherence, accuracy, and efficiency in document understanding tasks and multi-turn interactions, setting a new baseline for document-level multimodal understanding. Data, code, and models are released at https://github.com/OpenGVLab/Docopilot.

Vision-Language Models (VLMs) excel in diverse visual tasks but face challenges in document understanding, which requires fine-grained text processing. While typical visual tasks perform well with low-resolution inputs, reading-intensive applications demand high-resolution, resulting in significant computational overhead. Using OCR-extracted text in VLM prompts partially addresses this issue but underperforms compared to full-resolution counterpart, as it lacks the complete visual context needed for optimal performance.We introduce DocVLM, a method that integrates an OCR-based modality into VLMs to enhance document processing while preserving original weights. Our approach employs an OCR encoder to capture textual content and layout, compressing these into a compact set of learned queries incorporated into the VLM. Comprehensive evaluations across leading VLMs show that DocVLM significantly reduces reliance on high-resolution images for document understanding.In limited-token regimes (448x448), DocVLM with 64 learned queries improves DocVQA results from 56.0% to 86.6% when integrated with InternVL2 and from 84.4% to 91.2% with Qwen2-VL. In LLaVA-OneVision, DocVLM achieves improved results while using 80% less image tokens. The reduced token usage allows processing multiple pages effectively, showing impressive zero-shot results on DUDE and state-of-the-art performance on MP-DocVQA, highlighting DocVLM's potential for applications requiring high-performance and efficiency.

Document image segmentation is crucial in document analysis and recognition but remains challenging due to the heterogeneity of document formats and diverse segmentation tasks. Existing methods often treat these tasks separately, leading to limited generalization and resource wastage.This paper introduces DocSAM, a transformer-based unified framework for various document image segmentation tasks, including document layout analysis, multi-granularity text segmentation, and table structure recognition by modelling these tasks as a combination of instance and semantic segmentation.Specifically, DocSAM uses a Sentence BERT to map category names from each dataset into semantic queries of the same dimension as instance queries. These queries interact through attention mechanisms and are cross-attended with image features to predict instance and semantic segmentation masks. To predict instance categories, instance queries are dot-producted with semantic queries, and scores are normalized using softmax.As a result, DocSAM can be jointly trained on heterogeneous datasets, enhancing robustness and generalization while reducing computing and storage resources. Comprehensive evaluations show that DocSAM outperforms existing methods in accuracy, efficiency, and adaptability, highlighting its potential for advancing document image understanding and segmentation in various applications.

Text-rich document understanding (TDU) requires comprehensive analysis of documents containing substantial textual content and complex layouts. While Multimodal Large Language Models (MLLMs) have achieved fast progress in this domain, existing approaches either demand significant computational resources or struggle with effective multi-modal integration. In this paper, we introduce DocLayLLM, an efficient multi-modal extension of LLMs specifically designed for TDU. By lightly integrating visual patch tokens and 2D positional tokens into LLMs' input and encoding the document content using the LLMs themselves, we fully take advantage of the document comprehension capability of LLMs and enhance their perception of OCR information. We have also deeply considered the role of chain-of-thought (CoT) and innovatively proposed the techniques of CoT Pre-training and CoT Annealing. Our DocLayLLM can achieve remarkable performances with lightweight training settings, showcasing its efficiency and effectiveness. Experimental results demonstrate that our DocLayLLM outperforms existing OCR-dependent methods and OCR-free competitors. Code and model are available at https://github.com/whlscut/DocLayLLM.

Recent advances in 3D Gaussian Splatting (3D-GS) have shown remarkable success in representing 3D scenes and generating high-quality, novel views in real-time. However, 3D-GS and its variants assume that input images are captured based on pinhole imaging and are fully in focus. This assumption limits their applicability, as real-world images often feature shallow depth-of-field (DoF). In this paper, we introduce DoF-Gaussian, a controllable depth-of-field method for 3D-GS. We develop a lens-based imaging model based on geometric optics principles to control DoF effects. To ensure accurate scene geometry, we incorporate depth priors adjusted per scene, and we apply defocus-to-focus adaptation to minimize the gap in the circle of confusion. We also introduce a synthetic dataset to assess refocusing capabilities and the model's ability to learn precise lens parameters. Our framework is customizable and supports various interactive applications. Extensive experiments confirm the effectiveness of our method. Our project is available at https://dof-gaussian.github.io/.

According to the forgetting curve theory, we can enhance memory retention by learning extensive data and taking adequate rest. This means that in order to effectively retain new knowledge, it is essential to learn it thoroughly and ensure sufficient rest so that our brain can memorize without forgetting. The main takeaway from this theory is that learning extensive data at once necessitates sufficient rest before learning the same data again. This aspect of human long-term memory retention can be effectively utilized to address the continual learning of neural networks. Retaining new knowledge for a long period of time without catastrophic forgetting is the critical problem of continual learning. Therefore, based on Ebbinghaus' theory, we introduce the view-batch model that adjusts the learning schedules to optimize the recall interval between retraining the same samples. The proposed view-batch model allows the network to get enough rest to learn extensive knowledge from the same samples with a recall interval of sufficient length. To this end, we specifically present two approaches: 1) a replay method that guarantees the optimal recall interval, and 2) a self-supervised learning that acquires extensive knowledge from a single training sample at a time. We empirically show that these approaches of our method are aligned with the forgetting curve theory, which can enhance long-term memory. In our experiments, we also demonstrate that our method significantly improves many state-of-the-art continual learning methods in various protocols and scenarios. We open-source this project at https://github.com/hankyul2/ViewBatchModel.

Multi-modal large language models (MLLMs) have made significant progress, yet their safety alignment remains limited. Typically, current open-source MLLMs rely on the alignment inherited from their language module to avoid harmful generations. However, the lack of safety measures specifically designed for multi-modal inputs creates an alignment gap, leaving MLLMs vulnerable to vision-domain attacks such as typographic manipulation. Current methods utilize a carefully designed safety dataset to enhance model defense capability.However, it is unknown what is actually learned in the high-quality dataset.Through comparison experiments, we find that the alignment gap primarily arises from data distribution biases, while image content, response quality, or the contrastive behavior of the dataset makes little contribution to boosting multi-modal safety. To further investigate this and identify the key factors in improving MLLM safety, we propose finetuning MLLMs on a small set of benign instruct-following data with responses replaced by simple, clear rejection sentences.Experiments show that, without the need for labor-intensive collection of high-quality malicious data, model safety can still be significantly improved, as long as a specific fraction of rejection data exists in the finetuning set, indicating the security alignment is not lost but rather obscured during multi-modal pretraining or instruction fine-tuning. Simply correcting the underlying data bias is enough to address the vision domain safety gap.

Common choices of architecture give neural networks a preference for fitting data with simple functions. This simplicity bias is known as key to their success. This paper explores the limits of this assumption. Building on recent work that showed that activation functions are the origin of the simplicity bias (Teney, 2024), we introduce a method to meta-learn activation functions to modulate this bias.**Findings.** We discover multiple tasks where the assumption of simplicity is inadequate, and standard ReLU architectures are therefore suboptimal. In these cases, we find activation functions that perform better by inducing a prior of higher complexity. Interestingly, these cases correspond to domains where neural networks have historically struggled: tabular data, regression tasks, cases of shortcut learning, and algorithmic grokking tasks. In comparison, the simplicity bias proves adequate on image tasks, where learned activations are nearly identical to ReLUs and GeLUs.**Implications.** (1) Contrary to common belief, the simplicity bias is not universally useful. There exist real tasks where it is suboptimal. (2) The suitability of ReLU models for image classification is not accidental. (3) The success of ML ultimately depends on the adequacy between data and architectures, and there may be benefits for architectures tailored to specific distributions of tasks.

Vision-and-Language Navigation (VLN) agents are tasked with navigating an unseen environment using natural language instructions. In this work, we study if visual representations of sub-goals implied by the instructions can serve as navigational cues and lead to increased navigation performance. To synthesize these visual representations or "imaginations", we leverage a text-to-image diffusion model on landmark references contained in segmented instructions. These imaginations are provided to VLN agents as an added modality to act as landmark cues and an auxiliary loss is added to explicitly encourage relating these with their corresponding referring expressions. Our findings reveal an increase in success rate (SR) of ~1 point and up to ~0.5 points in success scaled by inverse path length (SPL) across agents. These results suggest that the proposed approach reinforces visual understanding compared to relying on language instructions alone.

Frequency shortcuts refer to specific frequency patterns that models heavily rely on for correct classification. Previous studies have shown that models trained on small image datasets often exploit such shortcuts, potentially impairing their generalization performance. However, existing methods for identifying frequency shortcuts require expensive computations and become impractical for analyzing models trained on large datasets. In this work, we propose the first approach to more efficiently analyze frequency shortcuts at a large scale. We show that both CNN and transformer models learn frequency shortcuts on ImageNet. We also expose that frequency shortcut solutions can yield good performance on out-of-distribution (OOD) test sets which largely retain texture information. However, these shortcuts, mostly aligned with texture patterns, hinder model generalization on rendition-based OOD test sets. These observations suggest that current OOD evaluations often overlook the impact of frequency shortcuts on model generalization. Future benchmarks could thus benefit from explicitly assessing and accounting for these shortcuts to build models that generalize across a broader range of OOD scenarios.

While deep neural networks have revolutionized image denoising capabilities, their deployment on edge devices remains challenging due to substantial computational and memory requirements. To this end, we present DnLUT, an ultra-efficient lookup table-based framework that achieves high-quality color image denoising with minimal resource consumption. Our key innovation lies in two complementary components: a Pairwise Channel Mixer (PCM) that effectively captures inter-channel correlations and spatial dependencies in parallel, and a novel L-shaped convolution design that maximizes receptive field coverage while minimizing storage overhead. By converting these components into optimized lookup tables post-training, DnLUT achieves remarkable efficiency - requiring only 500KB storage and 0.1% energy consumption compared to its CNN contestant DnCNN, while delivering 20x faster inference. Extensive experiments demonstrate that DnLUT outperforms all existing LUT-based methods by over 1dB in PSNR, establishing a new state-of-the-art in resource-efficient color image denoising.

In recent years, there has been a significant surge of interest in unifying image comprehension and generation within Large Language Models (LLMs). This growing interest has prompted us to explore extending this unification to videos. The core challenge lies in developing a versatile video tokenizer that captures both the spatial characteristics and temporal dynamics of videos to obtain representations for LLMs, and the representations can be further decoded into realistic video clips to enable video generation. In this work, we introduce Divot, a Diffusion-Powered Video Tokenizer, which leverages the diffusion process for self-supervised video representation learning. We posit that if a video diffusion model can effectively de-noise video clips by taking the features of a video tokenizer as the condition, then the tokenizer has successfully captured robust spatial and temporal information. Additionally, the video diffusion model inherently functions as a de-tokenizer, decoding videos from their representations. Building upon the Divot tokenizer, we present Divot-LLM through video-to-text autoregression and text-to-video generation by modeling the distributions of continuous-valued Divot features with a Gaussian Mixture Model. Experimental results demonstrate that our diffusion-based video tokenizer, when integrated with a pre-trained LLM, achieves competitive performance across various video comprehension and generation benchmarks. The instruction tuned Divot-LLM also excels in video storytelling, generating interleaved narratives and corresponding videos. Models and codes are available at https://github.com/TencentARC/Divot.

Existing diffusion-based purification methods aim to disrupt adversarial perturbations by introducing a certain amount of noise through a forward diffusion process, followed by a reverse process to recover clean examples. However, this approach is fundamentally flawed: the uniform operation of the forward process across all pixels compromises normal pixels while attempting to combat adversarial perturbations, resulting in the target model producing incorrect predictions. Simply relying on low-intensity noise is insufficient for effective defense. To address this critical issue, we implement a heterogeneous purification strategy grounded in the interpretability of neural networks. Our method decisively applies higher-intensity noise to specific pixels that the target model focuses on while the remaining pixels are subjected to only low-intensity noise. This requirement motivates us to redesign the sampling process of the diffusion model, allowing for the effective removal of varying noise levels. Furthermore, to evaluate our method against strong adaptative attack, our proposed method sharply reduces time cost and memory usage through a single-step resampling. The empirical evidence from extensive experiments across three datasets demonstrates that our method outperforms most current adversarial training and purification techniques by a substantial margin.

Many real-world applications of flow-based generative models desire a diverse set of samples that cover multiple modes of the target distribution. However, the predominant approach for obtaining diverse sets is not sample-efficient, as it involves independently obtaining many samples from the source distribution and mapping them through the flow until the desired mode coverage is achieved. As an alternative to repeated sampling, we introduce DiverseFlow: a training-free approach to improve the diversity of flow models. Our key idea is to employ a determinantal point process to induce a coupling between the samples that drives diversity under a fixed sampling budget. In essence, DiverseFlow allows exploration of more variations in a learned flow model with fewer samples. We demonstrate the efficacy of our method for tasks where sample-efficient diversity is desirable, such as text-guided image generation with polysemous words, inverse problems like large-hole inpainting, and class-conditional image synthesis.

Large Multimodal Models (LMMs) have emerged as powerful models capable of understanding various data modalities, including text, images, and videos. LMMs encode both text and visual data into tokens that are then combined and processed by an integrated Large Language Model (LLM). Including visual tokens substantially increases the total token count, often by thousands. The increased input length for LLM significantly raises the complexity of inference, resulting in high latency in LMMs. To address this issue, token pruning methods, which remove part of the visual tokens, are proposed. The existing token pruning methods either require extensive calibration and fine-tuning or rely on suboptimal importance metrics which results in increased redundancy among the retained tokens. In this paper, we first formulate token pruning as Max-Min Diversity Problem (MMDP) where the goal is to select a subset such that the diversity among the selected tokens is maximized. Then, we solve the MMDP to obtain the selected subset and prune the rest. The proposed method, DivPrune, reduces redundancy and achieves the highest diversity of the selected tokens. By ensuring high diversity, the selected tokens better represent the original tokens, enabling effective performance even at high pruning ratios without requiring fine-tuning. Extensive experiments with various LMMs show that DivPrune achieves state-of-the-art accuracy over 16 image- and video-language datasets. Additionally, DivPrune reduces both the end-to-end latency and GPU memory usage for the tested models. The code is available here.

In Open-set Supervised Anomaly Detection (OSAD), the existing methods typically generate pseudo anomalies to compensate for the scarcity of observed anomaly samples, while overlooking critical priors of normal samples, leading to less effective discriminative boundaries. To address this issue, we propose a Distribution Prototype Diffusion Learning (DPDL) method aimed at enclosing normal samples within a compact and discriminative distribution space. Specifically, we construct multiple learnable Gaussian prototypes to create a latent representation space for abundant and diverse normal samples and learn a Schr"odinger bridge to facilitate a diffusive transition toward these prototypes for normal samples while steering anomaly samples away. Moreover, to enhance inter-sample separation, we design a dispersion feature learning way in hyperspherical space, which benefits the identification of out-ofdistribution anomalies. Experimental results demonstrate the effectiveness and superiority of our proposed DPDL, achieving state-of-the-art performance on 9 public datasets.

Nowadays, domain adaptation techniques have been widely investigated for knowledge sharing from labeled source domain to unlabeled target domain. However, target domain may include some data samples that belong to unknown categories in real-world scenarios. Moreover, the target domain cannot access the source data samples due to privacy-preserving restrictions. In this paper, we focus on the source-free open set domain adaptation problem which includes two main challenges, i.e.,how to distinguish known and unknown target samples and how to exploit useful source information to provide trustworthy pseudo labels for known target samples. Existing approaches that directly applying conventional domain alignment methods could lead to sample mismatch and misclassification in this scenario.To overcome these issues, we propose Distinguish Then Exploit model (DTE) with two components, i.e., weight barcode estimation and sparse label assignment. Weight barcode estimation first calculate marginal probability of target samples via partially unbalanced optimal transport then quantize barcode results to distinguish unknown target sample. Sparse label assignment utilizes sparse sample-label matching via proximal term to fully exploit useful source information. Our empirically study on several datasets shows that DTE outperforms the state-of-the-art models on tackling the source-free open set domain adaptation problem.

Open-Vocabulary Semantic Segmentation (OVSS) has advanced with recent vision-language models (VLMs), enabling segmentation beyond predefined categories through various learning schemes. Notably, training-free methods offer scalable, easily deployable solutions for handling unseen data, a key goal of OVSS. Yet, a critical issue persists: lack of object-level context consideration when segmenting complex objects in the challenging environment of OVSS based on arbitrary query prompts. This oversight limits models' ability to group semantically consistent elements within object and map them precisely to user-defined arbitrary classes. In this work, we introduce a novel approach that overcomes this limitation by incorporating object-level contextual knowledge within images. Specifically, our model enhances intra-object consistency by distilling spectral-driven features from vision foundation models into the attention mechanism of the visual encoder, enabling semantically coherent components to form a single object mask. Additionally, we refine the text embeddings with zero-shot object presence likelihood to ensure accurate alignment with the specific objects represented in the images. By leveraging object-level contextual knowledge, our proposed approach achieves state-of-the-art performance with strong generalizability across diverse datasets.