CVPR 2025 论文列表

When a vision-language model (VLM) is prompted to identify an entity depicted in an image, it may answer "I see a conifer," rather than the specific label "Norway spruce". This raises two issues for evaluation: Firstly, the unconstrained generated text needs to be mapped to the evaluation label space (i.e., "conifer"). Secondly, a useful classification measure should give partial credit to less specific, but not incorrect, answers ("Norway spruce" being a type of "conifer"). To meet these requirements, we propose a framework for evaluating unconstrained text predictions such as those generated from a vision-language model against a taxonomy. Specifically, we propose the use of hierarchical precision and recall measures to assess the level of correctness and specificity of predictions with regard to a taxonomy. Experimentally, we first show that existing text similarity measures do not capture taxonomic similarity well. We then develop and compare different methods to map textual VLM predictions onto a taxonomy. This allows us to compute hierarchical similarity measures between the generated text and the ground truth labels. Finally, we analyze modern VLMs on fine-grained visual classification tasks based on our proposed taxonomic evaluation scheme.

Prompt learning has attracted widespread attention in adapting vision-language models to downstream tasks. Existing methods largely rely on optimization strategies to ensure the task-awareness of learnable prompts. Due to the scarcity of task-specific data, overfitting is prone to occur. The resulting prompts often do not generalize well or exhibit limited task-awareness. To address this issue, we propose a novel Task-Aware Clustering (TAC) framework for prompting vision-language models, which increases the task-awareness of learnable prompts by introducing task-aware pre-context. The key ingredients are as follows: (a) generating task-aware pre-context based on task-aware clustering that can preserve the backbone structure of a downstream task with only a few clustering centers, (b) enhancing the task-awareness of learnable prompts by enabling them to interact with task-aware pre-context via the well-pretrained encoders, and (c) preventing the visual task-aware pre-context from interfering the interaction between patch embeddings by masked attention mechanism. Extensive experiments are conducted on benchmark datasets, covering the base-to-novel, domain generalization, and cross-dataset transfer settings. Ablation studies validate the effectiveness of key ingredients. Comparative results show the superiority of our TAC over competitive counterparts. The code is available at https://github.com/FushengHao/TAC.

Current multimodal large language models (MLLMs) struggle with fine-grained or precise understanding of visuals although they give comprehensive perception and reasoning in a spectrum of vision applications. Recent studies either develop tool-using or unify specific visual tasks into the autoregressive framework, often at the expense of overall multimodal performance. To address this issue and enhance MLLMs with visual tasks in a scalable fashion, we propose Task Preference Optimization (TPO), a novel method that utilizes differentiable task preferences derived from typical fine-grained visual tasks. TPO introduces learnable task tokens that establish connections between multiple task-specific heads and the MLLM. By leveraging rich visual labels during training, TPO significantly enhances the MLLM's multimodal capabilities and task-specific performance. Through multi-task co-training within TPO, we observe synergistic benefits that elevate individual task performance beyond what is achievable through single-task training methodologies. Our instantiation of this approach with VideoChat and LLaVA demonstrates an overall 14.6% improvement in multimodal performance compared to baseline models. Additionally, MLLM-TPO demonstrates robust zero-shot capabilities across various tasks, performing comparably to state-of-the-art supervised models.

Despite recent advances in deep learning, most existing learning IMU odometry methods are trained on specific datasets, lack generalization, and are prone to overfitting, which limits their real-world application. To address these challenges, we present Tartan IMU, a foundation model designed for generalizable, IMU-based state estimation across diverse robotic platforms. Our approach consists of three-stage: First, a pre-trained foundation model leverages over 100 hours of multi-platform data to establish general motion knowledge, achieving 36% improvement in ATE over specialized models. Second, to adapt to previously unseen tasks, we employ the Low-Rank Adaptation (LoRA), allowing positive transfer with only 1.1 M trainable parameters. Finally, to support robotics deployment, we introduce online test-time adaptation, which eliminates the boundary between training and testing, allowing the model to continuously "learn as it operates" at 200 FPS in real-time.

Foundation models (FMs) such as CLIP have demonstrated impressive zero-shot performance across various tasks by leveraging large-scale, unsupervised pre-training. However, they often inherit harmful or unwanted knowledge from noisy internet-sourced datasets, compromising their reliability in real-world applications. Existing model unlearning methods either rely on access to pre-trained datasets or focus on coarse-grained unlearning (e.g., entire classes), leaving a critical gap for fine-grained unlearning. In this paper, we address the challenging scenario of selectively forgetting specific portions of knowledge within a class--without access to pre-trained data--while preserving the model's overall performance. We propose a novel three-stage approach that progressively unlearns targeted knowledge while mitigating over-forgetting. It consists of (1) a forgetting stage to fine-tune the CLIP on samples to be forgotten, (2) a reminding stage to restore performance on retained samples, and (3) a restoring stage to recover zero-shot capabilities using model souping. Additionally, we introduce knowledge distillation to handle the distribution disparity between forgetting/retaining samples and unseen pre-trained data. Extensive experiments on CIFAR-10, ImageNet-1K, and style datasets demonstrate that our approach effectively unlearns specific subgroups while maintaining strong zero-shot performance on semantically similar subgroups and other categories, significantly outperforming baseline unlearning methods, which lose effectiveness under the CLIP unlearning setting.

We introduce MAGI, a hybrid video generation framework that combines masked modeling for intra-frame generation with causal modeling for next-frame generation. Our key innovation, Complete Teacher Forcing (CTF), conditions masked frames on complete observation frames rather than masked ones (namely Masked Teacher Forcing, MTF), enabling a smooth transition from token-level (patch-level) to frame-level autoregressive generation. CTF significantly outperforms MTF, achieving a 23% improvement in FVD scores on first-frame conditioned video prediction. To address issues like exposure bias, we employ targeted training strategies, setting a new benchmark in autoregressive video generation. Experiments show that MAGI can generate long, coherent video sequences exceeding 100 frames, even when trained on as few as 16 frames, highlighting its potential for scalable, high-quality video generation.

Recently, Mamba-based frameworks have achieved substantial advancements across diverse computer vision and NLP tasks, particularly in their capacity for reasoning over long-range information with linear complexity. However, the fixed 2D-to-1D scanning pattern overlooks the local structures of an image, limiting its effectiveness in aggregating 2D spatial information. While stacking additional Mamba layers can partially address this issue, it increases parameter intensity and constrains real-time application. In this work, we reconsider the local optimal scanning path in Mamba, enhancing the rigid and uniform 1D scan through the local shortest path theory, thus creating a structure-aware Mamba suited for lightweight single-image super-resolution. Specifically, we draw inspiration from the Traveling Salesman Problem (TSP) to establish a local optimal scanning path for improved structural 2D information utilization. Here, local patch aggregation occurs in a content-adaptive manner with minimal propagation cost. TSP-Mamba demonstrates substantial improvements over existing Mamba-based and Transformer-based architectures. For example, TSP-Mamba surpasses MambaIR by up to 0.7dB in lightweight SISR, with comparable parameters and very slightly extra computational demands (1-2 GFlops for 720P images).

Referring audio-visual segmentation (Ref-AVS) aims to segment objects within audio-visual scenes using multimodal cues embedded in text expressions. While the Segment Anything Model (SAM) has revolutionized visual segmentation, its applicability to Ref-AVS, where multimodal cues act as novel prompts, remains unexplored. SAM's limitation to single-frame segmentation also hinders its ability to capture essential temporal context needed for multi-frame audio-visual segmentation. To address this gap, we propose TSAM, a novel extension of SAM designed to leverage multimodal cues for precise segmentation in dynamic audio-visual scenes. TSAM enhances SAM's image encoder with a temporal modeling branch, enabling spatio-temporal learning and deep multimodal fusion across video frames, while retaining SAM's pre-trained knowledge. Additionally, TSAM replaces SAM's user-interactive prompting mechanism with sparse and dense data-driven prompts, enabling more effective integration of audio-visual inputs and reference text expressions. Extensive experiments on the Ref-AVS dataset demonstrate TSAM's superiority over state-of-the-art methods. The results illustrate its effectiveness in segmenting objects in dynamic audio-visual scenes using text-based multimodal cues and its strong generalization to unseen objects.

Large pre-trained Vision-Language Models (VLMs) such as CLIP have demonstrated excellent zero-shot generalizability across various downstream tasks. However, recent studies have shown that the inference performance of CLIP can be greatly degraded by small adversarial perturbations, especially its visual modality, posing significant safety threats. To mitigate this vulnerability, in this paper, we propose a novel defense method called Test-Time Adversarial Prompt Tuning (TAPT) to enhance the inference robustness of CLIP against visual adversarial attacks. TAPT is a test-time defense method that learns defensive bimodal (textual and visual) prompts to robustify the inference process of CLIP. Specifically, it is an unsupervised method that optimizes the defensive prompts for each test sample by minimizing a multi-view entropy and aligning adversarial-clean distributions. We evaluate the effectiveness of TAPT on 11 benchmark datasets, including ImageNet and 10 other zero-shot datasets, demonstrating that it enhances the zero-shot adversarial robustness of the original CLIP by at least 48.9% against AutoAttack (AA), while largely maintaining performance on clean examples. Moreover, TAPT outperforms existing adversarial prompt tuning methods across various backbones, achieving an average robustness improvement of at least 36.6%. Code is available at https://github.com/xinwong/TAPT.

Large Language Models (LLMs) have demonstrated excellent capabilities in composing various modules together to create programs that can perform complex reasoning tasks on images. In this paper, we propose TANGO, an approach that extends the program composition via LLMs already observed for images, aiming to integrate those capabilities into embodied agents capable of observing and acting in the world. Specifically, by employing a simple PointGoal Navigation model combined with a memory-based exploration policy as a foundational primitive for guiding an agent through the world, we show how a single model can address diverse tasks without additional training. We task an LLM with composing the provided primitives to solve a specific task, using only a few in-context examples in the prompt. We evaluate our approach on three key Embodied AI tasks: Open-Set ObjectGoal Navigation, Multi-Modal Lifelong Navigation, and Open Embodied Question Answering, achieving state-of-the-art results without any specific fine-tuning in challenging zero-shot scenarios.

Text-to-video (T2V) generative models have advanced significantly, yet their ability to compose different objects, attributes, actions, and motions into a video remains unexplored. Previous text-to-video benchmarks also neglect this important ability for evaluation. In this work, we conduct the first systematic study on compositional text-to-video generation. We propose T2V-CompBench, the first benchmark tailored for compositional text-to-video generation. T2V-CompBench encompasses diverse aspects of compositionality, including consistent attribute binding, dynamic attribute binding, spatial relationships, motion binding, action binding, object interactions, and generative numeracy. We further carefully design evaluation metrics of multimodal large language model (MLLM)-based, detection-based, and tracking-based metrics, which can better reflect the compositional text-to-video generation quality of seven proposed categories with 1400 text prompts. The effectiveness of the proposed metrics is verified by correlation with human evaluations. We also benchmark various text-to-video generative models and conduct in-depth analysis across different models and various compositional categories. We find that compositional text-to-video generation is highly challenging for current models, and we hope our attempt could shed light on future research in this direction.

Understanding the traffic scenes and then generating high-definition (HD) maps present significant challenges in autonomous driving. In this paper, we defined a novel \underline T raffic \underline T opology \underline S cene \underline G raph (\text T ^2\text SG ), a unified scene graph explicitly modeling the lane, controlled and guided by different road signals ( e.g. , right turn), and topology relationships among them, which is always ignored by previous high-definition (HD) mapping methods. For the generation of \text T ^2\text SG , we propose TopoFormer, a novel one-stage \underline Topo logy Scene Graph Trans\underline Former with two newly-designed layers. Specifically, TopoFormer incorporates a Lane Aggregation Layer (LAL) that leverages the geometric distance among the centerline of lanes to guide the aggregation of global information. Furthermore, we proposed a Counterfactual Intervention Layer (CIL) to model the reasonable road structure ( e.g. , intersection, straight) among lanes under counterfactual intervention. Then the generated \text T ^2\text SG can provide a more accurate and explainable description of the topological structure in traffic scenes. Experimental results demonstrate that TopoFormer outperforms existing methods on the \text T ^2\text SG generation task, and the generated \text T ^2\text SG significantly enhances traffic topology reasoning in downstream tasks, achieving a state-of-the-art performance of 46.3 OLS on the OpenLane-V2 benchmark. Our source code is available at https://github.com/MICLAB-BUPT/T2SG.

Zero-shot object counting aims to count instances of arbitrary object categories specified by text descriptions. Existing methods typically rely on vision-language models like CLIP, but often exhibit limited sensitivity to text prompts. We present T2ICount, a diffusion-based framework that leverages rich prior knowledge and fine-grained visual understanding from pretrained diffusion models. While one-step denoising ensures efficiency, it leads to weakened text sensitivity. To address this challenge, we propose a Hierarchical Semantic Correction Module that progressively refines text-image feature alignment, and a Representational Regional Coherence Loss that provides reliable supervision signals by leveraging the cross-attention maps extracted from the denoising U-Net. Furthermore, we observe that current benchmarks mainly focus on majority objects in images, potentially masking models' text sensitivity. To address this, we contribute a challenging re-annotated subset of FSC147 for better evaluation of text-guided counting ability. Extensive experiments demonstrate that our method achieves superior performance across different benchmarks. Code is available at https://github.com/cha15yq/T2ICount

Pre-trained Vision-Language Models (VLMs) require Continual Learning (CL) to efficiently update their knowledge and adapt to various downstream tasks without retraining from scratch. However, for VLMs, in addition to the loss of knowledge previously learned from downstream tasks, pre-training knowledge is also corrupted during continual fine-tuning. This issue is exacerbated by the unavailability of original pre-training data, leaving VLM's generalization ability degrading. In this paper, we propose GIFT, a novel continual fine-tuning approach that utilizes synthetic data to overcome catastrophic forgetting in VLMs. Taking advantage of recent advances in text-to-image synthesis, we employ a pre-trained diffusion model to recreate both pre-training and learned downstream task data. In this way, the VLM can revisit previous knowledge through distillation on matching diffusion-generated images and corresponding text prompts. Leveraging the broad distribution and high alignment between synthetic image-text pairs in VLM's feature space, we propose a contrastive distillation loss along with an image-text alignment constraint. To further combat in-distribution overfitting and enhance distillation performance with limited amount of generated data, we incorporate adaptive weight consolidation, utilizing Fisher information from these synthetic image-text pairs and achieving a better stability-plasticity balance. Extensive experiments demonstrate that our method consistently outperforms previous state-of-the-art approaches across various settings.

Talking head video generation aims to generate a realistic talking head video that preserves the person's identity from a source image and the motion from a driving video. Despite the promising progress made in the field, it remains a challenging and critical problem to generate videos with accurate poses and fine-grained facial details simultaneously. Essentially, facial motion is often highly complex to model precisely, and the one-shot source face image cannot provide sufficient appearance guidance during generation due to dynamic pose changes. To tackle the problem, we propose to jointly learn motion and appearance codebooks and perform multi-scale codebook compensation to effectively refine both the facial motion conditions and appearance features for talking face image decoding. Specifically, the designed multi-scale motion and appearance codebooks are learned simultaneously in a unified framework to store representative global facial motion flow and appearance patterns. Then, we present a novel multi-scale motion and appearance compensation module, which utilizes a transformer-based codebook retrieval strategy to query complementary information from the two codebooks for joint motion and appearance compensation. The entire process produces motion flows of greater flexibility and appearance features with fewer distortions across different scales, resulting in a high-quality talking head video generation framework. Extensive experiments on various benchmarks validate the effectiveness of our approach and demonstrate superior generation results from both qualitative and quantitative perspectives when compared to state-of-the-art competitors. The project page is available at https://shaelynz.github.io/synergize-motion-appearance/.

The remarkable success of Large Language Models (LLMs) has extended to the multimodal domain, achieving outstanding performance in image understanding and generation. Recent efforts to develop unified Multimodal Large Language Models (MLLMs) that integrate these capabilities have shown promising results. However, existing approaches often involve complex designs in model architecture or training pipeline, increasing the difficulty of model training and scaling. In this paper, we propose SynerGen-VL, a simple yet powerful encoder-free MLLM capable of both image understanding and generation. To address challenges identified in existing encoder-free unified MLLMs, we introduce the token folding mechanism and the vision-expert-based progressive alignment pretraining strategy, effectively supporting high-resolution image understanding while reducing training complexity. After being trained on large-scale mixed image-text data with a unified next-token prediction objective, SynerGen-VL achieves or surpasses the performance of existing encoder-free unified MLLMs with comparable or smaller parameter sizes, and narrows the gap with task-specific state-of-the-art models, highlighting a promising path toward future unified MLLMs. Our code and models shall be released.

Due to the limited scale of multimodal table understanding (MTU) data, model performance is constrained. A straightforward approach is to use multimodal large language models to obtain more samples, but this may cause hallucinations, generate incorrect sample pairs, and cost significantly.To address the above issues, we design a simple yet effective synthesis framework that consists of two independent steps: table image rendering and table question and answer (Q&A) pairs generation.We use table codes (HTML, LaTeX, Markdown) to synthesize images and generate Q&A pairs with large language model (LLM).This approach leverages LLM's high concurrency and low cost to boost annotation efficiency and reduce expenses. By inputting code instead of images, LLMs can directly access the content and structure of the table, reducing hallucinations in table understanding and improving the accuracy of generated Q&A pairs. Finally, we synthesize a large-scale MTU dataset, SynTab, containing 636K images and 1.8M samples costing within \200 in US dollars. We further introduce a generalist tabular multimodal model, SynTab-LLaVA. This model not only effectively extracts local textual content within the table but also enables global modeling of relationships between cells.SynTab-LLaVA achieves SOTA performance on 21 out of 24 in-domain and out-of-domain benchmarks, demonstrating the effectiveness and generalization of our method. The Code is available at \href https://github.com/bang123-box/SynTab-LLaVA SynTab-LLaVA .

We propose a symbolic generative task description language and a corresponding inference engine that can represent arbitrary multimodal tasks as structured symbolic flows. Unlike conventional generative models, which rely on large-scale training and implicit neural representations to learn cross-modal mappings--often with high computational costs and limited flexibility--our framework introduces an explicit symbolic representation composed of three core primitives: functions, parameters, and topological logic. Using a pre-trained language model, our inference engine maps natural language instructions directly to symbolic workflows in a training-free manner. Our framework successfully performs over 12 diverse multimodal generative tasks, demonstrating strong performance and flexibility without requiring task-specific tuning. Experiments show that our method not only matches or outperforms existing state-of-the-art unified models in content quality but also offers greater efficiency, editability, and interruptibility. We believe symbolic task representations provide a cost-effective and extensible foundation for advancing the capabilities of generative AI.