CVPR 2025 论文列表

Existing unlearning algorithms in text-to-image generative models often fail to preserve the knowledge of semantically related concepts when removing specific target concepts--a challenge known as adjacency. To address this, we propose FADE (Fine-grained Attenuation for Diffusion Erasure), introducing adjacency-aware unlearning in diffusion models. FADE comprises two components: (1) the Concept Neighborhood, which identifies an adjacency set of related concepts, and (2) Mesh Modules, employing a structured combination of Expungement, Adjacency, and Guidance loss components. These enable precise erasure of target concepts while preserving fidelity across related and unrelated concepts. Evaluated on datasets like Stanford Dogs, Oxford Flowers, CUB, I2P, Imagenette, and ImageNet-1k, FADE effectively removes target concepts with minimal impact on correlated concepts, achieving at least a 12% improvement in retention performance over state-of-the-art methods. Our code and models are available on the project page: iab-rubric/unlearning/FG-Un.

Given that visual foundation models (VFMs) are trained on extensive datasets but often limited to 2D images, a natural question arises: how well do they understand the 3D world? With the differences in architecture and training protocols (i.e., objectives, proxy tasks), a unified framework to fairly and comprehensively probe their 3D awareness is urgently needed. Existing works on 3D probing suggest single-view 2.5D estimation (e.g., depth and normal) or two-view sparse 2D correspondence (e.g., matching and tracking). Unfortunately, these tasks ignore texture awareness, and require 3D data as ground-truth, which limits the scale and diversity of their evaluation set. To address these issues, we introduce Feat2GS, which readout 3D Gaussians attributes from VFM features extracted from unposed images. This allows us to probe 3D awareness for geometry and texture via novel view synthesis, without requiring 3D data. Additionally, the disentanglement of 3DGS parameters - geometry (x, a, S) and texture (c) - enables separate analysis of texture and geometry awareness. Under Feat2GS, we conduct extensive experiments to probe the 3D awareness of several VFMs, and investigate the ingredients that lead to a 3D aware VFM. Building on these findings, we develop several variants that achieve state-of-the-art across diverse datasets. This makes Feat2GS useful for probing VFMs, and as a simple-yet-effective baseline for novel-view synthesis. Code and data will be made available at fanegg.github.io/Feat2GS.

Parameter-efficient tuning (PET) aims to transfer pre-trained foundation models to downstream tasks by learning a small number of parameters. Compared to traditional fine-tuning, which updates the entire model, PET significantly reduces storage and transfer costs for each task regardless of exponentially increasing pre-trained model capacity. However, most PET methods inherit the inference latency of their large backbone models and often introduce additional computational overhead due to additional modules (e.g. adapters), limiting their practicality for compute-intensive applications. In this paper, we propose Faster Parameter-Efficient Tuning (FPET), a novel approach that enhances inference speed and training efficiency while maintaining high storage efficiency. Specifically, we introduce a plug-and-play token redundancy reduction module delicately designed for PET. This module refines tokens from the self-attention layer using an adapter to learn the accurate similarity between tokens and cuts off the tokens through a fully-differentiable token merging strategy, which uses a straight-through estimator for optimal token reduction. Experimental results prove that our FPET achieves faster inference and higher memory efficiency than the pre-trained backbone while keeping competitive performance on par with state-of-the-art PET methods.

Effective training of large Vision-Language Models (VLMs) on resource-constrained client devices in Federated Learning (FL) requires the usage of parameter-efficient fine-tuning (PEFT) strategies. To this end, we demonstrate the impact of two factors, viz., client-specific layer importance score that selects the most important VLM layers for fine-tuning and inter-client layer diversity score that encourages diverse layer selection across clients for optimal VLM layer selection. We first theoretically motivate and leverage the principal eigenvalue magnitude of layerwise Neural Tangent Kernels and show its effectiveness as client-specific layer importance score. Next, we propose a novel layer updating strategy dubbed F^3OCUS that jointly optimizes the layer importance and diversity factors by employing a data-free, multi-objective, meta-heuristic optimization on the server. We explore 5 different meta-heuristic algorithms and compare their effectiveness for selecting model layers and adapter layers towards PEFT-FL. Furthermore, we release a new MedVQA-FL dataset involving overall 707,962 VQA triplets and 9 modality-specific clients and utilize it to train and evaluate our method. Overall, we conduct more than 10,000 client-level experiments on 6 Vision-Language FL task settings involving 58 medical image datasets and 4 different VLM architectures of varying sizes to demonstrate the effectiveness of the proposed method. Project Page: https://pramitsaha.github.io/FOCUS/

This work asks: with abundant, unlabeled real faces, how to learn a robust and transferable facial representation that boosts various face security tasks with respect to generalization performance? We make the first attempt and propose a self-supervised pretraining framework to learn fundamental representations of real face images, FSFM, that leverages the synergy between masked image modeling (MIM) and instance discrimination (ID). We explore various facial masking strategies for MIM and present a simple yet powerful CRFR-P masking, which explicitly forces the model to capture meaningful intra-region Consistency and challenging inter-region Coherency. Furthermore, we devise an ID network that naturally couples with MIM to establish underlying local-to-global Correspondence through tailored self-distillation. These three learning objectives, namely 3C, empower encoding both local features and global semantics of real faces. After pretraining, a vanilla ViT serves as a universal vision Foundation Model for downstream Face Security tasks: cross-dataset deepfake detection, cross-domain face anti-spoofing, and unseen diffusion facial forgery detection. Extensive experiments on 10 public datasets demonstrate that our model transfers better than supervised pretraining, visual and facial self-supervised learning arts, and even outperforms task-specialized SOTA methods.

Visual question answering (VQA) systems face significant challenges when adapting to real-world data shifts, especially in multi-modal contexts. While robust fine-tuning strategies are essential for maintaining performance across in-distribution (ID) and out-of-distribution (OOD) scenarios, current evaluation settings are primarily unimodal or particular to some types of OOD, offering limited insight into the complexities of multi-modal contexts. In this work, we propose a new benchmark FRAMES-VQA (Fine-Tuning Robustness Across Multi-Modal Shifts in VQA) for evaluating robust fine-tuning for VQA tasks. We utilize ten existing VQA benchmarks, including VQAv2, IV-VQA, VQA-CP, OK-VQA and others, and categorize them into ID, near and far OOD datasets covering uni-modal, multi-modal and adversarial distribution shifts. We first conduct a comprehensive comparison of existing robust fine-tuning methods. We then quantify the distribution shifts by calculating the Mahalanobis distance using uni-modal and multi-modal embeddings extracted from various models. Further, we perform an extensive analysis to explore the interactions between uni- and multi-modal shifts as well as modality importance for ID and OOD samples. These analyses offer valuable guidance on developing more robust fine-tuning methods to handle multi-modal distribution shifts. The code is available at https://github.com/chengyuehuang511/FRAMES-VQA.

Test-time adaptation (TTA) aims to improve the performance of source-domain pre-trained models on previously unseen, shifted target domains. Traditional TTA methods primarily adapt model weights based on target data streams, making model performance sensitive to the amount and order of target data. The recently proposed diffusion-driven TTA methods mitigate this by adapting model inputs instead of weights, where an unconditional diffusion model, trained on the source domain, transforms target-domain data into a synthetic domain that is expected to approximate the source domain. However, in this paper, we reveal that although the synthetic data in diffusion-driven TTA seems indistinguishable from the source data, it is unaligned with, or even markedly different from the latter for deep networks. To address this issue, we propose a Synthetic-Domain Alignment (SDA) framework. Our key insight is to fine-tune the source model with synthetic data to ensure better alignment. Specifically, we first employ a conditional diffusion model to generate labeled samples, creating a synthetic dataset. Subsequently, we use the aforementioned unconditional diffusion model to add noise to and denoise each sample before fine-tuning. This Mix of Diffusion (MoD) process mitigates the potential domain misalignment between the conditional and unconditional models. Extensive experiments across classifiers, segmenters, and multimodal large language models (MLLMs, e.g., LLaVA) demonstrate that SDA achieves superior domain alignment and consistently outperforms existing diffusion-driven TTA methods. Our code is available at https://github.com/SHI-Labs/Diffusion-Driven-Test-Time-Adaptation-via-Synthetic-Domain-Alignment.

Pre-trained vision-language models provide a robust foundation for efficient transfer learning across various downstream tasks. In the field of video action recognition, mainstream approaches often introduce additional modules to capture temporal information. Although the additional modules increase the capacity of model, enabling it to better capture video-specific inductive biases, existing methods typically introduce a substantial number of new parameters and are prone to catastrophic forgetting of previously acquired generalizable knowledge. In this paper, we propose a parameter-efficient Multi-modal Spatio-Temporal Adapter (MSTA) to enhance the alignment between textual and visual representations, achieving a balance between generalizable knowledge and task-specific adaptation. Furthermore, to mitigate over-fitting and enhance generalizability, we introduce a spatio-temporal description-guided consistency constraint. This constraint involves providing template inputs (e.g., "a video of \ cls\ ") to the trainable language branch and LLM-generated spatio-temporal descriptions to the pre-trained language branch, enforcing output consistency between the branches. This approach reduces overfitting to downstream tasks and enhances the distinguishability of the trainable branch within the spatio-temporal semantic space. We evaluate the effectiveness of our approach across four tasks: zero-shot transfer, few-shot learning, base-to-novel generalization, and fully-supervised learning. Compared to many state-of-the-art methods, our MSTA achieves outstanding performance across all evaluations, while using only 2-7% of the trainable parameters in the original model.

Deep neural networks remain vulnerable to statistical variations in data, even with advances in normalization techniques. Existing methods use fixed-size normalization sets, restricting their adaptability to evolving and diverse data characteristics. We introduce Dynamic Group Normalization (DGN), a framework that dynamically adjusts channel grouping based on statistical awareness, ensuring a flexible and optimized formation of channel groups. By leveraging an efficient spatio-temporal mechanism, DGN continuously evaluates inter-channel relationships within layers and across training epochs, ensuring robust and responsive adaptation to the updated data statistics. Extensive evaluations of 34 architectures and 11 computer vision benchmarks show the consistent superiority of DGN over traditional normalization methods. It achieves significant accuracy gains in classification, detection, and segmentation tasks while maintaining computational efficiency. Moreover, it outperforms traditional methods in challenging scenarios, including out-of-distribution generalization, imbalanced long-tailed distributions, and corrupted data.

Depth completion involves predicting dense depth maps from sparse LiDAR inputs, a critical task for applications such as autonomous driving and robotics. However, sparse depth annotations from sensors limit the availability of dense supervision, which is necessary for learning detailed geometric features. To overcome this limitation, we propose a two-stage knowledge distillation framework that leverages powerful monocular foundation models to provide dense supervision for depth completion. In the first stage, we introduce a pre-training strategy that generates diverse training data from natural images to distill geometric knowledge to depth completion. Specifically, we simulate LiDAR scans by utilizing monocular depth and mesh reconstruction, thereby creating training data without requiring ground-truth depth. Nonetheless, monocular depth estimation suffers from inherent scale ambiguity in real-world settings. To address this, in the second stage, we employ a scale- and shift-invariant loss (SSI Loss) to learn real-world scales when fine-tuning on real-world datasets. Our two-stage distillation framework enables depth completion models to harness the strengths of monocular foundation models. Experimental results show that models trained with our two-stage distillation framework achieve top-ranked performance on the KITTI benchmark, demonstrating improvements in both quantitative and qualitative metrics.

Stereo matching is a key technique for metric depth estimation in computer vision and robotics. Real-world challenges like occlusion and non-texture hinder accurate disparity estimation from binocular matching cues. Recently, monocular relative depth estimation has shown remarkable generalization using vision foundation models. Thus, to facilitate robust stereo matching with monocular depth cues, we incorporate a robust monocular relative depth model into the recurrent stereo-matching framework, building a new framework for depth foundation model-based stereo-matching, DEFOM-Stereo. In the feature extraction stage, we construct the combined context and matching feature encoder by integrating features from conventional CNNs and DEFOM. In the update stage, we use the depth predicted by DEFOM to initialize the recurrent disparity and introduce a scale update module to refine the disparity at the correct scale. DEFOM-Stereo is verified to have much stronger zero-shot generalization compared with SOTA methods. Moreover, DEFOM-Stereo achieves top performance on the KITTI 2012, KITTI 2015, Middlebury, and ETH3D benchmarks, ranking 1^ st on many metrics. In the joint evaluation under the robust vision challenge, our model simultaneously outperforms previous models on the individual benchmarks, further demonstrating its outstanding capabilities.

We propose a novel knowledge distillation approach, CustomKD, that effectively leverages large vision foundation models (LVFMs) to enhance the performance of edge models (e.g., MobileNetV3). Despite recent advancements in LVFMs, such as DINOv2 and CLIP, their potential in knowledge distillation for enhancing edge models remains underexplored. While knowledge distillation is a promising approach for improving the performance of edge models, the discrepancy in model capacities and heterogeneous architectures between LVFMs and edge models poses a significant challenge. Our observation indicates that although utilizing larger backbones (e.g., ViT-S to ViT-L) in teacher models improves their downstream task performances, the knowledge distillation from the large teacher models fails to bring as much performance gain for student models as for teacher models due to the large model discrepancy. Our simple yet effective CustomKD customizes the well-generalized features inherent in LVFMs to a given student model in order to reduce model discrepancies. Specifically, beyond providing well-generalized original knowledge from teachers, CustomKD aligns the features of teachers to those of students, making it easy for students to understand and overcome the large model discrepancy overall. CustomKD significantly improves the performances of edge models in scenarios with unlabeled data such as unsupervised domain adaptation (e.g., OfficeHome and DomainNet) and semi-supervised learning (e.g., CIFAR-100 with 400 labeled samples and ImageNet with 1% labeled samples), achieving the new state-of-the-art performances.

The capacity of Vision transformers (ViTs) to handle variable-sized inputs is often constrained by computational complexity and batch processing limitations. Consequently, ViTs are typically trained on small, fixed-size images obtained through downscaling or cropping. While reducing computational burden, these methods result in significant information loss, negatively affecting tasks like image aesthetic assessment. We introduce Charm, a novel tokenization approach that preserves Composition, High-resolution, Aspect Ratio, and Multi-scale information simultaneously. Charm prioritizes high-resolution details in specific regions while downscaling others, enabling shorter fixed-size input sequences for ViTs while incorporating essential information. Charm is designed to be compatible with pre-trained ViTs and their learned positional embeddings. By providing multiscale input and introducing variety to input tokens, Charm improves ViT performance and generalizability for image aesthetic assessment. We avoid cropping or changing the aspect ratio to further preserve information. Extensive experiments demonstrate significant performance improvements on various image aesthetic and quality assessment datasets (up to 8.1%) using a lightweight ViT backbone. Code and pre-trained models are available at https://github.com/FBehrad/Charm.

The emergence of large multimodal models (LMMs) has brought significant advancements to pathology. Previous research has primarily focused on separately training patch-level and whole-slide image (WSI)-level models, limiting the integration of learned knowledge across patches and WSIs and resulting in redundant models. In this work, we introduce CPath-Omni, the first 15B parameter LMM that unifies patch and WSI analysis, consolidating a variety of tasks at both levels, including classification, visual question answering, captioning, and visual referring prompting. Extensive experiments demonstrate that CPath-Omni achieves state-of-the-art (SOTA) performance across seven diverse tasks on 39 out of 42 datasets, outperforming or matching task-specific models trained for individual tasks. Additionally, we develop a specialized pathology CLIP-based visual processor for CPath-Omni, CPath-CLIP, which, for the first time, integrates different vision models and incorporates a large language model as a text encoder to build a more powerful CLIP model, which achieves SOTA performance on nine zero-shot and four few-shot datasets. Our findings highlight CPath-Omni's ability to unify diverse pathology tasks, demonstrating its potential to streamline and advance the field of foundation model in pathology. The code and model are available at https://github.com/PathFoundation/CPath-Omni.

Recent vision-language models (VLMs) face significant challenges in test-time adaptation to novel domains. While cache-based methods show promise by leveraging historical information, they struggle with both caching unreliable feature-label pairs and indiscriminately using single-class information during querying, significantly compromising adaptation accuracy. To address these limitations, we propose COSMIC (\underline C lique-\underline O riented \underline S emantic \underline M ulti-space \underline I ntegration for \underline C LIP), a robust test-time adaptation framework that enhances adaptability through multi-granular, cross-modal semantic caching and graph-based querying mechanisms. Our framework introduces two key innovations: Dual Semantics Graph (DSG) and Clique Guided Hyper-class (CGH). The Dual Semantics Graph constructs complementary semantic spaces by incorporating textual features, coarse-grained CLIP features, and fine-grained DINOv2 features to capture rich semantic relationships. Building upon these dual graphs, the Clique Guided Hyper-class component leverages structured class relationships to enhance prediction robustness through correlated class selection. Extensive experiments demonstrate COSMIC's superior performance across multiple benchmarks, achieving significant improvements over state-of-the-art methods: 15.81% gain on out-of-distribution tasks and 5.33% on cross-domain generation with CLIP RN-50.

Class-Incremental Learning (CIL) aims to learn new classes sequentially while retaining the knowledge of previously learned classes. Recently, pre-trained models (PTMs) combined with parameter-efficient fine-tuning (PEFT) have shown remarkable performance in rehearsal-free CIL without requiring exemplars from previous tasks. However, existing adapter-based methods, which incorporate lightweight learnable modules into PTMs for CIL, create new adapters for each new task, leading to both parameter redundancy and failure to leverage shared knowledge across tasks. In this work, we propose ContinuaL Low-Rank Adaptation (CL-LoRA), which introduces a novel dual-adapter architecture combining task-shared adapters to learn cross-task knowledge and task-specific adapters to capture unique features of each new task. Specifically, the shared adapters utilize random orthogonal matrices and leverage knowledge distillation with gradient reassignment to preserve essential shared knowledge. In addition, we introduce learnable block-wise weights for task-specific adapters, which mitigate inter-task interference while maintaining the model's plasticity. We demonstrate CL-LoRA consistently achieves promising performance under multiple benchmarks with reduced training and inference computation, establishing a more efficient and scalable paradigm for continual learning with pre-trained models.

Depth estimation is a fundamental task in 3D vision. An ideal depth estimation model is expected to embrace meticulous detail, temporal consistency, and high efficiency. Although existing foundation models can perform well in certain specific aspects, most of them fall short of fulfilling all the above requirements simultaneously. In this paper, we present CH_3Depth, an efficient and flexible model for depth estimation with flow matching to address this challenge. Specifically, 1) we reframe the optimization objective of flow matching as the Inversion by Direct Iteration (InDI) to improve accuracy. 2) To enhance efficiency, we propose non-uniform sampling to achieve better prediction with fewer sampling steps. 3) We design the Latent Temporal Stabilizer (LTS) to enhance temporal consistency by aggregating latent codes of adjacent frames, enabling our method to be lightweight and compatible for video depth estimation. CH_3Depth achieves state-of-the-art performance in zero-shot evaluations across multiple image and video datasets, excelling in prediction accuracy, efficiency, and temporal consistency, highlighting its potential as the next foundation model for depth estimation.

While existing vision and multi-modal foundation models can handle multiple computer vision tasks, they often suffer from significant limitations, including huge demand for data and computational resources during training and inconsistent performance across vision tasks at deployment time. To address these challenges, we introduce Argus (The name comes from Argus Panoptes -- a hundred-eyed giant with "all-seeing" capability in Greek mythology), a compact and versatile vision foundation model designed to support a wide range of vision tasks through a unified multitask architecture. Argus employs a two-stage training strategy: (i) multitask pretraining over core vision tasks with a shared backbone that includes a lightweight adapter to inject task-specific inductive biases, and (ii) scalable and efficient adaptation to new tasks by fine-tuning only the task-specific decoders. Extensive evaluations demonstrate that Argus, despite its relatively compact and training-efficient design of merely 100M backbone parameters (only 13.6% of which are trained using 1.6M images), competes with and even surpasses much larger models. Compared to state-of-the-art foundation models, Argus not only covers a broader set of vision tasks but also matches or outperforms the models with similar sizes on 12 tasks. We expect that Argus will accelerate the real-world adoption of vision foundation models in resource-constrained scenarios.

Adapting Multi-modal Large Language Models (MLLMs) to target tasks often suffers from catastrophic forgetting, where acquiring new task-specific knowledge compromises performance on pre-trained tasks. In this paper, we introduce AdaDARE-\gamma, an efficient approach that alleviates catastrophic forgetting by controllably injecting new task-specific knowledge through adaptive parameter selection from fine-tuned models without requiring retraining procedures. This approach consists two key innovations: (1) an adaptive parameter selection mechanism that identifies and retains the most task-relevant parameters from fine-tuned models, and (2) a controlled task-specific information injection strategy that precisely balances the preservation of pre-trained knowledge with the acquisition of new capabilities. Theoretical analysis proves the optimality of our parameter selection strategy and establishes bounds for the task-specific information injection factor. Extensive experiments on InstructBLIP and LLaVA-1.5 across image captioning and visual question answering tasks demonstrate that AdaDARE-\gamma establishes new state-of-the-art results in balancing model performance. Specifically, it maintains 98.2% of pre-training effectiveness on original tasks while achieving 98.7% of standard fine-tuning performance on target tasks.