CVPR 2025 Papers

We investigate the generation of minority samples using pretrained text-to-image (T2I) latent diffusion models. Minority instances, in the context of T2I generation, can be defined as ones living on low-density regions of text-conditional data distributions. They are valuable for various applications of modern T2I generators, such as data augmentation and creative AI. Unfortunately, existing pretrained T2I diffusion models primarily focus on high-density regions, largely due to the influence of guided samplers (like CFG) that are essential for high-quality generation. To address this, we present a novel framework to counter the high-density-focus of T2I diffusion models. Specifically, we first develop an online prompt optimization framework that encourages emergence of desired properties during inference while preserving semantic contents of user-provided prompts. We subsequently tailor this generic prompt optimizer into a specialized solver that promotes generation of minority features by incorporating a carefully-crafted likelihood objective. Extensive experiments conducted across various types of T2I models demonstrate that our approach significantly enhances the capability to produce high-quality minority instances compared to existing samplers. Code is available at https://github.com/soobin-um/MinorityPrompt.

Traditional video instance segmentation (VIS) models rely on extensive per-frame video annotations, which are both time-consuming and costly. In this paper, we present MinMaxVIS, a novel VIS framework that reduces the dependency on fully labeled video datasets by utilizing a small set of labeled images from the target domain along with a large volume of general-domain, unlabeled images. MinMaxVIS operates in three stages: first, a preliminary segmentation model is trained on the small labeled set from the target domain; this model then retrieves relevant instances from the unlabeled dataset to build a high-quality pseudo-labeled set, ensuring a rich content alignment with the target domain while avoiding the inefficiencies of large-scale semi-supervised learning across the entire unlabeled set. Finally, we train MinMaxVIS on a combination of labeled and pseudo-labeled data, addressing challenges such as noise in pseudo-labels and instance association across frames. To simulate object continuity, we augment static images to create paired frames, allowing MinMaxVIS to capture instance associations effectively. MinMaxVIS outperforms the prior image-driven approach, MinVIS, achieving superior mAP scores with significantly reduced labeled data. For instance, MinMaxVIS with a Swin-L backbone attains 62.2 mAP on YouTube-VIS 2019 using only 2% labeled data and additional unlabeled images from SA-1B. This surpasses MinVIS, which uses the same backbone trained on fully labeled YouTube-VIS 2019, by 0.6 mAP.

The rapid scaling of large vision pretrained models makes fine-tuning tasks more and more difficult on devices with low computational resources. We explore a new visual adaptation paradigm called separated tuning, which treats large pretrained models as standalone feature extractors that run on powerful cloud servers. The fine-tuning carries out on devices which possess only low computational resources (slow CPU, no GPU, small memory, etc.) Existing methods that are potentially suitable for our separated tuning paradigm are discussed. But, three major drawbacks hinder their application in separated tuning: low adaptation capability, large adapter network, and in particular, high information transfer overhead. To address these issues, we propose Minimal Interaction Separated Tuning, or MIST, which reveals that the sum of intermediate features from pretrained models not only has minimal information transfer but also has high adaptation capability. With a lightweight attention-based adaptor network, MIST achieves information transfer efficiency, parameter efficiency, computational and memory efficiency, and at the same time demonstrates competitive results on various visual adaptation benchmarks.

Deep learning has achieved significant advancements in medical image segmentation, but existing models still face challenges in accurately segmenting lesion regions. The main reason is that some lesion regions in medical images have unclear boundaries, irregular shapes, and small tissue density differences, leading to label ambiguity. However, the existing model treats all data equally without taking quality differences into account in the training process, resulting in noisy labels negatively impacting model training and unstable feature representations. In this paper, a **d**ata-driven **a**lternating **le**arning (**DALE**) paradigm is proposed to optimize the model's training process, achieving stable and high-precision segmentation. The paradigm focuses on two key points: (1) reducing the impact of noisy labels, and (2) calibrating unstable representations. To mitigate the negative impact of noisy labels, a loss consistency-based collaborative optimization method is proposed, and its effectiveness is theoretically demonstrated. Specifically, the label confidence parameters are introduced to dynamically adjust the influence of labels of different confidence levels during model training, thus reducing the influence of noise labels. To calibrate the learning bias of unstable representations, a distribution alignment method is proposed. This method restores the underlying distribution of unstable representations, thereby enhancing the discriminative capability of fuzzy region representations. Extensive experiments on various benchmarks and model backbones demonstrate the superiority of the DALE paradigm, achieving an average performance improvement of up to 7.16%.

Real-world videos consist of sequences of events. Generating such sequences with precise temporal control is infeasible with existing video generators that rely on a single paragraph of text as input. When tasked with generating multiple events described using a single prompt, such methods often ignore some of the events or fail to arrange them in the correct order. To address this limitation, we present MinT, a multi-event video generator with temporal control. Our key insight is to bind each event to a specific period in the generated video, which allows the model to focus on one event at a time. To enable time-aware interactions between event captions and video tokens, we design a time-based positional encoding method, dubbed ReRoPE. This encoding helps to guide the cross-attention operation. By fine-tuning a pre-trained video diffusion transformer on temporally grounded data, our approach produces coherent videos with smoothly connected events. For the first time in the literature, our model offers control over the timing of events in generated videos. Extensive experiments demonstrate that MinT outperforms existing commercial and open-source models by a large margin.

Adversarial attacks remain a significant threat that can jeopardize the integrity of Machine Learning (ML) models. In particular, query-based black-box attacks can generate malicious noise without having access to the victim model's architecture, making them practical in real-world contexts. The community has proposed several defenses against adversarial attacks, only to be broken by more advanced and adaptive attack strategies. In this paper, we propose a framework that detects if an adversarial noise instance is being generated. Unlike existing stateful defenses that detect adversarial noise generation by monitoring the input space, our approach learns adversarial patterns in the input update similarity space. In fact, we propose to observe a new metric called Delta Similarity (DS), which we show it captures more efficiently the adversarial behavior.We evaluate our approach against 8 state-of-the-art attacks, including adaptive attacks, where the adversary is aware of the defense and tries to evade detection. We find that our approach is significantly more robust than existing defenses both in terms of specificity and sensitivity.

Federated Semi-Supervised Learning (FSSL) aims to leverage unlabeled data across clients with limited labeled data to train a global model with strong generalization ability. Most FSSL methods rely on consistency regularization with pseudo-labels, converting predictions from local or global models into hard pseudo-labels as supervisory signals. However, we discover that the quality of pseudo-label is largely deteriorated by data heterogeneity, an intrinsic facet of federated learning. In this paper, we study the problem of FSSL in-depth and show that (1) heterogeneity exacerbates pseudo-label mismatches, further degrading model performance and convergence, and (2) local and global models' predictive tendencies diverge as heterogeneity increases. Motivated by these findings, we propose a simple and effective method called Semi-supervised Aggregation for Globally-Enhanced Ensemble (SAGE), that can flexibly correct pseudo-labels based on confidence discrepancies. This strategy effectively mitigates performance degradation caused by incorrect pseudo-labels and enhances consensus between local and global models. Experimental results demonstrate that SAGE outperforms existing FSSL methods in both performance and convergence. Our code is available at \href https://github.com/Jay-Codeman/SAGE https://github.com/Jay-Codeman/SAGE .

Text serves as the key control signal in video generation due to its narrative nature. To render text descriptions into video clips, current video diffusion models borrow features from text encoders yet struggle with limited text comprehension. The recent success of large language models (LLMs) showcases the power of decoder-only transformers, which offers three clear benefits for text-to-video (T2V) generation, namely, precise text understanding resulting from the superior scalability, imagination beyond the input text enabled by next token prediction, and flexibility to prioritize user interests through instruction tuning. Nevertheless, the feature distribution gap emerging from the two different text modeling paradigms hinders the direct use of LLMs in established T2V models. This work addresses this challenge with Mimir, an end-to-end training framework featuring a carefully tailored token fuser to harmonize the outputs from text encoders and LLMs. Such a design allows the T2V model to fully leverage learned video priors while capitalizing on the text-related capability of LLMs. Extensive quantitative and qualitative results demonstrate the effectiveness of our approach in generating high-quality videos with excellent text comprehension, especially when processing short captions and managing shifting motions. The code and models will be made publicly available.

Recently, In-context Learning (ICL) has become a significant inference paradigm in Large Multimodal Models (LMMs), utilizing a few in-context demonstrations (ICDs) to prompt LMMs for new tasks. However, the synergistic effects in multimodal data increase the sensitivity of ICL performance to the configurations of ICDs, stimulating the need for a more stable and general mapping function. Mathematically, in Transformer-based models, ICDs act as "shift vectors" added to the hidden states of query tokens. Inspired by this, we introduce Mimic In-Context Learning (MimIC) to learn stable and generalizable shift effects from ICDs. Specifically, compared with some previous shift vector-based methods, MimIC more strictly approximates the shift effects by integrating lightweight learnable modules into LMMs with four key enhancements: 1) inserting shift vectors after attention layers, 2) assigning a shift vector to each attention head, 3) making shift magnitude query-dependent, and 4) employing a layer-wise alignment loss. Extensive experiments on two LMMs (Idefics-9b and Idefics2-8b-base) across three multimodal tasks (VQAv2, OK-VQA, Captioning) demonstrate that MimIC outperforms existing shift vector-based methods. The code is available at https://anonymous.4open.science/r/MimIC/.

Scientific research demands sophisticated reasoning over multimodal data, a challenge especially prevalent in biology. Despite recent advances in multimodal large language models (MLLMs) for AI-assisted research, existing multimodal reasoning benchmarks only target up to college-level difficulty, while research-level benchmarks emphasize lower-level perception, falling short of the complex multimodal reasoning needed for scientific discovery. To bridge this gap, we introduce MicroVQA, a visual-question answering (VQA) benchmark designed to assess three reasoning capabilities vital in research workflows: expert image understanding, hypothesis generation, and experiment proposal. MicroVQA consists of 1,042 multiple-choice questions (MCQs) curated by biology experts across diverse microscopy modalities, ensuring VQA samples represent real scientific practice. In constructing the benchmark, we find that standard MCQ generation methods induce language shortcuts, motivating a new two-stage pipeline: an optimized LLM prompt structures question-answer pairs into MCQs; then, an agent-based 'RefineBot' updates them to remove shortcuts. Benchmarking on state-of-the-art MLLMs reveal a peak performance of 53%; models with smaller LLMs only slightly underperform top models, suggesting that language-based reasoning is less challenging than multimodal reasoning; and tuning with scientific articles enhances performance. Expert analysis of chain-of-thought responses shows that perception errors are the most frequent, followed by knowledge errors and then overgeneralization errors. These insights highlight the challenges in multimodal scientific reasoning, showing MicroVQA is a valuable resource advancing AI-driven biomedical research. MicroVQA is available at https://huggingface.co/datasets/jmhb/microvqa and project at https://jmhb0.github.io/microvqa.

Recent advancements in handwritten text recognition (HTR) have enabled the effective conversion of handwritten text to digital formats. However, achieving robust recognition across diverse writing styles remains challenging. Traditional HTR methods lack writer-specific personalization at test time due to limitations in model architecture and training strategies. Existing attempts to bridge this gap, through gradient-based meta-learning, still require labeled examples and suffer from parameter-inefficient fine-tuning, leading to substantial computational and memory overhead. To overcome these challenges, we propose an efficient framework that formulates personalization as prompt tuning, incorporating an auxiliary image reconstruction task with a self-supervised loss to guide prompt adaptation with unlabeled test-time examples. To ensure self-supervised loss effectively minimizes text recognition error, we leverage meta-learning to learn the optimal initialization of the prompts. As a result, our method allows the model to efficiently capture unique writing styles by updating less than 1% of its parameters and eliminating the need for time-intensive annotation processes. We validate our approach on the RIMES and IAM Handwriting Database benchmarks, where it consistently outperforms previous state-of-the-art methods while using 20x fewer parameters. We believe this represents a significant advancement in personalized handwritten text recognition, paving the way for more reliable and practical deployment in resource-constrained scenarios.

Shadows are often underconsidered or even ignored in image editing applications, limiting the realism of the edited results. In this paper, we introduce MetaShadow, a three-in-one versatile framework that enables detection, removal, and controllable synthesis of shadows in natural images in an object-centered fashion. MetaShadow combines the strengths of two cooperative components: Shadow Analyzer, for object-centered shadow detection and removal, and Shadow Synthesizer, for reference-based controllable shadow synthesis. Notably, we optimize the learning of the intermediate features from Shadow Analyzer to guide Shadow Synthesizer to generate more realistic shadows that blend seamlessly with the scene. Extensive evaluations on multiple shadow benchmark datasets show significant improvements of MetaShadow over the existing state-of-the-art methods on object-centered shadow detection, removal, and synthesis. MetaShadow excels in supporting imageediting tasks such as object removal, relocation, and insertion, pushing the boundaries of object-centered image editing.

Articulated 3D object generation is fundamental for creating realistic, functional, and interactable virtual assets which are not simply static. We introduce MeshArt, a hierarchical transformer-based approach to generate articulated 3D meshes with clean, compact geometry, reminiscent of human-crafted 3D models. We approach articulated mesh generation in a part-by-part fashion across two stages.First, we generate a high-level articulation-aware object structure; then, based on this structural information, we synthesize each part's mesh faces. Key to our approach is modeling both articulation structures and part meshes as sequences of quantized triangle embeddings, leading to a unified hierarchical framework with transformers for autoregressive generation.Object part structures are first generated as their bounding primitives and articulation modes; a second transformer, guided by these articulation structures, then generates each part's mesh triangles.To ensure coherency among generated parts, we introduce structure-guided conditioning that also incorporates local part mesh connectivity.MeshArt shows significant improvements over state of the art, with 57.1% improvement in structure coverage and a 209-point improvement in mesh generation FID.

Mesh saliency enhances the adaptability of 3D vision by identifying and emphasizing regions that naturally attract visual attention. To investigate the interaction between geometric structure and texture in shaping visual attention, we establish a comprehensive mesh saliency dataset, which is the first to systematically capture the differences in saliency distribution under both textured and non-textured visual conditions. Furthermore, we introduce mesh Mamba, a unified saliency prediction model based on a state space model (SSM), designed to adapt across various mesh types. Mesh Mamba effectively analyzes the geometric structure of the mesh while seamlessly incorporating texture features into the topological framework, ensuring coherence throughout appearance-enhanced modeling. More importantly, by subgraph embedding and a bidirectional SSM, the model enables global context modeling for both local geometry and texture, preserving the topological structure and improving the understanding of visual details and structural complexity. Through extensive theoretical and empirical validation, our model not only improves performance across various mesh types but also demonstrates high scalability and versatility, particularly through cross validations of various visual features.

Masked Image Modeling (MIM) with Vector Quantization (VQ) has achieved great success in both self-supervised pre-training and image generation. However, most existing methods struggle to address the trade-off in the shared latent space for generation quality vs. representation learning and efficiency. To push the limits of this paradigm, we propose MergeVQ, which incorporates token merging techniques into VQ-based generative models to bridge the gap between image generation and visual representation learning in a unified architecture. During pre-training, MergeVQ decouples top-k semantics from latent space with the token merge module after self-attention blocks in the encoder for subsequent Look-up Free Quantization (LFQ) and global alignment and recovers their fine-grained details through cross-attention in the decoder for reconstruction. As for second-stage generation, we introduce MergeAR, which performs KV Cache compression for efficient raster-order prediction. Extensive experiments on ImageNet verify that MergeVQ as an AR generative model achieves competitive performance in both visual representation learning and image generation tasks while maintaining favorable token efficiency and inference speed. The code and model will be available at https://apexgen-x.github.io/MergeVQ.

We propose scaling up 3D scene reconstruction by training with synthesized data. At the core of our work is MegaSynth, a procedurally generated 3D dataset comprising 700K scenes - over 50 times larger than the prior real dataset DL3DV - dramatically scaling the training data. To enable scalable data generation, our key idea is eliminating semantic information, removing the need to model complex semantic priors such as object affordances and scene composition. Instead, we model scenes with basic spatial structures and geometry primitives, offering scalability. Besides, we control data complexity to facilitate training while loosely aligning it with real-world data distribution to benefit real-world generalization. We explore training LRMs with both MegaSynth and available real data. Experiment results show that joint training or pre-training with MegaSynth improves reconstruction quality by 1.2 to 1.8 dB PSNR across diverse image domains. Moreover, models trained solely on MegaSynth perform comparably to those trained on real data, underscoring the low-level nature of 3D reconstruction. Additionally, we provide an in-depth analysis of MegaSynth's properties for enhancing model capability, training stability, and generalization.

Deep multi-view clustering methods utilize information from multiple views to achieve enhanced clustering results and have gained increasing popularity in recent years. Most existing methods typically focus on either inter-view or intra-view relationships, aiming to align information across views or analyze structural patterns within individual views. However, they often incorporate inter-view complementary information in a simplistic manner, while overlooking the complex, high-order relationships within multi-view data and the interactions among samples, resulting in an incomplete utilization of the rich information available. Instead, we propose a multi-scale approach that exploits all of the available information. We first introduce a dual graph diffusion module guided by a consensus graph. This module leverages inter-view information to enhance the representation of both nodes and edges within each view. Secondly, we propose a novel contrastive loss function based on hypergraphs to more effectively model and leverage complex intra-view data relationships. Finally, we propose to adaptively learn fusion weights at the sample level, which enables a more flexible and dynamic aggregation of multi-view information. Extensive experiments on eight datasets show favorable performance of the proposed method compared to state-of-the-art approaches, demonstrating its effectiveness across diverse scenarios.

Despite significant progress in Vision-Language Pre-training (VLP), current approaches predominantly emphasize feature extraction and cross-modal comprehension, with limited attention to generating or transforming visual content. This gap hinders the model's ability to synthesize coherent and novel visual representations from textual prompts, thereby reducing the effectiveness of multi-modal learning. In this work, we propose MedUnifier, a unified VLP framework tailored for medical data. MedUnifier seamlessly integrates text-grounded image generation capabilities with multi-modal learning strategies, including image-text contrastive alignment, image-text matching and image-grounded text generation. Unlike traditional methods that reply on continuous visual representations, our approach employs visual vector quantization, which not only facilitates a more cohesive learning strategy for cross-modal understanding but also enhances multi-modal generation quality by effectively leveraging discrete representations. Our framework's effectiveness is evidenced by the experiments on established benchmarks, including uni-modal tasks, cross-modal tasks, and multi-modal tasks, where it achieves state-of-the-art performance across various tasks. MedUnifier also offers a highly adaptable tool for a wide range of language and vision tasks in healthcare, marking advancement toward the development of a generalizable AI model for medical applications.

Creating high-fidelity head avatars from multi-view videos is essential for many AR/VR applications. However, current methods often struggle to achieve high-quality renderings across all head components (e.g., skin vs. hair) due to the limitations of using one single representation for elements with varying characteristics. In this paper, we introduce a Hybrid Mesh-Gaussian Head Avatar (MeGA) that models different head components with more suitable representations. Specifically, we employ an enhanced FLAME mesh for the facial representation and predict a UV displacement map to provide per-vertex offsets for improved personalized geometric details. To achieve photorealistic rendering, we use deferred neural rendering to obtain facial colors and decompose neural textures into three meaningful parts. For hair modeling, we first build a static canonical hair using 3D Gaussian Splatting. A rigid transformation and an MLP-based deformation field are further applied to handle complex dynamic expressions. Combined with our occlusion-aware blending, MeGA generates higher-fidelity renderings for the whole head and naturally supports diverse downstream tasks. Experiments on the NeRSemble dataset validate the effectiveness of our designs, outperforming previous state-of-the-art methods and enabling versatile editing capabilities, including hairstyle alteration and texture editing.