ICCV 2025 论文列表

Multimodal large language models (MLLMs) have revolutionized cross-modal understanding but continue to struggle with hallucinations - fabricated content contradicting visual inputs. Existing hallucination mitigation methods either incur prohibitive computational costs or introduce distribution mismatches between training data and model outputs. We identify a critical insight: hallucinations predominantly emerge at the early stages of text generation and propagate through subsequent outputs. To address this, we propose SENTINEL (Sentence-level Early iNtervention Through IN-domain prEference Learning), a framework that eliminates dependency on human annotations. Specifically, we first bootstrap high-quality in-domain preference pairs by iteratively sampling model outputs, validating object existence through cross-checking with two open-vocabulary detectors, and classifying sentences into hallucinated/non-hallucinated categories. Subsequently, we use context-coherent positive samples and hallucinated negative samples to iteratively build context-aware preference data. Finally, we train models using a context-aware preference loss (C-DPO) that emphasizes discriminative learning at the sentence level where hallucinations initially manifest. Experimental results show that SENTINEL can reduce hallucinations by 90% over the original model and outperforms the previous state-of-the-art method on both the hallucination benchmarks and general capabilities benchmarks, manifesting its superiority and generalization ability. The models, datasets, and code are available at https://github.com/pspdada/SENTINEL.

Farthest Point Sampling (FPS) is widely used in existing point-based models because it effectively preserves structural integrity during downsampling. However, it incurs significant computational overhead, severely impacting the model's inference efficiency. Random sampling or grid sampling is considered faster downsampling methods; however, these fast downsampling methods may lead to the loss of geometric information during the downsampling process due to their overly simplistic and fixed rules, which can negatively affect model performance. To address this issue, we propose FastAdapter, which aggregates local contextual information through a small number of anchor points and facilitates interactions across spatial and layer dimensions, ultimately feeding this information back into the downsampled point cloud to mitigate the information degradation caused by fast downsampling methods. In addition to using FastAdapter to enhance model performance in methods that already employ fast downsampling, we aim to explore a more challenging yet valuable application scenario. Specifically, we focus on pre-trained models that utilize FPS, embedding FastAdapter and replacing FPS with random sampling for lightweight fine-tuning. This approach aims to significantly improve inference speed while maintaining relatively unchanged performance. Experimental results on ScanNet, S3DIS, and SemanticKITTI demonstrate that our method effectively mitigates the geometric information degradation issues caused by fast downSampling.

Fast Adversarial Training (FAT) employs the single-step Fast Gradient Sign Method (FGSM) to generate adversarial examples, reducing the computational costs of traditional adversarial training. However, FAT suffers from Catastrophic Overfitting (CO), where models' robust accuracy against multi-step attacks plummets to zero during training. Recent studies indicate that CO occurs because single-step adversarial perturbations contain label information that models exploit for prediction, leading to overfitting and diminished robustness against more complex attacks. In this paper, we discover that after CO occurs, the label information of certain samples can transfer across different samples, significantly increasing the likelihood of modified images being classified as the intended label. This discovery offers a new perspective on why various adversarial initialization strategies are effective. To address this issue, we introduce an innovative FAT strategy that leverages special samples to capture transferable label information and proactively removes potential label information during training, complemented by a non-uniform label smoothing technique to further eliminate label information. Experimental results across three datasets demonstrate that our method maintains competitive robustness against several attacks compared to other FAT approaches, with ablation studies confirming the effectiveness of our methodology. The code is available at https://github.com/fzjcdt/LIET.

Online mistake detection is crucial across various domains, ranging from industrial automation to educational applications, as mistakes can be corrected by the human operator after their detection due to the continuous inference on a video stream. While prior research mainly addresses procedural errors that often relate to temporal and ordering information, identifying a broader range of error types is essential for real-world implementation. In this work, we present MistSense, an approach for online mistake identification that includes this versatility by considering both procedural errors, which involve incorrect action sequences, and execution errors, such as motor inaccuracies or improper equipment use. Our method integrates RGB and hand pose features to capture fine-grained contextual cues in order to detect a mistake. By jointly modeling spatial and sequential aspects of human actions, our framework enables robust and adaptive error detection in dynamic environments. Once a mistake has been detected, we leverage a large language model (LLM) which provides an error explanation that gives the user further insights into why an action has been identified as a mistake. The evaluation on common mistake detection benchmarks shows the effectiveness of our approach.

Recently, Multimodal Large Language Models (MLLMs) have emerged as a leading framework for enhancing the ability of Large Language Models (LLMs) to interpret non-linguistic modalities. Despite their impressive capabilities, the robustness of MLLMs under conditions where one or more modalities are missing remains largely unexplored. In this paper, we investigate the extent to which MLLMs can maintain performance when faced with missing modality inputs. Moreover, we propose a novel framework to mitigate the aforementioned issue called retrieval-augmented generation for missing modalities (MissRAG). It consists of a novel multimodal RAG technique alongside a tailored prompt engineering strategy designed to enhance model robustness by mitigating the impact of absent modalities while preventing the burden of additional instruction tuning. To demonstrate the effectiveness of our techniques, we conduct comprehensive evaluations across five diverse datasets, covering tasks such as audio-visual question answering, audio-visual captioning, and multimodal sentiment analysis. Our source code is available at https://github.com/aimagelab/MissRAG.

Continual learning aims to enable models to learn sequentially from continuously incoming data while retaining performance on previously learned tasks. With the Contrastive Language-Image Pre-trained model (CLIP) exhibiting strong capabilities across various downstream tasks, there has been growing interest in leveraging CLIP for continual learning in such scenarios. Most existing works overlook the inherent modality gap in CLIP, a key factor in its generalization and adaptability. In this paper, we analyze the variations in the modality gap during the fine-tuning of vision-language pre-trained models. Our observations reveal that the modality gap effectively reflects the extent to which pre-trained knowledge is preserved. Based on these insights, we propose a simple yet effective method, MG-CLIP, that improves CLIP's performance in class-incremental learning. Our approach leverages modality gap preservation to mitigate forgetting and modality gap compensation to enhance the capacity for new data, introducing a novel modality-gap-based perspective for continual learning. Extensive experiments on multiple benchmarks demonstrate that our method outperforms existing approaches without requiring additional replay data. Our code is available at https://github.com/linlany/MindtheGap.

Leveraging the vision foundation models has emerged as a mainstream paradigm that improves the performance of image feature matching. However, previous works have ignored the misalignment when introducing the foundation models into feature matching. The misalignment arises from the discrepancy between the foundation models focusing on single-image understanding and the cross-image understanding requirement of feature matching. Specifically, 1) the embeddings derived from commonly used foundation models exhibit discrepancies with the optimal embeddings required for feature matching; 2) lacking an effective mechanism to leverage the single-image understanding ability into cross-image understanding. A significant consequence of the misalignment is they struggle when addressing multi-instance feature matching problems. To address this, we introduce a simple but effective framework, called IMD (Image feature Matching with a pre-trained Diffusion model) with two parts: 1) Unlike the dominant solutions employing contrastive-learning based foundation models that emphasize global semantics, we integrate the generative-based diffusion models to effectively capture instance-level details. 2) We leverage the prompt mechanism in generative model as a natural tunnel, propose a novel cross-image interaction prompting module to facilitate bidirectional information interaction between image pairs. To more accurately measure the misalignment, we propose a new benchmark called IMIM, which focuses on multi-instance scenarios. Our proposed IMD establishes a new state-of-the-art in commonly evaluated benchmarks, and the superior improvement 12% in IMIM indicates our method efficiently mitigates the misalignment.

By using a control variate to calibrate the local gradient of each client, Scaffold has been widely known as a powerful solution to mitigate the impact of data heterogeneity in Federated Learning. Although Scaffold achieves significant performance improvements, we show that this superiority is at the cost of increased security vulnerabilities. Specifically, this paper presents BadSFL, the first backdoor attack targeting Scaffold, which turns benign clients into accomplices to amplify the attack effect. The core idea of BadSFL is to uniquely tamper with the control variate to subtly steer benign clients' local gradient updates towards the attacker's poisoned direction, effectively turning them into unwitting accomplices, significantly enhancing the backdoor persistence. Additionally, BadSFL leverages a GAN-enhanced poisoning strategy to enrich the attacker's dataset, maintaining high accuracy on both benign and backdoored samples while remaining stealthy. Extensive experiments demonstrate that BadSFL achieves superior attack durability, maintaining effectiveness for over 60 global rounds--lasting up to three times longer than existing baselines even after ceasing malicious model injections.

Image-to-point-cloud (I2P) registration is a fundamental problem in computer vision, focusing on establishing 2D-3D correspondences between an image and a point cloud. Recently, the differentiable perspective-n-point (PnP) has been widely used to supervise I2P registration networks by enforcing projective constraints on 2D-3D correspondences. However, differentiable PnP is highly sensitive to noise and outliers in the predicted correspondences, which hinders the effectiveness of correspondence learning. Inspired by the robustness of blind PnP to noise and outliers in correspondences, we propose an approximate blind PnP-based correspondence learning approach. To mitigate the high computational cost of blind PnP, we reformulate it as a more tractable problem: minimizing the Chamfer distance between learned 2D and 3D keypoints, referred to as MinCD-PnP. To effectively solve MinCD-PnP, we introduce a lightweight multi-task learning module, MinCD-Net, which can be easily integrated into the existing I2P registration architectures. Extensive experiments on 7-Scenes, RGBD-V2, ScanNet, self-collected, and KITTI datasets demonstrate that MinCD-Net outperforms state-of-the-art methods and achieves higher inlier ratio and registration recall in both cross-scene and cross-dataset settings. The source code: https://github.com/anpei96/mincd-pnp-demo.

We propose MikuDance, a diffusion-based pipeline incorporating mixed motion dynamics to animate stylized character art. MikuDance consists of two key techniques: Mixed Motion Modeling and Mixed-Control Diffusion, to address the challenges of high-dynamic motion and reference-guidance misalignment in character art animation. Specifically, a Scene Motion Tracking strategy is presented to explicitly model the dynamic camera in pixel-wise space, enabling unified character-scene motion modeling. Building on this, the Mixed-Control Diffusion implicitly aligns the scale and body shape of diverse characters with motion guidance, allowing flexible control of local character motion. Subsequently, a Motion-Adaptive Normalization module is incorporated to effectively inject global scene motion, paving the way for comprehensive character art animation. Through extensive experiments, we demonstrate the effectiveness and generalizability of MikuDance across various character art and motion guidance, consistently producing high-quality animations with remarkable motion dynamics.

We present MiDSummer, a two-stage framework for generating immersive Gaussian Splatting scenes that leverages multiple diffusion guidance signals to enable structured layout control, enhanced physical realism, and improved visual quality.While 3D scene generation has seen significant recent advances, current approaches could benefit from: (1) achieving precise, reliable layout control while preserving open-world generalization and physical plausibility, (2) balancing high-level semantic reasoning with low-level, directly controllable geometric constraints, and (3) effectively utilizing layout knowledge for visual refinement. Our work addresses these challenges through a structured two-stage planning-assembly framework. For planning, we introduce a dual layout diffusion guidance approach to bridge semantic reasoning and geometric controllability. Our approach uniquely integrates LLMs' open-vocabulary reasoning with Graph Diffusion Models' (GDM) geometric precision by incorporating multi-level self-consistency scores over scene graph structures and layout bounding box parameters. This fusion enables fine-grained control over scene composition while ensuring physical plausibility and faithful prompt interpretation. For assembly, we propose a layout-guided optimization technique for scene refinement. We effectively incorporate layout priors obtained during the planning stage into a Stable Diffusion (SD)-based refinement process that jointly optimizes camera trajectories and scene splats. This layout-aware joint optimization, constrained by multi-view consistency, produces visually compelling immersive scenes that are structurally coherent and controllable.

Standard convolutions are prevalent in image processing and deep learning, but their fixed kernels limits adaptability. Several deformation strategies of the reference kernel grid have been proposed. Yet, they lack a unified theoretical framework. By returning to a metric perspective for images, now seen as two-dimensional manifolds equipped with notions of local and geodesic distances, either symmetric (Riemannian) or not (Finsler), we provide a unifying principle: the kernel positions are samples of unit balls of implicit metrics. With this new perspective, we also propose metric convolutions, a novel approach that samples unit balls from explicit signal-dependent metrics, providing interpretable operators with geometric regularisation. This framework, compatible with gradient-based optimisation, can directly replace existing convolutions applied to either input images or deep features of neural networks. Metric convolutions typically require fewer parameters and provide better generalisation. Our approach shows competitive performance in standard denoising and classification tasks.

Miniaturized endoscopy has advanced accurate visual perception within the human body. Prevailing research remains limited to conventional cameras employing convex lenses, where the physical constraints with millimetre-scale thickness impose serious impediments on the micro-level clinical. Recently, with the emergence of meta-optics, ultra-micro imaging based on metalenses (micron-scale) has garnered great attention, serving as a promising solution. However, due to the physical difference of metalens, there is a large gap in data acquisition and algorithm research. In light of this, we aim to bridge this unexplored gap, advancing the novel metalens endoscopy. First, we establish datasets for metalens endoscopy and conduct preliminary optical simulation, identifying two derived optical issues that physically adhere to strong optical priors. Second, we propose MetaScope, a novel optics-driven neural network tailored for metalens endoscopy driven by physical optics. MetaScope comprises two novel designs: Optics-informed Intensity Adjustment (OIA), rectifying intensity decay by learning optical embeddings, and Optics-informed Chromatic Correction (OCC), mitigating chromatic aberration by learning spatial deformations informed by learned Point Spread Function (PSF) distributions. To enhance joint learning, we further deploy a gradient-guided distillation to transfer knowledge from the foundational model adaptively. Extensive experiments demonstrate that MetaScope not only outperforms state-of-the-art methods in both metalens segmentation and restoration but also achieves impressive generalized ability in real biomedical scenes.

In this work, we propose Visual-Predictive Instruction Tuning (VPiT) - a simple and effective extension to visual instruction tuning that enables a pretrained LLM to quickly morph into an unified autoregressive model capable of generating both text and visual tokens. VPiT teaches an LLM to predict discrete text tokens and continuous visual tokens from any input sequence of image and text data curated in an instruction-following format. Our empirical investigation reveals several intriguing properties of VPiT: (1) visual generation ability emerges as a natural byproduct of improved visual understanding, and can be unlocked efficiently with a small amount of generation data; (2) while we find understanding and generation to be mutually beneficial, understanding data contributes to both capabilities more effectively than generation data. Building upon these findings, we train our MetaMorph model and achieve competitive performance on both visual understanding and generation. In visual generation, MetaMorph can leverage the world knowledge and reasoning abilities gained from LLM pretraining, and overcome common failure modes exhibited by other generation models. Our results suggest that LLMs may have strong "prior" vision capabilities that can be efficiently adapted to both visual understanding and generation with a relatively simple instruction tuning process.

With the rapid progress of diffusion models (DMs), significant efforts are being made to unlearn harmful or copyrighted concepts from pretrained DMs to prevent potential model misuse. However, it is observed that even when DMs are properly unlearned before release, malicious finetuning can compromise this process, causing DMs to relearn the unlearned concepts. This occurs partly because certain benign concepts (e.g., "skin") retained in DMs are related to the unlearned ones (e.g., "nudity"), facilitating their relearning via finetuning. To address this, we propose meta-unlearning on DMs. Intuitively, a meta-unlearned DM should behave like an unlearned DM when used as is; moreover, if the meta-unlearned DM undergoes malicious finetuning on unlearned concepts, the related benign concepts retained within it will be triggered to self-destruct, hindering the relearning of unlearned concepts. Our meta-unlearning framework is compatible with most existing unlearning methods, requiring only the addition of an easy-to-implement meta objective. We validate our approach through empirical experiments on meta-unlearning concepts from Stable Diffusion models (SD-v1-4 and SDXL), supported by extensive ablation studies.

The sample selection approach is a widely adopted strategy for learning with noisy labels, where examples with lower losses are effectively treated as clean during training. However, this clean set often becomes dominated by easy examples, limiting the model's meaningful exposure to more challenging cases and reducing its expressive power. To overcome this limitation, we introduce a novel metric called Dynamic Center Distance (DCD), which can quantify sample difficulty and provide information that critically complements loss values. Unlike approaches that rely on predictions, DCD is computed in feature space as the distance between sample features and a dynamically updated center, established through a proposed meta-learning framework. Building on preliminary semi-supervised training that captures fundamental data patterns, we incorporate DCD to further refine the classification loss, down-weighting well-classified examples and strategically focusing training on a sparse set of hard instances. This strategy prevents easy examples from dominating the classifier, leading to more robust learning. Extensive experiments across multiple benchmark datasets, including synthetic and real-world noise settings, as well as natural and medical images, consistently demonstrate the effectiveness of our method.

The increasing frequency of extreme weather events due to global climate change urges accurate weather prediction. Recently, great advances are made by the end-to-end methods, thanks to deep learning techniques, but they face limitations of representation inconsistency in multivariable integration and struggle to effectively capture the dependency between variables, which is required in complex weather systems. Treating different variables as distinct modalities and applying a two-stage training approach from multimodal models can partially alleviate this issue, but due to the inconformity in training tasks between the two stages, the results are often suboptimal. To address these challenges, we propose an implicit two-stage training method, configuring separate encoders and decoders for each variable. In detailed, in the first stage, the Translator is frozen while the Encoders and Decoders learn a shared latent space, in the second stage, the Encoders and Decoders are frozen, and the Translator captures inter-variable interactions for prediction. Besides, by introducing a self-attention mechanism for multivariable fusion in the latent space, the performance achieves further improvements. Empirically, extensive experiments shows state-of-the-art performance of our method. In specific, it reduces the MSE for near-surface air temperature and relative humidity predictions by 28.82% and 23.39%, respectively.

We introduce MeshPad, a generative approach that creates 3D meshes from sketch inputs. Building on recent advances in artist-reminiscent triangle mesh generation, our approach addresses the need for interactive mesh creation. To this end, we focus on enabling consistent edits by decomposing editing into 'deletion' of regions of a mesh, followed by 'addition' of new mesh geometry. Both operations are invoked by simple user edits of a sketch image, facilitating an iterative content creation process and enabling the construction of complex 3D meshes. Our approach is based on a triangle sequence-based mesh representation, exploiting a large Transformer model for mesh triangle addition and deletion. In order to perform edits interactively, we introduce a vertex-aligned speculative prediction strategy on top of our additive mesh generator. This speculator predicts multiple output tokens corresponding to a vertex, thus significantly reducing the computational cost of inference and accelerating the editing process, making it possible to execute each editing step in only a few seconds. Comprehensive experiments demonstrate that MeshPad outperforms state-of-the-art sketch-conditioned mesh generation methods, achieving more than 22% mesh quality improvement in Chamfer distance, and being preferred by 90% of participants in perceptual evaluations.

In this paper, we introduce MeshMamba, a neural network model for learning 3D articulated mesh models by employing the recently proposed Mamba State Space Models (Mamba-SSMs). MeshMamba is efficient and scalable in handling a large number of input tokens, enabling the generation and reconstruction of body mesh models with more than 10,000 vertices, capturing clothing and hand geometries. The key to effectively learning MeshMamba is the serialization technique of mesh vertices into orderings that are easily processed by Mamba. This is achieved by sorting the vertices based on body part annotations or the 3D vertex locations of a template mesh, such that the ordering respects the structure of articulated shapes. Based on MeshMamba, we design 1) MambaDiff3D, a denoising diffusion model for generating 3D articulated meshes and 2) Mamba-HMR, a 3D human mesh recovery model that reconstructs a human body shape and pose from a single image. Experimental results showed that MambaDiff3D can generate dense 3D human meshes in clothes, with grasping hands, etc., and outperforms previous approaches in the 3D human shape generation task. Additionally, Mamba-HMR extends the capabilities of previous non-parametric human mesh recovery approaches, which were limited to handling body-only poses using around 500 vertex tokens, to the whole-body setting with face and hands, while achieving competitive performance in (near) real-time.