ICCV 2025 Papers

Video large language models (ViLLMs) excel in general video understanding, e.g., recognizing activities like talking and eating, but struggle with identity-aware comprehension, such as "Wilson is receiving chemotherapy" or "Tom is discussing with Sarah", limiting their applicability in smart healthcare and smart home environments. To address this limitation, we propose a one-shot learning framework PVChat, the first personalized ViLLM that enables subject-aware question answering (QA) from a single video for each subject. Our approach optimizes a Mixture-of-Heads (MoH) enhanced ViLLM on a synthetically augmented video-QA dataset, leveraging a progressive image-to-video learning strategy. Specifically, we introduce an automated augmentation pipeline that synthesizes identity-preserving positive samples and retrieves hard negatives from existing video corpora, generating a diverse training dataset with four QA types: existence, appearance, action, and location inquiries. To enhance subject-specific learning, we propose a ReLU Routing MoH attention mechanism, alongside two novel objectives: (1) Smooth Proximity Regularization for progressive learning through exponential distance scaling and (2) Head Activation Enhancement for balanced attention routing. Finally, we adopt a two-stage training strategy, transitioning from image pre-training to video fine-tuning, enabling a gradual learning process from static attributes to dynamic representations. We evaluate PVChat on diverse datasets covering medical scenarios, TV series, anime, and real-world footage, demonstrating its superiority in personalized feature understanding after learning from a single video, compared to state-of-the-art ViLLMs.

Motion skeletons drive 3D character animation by transforming bone hierarchies, but differences in proportions or structure make motion data hard to transfer across skeletons, posing challenges for data-driven motion synthesis. Temporal Point Clouds (TPCs) offer an unstructured, cross-compatible motion representation. Though reversible with skeletons, TPCs mainly serve for compatibility, not for direct motion task learning. Doing so would require data synthesis capabilities for the TPC format, which presents unexplored challenges regarding its unique temporal consistency and point identifiability. Therefore, we propose PUMPS, the primordial autoencoder architecture for TPC data. PUMPS independently reduces frame-wise point clouds into sampleable feature vectors, from which a decoder extracts distinct temporal points using latent Gaussian noise vectors as sampling identifiers. We introduce linear assignment-based point pairing to optimise the TPC reconstruction process, and negate the use of expensive point-wise attention mechanisms in the architecture. Using these latent features, we pre-train a motion synthesis model capable of performing motion prediction, transition generation, and keyframe interpolation. For these pre-training tasks, PUMPS performs remarkably well even without native dataset supervision, matching state-of-the-art performance. When fine-tuned for motion denoising or estimation, PUMPS outperforms many respective methods without deviating from its generalist architecture.

Recent advancements in multimodal foundation models have yielded significant progress in vision-language understanding. Initial attempts have also explored the potential of multimodal large language models for visual content generation. However, existing approaches face a trade-off between generation diversity and controllability, struggling to meet the varying granularity demands of different image generation tasks within a unified MLLM framework. In this work, we propose PUMA, emPowering Unified MLLM with Multi-grAnular visual generation, a novel paradigm that tackles the diversity-controllability trade-off. PUMA achieves this by unifying multi-granular visual features as both inputs and outputs of MLLMs, thus effectively meeting the distinct granularity needs for diverse generation and precise manipulation within a single framework. Following multimodal pretraining and instruction tuning, PUMA demonstrates remarkable capabilities in a wide range of multimodal tasks, including image understanding, diverse text-to-image generation, editing, inpainting, colorization, and conditional generation. This work marks a significant stride towards realizing truly unified MLLMs capable of seamlessly adapting to the diverse granularity demands and task requirements inherent in various visual tasks. The code and model will be released upon acceptance.

Current multimodal fusion approaches in computational oncology primarily focus on integrating multi-gigapixel histology whole slide images (WSIs) with genomic or transcriptomic data, demonstrating improved survival prediction. We hypothesize that incorporating pathology reports can further enhance prognostic performance. Pathology reports, as essential components of clinical workflows, offer readily available complementary information by summarizing histopathological findings and integrating expert interpretations and clinical context. However, fusing these modalities poses challenges due to their heterogeneous nature. WSIs are high-dimensional, each containing several billion pixels, whereas pathology reports consist of concise text summaries of varying lengths, leading to potential modality imbalance. To address this, we propose a prototype-based approach to generate balanced representations, which are then integrated using a Transformer-based fusion model for survival prediction that we term PS3 (Predicting Survival from Three Modalities). Specifically, we present: (1) Diagnostic prototypes from pathology reports, leveraging self-attention to extract diagnostically relevant sections and standardize text representation; (2) Histological prototypes to compactly represent key morphological patterns in WSIs; and (3) Biological pathway prototypes to encode transcriptomic expressions, accurately capturing cellular functions. PS3, the three-modal transformer model, processes the resulting prototype-based multimodal tokens and models intra-modal and cross-modal interactions across pathology reports, WSIs and transcriptomic data. The proposed model outperforms state-of-the-art methods when evaluated against clinical, unimodal and multimodal baselines on six datasets from The Cancer Genome Atlas (TCGA). The code will be released upon publication.

Human pose sequence refinement plays a crucial role in improving the temporal coherence of pose estimation across the sequence of frames. Despite its importance in real-world applications, human pose sequence refinement has received less attention than human pose estimation. In this paper, we propose PS-Mamba, a novel framework that refines human pose sequences by integrating spatial-temporal graph learning with state space modeling. Specifically, we introduce the Spatial-Temporal Graph State Space (ST-GSS) block, which captures spatial and temporal dependencies across joints to smooth pose sequences while preserving structural integrity. The spatial-temporal graph learns intricate joint interactions, while the state space component effectively manages temporal dynamics, reducing both short- and long-term pose instability. Besides, we incorporate a dynamic graph weight matrix to adaptively model the relative influence of joint interactions, further mitigating pose ambiguity. Experiments on challenging benchmarks show that our PS-Mamba outperforms SOTAs, achieving -14.21 mm MPJPE (+18.5%\uparrow), -13.59 mm PA-MPJPE (+22.1%\uparrow), and -0.42 mm/s2 ACCEL (+9.7%\uparrow) compared to SynSP on AIST++, significantly reducing jitters and enhancing pose stability.

The data privacy constraint in online continual learning (OCL), where the data can be seen only once, complicates the catastrophic forgetting problem in streaming data. A common approach applied by the current SOTAs in OCL is the use of memory-saving exemplars or features from previous classes to be replayed in the current task. On the other hand, the prompt-based approach performs excellently in continual learning, but at the cost of a growing number of trainable parameters. The first approach may not be applicable in practice due to the data openness policy, while the second approach has the issue of throughput associated with the streaming data. In this study, we propose a novel prompt-based method for online continual learning that includes 4 main components: (1) single light-weight prompt generator as a general knowledge, (2) trainable scaler-and-shifter as specific knowledge, (3) pre-trained model (PTM) generalization preserving, and (4) hard-soft updates mechanism. Our proposed method achieves significantly higher performance than the current SOTAs in CIFAR100, ImageNet-R, ImageNet-A, and CUB dataset. Our complexity analysis shows that our method requires a relatively smaller number of parameters and achieves moderate training time, inference time, and throughput. For further study, the source code of our method is available at https://github.com/anwarmaxsum/PROL.

We present PROGRESSOR, a novel framework that learns a task-agnostic reward function from videos, enabling policy training through goal-conditioned reinforcement learning (RL) without manual supervision. Underlying this reward is an estimate of the distribution over task progress as a function of the current, initial, and goal observations that is learned in a self-supervised fashion. Crucially, PROGRESSOR refines rewards adversarially during online RL training by pushing back predictions for out-of-distribution observations in order to mitigate distribution shift inherent in non-expert observations. Utilizing this progress prediction as a dense reward together with an adversarial push-back, we show that PROGRESSOR enables robots to learn complex behaviors without any external supervision. Pretrained on large-scale egocentric human video from EPIC-KITCHENS, PROGRESSOR requires no fine-tuning on in-domain task-specific data for generalization to real-robot offline RL under noisy demonstrations, outperforming contemporary methods that provide dense visual reward for robotic learning. Our findings highlight the potential of PROGRESSOR for scalable robotic applications where direct action labels and task-specific rewards are not readily available.

We propose PRM, a novel photometric stereo based large reconstruction model to reconstruct high-quality meshes with fine-grained details. Previous large reconstruction models typically prepare training images under fixed and simple lighting, offering minimal photometric cues for precise reconstruction. Furthermore, images containing specular surfaces are treated as out-of-distribution samples, resulting in degraded reconstruction quality. To handle these challenges, PRM renders images by varying materials and lighting, which not only improves the local details by providing rich photometric cues but also increases the model's robustness to variations in the appearance of input images. To offer enhanced flexibility, we incorporate a real-time physically-based rendering (PBR) method and mesh rasterization for ground-truth rendering. By using an explicit mesh as 3D representation, PRM ensures the application of differentiable PBR for predicted rendering. This approach models specular color more accurately for images with varying materials and illumination than previous neural rendering methods and supports multiple supervisions for geometry optimization. Extensive experiments demonstrate that PRM significantly outperforms other models.

We introduce Projection-based Reduction of Implicit Spurious bias in vision-language Models (PRISM), a new data-free and task-agnostic solution for bias mitigation in VLMs like CLIP. VLMs often inherit and amplify biases in their training data, leading to skewed predictions.PRISM is designed to debias VLMs without relying on predefined bias categories or additional external data. It operates in two stages: first, an LLM is prompted with simple class prompts to generate scene descriptions that contain spurious correlations. Next, PRISM uses our novel contrastive-style debiasing loss to learn a projection that maps the embeddings onto a latent space that minimizes spurious correlations while preserving the alignment between image and text embeddings. Extensive experiments demonstrate that PRISM outperforms current debiasing methods on the commonly used Waterbirds and CelebA datasets We make our code public at: https://github.com/MahdiyarMM/PRISM.

We formulate the motor system of an interactive avatar as a generative motion model that can drive the body to move through 3D space in a perpetual, realistic, controllable, and responsive manner. Although human motion generation has been extensively studied, many existing methods lack the responsiveness and realism of real human movements. Inspired by recent advances in foundation models, we propose PRIMAL, which is learned with a two-stage paradigm. In the pretraining stage, the model learns body movements from a large number of sub-second motion segments, providing a generative foundation from which more complex motions are built. This training is fully unsupervised without annotations. Given a single-frame initial state during inference, the pretrained model not only generates unbounded, realistic, and controllable motion, but also enables the avatar to be responsive to induced impulses in real time. In the adaptation phase, we employ a novel ControlNet-like adaptor to fine-tune the base model efficiently, adapting it to new tasks such as few-shot personalized action generation and spatial target reaching. Evaluations show that our proposed method outperforms state-of-the-art baselines. We leverage the model to create a real-time character animation system in Unreal Engine that feels highly responsive and natural.

Event cameras excel in high temporal resolution and dynamic range but suffer from dense noise in rainy conditions. Existing event deraining methods face trade-offs between temporal precision, deraining effectiveness, and computational efficiency. In this paper, we propose PRE-Mamba, a novel point-based event camera deraining framework that fully exploits the spatiotemporal characteristics of raw event and rain. Our framework introduces a 4D event cloud representation that integrates dual temporal scales to preserve high temporal precision, a Spatio-Temporal Decoupling and Fusion module (STDF) that enhances deraining capability by enabling shallow decoupling and interaction of temporal and spatial information, and a Multi-Scale State Space Model (MS3M) that captures deeper rain dynamics across dual-temporal and multi-spatial scales with linear computational complexity. Enhanced by frequency-domain regularization, PRE-Mamba achieves superior performance (0.95 SR, 0.91 NR, and 0.4s/M events) with only 0.26M parameters on EventRain-27K, a comprehensive dataset with labeled synthetic and real-world sequences. Moreover, our method generalizes well across varying rain intensities, viewpoints, and even snowy conditions. Code and dataset: https://github.com/softword-tt/PRE-Mamba.

Recent approaches to 3D reconstruction in dynamic scenes primarily rely on the integration of separate geometry estimation and matching modules, where the latter plays a critical role in distinguishing dynamic regions and mitigating the interference caused by moving objects. Furthermore, the matching module explicitly models object motion, enabling the tracking of specific targets and advancing motion understanding in complex scenarios. Recently, the proposed representation of pointmap in DUSt3R suggests a potential solution to unify both geometry estimation and matching in 3D space, effectively reducing computational overhead by eliminating the need for redundant auxiliary modules. However, it still struggles with ambiguous correspondences in dynamic regions, which limits reconstruction performance in such scenarios. In this work, we present POMATO, a unified framework for dynamic 3D reconstruction by marrying POintmap MAtching with Temporal mOtion. Specifically, our method first learns an explicit matching relationship by mapping RGB pixels across different views to 3D pointmaps within a unified coordinate system. Furthermore, we introduce a temporal motion module for dynamic motions that ensures scale consistency across different frames and enhances performance in 3D reconstruction tasks requiring both precise geometry and reliable matching, most notably 3D point tracking. We show the effectiveness of our proposed POMATO by demonstrating the remarkable performance across multiple downstream tasks, including video depth estimation, 3D point tracking, and pose estimation. Code and models are publicly available at https://github.com/wyddmw/POMATO.

We completely classify all minimal problems for Structure-from-Motion (SfM) where arrangements of points and lines are fully observed by multiple uncalibrated pinhole cameras. We find 291 minimal problems, 73 of which have unique solutions and can thus be solved linearly.Two of the linear problems allow an arbitrary number of views, while all other minimal problems have at most 9 cameras. All minimal problems have at most 7 points and at most 12 lines. We compute the number of solutions of each minimal problem, as this gives a measurement of the problem's intrinsic difficulty, and find that these number are relatively low (e.g., when comparing with minimal problems for calibrated cameras). Finally, by exploring stabilizer subgroups of subarrangements, we develop a geometric and systematic way to 1) factorize minimal problems into smaller problems, 2) identify minimal problems in underconstrained problems, and 3) formally prove non-minimality.

Continual learning (CL) requires models to continuously adapt to new tasks without forgetting past knowledge. In this work, we propose \underline P roactive \underline L ow-rank \underline A llocatio\underline N (PLAN), a framework that extends Low-Rank Adaptation (LoRA) to enable efficient and interference-aware fine-tuning of large pre-trained models in CL settings. PLAN proactively manages the allocation of task-specific subspaces by introducing orthogonal basis vectors for each task and optimizing them through a perturbation-based strategy that minimizes conflicts with previously learned parameters. Furthermore, PLAN incorporates a novel selection mechanism that identifies and assigns basis vectors with minimal sensitivity to interference, reducing the risk of degrading past knowledge while maintaining efficient adaptation to new tasks. Empirical results on standard CL benchmarks demonstrate that PLAN consistently outperforms existing methods, establishing a new state-of-the-art for continual learning with foundation models.

Diffusion models have shown impressive results in generating high-quality conditional samples using guidance techniques such as Classifier-Free Guidance (CFG). However, existing methods often require additional training or neural function evaluations (NFEs), making them incompatible with guidance-distilled models. Also, they rely on heuristic approaches that need identifying target layers. In this work, we propose a novel and efficient method, termed PLADIS, which boosts pre-trained models (U-Net/Transformer) by leveraging sparse attention. Specifically, we extrapolate query-key correlations using softmax and its sparse counterpart in the cross-attention layer during inference, without requiring extra training or NFEs. By leveraging the noise robustness of sparse attention, our PLADIS unleashes the latent potential of text-to-image diffusion models, enabling them to excel in areas where they once struggled with newfound effectiveness. It integrates seamlessly with guidance techniques, including guidance-distilled models. Extensive experiments show notable improvements in text alignment and human preference, offering a highly efficient and universally applicable solution. See our project page: https://github.com/cubeyoung/PLADIS

Text-to-Image (T2I) models have gained widespread adoption across various applications. Despite the success, the potential misuse of T2I models poses significant risks of generating Not-Safe-For-Work (NSFW) content. To investigate the vulnerability of T2I models, this paper delves into adversarial attacks to bypass the safety mechanisms under black-box settings. Most previous methods rely on word substitution to search adversarial prompts. Due to limited search space, this leads to suboptimal performance compared to gradient-based training. However, black-box settings present unique challenges to training gradient-driven attack methods, since there is no access to the internal architecture and parameters of T2I models. To facilitate the learning of adversarial prompts in black-box settings, we propose a novel prompt learning attack framework (PLA), where insightful gradient-based training tailored to black-box T2I models is designed by utilizing multimodal similarities. Experiments show that our new method can effectively attack the safety mechanisms of black-box T2I models including prompt filters and post-hoc safety checkers with a high success rate compared to state-of-the-art methods. Warning: This paper may contain offensive model-generated content.

Generating realistic group interactions involving multiple characters remains challenging due to increasing complexity as group size expands. While existing conditional diffusion models incrementally generate motions by conditioning on previously generated characters, they rely on single shared prompts, limiting nuanced control and leading to overly simplified interactions. In this paper, we introduce Person-Interaction Noise Optimization (PINO), a novel, training-free framework designed for generating realistic and customizable interactions among groups of arbitrary size. PINO decomposes complex group interactions into semantically relevant pairwise interactions, and leverages pretrained two-person interaction diffusion models to incrementally compose group interactions. To ensure physical plausibility and avoid common artifacts such as overlapping or penetration between characters, PINO employs physics-based penalties during noise optimization. This approach allows precise user control over character orientation, speed, and spatial relationships without additional training. Comprehensive evaluations demonstrate that PINO generates visually realistic, physically coherent, and adaptable multi-person interactions suitable for diverse animation, gaming, and robotics applications.

We introduce PHD, a novel approach for personalized 3D human mesh recovery (HMR) and body fitting that leverages user-specific shape information to improve pose estimation accuracy from videos. Traditional HMR methods are designed to be user-agnostic and optimized for generalization. While these methods often refine poses using constraints derived from the 2D image to improve alignment, this process compromises 3D accuracy by failing to jointly account for person-specific body shapes and the plausibility of 3D poses. In contrast, our pipeline decouples this process by first calibrating the user's body shape and then employing a personalized pose fitting process conditioned on that shape. To achieve this, we develop a body shape-conditioned 3D pose prior, implemented as a Point Diffusion Transformer, which iteratively guides the pose fitting via a Point Distillation Sampling loss. This learned 3D pose prior effectively mitigates errors arising from an over-reliance on 2D constraints. Consequently, our approach improves not only pelvis-aligned pose accuracy but also absolute pose accuracy -- an important metric often overlooked by prior work. Furthermore, our method is highly data-efficient, requiring only synthetic data for training, and serves as a versatile plug-and-play module that can be seamlessly integrated with existing 3D pose estimators to enhance their performance. Code and models are available at https://PHD-Pose.github.io.