CVPR 2026 论文列表

Machine unlearning (MU) aims to remove sensitive or undesired content from pre-trained models. Existing MU methods are commonly characterized as gradually degrading model performance on undesired data to realize approximate forgetting. Despite their successes, the effectiveness in multimodal unlearning tasks remains largely unexplored. In this paper, we first conduct an in-depth analysis and reveal that traditional MU methods tend to disrupt cross-modal alignment, leading to incomplete forgetting in multimodal scenarios. To tackle this challenge, we propose VL-Eraser, a novel unlearning paradigm for VLM unlearning. VL-Eraser reformulates unlearning in VLMs as a two-stage process: distillation and deletion. Specifically, VL-Eraser first introduces a vacuum distillation that disentangles undesired knowledge from the intricate parameters of VLMs and transfers it into low-rank adapters (LoRA). After distillation, unlearning is efficiently achieved by deleting the LoRA parameters from the original model. Extensive experiments across multiple benchmarks demonstrate that VL-Eraser achieves superior unlearning performance while preserving utility compared to the state-of-the-art baselines.

Language-driven dexterous grasp generation requires the models to understand task semantics, 3D geometry, and complex hand-object interactions. While vision-language models have been applied to this problem, existing approaches directly map observations to grasp parameters without intermediate reasoning about physical interactions. We present DextER, Dexterous Grasp Generation with Embodied Reasoning, which introduces contact-based embodied reasoning for multi-finger manipulation. Our key insight is that predicting which hand links contact where on the object surface provides an embodiment-aware intermediate representation, bridging task semantics with physical constraints. DextER autoregressively generates embodied contact tokens specifying which finger links contact where on the object surface, followed by grasp tokens encoding the hand configuration. On DexGYS, DextER achieves 67.14% success rate, outperforming state-of-the-art by 3.83 p.p. with 96.4% improvement in intention alignment. We also demonstrate steerable generation through partial contact specification, providing fine-grained control over grasp synthesis.

Large-scale vision-language models (VLMs) have recently achieved remarkable multimodal understanding, but their massive size makes them impractical for deployment on mobile or edge devices. This raises the need for compact yet capable VLMs that can efficiently learn from powerful large teacher. However, distilling knowledge from large teacher to small student remains challenging due to their large size gap: the student often fails to reproduce the teacher's complex, high-dimensional representations, leading to unstable learning and degraded performance. To address this, we propose Masters (Masking teacher and reinforcing student), a mask-progressive reinforcement learning (RL) distillation framework. Masters first masks and non-dominant weights of the teacher to reduce unnecessary complexity, then progressively restores the teacher from mask to gradually increase the teacher capacity during training. This strategy allows the student to learn richer representations of teacher in a smooth and stable manner. To further refine knowledge transfer, Masters integrates an offline RL stage with two complementary rewards: an accuracy reward that measures the correctness of the generated responses, and a distillation reward that quantifies the ease of their responses' transferability from teacher to student. Unlike online think-answer RL paradigms that are computationally expensive and generate lengthy responses, our offline RL leverages pre-generated responses from masked teachers. These provide rich yet efficient guidance, enabling the students to achieve strong performance without requiring the think-answer process. Extensive experiments across diverse VLM benchmarks demonstrate that Masters outperforms existing compact VLMs and partially surpasses large ones, while being far more efficient. Moreover, gradually increasing the teacher sizes during distillation (e.g., from 14B to 38B) yields smoother convergence and stronger generalization than one-shot distillation (e.g., 38B), revealing a scalable path toward efficient and deployable VLMs.

Scalable Vector Graphics (SVG) are central to modern web design, and the demand to animate them continues to grow as web environments become increasingly dynamic.Yet automating the animation of vector graphics remains challenging for vision-language models (VLMs) despite recent progress in code generation and motion planning.VLMs routinely mis-handle SVGs, since visually coherent parts are often fragmented into low-level shapes that offer little guidance of which elements should move together. In this paper, we introduce a framework that recovers the semantic structure required for reliable SVG animation and reveals the missing layer that current VLM systems overlook. This is achieved through a statistical aggregation of multiple weak part predictions, allowing the system to stably infer semantics from noisy predictions.By reorganizing SVGs into semantic groups, our approach enables VLMs to produce animations with far greater coherence. Our experiments demonstrate substantial gains over existing approaches, suggesting that semantic recovery is the key step that unlocks robust SVG animation and supports more interpretable interactions between VLMs and vector graphics.

High-fidelity generative video editing has seen significant quality improvements by leveraging pre-trained video foundation models. However, their computational cost is a major bottleneck, as they are often designed to inefficiently process the full video context regardless of the inpainting mask's size, even for sparse, localized edits. In this paper, we introduce EditCtrl, an efficient video inpainting control framework that focuses computation only where it is needed. Our approach features a novel local video context module that operates solely on masked tokens, yielding a computational cost proportional to the edit size. This local-first generation is then guided by a lightweight temporal global context embedder that ensures video-wide context consistency with minimal overhead. Not only is EditCtrl 10xmore compute efficient than state-of-the-art generative editing methods, it even improves editing quality compared to methods designed with full-attention. Finally, we showcase how EditCtrl unlocks new capabilities, including multi-region editing with text prompts and real-time autoregressive content propagation.

Realistic and efficient 3D garment generation remains a longstanding challenge in computer vision and digital fashion. Existing methods typically rely on large vision- language models to produce serialized representations of 2D sewing patterns, which are then transformed into simulation-ready 3D meshes using garment modeling framework such as GarmentCode. Although these approaches yield high-quality results, they often suffer from slow inference times, ranging from 30 seconds to a minute. In this work, we introduce SwiftTailor, a novel two-stage framework that unifies sewing-pattern reasoning and geometry-based mesh synthesis through a compact geometry image representation. SwiftTailor comprises two lightweight modules: PatternMaker, an efficient vision-language model that predicts sewing patterns from diverse input modalities, and GarmentSewer, an efficient dense prediction transformer that converts these patterns into a novel Garment Geometry Image, encoding the 3D surface of all garment panels in a unified UV space. The final 3D mesh is reconstructed through an efficient inverse mapping process that incorporates remeshing and dynamic stitching algorithms to directly assemble the garment, thereby amortizing the cost of physical simulation. Extensive experiments on the Multimodal GarmentCodeData demonstrate that SwiftTailor achieves state-of-the-art accuracy and visual fidelity while significantly reducing inference time. This work offers a scalable, interpretable, and high-performance solution for next-generation 3D garment generation.

Vision foundation models have shown great promise for open-set 3D object retrieval (3DOR) through efficient adaptation to multi-view images. Leveraging semantically aligned latent space, previous work typically adapts the CLIP encoder to build view-based 3D descriptors. Despite CLIP's strong generalization ability, its lack of fine-grainedness prompted us to explore the potential of a more recent self-supervised encoder--DINO. To address this, we propose DINO Eats CLIP (DEC), a novel framework for dynamic multi-view integration that is regularized by synthesizing data for unseen classes. We first find that simply mean-pooling over view features from a frozen DINO backbone gives decent performance. Yet, further adaptation causes severe overfitting on average view patterns of known classes. To combat it, we then design a module named Chunking and Adapting Module (CAM). It segments multi-view images into chunks and dynamically integrates local view relations, yielding more robust features than the standard pooling strategy. Finally, we propose Virtual Feature Synthesis (VFS) module to mitigate bias towards known categories explicitly. Under the hood, VFS leverages CLIP's broad, pre-aligned vision-language space to synthesize virtual features for unseen classes. By exposing DEC to these virtual features, we greatly enhance its open-set discrimination capacity. Extensive experiments on standard open-set 3DOR benchmarks demonstrate its superior efficacy.

Graphical user interface (GUI) agents powered by large vision-language models (VLMs) have shown remarkable potential in automating digital tasks, highlighting the need for high-quality trajectory data to support effective agent training. Yet existing trajectory synthesis pipelines often yield agents that fail to generalize beyond simple interactions. We identify this limitation as stemming from the neglect of semantic-ambiguous actions--interactions whose meanings are context-dependent, sequentially dependent, or visually ambiguous. Such actions are crucial for real-world robustness but are under-represented and poorly processed in current datasets, leading to semantic misalignment between task instructions and execution. To address these issues, we propose HATS, a Hardness-Aware Trajectory Synthesis framework designed to mitigate the impact of semantic ambiguity. We define hardness as the degree of semantic ambiguity associated with an action and develop two complementary modules: (1) a hardness-driven exploration that guides data collection toward ambiguous yet informative interactions, and (2) an alignment-guided refinement that iteratively validates and repairs instruction-execution consistency. The two modules operate in a closed-loop manner--exploration supplies refinement with challenging trajectories, while refinement feedback updates the hardness signal to guide future exploration. Extensive experiments show that agents trained with HATS consistently outperform state-of-the-art baselines across benchmark GUI environments.

Fine-grained anomaly detection is crucial in industrial and medical applications, but labeled anomalies are often scarce, making zero-shot detection challenging. While vision-language models like CLIP offer promising solutions, they struggle with foreground-background feature entanglement and coarse textual semantics. We propose FB-CLIP, a framework that enhances anomaly localization via multi-strategy textual representations and foreground-background separation. In the textual modality, it combines End-of-Text features, global-pooled representations, and attention-weighted token features for richer semantic cues. In the visual modality, multi-view soft separation along identity, semantic, and spatial dimensions, together with background suppression, reduces interference and improves discriminability. Semantic Consistency Regularization (SCR) aligns image features with normal and abnormal textual prototypes, suppressing uncertain matches and enlarging semantic gaps. Experiments show that FB-CLIP effectively distinguishes anomalies from complex backgrounds, achieving accurate fine-grained anomaly detection and localization under zero-shot settings. Code : \href https://github.com/Xi-Mu-Yu/FB-CLIP https://github.com/Xi-Mu-Yu/FB-CLIP .

We present PosterIQ, a design-driven benchmark for poster understanding and generation, annotated across composition structure, typographic hierarchy, and semantic intent. It includes 7,765 image-annotation instances and 822 generation prompts spanning real, professional, and synthetic cases. To bridge visual design cognition and generative modeling, we define tasks for layout parsing, text-image correspondence, typography/readability and font perception, design quality assessment, and controllable, composition-aware generation with metaphor. We evaluate state-of-the-art MLLMs and diffusion-based generators, finding persistent gaps in visual hierarchy, typographic semantics, saliency control, and intention communication; commercial models lead on high-level reasoning but act as insensitive automatic raters, while generators render text well yet struggle with composition-aware synthesis. Extensive analyses show PosterIQ is both a quantitative benchmark and a diagnostic tool for design reasoning, offering reproducible, task-specific metrics. We aim to catalyze models' creativity and integrate human-centred design principles into generative vision-language systems.

Whole Slide Images (WSIs) exhibit hierarchical structure, where diagnostic information emerges from cellular morphology, regional tissue organization, and global context. Existing Computational Pathology (CPath) Multimodal Large Language Models (MLLMs) typically compress an entire WSI into a single embedding, which hinders fine-grained grounding and ignores how pathologists synthesize evidence across different scales. We introduce MLLM-HWSI, a Hierarchical WSI-level MLLM that aligns visual features with pathology language at four distinct scales, cell as word, patch as phrase, region as sentence, and WSI as paragraph to support interpretable evidence-grounded reasoning. MLLM-HWSI decomposes each WSI into multi-scale embeddings with scale-specific projectors and jointly enforces (i) a hierarchical contrastive objective and (ii) a cross-scale consistency loss, preserving semantic coherence from cells to the WSI. We compute diagnostically relevant patches and aggregate segmented cell embeddings into a compact cellular token per-patch using a lightweight Cell-Cell Attention Fusion (CCAF) transformer. The projected multi-scale tokens are fused with text tokens and fed to an instruction-tuned LLM for open-ended reasoning, VQA, report, and caption generation tasks. Trained in three stages, MLLM-HWSI achieves new SOTA results on 13 WSI-level benchmarks across six CPath tasks. By aligning language with multi-scale visual evidence, MLLM-HWSI provides accurate, interpretable outputs that mirror diagnostic workflows and advance holistic WSI understanding. Code is available at: \href https://github.com/BasitAlawode/HWSI-MLLM Github .

Vision-Language Models (VLMs), such as CLIP, have achieved impressive zero-shot recognition performance but remain highly susceptible to adversarial perturbations, posing significant risks in safety-critical scenarios. Previous training-time defenses rely on adversarial fine-tuning, which requires labeled data and costly retraining, while existing test-time strategies fail to reliably distinguish between clean and adversarial inputs, thereby preventing both adversarial robustness and clean accuracy from reaching their optimum. To address these limitations, we propose Test-Time Padding (TTP), a lightweight defense framework that performs adversarial detection followed by targeted adaptation at inference. TTP identifies adversarial inputs via the cosine similarity shift between CLIP feature embeddings computed before and after spatial padding, yielding a universal threshold for reliable detection across architectures and datasets. For detected adversarial cases, TTP employs trainable padding to restore disrupted attention patterns, coupled with a similarity-aware ensemble strategy for a more robust final prediction. For clean inputs, TTP leaves them unchanged by default or optionally integrates existing test-time adaptation techniques for further accuracy gains. Comprehensive experiments on diverse CLIP backbones and fine-grained benchmarks show that TTP consistently surpasses state-of-the-art test-time defenses, delivering substantial improvements in adversarial robustness without compromising clean accuracy. The code for this paper will be released soon.

Despite the inherent advantages of thermal infrared(TIR) imaging, large-scale data collection and annotation remain a major bottleneck for TIR-based perception. A practical alternative is to synthesize pseudo TIR data via image translation; however, most RGB-to-TIR approaches heavily rely on RGB-centric priors that overlook thermal physics, yielding implausible heat distributions. In this paper, we introduce TherA, a controllable RGB-to-TIR translation framework that produces diverse and thermally plausible images at both scene and object level. TherA couples TherA-VLM with a latent-diffusion-based translator. Given a single RGB image and a user-prompted condition pair, TherA-VLM yields a thermal-aware embedding that encodes scene, object, material, and heat-emission context reflecting the input scene-condition pair. Conditioning the diffusion model on this embedding enables realistic TIR synthesis and fine-grained control across time of day, weather, and object state. Compared to other baselines, TherA achieves state-of-the-art translation performance, demonstrating improved zero-shot translation performance up to 33% increase averaged across all metrics.

Rare diseases often manifest with distinctive facial phenotypes in children, offering valuable diagnostic cues for clinicians and AI-assisted screening systems. However, progress in this field is severely limited by the scarcity of curated, ethically sourced facial data and the high similarity among phenotypes across different conditions. To address these challenges, we introduce RDFace, a curated benchmark dataset comprising 456 pediatric facial images spanning 103 rare genetic conditions (average 4.4 samples per condition). Each ethically verified image is paired with standardized metadata. RDFace enables the development and evaluation of data-efficient AI models for rare disease diagnosis under real-world low-data constraints. We benchmark multiple pretrained vision backbones using cross-validation and explore synthetic augmentation with DreamBooth and FastGAN. Generated images are filtered via facial landmark similarity to maintain phenotype fidelity and merged with real data, improving diagnostic accuracy by up to 13.7% in ultra-low-data regimes. To assess semantic validity, phenotype descriptions generated by a vision-language model from real and synthetic images achieve a report similarity score of 0.84. RDFace establishes a transparent, benchmark-ready dataset for equitable rare disease AI research and presents a scalable framework for evaluating both diagnostic performance and the integrity of synthetic medical imagery.

Multimodal Large Language Models (MLLMs) have shown remarkable progress in visual reasoning and understanding tasks but still struggle to capture the complexity and subjectivity of human emotions. Existing approaches based on supervised fine-tuning often suffer from limited generalization and poor interpretability, while reinforcement learning methods such as Group Relative Policy Optimization fail to align with the intrinsic characteristics of emotional cognition.To address these challenges, we propose Reflective Reinforcement Learning for Emotional Reasoning (EMO-R3), a framework designed to enhance the emotional reasoning ability of MLLMs. Specifically, we introduce Structured Emotional Thinking to guide the model to perform step-by-step emotional reasoning in a structured and interpretable manner, and design a Reflective Emotional Reward that enables the model to re-evaluate its reasoning based on visual-text consistency and emotional coherence. Extensive experiments demonstrate that EMO-R3 significantly improves both the interpretability and emotional intelligence of MLLMs, achieving superior performance across multiple visual emotional understanding benchmarks.

Real-world reasoning often requires combining information across modalities, connecting textual context with visual cues in a multi-hop process. Yet, most multimodal benchmarks fail to capture this ability: they typically rely on single images or set of images, where answers can be inferred from a single modality alone. This limitation is mirrored in the training data, where interleaved image-text content rarely enforces complementary, multi-hop reasoning. As a result, Vision-Language Models (VLMs) frequently hallucinate and produce reasoning traces poorly grounded in visual evidence. To address this gap, we introduce CRIT, a new dataset and benchmark built with a graph-based automatic pipeline for generating complex cross-modal reasoning tasks. CRIT consists of diverse domains ranging from natural images, videos, and text-rich sources, and includes a manually verified test set for reliable evaluation. Experiments on this benchmark reveal that even state-of-the-art models struggle on such reasoning tasks. Models trained on CRIT show significant gains in cross-modal multi-hop reasoning, including strong improvements on SPIQA and other standard multimodal benchmarks.

Multimodal Large Language Models (MLLMs) have achieved remarkable progress in vision-language understanding, yet processing long visual token sequences remains computationally expensive. Existing approaches mitigate this cost by reducing image tokens, either by discarding them after the visual encoder or by pruning them in the early Transformer layers of the LLM. While these strategies improve efficiency, they inevitably discard informative visual content and risk degrading downstream reasoning performance. To address this challenge, we introduce Hi-Lo Prune, a training-free pruning strategy that is built on a simple principle: look at what you will lose before you remove it. Instead of directly dropping tokens, Hi-Lo Prune first identifies which tokens are safe to prune through a coarse-to-fine selection process, and then encourages the model to absorb their information before pruning occurs. Specifically, the framework consists of three stages: (1) Hierarchical Pruning Token Selection. After visual encoding, we apply a coarse-to-fine process that identifies tokens to retain and selects a critical set of pruning candidates from redundant ones. (2) Attention-Guided Candidate Token Merge. Before removing selected tokens, an attention mechanism is applied to the early LLM layers, which explicitly transfers information from these candidates to the retained tokens. (3) Low-informative Candidate Token Removal. At a designated Transformer layer, the pruned tokens are removed, reducing computation for all subsequent layers. This design enables aggressive early-layer pruning while preserving critical visual cues. Experiments on Qwen2-VL, Qwen2.5-VL, and Qwen3-VL demonstrate that Hi-Lo Prune consistently outperforms existing pruning methods across multiple benchmarks, achieving strong performance even under high pruning ratios without any fine-tuning.

Accurate trajectory prediction is vital for safe autonomous driving, yet existing approaches struggle to balance modeling power and computational efficiency. Attention-based architectures incur quadratic complexity with increasing agents, while recurrent models struggle to capture long-range dependencies and fine-grained local dynamics. Building upon this, we present FoSS, a dual-branch framework that unifies frequency-domain reasoning with linear-time sequence modeling. The frequency-domain branch performs a discrete Fourier transform to decompose trajectories into amplitude components encoding global intent and phase components capturing local variations, followed by a progressive helix reordering module that preserves spectral order; two selective state-space submodules, Coarse2Fine-SSM and SpecEvolve-SSM, refine spectral features with \mathcal O (N) complexity. In parallel, a time-domain dynamic selective SSM reconstructs self-attention behavior in linear time to retain long-range temporal context. A cross-attention layer fuses temporal and spectral representations, while learnable queries generate multiple candidate trajectories, and a weighted fusion head expresses motion uncertainty. Experiments on Argoverse 1 and Argoverse 2 benchmarks demonstrate that FoSS achieves state-of-the-art accuracy while reducing computation by 22.5% and parameters by over 40%. Comprehensive ablations confirm the necessity of each component.

One of the most exciting applications of vision models involve pixel-level reasoning. Despite the abundance of vision foundation models, we still lack representations that effectively embed spatio-temporal properties of visual scenes at the pixel level. Existing frameworks either train on image-based pretext tasks, which do not account for dynamic elements, or on video sequences for action-level reasoning, which does not scale to dense pixel-level prediction. We present a framework that learns pixel-accurate feature descriptors from videos, LILA. The core element of our training framework is linear in-context learning. LILA leverages spatio-temporal cue maps--depth and motion--estimated with off-the-shelf networks. Despite the noisy nature of those cues, LILA trains effectively on uncurated video datasets, embedding semantic and geometric properties in a temporally consistent manner. We demonstrate compelling empirical benefits of the learned representation across a diverse suite of vision tasks: video object segmentation, surface normal estimation and semantic segmentation.