CVPR 2026 论文列表

Large-scale video-text pretraining achieves strong performance but depends on noisy, synthetic captions with limited semantic coverage, often overlooking implicit world knowledge such as object motion, 3D geometry, and physical cues. In contrast, masked video modeling (MVM) directly exploits spatiotemporal structures but trails text-supervised methods on general tasks.We find this gap arises from overlooked architectural issues: pixel-level reconstruction struggles with convergence and its low-level requirement often conflicts with semantics, while latent prediction often encourages shortcut learning.To address these, we disentangle the traditional encoder-decoder design into an Encoder-Predictor-Decoder (EPD) framework, where the predictor acts as a latent world model, and propose InternVideo-Next, a two-stage pretraining scheme that builds a semantically consistent yet detail-preserving latent space for this world model.First, conventional linear decoder in pixel MVM enforces the predictor's output latent to be linearly projected to, thus separable in pixel space, causing the conflict with semantic abstraction.Our Stage 1 proposes a conditional diffusion decoder and injects clean image-level semantic priors to enhance semantics and convergence, thus bridging pixel-level fidelity with high-level semantic abstraction.Stage 2 further learns world knowledge by predicting frozen Stage 1 targets within this space, mitigating shortcut learning.Trained on public, unlabeled videos, InternVideo-Next achieves state-of-the-art results across benchmarks and provides a scalable path toward general video representation learning.

As powerful generative models, text-to-image diffusion models have recently been explored for discriminative tasks. A line of research focuses on adapting a pre-trained diffusion model to semantic segmentation without any further training, leading to training-free diffusion segmentors. These methods typically rely on cross-attention maps from the model's attention layers, which are assumed to capture semantic relationships between image pixels and text tokens. Ideally, such approaches should benefit from more powerful diffusion models, i.e., stronger generative capability should lead to better segmentation. However, we observe that existing methods often fail to scale accordingly. To understand this issue, we identify two underlying gaps: (i) Cross-attention is computed across multiple heads and layers, but there exists a discrepancy between these individual attention maps and a unified global representation. (ii) Even when a global map is available, it does not directly translate to accurate semantic correlation for segmentation, due to score imbalances among different text tokens. To bridge these gaps, we propose two techniques: auto aggregation and per-pixel rescaling, which together enable training-free segmentation to better leverage generative capability. We evaluate our approach on standard semantic segmentation benchmarks and further integrate it into a generative technique, demonstrating both improved performance and broad applicability. Codes are at https://github.com/Darkbblue/goca.

Recent Tracking-by-Query-Propagation (TBP) methods have advanced Multi-Object Tracking (MOT) by enabling end-to-end (E2E) pipelines with long-range temporal modeling. However, this reliance on query propagation introduces unexplored architectural vulnerabilities to adversarial attacks. We present FADE, a novel attack framework designed to exploit these specific vulnerabilities. FADE employs two attack strategies targeting core TBP mechanisms: (i) Temporal Query Flooding: Generates spurious temporally consistent track queries to exhaust the tracker's limited query budget, forcing it to terminate valid tracks. (ii) Temporal Memory Corruption: Directly attacks the query updater's memory by severing temporal links via state de-correlation and erasing the learned feature identity of matched tracks. Furthermore, we introduce a differentiable pipeline to optimize these attacks for physical-world realizability by leveraging simulations of advanced perception sensor spoofing. Experiments on MOT17 and MOT20 benchmarks demonstrate that FADE is highly effective against state-of-the-art TBP trackers, causing significant identity switches and track terminations.

Humans learn by observing, interacting with environments, and internalizing physics and causality. We explore whether agents can similarly acquire human-like reasoning through interaction and experience. To study this, we introduce a Game-to-Unseen (G2U) benchmark of 1,000+ heterogeneous games that exhibit significant visual domain gaps. Existing approaches, including VLMs and world models, struggle with underlying physics and causality due to overfitting on visual details. VLM/VLA agents reason but lack look-ahead in interactive settings, while world models imagine but imitate visual patterns rather than analyze physics and causality. We therefore propose IPR (Interactive Physical Reasoner), using world-model rollouts to score and reinforce a VLM's policy, and introduce PhysCode, a physics-centric action code aligning semantic intent with dynamics to provide a shared action space for prediction and reasoning. Pretrained on 1,000+ games, IPR performs robustly on levels from primitive intuition to goal-driven reasoning, and even surpasses GPT-5 overall. We find that performance improves with more training games and interaction steps, and that the model also zero-shot transfers to unseen games. These results support physics-centric interaction as a path to steadily improving physical reasoning. Further demos and details can be found at https://mybearyzhang.github.io/ipr-1.

We introduce AvatarPointillist, a novel framework for generating dynamic 4D Gaussian avatars from a single portrait image. At the core of our method is a decoder-only Transformer that autoregressively generates a point cloud for 3D Gaussian Splatting. This sequential approach allows for precise, adaptive construction, dynamically adjusting point density and the total number of points based on the subject's complexity. During point generation, the AR model also jointly predicts per-point binding information, enabling realistic animation. After generation, a dedicated Gaussian decoder converts the points into complete, renderable Gaussian attributes. We demonstrate that conditioning the decoder on the latent features from the AR generator enables effective interaction between stages and markedly improves fidelity. Extensive experiments validate that AvatarPointillist produces high-quality, photorealistic, and controllable avatars. We believe this autoregressive formulation represents a new paradigm for avatar generation, and we will release our code to inspire future research.

Reducing the token count is crucial for efficient training and inference of latent diffusion models, especially at high resolution. A common approach is to build high-compression image tokenizers that store more information by allocating more channels per token. However, when trained solely with reconstruction objectives, high-dimensional latent spaces often fail to maintain meaningful structure, which complicates diffusion training. Existing methods introduce additional objectives, such as semantic alignment or selective dropout, to enforce structure, but typically require costly retraining of the diffusion model. In contrast, pretrained diffusion models already exhibit a structured, lower-dimensional latent space, suggesting a simpler strategy: expand the latent dimensionality while preserving this structure. To this end, we propose Detail-Aligned VAE (DA-VAE), which increases the compression ratio of a pretrained VAE with only lightweight adaptation of the diffusion backbone. DA-VAE imposes an explicit latent layout: the first C channels are inherited from the pretrained VAE at a base resolution, while additional D channels encode high-resolution details. We introduce a simple yet effective detail-alignment mechanism that explicitly regularizes the expanded latent space to preserve the structural properties of the original space defined by the first C channels. Finally, we present a warm-start fine-tuning strategy that enables 1024 x1024 image generation with Stable Diffusion 3.5 using only 32 x32 tokens, 4xfewer than the original model, within a compute budget of 5 H100-days. It further enables 2048 x2048 generation with SD3.5, achieving a 6xspeedup while preserving image quality. We also validate the method and its design choices quantitatively on ImageNet.

Scaling vision foundation models is constrained by the quadratic complexity of self-attention. Although subquadratic attention alternatives like linear attention variants and state-space models successfully reduce the model complexity, they typically serialize images into 1D token sequences, compromising spatial coherence and efficiency. Generalized Spatial Propagation Networks (GSPN) offer a linear-time alternative that propagates context directly on the 2D grid via line-scan propagation and removes positional embeddings, yet the original design hits GPU-scaling limits: growing batch/channels saturate SM concurrency, serializing scans, and spiking latency. We introduce Compact GSPN (C-GSPN), a ViT block that compresses the propagation space to preserve accuracy while cutting propagation latency by nearly 10x. We further improve efficiency with lightweight projections and fused CUDA kernels. To enable large-scale pretraining, we adopta two-stage cross-operator distillation strategy that combines layer-wise supervision with end-to-end alignment. In a representative 1K configuration (batch 32, C=1152), C-GSPN achieves up to 2x speedup, maintains competitive zero-shot accuracy, and improves segmentation by +2.1%. Extensive experiments and ablations show that the proposed compression and two-stage distillation are criticalfor strong transfer while substantially reducing compute, enabling the first extension of a subquadratic operator to foundation-scale (CLIP-style) vision pretraining.

Effectively grounding complex language to pixels in remote sensing (RS) images is a critical challenge for applications like disaster response and environmental monitoring. Current models can parse simple, single-target commands but fail when presented with complex geospatial scenarios, e.g., segmenting objects at various granularities, executing multi-target instructions, and interpreting implicit user intent. To drive progress against these failures, we present LaSeRS, the first large-scale dataset built for comprehensive training and evaluation across four critical dimensions of language-guided segmentation: hierarchical granularity, target multiplicity, reasoning requirements, and linguistic variability. By capturing these dimensions, LaSeRS moves beyond simple commands, providing a benchmark for complex geospatial reasoning. This addresses a critical gap: existing datasets oversimplify, leading to sensitivity-prone real-world models. We also propose SegEarth-R2, an MLLM architecture designed for comprehensive language-guided segmentation in RS, which directly confronts these challenges. The model's effectiveness stems from two key improvements: (1) a spatial attention supervision mechanism specifically handles the localization of small objects and their components, and (2) a flexible and efficient segmentation query mechanism that handles both single-target and multi-target scenarios. Experimental results demonstrate that our SegEarth-R2 achieves outstanding performance on LaSeRS and other benchmarks, establishing a powerful baseline for the next generation of geospatial segmentation. All data and code will be released.

User Interface (UI) display defect detection poses challenges far beyond UI understanding, requiring fine-grained element boundary understanding, missing-content detection, and reasoning about sequential interface semantic consistency. However, the capabilities of multimodal large language models (MLLMs) and vision-language models (VLMs) for detecting UI defects in realistic, complex interfaces have not been systematically validated. To fill this gap, we present UI-Lens, the first multi-dimensional UI display detection benchmark for Chinese-language UI scenarios. The dataset comprises 4,759 pages meticulously annotated by design experts, covering six core display defect categories. We conduct a systematic evaluation of 10 mainstream models (8 closed-source, 2 open-source). Results show clear shortcomings in current models: for tasks requiring fine-grained element boundary understanding, performance is near random, with task-average F1 scores of 20.36% and 31.21% on Text Overflow and Container Overlap, respectively; for sequential interface semantic consistency (e.g., Text Inconsistency), the task-average F1 score is only 10.61%, indicating severe underperformance. We release UI-Lens to catalyze research toward more robust UI display defect detection with fine-grained boundary awareness in realistic, complex interfaces.

With the rapid development of large multimodal models, reliable judge and critic models have become essential for open-ended evaluation and preference alignment, providing pairwise preferences, numerical scores, and explanatory justifications for assessing model-generated responses. However, existing critics are primarily trained in general visual domains such as captioning or image question answering, leaving physical AI tasks involving perception, causal reasoning, and planning largely underexplored. We introduce PhyCritic, a multimodal critic model optimized for physical AI through a two-stage RLVR pipeline: a physical skill warmup stage that enhances physically oriented perception and reasoning, followed by self-referential critic finetuning, where the critic generates its own prediction as an internal reference before judging candidate responses, improving judgment stability and physical correctness. Across both physical and general-purpose multimodal judge benchmarks, PhyCritic achieves strong performance gains over open-source baselines and, when applied as a policy model, further improves perception and reasoning in physically grounded tasks.

Sketch-based caricature synthesis suffers from a fundamental failure mode: when identity and shape conditions are combined in diffusion models, they create destructive interference that causes inevitable collapse toward either bland portraits or unrecognizable distortions. We identify the root cause as condition signal contamination -- competing probability distributions in the denoising trajectory that make balanced generation impossible. We present CaricHarmony, the first training-free method that explicitly resolves this contamination through parallel uncontaminated diffusion paths. During inference, we maintain three paths: \mathcal P ^ \mathrm i (pure identity), \mathcal P ^ \mathrm s (pure shape), and \mathcal P ^ \mathrm i+s (harmonized output). Novel energy functions operating on cross-attention features provide gradient guidance that steers \mathcal P ^ \mathrm i+s toward optimal balance: \mathcal E _ \mathrm shape ensures sketch fidelity through layout and semantic alignment, while \mathcal E _ \mathrm id employs token-level correspondence matching robust to extreme distortions. Unlike DemoCaricature requiring 70 seconds per-identity fine-tuning or CaricatureBooth constrained to Bezier curves, CaricHarmony accepts any sketch format and generates in under 16 seconds. Experiments demonstrate state-of-the-art performance: 0.8615 shape CLIP score (vs. 0.8450) under comparable identity consistency score, with 7.81 overall user preference score (vs. 6.06). Our method fundamentally reconceptualizes the ID-shape conflict as conditioning signal contamination for diffusion models, enabling unprecedented creative control while preserving recognition.

Frozen visual embeddings (e.g., CLIP, DINOv2/v3, SSCD) power retrieval and integrity systems, yet their use on face-containing data is constrained by unmeasured identity leakage and a lack of deployable mitigations. We take an attacker-aware view and contribute: (i) a benchmark of visual embeddings that reports open-set verification at low false-accept rates, a calibrated diffusion-based template inversion check, and face-context attribution with equal-area perturbations; and (ii) propose a one-shot linear projector that removes an estimated identity subspace while preserving the complementary space needed for utility, which for brevity we denote as the identity sanitization projection ISP. Across CelebA-20 and VGGFace2, we show that these encoders are robust under open-set linear probes, with CLIP exhibiting relatively higher leakage than DINOv2/v3 and SSCD, robust to template inversion, and are context-dominant. In addition, we show that ISP drives linear access to near-chance while retaining high non-biometric utility, and transfers across datasets with minor degradation. Our results establish the first attacker-calibrated facial privacy audit of non-FR encoders and demonstrate that linear subspace removal achieves strong privacy guarantees while preserving utility for visual search and retrieval.

Existing image fusion methods face difficulties in adapting to unseen fusion tasks and have limitations in balancing semantic information with pixel-level details. This limitation can be attributed to three key challenges: (1) the lack of a unified, task-agnostic optimization objective; (2) the inherent difficulty in balancing semantic fidelity and pixel-level richness; and (3) an over-reliance on supervised learning, which limits transferability across tasks. To overcome these issues, this work proposes a unified fusion framework that generalizes to diverse fusion tasks even when trained solely on infrared-visible image pairs. Specifically, inspired by the free-energy principle, we introduce a fusion paradigm that combines high pixel-entropy expectation with low semantic-entropy expectation, and we design a frequency-aware feature decoupling mechanism to balance semantic content and pixel detail. Furthermore, an unsupervised dual-path trade-off strategy provides collaborative constraints at both semantic and pixel levels. Experiments show that our method significantly outperforms existing state-of-the-art methods in visual quality and downstream-task performance. It not only handles trained tasks efficiently but also generalizes well to unseen fusion tasks, while featuring lightweight model parameters and strong practical applicability. The code is available at https://github.com/XiaoW-Liu/DECC.

Adapting image-pretrained backbones to video typically relies on time-domain adapters tuned to a single temporal scale. Our experiments show that these modules pick up static image cues and very fast flicker changes, while overlooking medium-speed motion. Capturing dynamics across multiple time-scales is, however, crucial for fine-grained temporal analysis (i.e., opening vs. closing bottle).To address this, we introduce Frame2Freq -- a family of frequency-aware adapters that perform spectral encoding during image-to-video adaptation of pretrained Vision Foundation Models (VFMs), improving fine-grained action recognition. Frame2Freq uses Fast Fourier Transform (FFT) along time and learns frequency-band specific embeddings that adaptively highlight the most discriminative frequency ranges. Across five fine-grained activity recognition datasets, Frame2Freq outperforms prior PEFT methods and even surpasses fully fine-tuned models on four of them. These results provide encouraging evidence, that frequency analysis methods are a powerful tool for modeling temporal dynamics in image-to-video transfer.

3D shape generation has become increasingly important for graphics and vision applications. Current part-aware 3D generation usually overlooks hierarchical part relations or inefficiently encodes multi-level semantics in Euclidean space. Thus we propose a novel framework for hierarchical and efficient part-aware 3D generation in hyperbolic space. Our contributions are three-fold: (1) Hierarchical Hyperbolic Mixture Model (H^2MM): We propose part-aware semantic representation of objects within a hyperbolic manifold, providing a high-fidelity hierarchical part-aware representation of object details and semantics. (2) Hyperbolic Semantically Consistent Diffusion Model: We design the geodesic diffusion process that preserves the hierarchical and semantic structure of H^ 2 MM, and progressively generates semantics from conditions and generates object under their joint guidance. We use an adaptive tree-structured neural network to loosen the constraint of jointly generating nodes and edges in previous hyperbolic diffusion. (3) Hyperbolic Diffusion Model Solver: We leverage higher-order Riemannian gradient on hyperbolic manifolds for designing a fast dedicated high-order solver for diffusion ODEs with the convergence order guarantee. Extensive experiments demonstrate that our method achieves superior quality and efficiency.

We propose a novel learning-based task named fractured object recovery. Unlike the previous fractured object reassembly task that only aligns existing parts with overlaps, our task aims to recover the complete shape by not only reassembling irrelevant parts but also predicting missing parts. Our task coincides with practical experiences, where the prior knowledge of similar shapes can be leveraged such that even non-overlapping parts can be reasoned into adequate locations. We also present the first learning model for the proposed task by correlating features of both existing and missing parts using a transformer, where the latter is naturally represented as missing tokens. To facilitate the task, we introduce a new dataset based on the existing fractured object benchmark by imposing different configurations of missing parts. We perform extensive evaluations to demonstrate the performance of the proposed model over baselines. The results show that joint part reassembly and prediction can be made possible and also have mutual benefits, which we believe can inspire future research and favor real applications.

Vision Language Models (VLMs) exhibit a fundamental semantic-to-geometric gap in spatial reasoning: they excel at qualitative semantic inference but their reasoning operates within a lossy semantic space, misaligned with high-fidelity geometry. Current paradigms fail to bridge this gap. Training-based methods suffer from an "oracle paradox," learning flawed spatial logic from imperfect oracles. Tool-integrated methods constrain the final computation but critically leave the VLM's planning process unconstrained, resulting in geometrically flawed plans. In this work, we propose Geometrically-Constrained Agent (GCA), a training-free agentic paradigm that resolves this gap by introducing a formal task constraint. Specifically, we strategically decouples the VLM's role into two stages. First, acting as a semantic analyst, the VLM translates the user's ambiguous query into the formal, verifiable task constraint, which defines the reference frame and objective. Second, acting as a task solver, the VLM generates and executes tool calls strictly within the deterministic bounds defined by the constraint. This geometrically-constrained strategy successfully resolve the semantic-to-geometric gap, yielding a robust and verifiable reasoning pathway for complex spatial tasks. Comprehensive experiments demonstrate that GCA achieves SOTA performance on multiple spatial reasoning benchmarks, surpassing existing training-based and tool-integrated methods by 27%.

Privacy-preserving Person Re-Identification (PP-ReID) addresses the core privacy-utility trade-off in Re-ID by retrieving a person across multiple non-overlapping cameras while applying anonymization techniques to protect sensitive information. However, prior PP-ReID studies are confined to single-modality visible scenarios, as 24-hour surveillance systems require robust cross-modal visible-infrared (VI) capabilities. Extending PP-ReID to the cross-modal VI setting is therefore crucial. Accordingly, we introduce a new task: Privacy-Preserving Visible-Infrared Person Re-Identification (PP-VI-ReID). This task presents two severe challenges: 1) Crude anonymization strategies destroy identity-critical information and disrupt cross-modal alignment 2) The anonymization process creates inconsistent distortions across modalities. It disrupts color-based textures in visible images while obscuring thermal contours in infrared images. This inconsistency with modality gap forms a Mixed Gap. To overcome these challenges, we propose a framework, the Precise Privacy-preserving and Alignment Network (PPA) with two components: 1) A Keypoint-Preserving Regularization (KPR) module leverages human pose as a prior to guide structure-aware anonymization, preserving essential body features. 2) A Differential Consistency-guided Modality Alignment (DCMA) treats anonymization perturbations not as varying noise, but as a stable, learnable offset, facilitating robust alignment between raw and anonymized features across modalities. Experiments on SYSU-MM01 and RegDB validate our framework, establishing a strong baseline.

Transformer architecture search (TAS) discovers optimal vision transformer (ViT) architectures automatically, reducing human effort to manually design ViTs. However, existing TAS methods suffer from the feature collapse problem, where subnets within a supernet fail to learn subnet-specific features, mainly due to the shared weights in a supernet, limiting the performance of individual subnets. To address this, we propose TAS-LoRA, a novel method that introduces parameter-efficient low-rank adaptation (LoRA) to enable subnet-specific feature learning, while maintaining computational efficiency. TAS-LoRA incorporates a Mixture-of-LoRA-Experts (MoLE) strategy, where a lightweight router dynamically assigns LoRA experts based on subnet architectures, and introduces a group-wise router initialization technique to encourage diverse feature learning across experts early in training. Extensive experiments on ImageNet and several transfer learning benchmarks, including CIFAR-10/100, Flowers, CARS, and INAT-19, demonstrate that TAS-LoRA mitigates feature collapse effectively, improving performance over state-of-the-art TAS methods significantly.