CVPR 2026 论文列表

This paper studies the problem of universal test-time prompt learning for vision-language models (VLMs) which aims to enhance prompt learning for a pre-trained VLM via unlabeled target data containing out-of-distribution (OOD) samples. However, existing test-time adaptation approaches often overlook class-specific diversity in the target domain and rely on unreliable pseudo-labels due to inadequate uncertainty estimation, which may result in additional adaptation bias during test time. Towards this end, we propose a novel framework named Separability-aware Conjugate Optimization with Prototypical Retrieval (STAR) for universal test-time prompt learning of VLMs. The core of our STAR is to incorporate a separability-aware gating mechanism into conjugate optimization for reliable pseudo-learning with OOD samples. In particular, we first compute the Fisher score to quantify the separability between in-distribution (ID) and OOD samples, which guides our soft gating mechanism for divided training. Then, we employ conjugate optimization to derive reliable pseudo-labels of unlabeled data for test-time adaptation. To further mitigate biases in OOD detection, we maintain a dynamic memory bank which stores high-confidence samples to build class-wise prototypes, which would serve as queries for prototypical retrieval to calibrate OOD detection. Extensive experiments on multiple benchmarks demonstrate that STAR consistently outperforms competing baseline methods.

Exemplar replay has become an effective strategy for mitigating catastrophic forgetting in federated continual learning (FCL) by retaining representative samples from past tasks. Existing studies focus on designing sample-importance estimation mechanisms to identify information-rich samples. However, they typically overlook strategies for effectively utilizing the selected exemplars, which limits their performance under continual dynamic heterogeneity across clients and tasks. To address this issue, this paper proposes a federated geometry-aware correction method, termed FEAT, which alleviates imbalance-induced representation collapse that drags rare-class features toward frequent classes across clients. Specifically, it consists of two key modules: 1) the Geometric Structure Alignment module performs structural knowledge distillation by aligning the pairwise angular similarities between feature representations and their corresponding Equiangular Tight Frame prototypes, which are fixed and shared across clients to serve as a class-discriminative reference structure. This encourages geometric consistency across tasks and helps mitigate representation drift; 2) the Energy-based Geometric Correction module removes task-irrelevant directional components from feature embeddings, which reduces prediction bias toward majority classes. This improves sensitivity to minority classes and enhances the model's robustness under class-imbalanced data distributions. Extensive experiments on three benchmark datasets demonstrate that FEAT substantially achieves a 4%-8% improvement in Top-1 accuracy compared to nine state-of-the-art methods.

Diffusion models have achieved remarkable success in generating high-fidelity content but suffer from slow, iterative sampling, resulting in high latency that limits their use in interactive applications. We introduce DRiffusion, a parallel sampling framework that parallelizes diffusion inference through a draft-and-refine process. DRiffusion employs skip transitions to generate multiple draft states for future timesteps and computes their corresponding noises in parallel, which are then used in the standard denoising process to produce refined results. Theoretically, our method achieves an acceleration rate of \tfrac 1 n or \tfrac 2 n+1 , depending on whether the conservative or aggressive mode is used, where n denotes the number of devices. Empirically, DRiffusion attains 1.4x-3.7x speedup across multiple diffusion models while incur minimal degradation in generation quality: on MS-COCO dataset, both FID and CLIP remain largely on par with those of the original model, while PickScore and HPSv2.1 show only minor average drops of 0.17 and 0.43, respectively. These results verify that DRiffusion delivers substantial acceleration and preserves perceptual quality.

Document image dewarping remains a challenging task in the deep learning era. While existing methods have improved by leveraging text line awareness, they typically focus only on a single horizontal dimension. In this paper, we propose a fine-grained deformation perception model that focuses on Dual Dimensions of document horizontal-vertical-lines to improve document Dewarping called D2Dewarp. It can perceive distortion trends in different directions across document details. To combine the horizontal and vertical granularity features, an effective fusion module based on X and Y coordinate is designed to facilitate interaction and constraint between the two dimensions for feature complementarity. Due to the lack of annotated line features in current public dewarping datasets, we also propose an automatic fine-grained annotation method using public document texture images and automatic rendering engine to build a new large-scale distortion training dataset named DocDewarpHV. The code and dataset will be publicly released. On three public Chinese and English benchmarks, both quantitative and qualitative results show that our method achieves better rectification results compared with the state-of-the-art methods. The code and dataset are available at https://github.com/xiaomore/D2Dewarp.

Radiology report generation (RRG) aims to automatically describe medical images via free-text reports. In clinical practice, comparing current and prior chest X-rays is essential for assessing disease progression, motivating the development of longitudinal RRG methods. However, most existing approaches often struggle to capture fine-grained temporal changes, as they often rely on unidirectional alignments or static reasoning pipelines, overlooking the bidirectional and asymmetric nature of disease evolution. To tackle these challenges, we propose BiOTPrompt, a novel framework for disease evolution-aware radiology report generation, which introduces a Bidirectional Optimal Transport (BiOT) mechanism to explicitly model progression dynamics between historical and current chest X-rays. By analyzing the asymmetry between bidirectional transport plans, BiOTPrompt can identify newly emerged and resolved regions, which are then used to construct dynamic prompts that guide large language models (LLMs) in generating clinically relevant diagnostic reports. Furthermore, we incorporate a vision-language consistency constraint to ensure alignment between visual evidence and textual descriptions, mitigating hallucinations and enhancing factual correctness. Extensive experiments on the Longitudinal-MIMIC dataset demonstrate that BiOTPrompt achieves state-of-the-art performance in both language metrics and clinical relevance, setting a new standard for longitudinal radiology report generation.

Instruction-based image editing with diffusion models has achieved impressive results, yet existing methods struggle with fine-grained instructions specifying precise attributes such as colors, positions, and quantities. While recent approaches employ Group Relative Policy Optimization (GRPO) for alignment, they optimize only at individual sampling steps, providing sparse feedback that limits trajectory-level control. We propose a unified framework **CogniEdit**, combining multi-modal reasoning with dense reward optimization that propagates gradients across consecutive denoising steps, enabling trajectory-level gradient flow through the sampling process. Our method comprises three components: (1) Multi-modal Large Language Models for decomposing complex instructions into actionable directives, (2) Dynamic Token Focus Relocation that adaptively emphasizes fine-grained attributes, and (3) Dense GRPO-based Optimization that propagates gradients across consecutive steps for trajectory-level supervision. Extensive experiments on benchmark datasets demonstrate that our CogniEdit achieves state-of-the-art performance in balancing fine-grained instruction following with visual quality and editability preservation. Our code is available at https://github.com/yl4467/CogniEdit.

Cross-degradation generalization remains a critical challenge for RGB-infrared multimodal object detection, especially when training data covers limited degradation types. This paper presents a distribution alignment framework with a key insight: aligning fused features to the pretrained distribution where the frozen detector performs optimally, rather than adapting to training-specific degradations. By freezing the pretrained detector and training only a lightweight fusion module, our approach leverages complementary infrared information to reduce distribution shift while maintaining computational efficiency. The method achieves state-of-the-art results on three benchmarks with 4x faster training. Critically, we demonstrate that aligning to the pretrained distribution substantially outperforms aligning to training degradations when generalizing to unseen scenarios.

In 3D reconstruction, the problem of inverse rendering, namely recovering the illumination of the scene and the material properties, is fundamental. Existing Gaussian Splatting-based methods primarily target static scenes and often assume simplified or moderate lighting to avoid entan- gling shadows with surface appearance. This limits their ability to accurately separate lighting effects from mate- rial properties, particularly in real-world conditions. We address this limitation by leveraging dynamic elements-- regions of the scene that undergo motion--as a supervisory signal for inverse rendering. Motion reveals the same sur- faces under varying lighting conditions, providing stronger cues for disentangling material and illumination. This the- sis is supported by our experimental results which show we improve LPIPS by 23% for albedo estimation and by 15% for scene relighting relative to next-best baseline. To this end, we introduce LumiMotion, the first Gaussian-based approach that leverages dynamics for inverse rendering and operates in arbitrary dynamic scenes. Our method learns a dynamic 2D Gaussian Splatting representation that em- ploys a set of novel constraints which encourage the dy- namic regions of the scene to deform, while keeping static regions stable. As we demonstrate, this separation is crucial for correct optimization of the albedo. Finally, we release a new synthetic benchmark comprising five scenes under four lighting conditions, each in both static and dynamic variants, for the first time enabling systematic evaluation of inverse rendering methods in dynamic environments and challenging lighting.

Vision-Language Models (VLMs) are increasingly adopted for medical applications, but their clinical utility is limited by a core weakness in quantitative reasoning. This limitation affects tasks ranging from regression of lesion sizes to prediction of bounding-box coordinates and stems from the discrete tokenization schemes underlying Large Language Models (LLMs). To address this, we propose Rotary Number Encoding and Decoding (RNED), a principled method for embedding continuous numerical values directly in the representation space of a VLM. Analogous to rotary position encoding, RNED represents a scalar by applying a number-specific rotation matrix to a dedicated numeric token embedding. This norm-preserving transformation maintains ordinal structure over a wide numerical range and integrates seamlessly with pretrained model weights. For decoding, we introduce a robust score-matching-based scheme to recover continuous values from hidden states in the presence of stochastic noise. We evaluate RNED on two quantitative tasks: radiological measurement estimation and medical visual grounding. On both internal and public benchmarks, RNED consistently outperforms existing VLM baselines. Together, these results show that RNED offers a robust, generalizable solution for numerical reasoning in medical VLMs, enabling models that are both quantitatively reliable and clinically applicable.

Due to the prohibitive cost of data annotation and the inability to obtain sufficient sample data for all defect categories, municipal sewer pipe defect detection poses significant generalization challenges for traditional models. Multi-Label Zero-Shot Learning (ML-ZSL) offers a viable solution to address this challenge. However, existing methods struggle to establish robust and fine-grained visual-semantic alignment between the complex visual environment inside the pipes and the often sparse semantic descriptions, leading to a critical issue: Alignment Ambiguity. To mitigate this, we propose a novel Steering-Fusion-Refining Network (SFR-Net) that follows a three-stage paradigm to progressively dissolve this ambiguity. This is achieved as the Representation Steering (RS) module first integrates a parameter-efficient feature steering mechanism to continuously adapt the representation to the pipe scene; the Multi-Granularity Evidence Fusion (MEF) module subsequently aggregates unambiguous multi-granularity visual evidence through decoupled global and local paths; and the Generalized Relational Score Refining (GR) module ultimately learns and transfers relational logic from seen defects to gain a universal score correction ability, directly refining preliminary prediction scores and significantly boosting the model's zero-shot generalization and prediction consistency. Extensive experiments on the public Sewer-ML dataset and our private WZ-Pipe dataset demonstrate that the proposed SFR-Net achieves state-of-the-art (SOTA) performance in multi-label zero-shot learning task.

Can a diffusion model produce its own "mental average" of a concept--one that is as sharp and realistic as a typical sample? We introduce Diffusion Mental Averages (DMA), a model-centric answer to this question. While prior methods aim to average image collections, they produce blurry results when applied to diffusion samples from the same prompt. These data-centric techniques operate outside the model, ignoring the generative process. In contrast, DMA averages within the diffusion model's semantic space, as discovered by recent studies. Since this space evolves across timesteps and lacks a direct decoder, we cast averaging as trajectory alignment: optimize multiple noise latents so their denoising trajectories progressively converge toward shared coarse-to-fine semantics, yielding a single sharp prototype. We extend our approach to multimodal concepts (e.g., dogs with many breeds) by clustering samples in semantically-rich spaces such as CLIP and applying Textual Inversion or LoRA to bridge CLIP clusters into diffusion space. This is, to our knowledge, the first approach that delivers consistent, realistic averages, even for abstract concepts, serving as a concrete visual summary and a lens into model biases and concept representation.

We train a language model to generate LEGO(r)-brick build sequences. While prior work has been restricted to discrete, voxel-like towers, we consider a much broader set of pieces, encompassing thousands of part types with diverse connection semantics. To enable this, we first collect a large-scale dataset of over 100,000 human-designed LDraw brick objects and scenes. The complexity of our setting makes it challenging to autoregressively assemble structures that satisfy physical constraints. When predicting block pose directly, build sequences quickly become invalid after a small number of steps. Although pieces are placed in 3D space, it is the spatial relationships of the parts which define the whole. With this in mind, we design a graph-based program representation that parametrizes structure through connectivity, improving the physical grounding of generated sequences. To enable future applications, we make our dataset and models available for research purposes. https://kulits.github.io/BrickNet

We introduce **DiffBMP**, a scalable and efficient differentiable rendering engine for a collection of bitmap images. Our work addresses a limitation that traditional differentiable renderers are constrained to vector graphics, given that most images in the world are bitmaps. Our core contribution is a highly parallelized rendering pipeline, featuring a custom CUDA implementation for calculating gradients. This system can, for example, optimize the position, rotation, scale, color, and opacity of thousands of bitmap primitives all in under 1 min using a consumer GPU. We employ and validate several techniques to facilitate the optimization: soft rasterization via Gaussian blur, structure-aware initialization, noisy canvas, and specialized losses/heuristics for videos or spatially constrained images. We demonstrate DiffBMP is not just an isolated tool, but a practical one designed to integrate into creative workflows. It supports exporting compositions to a native, layered file format, and the entire framework is publicly accessible via an easy-to-hack Python package.

Atmospheric turbulence significantly degrades long-range imaging by introducing geometric warping and exposure-time-dependent blur, which adversely affects both visual quality and the performance of high-level vision tasks. Existing methods for synthesizing turbulence effects often oversimplify the relationship between blur and exposure-time, typically assuming fixed or binary exposure settings. This leads to unrealistic synthetic data and limited generalization capability of trained models. To address this gap, we revisit the modulation transfer function (MTF) formulation and propose a novel Exposure-Time-dependent MTF (ET-MTF) that models blur as a continuous function of exposure-time. For blur synthesis, we derive a tilt-invariant point spread function (PSF) from the ET-MTF, which, when integrated with a spatially varying blur-width field, provides a comprehensive and physically accurate characterization of turbulence-induced blur. Building on this synthesis pipeline, we construct ET-Turb, a large-scale synthetic turbulence dataset that explicitly incorporates continuous exposure-time modeling across diverse optical and atmospheric conditions. The dataset comprises 5,083 videos (2,005,835 frames), partitioned into 3,988 training and 1,095 test videos. Extensive experiments demonstrate that models trained on ET-Turb produce more realistic restorations and achieve superior generalization on real-world turbulence data compared to those trained on other datasets. The dataset is publicly available at: github.com/Jun-Wei-Zeng/ET-Turb.

Sketch editing requires jointly handling high-level semantic changes and precise local redrawing, a combination that is particularly challenging for sparse, style-sensitive line art. Unlike natural images, sketches rely on minimal visual cues, making it difficult for existing methods to reconcile global semantic modifications with fine-grained structural control while preserving overall coherence. We present SketchAssist, an interactive sketch assistant that unifies instruction-guided editing with line-guided region redrawing, enabling efficient and controllable sketch manipulation while preserving overall composition. To support this task, we introduce a controllable data generation pipeline that constructs structured edit sequences with precise attribute variations and maintains structural alignment across multi-step modifications, while expanding stylistic diversity via style-preserving transformations. Building on this data, SketchAssist adopts a unified framework based on DiT, using a multi-channel input representation to encode sketches, masks, and guidance signals within a single interface. To further handle different editing modes, we integrate a Task-guided Mixture-of-Experts (T-MoE) into LoRA layers, enabling adaptive control over semantic and structural guidance. Extensive experiments demonstrate state-of-the-art performance on both tasks, achieving strong instruction adherence and improved structural and style consistency compared to recent methods. Together, our method provide a practical and controllable solution for sketch editing.

Reinforcement Learning with Verifiable Rewards (RLVR) has recently demonstrated remarkable potential in enhancing the reasoning capability of Large Reasoning Models (LRMs). However, RLVR often drives the policy toward over-determinism, resulting in ineffective exploration and premature policy convergence. While promoting token-level diversity has shown promise in mitigating entropy collapse, we argue that the latent dynamics underlying token generation encode a far richer computational structure for steering policy optimization toward a more effective exploration-exploitation tradeoff. To enable tractable analysis and intervention of the latent dynamics of LRMs, we leverage Koopman operator theory to obtain a linearized representation of their hidden state dynamics. This enables us to introduce Dynamic Spectral Dispersion (DSD), a new metric to quantify the heterogeneity of the model's latent dynamics, serving as a direct indicator of policy exploration. Building upon these foundations, we propose Reasoning with Latent eXploration (ReLaX), a framework that explicitly incorporates latent dynamics to regulate exploration and exploitation during policy optimization. Comprehensive experiments across a wide range of multimodal and text-only reasoning benchmarks show that ReLaX consistently incentivizes reasoning capability and outperforms existing token-level methods. Our project is available at https://github.com/ZhangShimin1/ReLaX.

Multimodal tiny object detection plays a critical role in real-world applications, yet remains highly challenging due to weak target representations and complex cross-modal interference. Existing frequency-domain methods for tiny object detection are still largely limited to the visible modality and overlook complementary cross-modal frequency cues in multimodal scenes. In this paper, we investigate cross-modal frequency learning for RGBT tiny object detection. Through frequency characteristic analysis, we find that tiny objects in both RGB and infrared modalities contain richer mid- and high-frequency components as object size decreases. Motivated by this observation, we propose a Dynamic Frequency-Decoupled Cross-Modal Learning Transformer (DyFCLT). Specifically, DyFCLT introduces a Dynamic Frequency-Band Decoupled Cross-Modal Attention (DFCA) mechanism to perform fine-grained cross-modal interaction across dynamic frequency sub-bands, and a Selective Smoothing Enhancement (SSE) module to suppress background noise and enhance foreground responses during multi-scale fusion. Extensive experiments on two RGBT tiny object detection benchmarks and one general-scale benchmark demonstrate that DyFCLT achieves state-of-the-art performance with strong generalization across different scales and scenes.

Intrinsic image decomposition aims to separate images into physical components such as albedo, depth, normals, and illumination. While recent diffusion- and transformer-based models benefit from paired supervision from synthetic datasets, their generalization to diverse, real-world scenarios remains challenging. We propose ReasonX, a novel framework that leverages a multimodal large language model (MLLM) as a perceptual judge providing relative intrinsic comparisons, and uses these comparisons as GRPO rewards for fine-tuning intrinsic decomposition models on unlabeled, in-the-wild images. Unlike RL methods for generative models, our framework aligns conditional intrinsic predictors by rewarding agreement between the judge's relational assessments and analytically derived relations from the model's outputs. ReasonX is model-agnostic and can be applied to different intrinsic predictors. Across multiple base architectures and modalities, ReasonX yields significant improvements, including 9-25% WHDR reduction on IIW albedo and up to 46% depth accuracy gains on ETH3D, highlighting the promise of MLLM-guided comparative supervision to bridge low- and high-level vision reasoning.

Diffusion models face a fundamental trade-off between generation quality and computational efficiency. Latent Diffusion Models (LDMs) offer an efficient solution but suffer from potential information loss and non-end-to-end training. In contrast, existing pixel space models bypass VAEs but are computationally prohibitive for high-resolution synthesis. To resolve this dilemma, we propose DiP, an efficient pixel space diffusion framework. DiP decouples generation into a global and a local stage: a Diffusion Transformer (DiT) backbone operates on large patches for efficient global structure construction, while a co-trained lightweight Patch Detailer Head leverages contextual features to restore fine-grained local details. This synergistic design achieves computational efficiency comparable to LDMs without relying on a VAE. DiP is accomplished with up to 10xfaster inference speeds than previous method while increasing the total number of parameters by only 0.3%, and achieves an 1.79 FID score on ImageNet 256x256.