Text-Video Retrieval (TVR) retrieves videos that match a natural-language query, but extending image-text models such as CLIP to videos is fundamentally limited by the lack of temporal modeling. Videos exhibit frame-wise heterogeneity in appearance and motion, and compressing all frames into a single representation often obscures temporal structure and semantic transitions. To address this, we propose Temporal-Aware Mixture-of-Experts for Text-Video Retrieval (TAME), a CLIP-based framework that jointly models frame-level structure and temporal relations. First, we integrate sparse Mixture-of-Experts (MoE) layers into both CLIP encoders and apply frame-consistent routing on the vision branch so that experts specialize according to frame-level visual patterns while preserving the original vision-language alignment. Second, we introduce Frame-Temporal (FT) tokens that aggregate global cross-frame information and feed it back to each frame, enabling the visual encoder to capture long-range temporal dependencies without harming local details. Third, we design a Cross-Temporal Interaction and Aggregation (CTIA) module that refines frame-wise sentence-video similarities through staged temporal filtering and fusion. Experiments on standard TVR benchmarks show that TAME consistently improves over CLIP-based baselines. On MSR-VTT, it improves R@1 by 4.0 over CLIP4Clip, and also achieves consistent gains on DiDeMo, MSVD, LSMDC, and ActivityNet. The code is available at https://github.com/sejong-rcv/TAME.
Pedram MohajerAnsari, Amir Salarpour, Run Wang +1cs.CV
Vision-language models such as CLIP and SigLIP provide strong zero-shot recognition, but their predictions can degrade when deployed on target data that differ from the pretraining distribution. Test-time adaptation offers a practical way to improve robustness without source data or target labels, yet existing methods often rely on either prompt-side adaptation or image-side target evidence alone. In this work, we introduce GATE, a training-free two-pass transductive test-time adaptation framework that uses the unlabeled target set while keeping the image encoder, text encoder, and prompt parameters fully frozen. Instead of representing each class with a single prototype, GATE builds two complementary Gaussian sources of evidence in the shared vision-language feature space: a text Gaussian estimated from multiple language descriptions and an image Gaussian estimated from reliable unlabeled target samples. A class-wise reliability gate controls the influence of image-derived pseudo-evidence, and a score-level generalized Product-of-Experts fusion produces a normalized residual correction to the original zero-shot logits. Across fine-grained recognition datasets, ImageNet-family distribution shifts, multiple CLIP backbones, and SigLIP-B/16, GATE achieves the best average accuracy in every benchmark/backbone group. It improves zero-shot performance by an average of 5.41 points and outperforms the strongest non-GATE baseline by 1.94 points, demonstrating the benefit of reliability-gated distributional evidence for frozen VLM adaptation.
Zero-shot classification needs efficient label retrieval and fine-grained visual reasoning, yet discriminative and generative vision-language models fail in complementary ways.When CLIP's top-1 prediction is wrong, the correct label often remains in its top-$K$ shortlist, making disambiguation rather than recall the key challenge.Standalone generative models, however, are hindered by large label spaces and unconstrained outputs.This complementarity motivates separating broad candidate retrieval from fine-grained, image-grounded verification.We propose G2D, a training-free framework that uses a generative VLM to verify CLIP-retrieved candidates against the image.Candidate names and CLIP probabilities provide a structured prior for resolving visually similar classes.Fixed confidence routing, entropy-adaptive candidate sizing, and trie-constrained decoding focus generative reasoning on uncertain samples and ensure one valid output for each input at test time.Across eight benchmarks, G2D achieves 68.85% average accuracy, versus 59.35% for CLIP and 63.11% for the standalone VLM.Across seven generator configurations, candidate-set verification improves average accuracy by 1.08--27.42 percentage points.G2D also transfers to DCLIP, WaffleCLIP, and CuPL, supporting a practical interface between discriminative proposal and generative visual reasoning. Code: https://github.com/Harzva/G2D
Diego Castro Elvira, Navil Pineda Rugerio, Jesús García-Ramírez +2cs.CV cs.AI
AI-generated art has sparked debates around potential plagiarism, as these images may closely resemble existing artworks. This research quantifies the similarity between original pieces and AI-generated counterparts, particularly those produced by the Stable Diffusion XL Refiner 1.0. We use Siamese Networks with frozen CLIP encoders and cosine similarity optimized through triplet loss. A dataset of paired original and generated images was built using image-to-image generation and custom prompts, enriched with semantic descriptors and BLIP-2 captions. Prior studies report up to 81\% style replication and 90\% visual similarity. Our results show high discriminative performance: training accuracy reached 99.9\%, and the best model configuration achieved 99.4\% test accuracy with strong inter-class separation ($δμ$ = 0.677), demonstrating the effectiveness of our semantic-visual embeddings.
Visual prompting (VP) has emerged as a parameter-efficient method for adapting pre-trained models to downstream tasks. However, existing approaches encounter a trade-off between flexibility and efficiency. Some methods apply a fixed prompt to all images, ignoring individual image characteristics, while others introduce auxiliary networks to generate diverse prompts. Although the latter can improve performance, it also significantly increases parameter usage and the potential for overfitting to specific datasets. Furthermore, the auxiliary networks, combined with inherent biases in pre-trained models, limit scalability and generalization. In this paper, we propose Energy-Shaped Visual Prompting (ES-VP), a novel approach that generates image-specific prompts using low-rank initialization and energy-guided dynamic adaptation, achieving superior performance with fewer parameters compared to single-prompt methods. ES-VP directly utilizes the pre-trained model for adaptive prompt generation, ensuring both parameter efficiency and improved generalization. Extensive experiments conducted on five architectures across fifteen datasets demonstrate that ES-VP consistently outperforms current state-of-the-art (SOTA) single and diverse VP methods. For instance, using the CLIP architecture across four datasets, ES-VP outperforms the SOTA method DAM-VP by an average of 2.6\% in accuracy while utilizing 590$\times$ fewer VP parameters, thereby establishing a new benchmark for efficient and generalizable model adaptation.
Laura C. Diaz-Delgado, Emmanuel Martinez, Henry Arguelloeess.IV cs.CV
Self-supervised learning for imaging inverse problems is increasingly important in photon-limited settings, where acquiring clean ground truth is impractical and reconstruction must remain stable under dataset and acquisition shifts. This challenge is amplified under Poisson noise, whose signal-dependent statistics interact with sampling operators (e.g., CFA mosaicing). Meanwhile, foundation vision encoders trained at web scale offer distortion-invariant, content-related representations that generalize well across domains, suggesting a promising route to build priors that transfer beyond the training distribution without expensive fine-tuning. This paper proposes an ADMM-inspired unrolled plug-and-play solver for Poisson inverse problems that decouples a closed-form data-consistency update from a parameter-efficient prior. The prior is implemented as a lightweight decoder operating on frozen CLIP RN50 dense multi-scale features, adapting foundation representations with less trainable parameters. For self-supervision, the method integrates GR2R measurement-domain re-corruption with an Equivariant Imaging regularizer via virtual acquisitions. Experiments on Poisson CFA demosaicing and deblurring show competitive quality, improved robustness under shifts, and self-supervised performance approaching supervised training.
Compositional Zero-Shot Learning (CZSL) aims to recognize unseen attribute-object compositions by leveraging knowledge of primitive concepts learned from seen compositions. Although recent works achieve impressive performance in CZSL by leveraging large vision-language models, they primarily rely on discriminative representations that may not explicitly preserve the structured relationships between primitive concepts and their compositions. Motivated by the recent success of diffusion-based classifiers and their competitive performance relative to discriminative models, we investigate whether intermediate diffusion representations can provide complementary cues for CZSL. To this end, we propose DIFFCZSL, a diffusion-augmented framework that injects generative priors from pre-trained diffusion models into CLIP-based CZSL pipelines. We extract intermediate diffusion representations and project them into the CLIP embedding space to provide auxiliary supervision on both image and text modalities. Through contrastive alignment between CLIP embeddings and diffusion features during training, our method encourages the embedding geometry toward richer composition-aware semantics, while introducing no additional cost at inference time. Extensive experiments on three public CZSL benchmarks demonstrate consistent improvements over strong CLIP-based baselines under both closed-world and open-world settings. Our results highlight the complementary strengths of generative diffusion representations and discriminative vision-language models for compositional generalization.
Dual-encoder models such as CLIP score an image-caption pair by a single inner product of two independently computed unit vectors, and fail at binding, often scoring near chance when asked to distinguish "a red car and a blue dog" from "a blue car and a red dog". We give a mathematical account of when this failure is necessary and when it is contingent. Working within the ideal-encoder framework proposed by Kang et al., we first show the relevant axioms are satisfiable, so every impossibility must enter through an added, checkable hypothesis. We then prove three such obstructions. Depth: for recursive role-binding codes the swap margin obeys an exact law $m(D) = 2b^{-D}$ in the nesting depth D, with a finite-dimension version holding up to one explicitly flagged concentration estimate; the resolvable depth grows only logarithmically in the dimension and is single-digit at CLIP scale, the nesting depth of ordinary language. Objective: architecture-free throttle theorems showing that the contrastive objective's entire reward for binding is bounded by the rate at which training contrasts a caption against its own swap, a rate that vanishes at web scale, and that exactly reversed binding costs only that rate times the mean binding margin; both are verified in simulation. Geometry: a tight smoothness-binding frontier: the closer the two swap-related captions must embed to a shared paraphrase anchor, the smaller the binding margin can be, with an exact constant. Measuring its text-only diagnostic across 18 deployed text encoders, every model sits at roughly 25-35% of its ceiling, and the induced per-item ceiling tracks SugarCrepe's subset difficulty at r = 0.99. Binding failure in deployed dual encoders is thus not a dimension or smoothness limit today, but an incentive and code-structure limit, with a proved depth ceiling that remains once those are fixed.
A multiplicative dual-encoder network computes a real-valued output for a pair of inputs as the inner product of their separate encodings. This architecture has been developed independently in operator learning, bipartite matching, contrastive vision-language models, retrieval, and other areas, yet no unified theory guides the basic design decisions: how many interaction modes to represent, how to normalize the encoders, and when the architecture should be avoided. We provide such a foundation by introducing the class of functions of low interaction rank, a class whose intrinsic complexity is measured by its interaction spectrum. Within this framework, approximation error decomposes into a spectral truncation term and an encoder-realization term; sample complexity is governed by the sum of the two encoder complexities rather than their product; and a usability criterion based on spectral decay determines when the architecture can succeed. The same framework exposes a central identifiability problem: the encoders are defined only up to a linear gauge symmetry that leaves the learned coordinates arbitrary. We show that normalization is gauge fixing and that whitening pins the interaction modes up to permutation and sign, thereby explaining the uninterpretability of contrastive dimensions and providing a constructive remedy. Experiments on synthetic kernels, operator learning, and CLIP models validate the theoretical predictions: spectral decay rates match the predicted scaling, whitening recovers the true modes, and independently trained CLIP models are related by a single rotation which, after removal by whitening, exposes interpretable concept axes. The code of this paper is provided at https://github.com/RS2002/Mul-Net .
Fine-tuning vision-language models such as CLIP typically requires backpropagation (BP) through the full model, which is infeasible when only forward-pass access is available, as is common for memory-constrained edge devices and proprietary model deployments. Prior BP-free, zeroth-order prompt-tuning methods avoid this requirement but often tune prompts in a single modality or optimize over a search space large enough that convergence requires thousands of forward passes, which is impractical under realistic query budgets. We propose ZOMP (Zeroth-Order Multimodal Prompt tuning), a query-efficient, fully forward-only method that tunes deep prompts in both the vision and text branches of a frozen CLIP model using simultaneous perturbation stochastic approximation. ZOMP combines three ingredients: a cross-modal low-rank reparameterization that ties the two branches through a shared factor and keeps the effective search dimensionality small, a gradient-correction momentum term that stabilizes the noisy zeroth-order estimate, and a budget-indexed rank schedule that unlocks capacity as the query budget is spent. Across 13 vision-language benchmarks under a matched 5,000-query budget, ZOMP consistently outperforms prior BP-free prompt-tuning methods in both few-shot accuracy and query efficiency, and it generalizes better across base-to-new, cross-dataset transfer, and out-of-distribution settings. Our results show that jointly exploiting multimodality and low-rank structure is an effective route to practical, query-efficient BP-free prompt tuning.
Yasmin Yasin, Muhammad Usman, Ibrahim Radwan +1cs.CV
Backlit and low-light images often suffer from severe exposure imbalance or global underexposure, presenting significant challenges for both visual perception and downstream computer vision tasks. In this paper, we propose a unified, unsupervised enhancement framework that addresses both types of degradation without relying on paired ground-truth data. Our approach builds on CLIP-guided prompt learning to semantically supervise enhancement using learned positive and negative textual prompts. To improve the quality of our improvements over prior work, we design a symmetric residual U-Net backbone augmented with an Atrous Spatial Pyramid Pooling module. This architecture captures multi-scale contextual information, enabling adaptive correction under spatially heterogeneous illumination. During training, the enhancement network is guided by CLIP-based semantic similarity losses and refined via an iterative prompt optimization mechanism. Extensive experiments on both paired and unpaired datasets, including BAID, Backlit300, LOL, and VE-LOL-L, demonstrate that our framework consistently outperforms state-of-the-art supervised and unsupervised methods in terms of fidelity, perceptual quality, and generalization. Furthermore, our work emphasizes the need for stronger benchmarking protocols for backlit enhancement, a relatively underexplored area. The proposed framework provides a robust, scalable solution for real-world illumination enhancement across diverse lighting conditions.
Zero-shot anomaly detection (ZSAD) aims to detect anomalies and localize defective regions in unseen target domains without target training data. Recent ZSAD methods build on pretrained vision models, particularly CLIP, and construct normal and anomaly references from textual prompts or learnable visual representations. These methods perform anomaly discrimination primarily in spatial feature spaces, where subtle changes in texture, boundaries, and local structures can be confused with normal appearance variations. Although inconspicuous spatially, such defects can disrupt local texture regularity or boundary continuity, inducing response deviations across frequency bands. However, existing ZSAD methods do not explicitly model these frequency-dependent characteristics. Our image-domain analysis reveals that local defects exhibit spatial-frequency deviations from normal references across low-, middle-, and high-frequency bands, indicating that anomaly evidence is not universally dominated by high-frequency responses. Motivated by this observation, we propose FreqAnchorAD, a frequency-aware framework that organizes frequency-enhanced responses for anchor-relative anomaly discrimination. Specifically, the Local Frequency Compensation Module (LFCM) enhances intermediate patch tokens with local spatial-frequency cues. The Frequency-Deviation Anchor Projector (FDAP), our core discrimination module, organizes enhanced responses along a source-derived channel coordinate and measures anomaly evidence through relative similarity to normal and anomaly anchors. Finally, Asymmetric Anchor Supervision (AAS) stabilizes normal-anchor alignment while preserving diverse anomaly patterns. Experiments on thirteen industrial and medical benchmarks show that FreqAnchorAD achieves state-of-the-art mean performance in image-level anomaly recognition and pixel-level defect localization.
Prompt tuning adapts vision--language models with few trainable parameters, but existing approaches trade off efficiency and adaptation: static textual prompts can overfit source classes, image-conditioned prompts add per-instance computation, and multimodal tuning modifies the visual branch. We propose VDA (Visual Distribution Anchoring), a training-free target adaptation framework that augments a frozen semantic classifier with class-level visual prototypes estimated offline from an unlabeled target pool. We first ask whether prototypes can be synthesized from class names. A text-to-centroid mapper reconstructs held-out source prototypes but fails under dataset shift because class names specify semantic identity, not target-domain appearance. An oracle analysis confirms that true target prototypes are highly discriminative. VDA therefore uses frozen semantic and domain-template classifiers to partition unlabeled target images into class-correlated groups. Confidence-ranked image features form normalized prototypes, fused with the semantic classifier using one global weight. Adaptation requires no target labels, target-side optimization, uniform class-prior assumption, iterative refinement, or test-query access, and yields a fixed, cacheable classifier. Controlled experiments show that class-specific partitioning drives gains and that visually local pseudo-label errors can remain useful despite being class-incorrect. Across ten ImageNet-to-target transfers, the same frozen design improves zero-shot CLIP, TCP, and MaPLe by 3.22, 3.39, and 3.35 points, respectively, improving nine of ten targets in every setting. Its visual correction further improves leakage-free PromptKD by 2.79 points, complementing zero-shot, source-prompted, multimodal-prompted, and target-distilled classifiers.
While text-to-image diffusion models achieve impressive visual quality, they frequently struggle to maintain precise alignment with complex compositional prompts. An effective strategy is to improve the inference process of diffusion models, thereby better leveraging their pretrained priors to address misalignment. Existing training-free methods can be divided into two categories. The first category focuses on improving the randomly sampled initial noise, either performing costly search over noise pools or manipulating sampled noise without ensuring reliable semantic injection. The second category focuses on improving the denoising trajectory, lacking explicit mechanisms to timely diagnose and correct semantic errors. we propose \textbf{AnchorSteer}, a training-free framework that exerts fine-grained control over \textbf{both initialization} and \textbf{the denoising trajectory}. AnchorSteer consists of two synergistic components: \textbf{Semantic Anchoring} replaces uninformative Gaussian noise with text-aligned initializations via CLIP-based prior extraction and a novel Latent-Prior Score Distillation Sampling (LP-SDS) objective. Specifically, LP-SDS distills CLIP visual priors into the knowledge distribution of diffusion models, mitigating the domain gap between CLIP-based priors and diffusion-based priors. \textbf{Reflective Steering} transforms passive denoising with an active Think--Erase--Retouch loop that enables mid-generation self-correction. It leverages VLM-based diagnosis to detect semantic deviations and performs targeted latent refinement to suppress erroneous content and recover missing attributes. Extensive experiments on GenEval and T2I-CompBench++ demonstrate that AnchorSteer consistently outperforms existing baselines in text--image alignment while preserving high visual quality.
Open-world object detection (OWOD) requires a detector to recognize known categories, discover unnamed objects from unseen categories, and incrementally learn newly annotated classes. PROB improves unknown discovery by modeling class-agnostic probabilistic objectness in the decoder-query space. However, visual objectness alone cannot determine whether an object-like query corresponds to a hard known instance, an unseen-category object, or background clutter, resulting in an ambiguous known-unknown decision boundary. We propose MSPO, a lightweight semantic calibration framework that augments PROB with task-aware known-category language priors while preserving its detector architecture and incremental learning protocol. For each currently known category, MSPO constructs an extended text description covering category attributes, visual appearance, typical scenes, and functional usage, and encodes it using a frozen CLIP text encoder. Decoder query features are projected into the same semantic space to estimate their support from the current known-category semantics. This semantic evidence is fused with PROB's visual objectness to calibrate known and unknown predictions without turning OWOD into open-vocabulary classification. Importantly, MSPO never uses future-category names, and all unseen categories remain unnamed during evaluation. Experiments on M-OWODB and S-OWODB show that MSPO improves the strong PROB baseline on the main aggregate metrics while retaining competitive unknown recall. It also improves early unknown-confusion metrics and raises PASCAL VOC final mAP by up to 2.7 points. These results demonstrate that known-category language semantics provide an effective calibration signal for probabilistic objectness under the standard OWOD setting.
Joseph Fioresi, Fabian Caba Heilbron, Pankaj Nathani +2cs.CV cs.AI
Multimodal embedding spaces in models like CLIP enable powerful capabilities such as semantic similarity retrieval and cross-modal zero-shot classification. These embeddings compress high-level semantics into a single vector, which comes at the cost of primarily expressing a dominant semantics like main object while suppressing other important attributes such as camera angle or color tone. We propose a text-conditioned transformation of visual embeddings that makes such attributes explicitly accessible. Given a natural language description of an attribute category (e.g., "color" or "art style"), a network generates an affine transformation that emphasizes the specified attribute. Conditioning on text enables it to learn many attributes simultaneously, accessing them at inference time through an intuitive interface. The network is trained to align transformed embeddings with the frozen latent space, enabling retrieval using existing large-scale embeddings without any re-encoding. When applied to a full set, the same mechanism transforms the latent space for attribute disentanglement tasks such as multi-clustering. By operating directly in latent space, our method provides a unified and efficient framework for controlling embedding spaces, demonstrating state-of-the-art performance across both attribute-based retrieval and multi-attribute organization tasks with near-zero inference cost. Project page: https://joefioresi718.github.io/ControlEmbed_webpage/
Hoang Le, Joonwoo Kwon, Elkhan Ismayilzada +2cs.CV cs.AI
Inferring physical properties such as mass, stiffness, and elasticity from a single image is essential for simulation and embodied AI, yet most existing approaches rely on multi-view reconstruction or physics-based supervision. We introduce SiPhy, a unified framework for single-image physical property reasoning that aligns 3D-aware visual cues, depth with language-based material knowledge. From one RGB image, SiPhy samples pseudo-voxel points, extracts CLIP features, and grounds them to material candidates proposed by a VLM. A part-based contrastive aggregator enforces region consistency, while a heaviness-aware refinement improves thickness and volume estimation for dense objects. Across ABO-500, MVImgNet-100, and PhysXNet-100, SiPhy achieves state-of-the-art single-image performance, surpassing multi-view reconstruction methods by improving mass MnRE by up to 93% (vs. PUGS), reducing density MAE by 35.5% (vs. NeRF2Physics), and lowering Young's modulus error by 23.5%. We further validate SiPhy on real hand-object interaction datasets, demonstrating its potential as a data annotation engine for physical understanding from single-view imagery.
Hateful optical illusions expose a serious gap in current multimodal safety systems. On original-view hateful illusions, previous work shows that six moderation classifiers achieve at most 20.9 to 24.5% accuracy and nine state-of-the-art VLMs remain at or below 10.2% with illusion-aware prompting, leaving most hidden hate undetected. We formulate hidden hateful illusion detection as a perceptual retrieval problem and propose Adaptive View Retrieval. This retrieve-and-calibrate framework assembles a complementary view bank for the image and hidden-message templates, adaptively selects which views to trust, retrieves hidden-message identities, and calibrates whether the recovered evidence is harmful. On HatefulIllusion with a frozen CLIP encoder, Adaptive View Retrieval reaches 93.2% balanced accuracy on the held-out test split. It substantially outperforms original-view baselines and fixed single-transform filters across hate slangs, hate symbols, and visibility levels. The same design also surpasses official fine-tuned CLIP baselines, matches or exceeds human performance on IllusionMNIST, IllusionFashionMNIST, and IllusionAnimals, and outperforms zoom-out preprocessing on HC-Bench under the SemVink protocol. Together, these results show that robust multimodal moderation requires recovering hidden meaning before deciding whether it is harmful.
Few-shot Test-Time Domain Adaptation (FSTT-DA) seeks to adapt models to novel domains using only a handful of unlabeled target samples. This setting is more realistic than typical domain adaptation setups, which assume access to target data during source training. However, prior FSTT-DA approaches fail to effectively leverage source domain-specific knowledge, relying on shallow batch normalization updates, prompt-based methods that treat the model as a black box, or ensembling strategies that do not capture cross-domain relationships. To address these limitations, we introduce a new FSTT-DA framework that integrates LoRA fine-tuning with model merging. In our approach, separate LoRA modules are fine-tuned on CLIP's vision encoder for each source domain. Since LoRA modifies only a small fraction of the model's parameters, it retains the base model's generalized knowledge while internally learning domain-specific features. To adapt the learned knowledge to a specific target domain, we propose a hypernetwork trained via meta-learning that generates per-column merging factors to combine LoRA modules. Given a small batch of target images, the hypernetwork produces merging weights that fuse source LoRA modules into a single adapted representation. Our results demonstrate state-of-the-art performance across various domain adaptation datasets. Our code is publicly available at https://github.com/nahbois4321/DA-MergeLoRA.
Multimodal models such as CLIP learn a shared embedding space for cross-modal retrieval, but continual adaptation to sequentially arriving data can disrupt the cross-modal alignment acquired from earlier phases. Conventional continual-learning methods return a single checkpoint, which commits every retrieval direction to the same stability-plasticity trade-off. We propose AlphaWiSE, a post-hoc weight-space interpolation method that composes two frozen source checkpoints. For each aligned parameter tensor identified by its checkpoint key, AlphaWiSE fits one scalar interpolation coefficient shared by all tensor entries. The coefficients are fitted on a smaller exemplar memory and used to materialize one interpolated checkpoint. The deployed model has the same architecture and parameter count as either source checkpoint, which does not require additional inference time. Extensive experiments on audio-image-text retrieval show consistent improvements over strong continual-learning baselines across multiple retrieval directions and evaluation metrics.
Contrastive Language-Image Pretraining (CLIP) representations form a semantic embedding space governed by cosine similarity, reflecting an intrinsic hyperspherical geometry. However, existing probabilistic interpretations typically rely on Gaussian assumptions, which fail to capture this directional and multimodal structure. We propose a principled density model for the CLIP latent space based on Mixtures of von Mises-Fisher (MovMF) distributions defined on the unit hypersphere. Using the Expectation-Maximization (EM) algorithm, we efficiently learn a probabilistic model in which each mixture component corresponds to a coherent semantic concept. This formulation yields a closed-form likelihood naturally aligned with hyperspherical geometry, enabling accurate and interpretable density estimation. Empirically, our model significantly improves long-tailed and out-of-distribution detection and provides a natural semantic decomposition, representing each embedding as a sparse probabilistic combination of interpretable concepts. These results suggest that CLIP latent space is more faithfully characterized as a hyperspherical semantic mixture rather than an isotropic Gaussian, establishing a simple and geometrically consistent probabilistic framework for modeling and understanding multimodal representations. Project page is available at https://xiaoyuzhizi.github.io/movmf-clip/.
Image geolocation aims to infer the geographic origin of an image from visual content alone. However, this task remains challenging in regions where countries share similar urban, roadside, architectural, and environmental characteristics. Many existing geolocation models focus on coordinate level prediction or classification performance while providing limited insight into how visual evidence contributes to location predictions. This study presents an explainable country level image geolocation pipeline for 11 ASEAN countries. First, we collected 4,850 images from GeoGuessr style sources, Google Images, and additional street level imagery. We then evaluated three approaches on this dataset: CLIP zero shot classification, a LightGBM classifier, and an MLP classifier. The MLP achieved the best test performance, attaining an accuracy and F1 score of 85.91%. For explainability, predictions generated by the MLP classifier were analyzed post hoc using CLIP attention rollout, YOLO26 object detection on the original images, and Energy Based Pointing Game (EBPG) overlap metrics. Object level analysis indicates that frequently detected objects are not necessarily associated with the highest attention density, suggesting that object frequency and attention based visual evidence capture different aspects of a scene. These results demonstrate that the proposed model can support accurate regional image geolocation while enabling object level inspection of the visual cues underlying its predictions.
Ahsan Habib Akash, Dipkamal Bhusal, Stacey Jones +3cs.CV
Deep neural networks are widely deployed in high-stakes visual applications where interpretability is critical, yet existing explanations face a trade-off: post-hoc concept methods recover factors that are faithful to a model's behavior but unnamed, while naming and by-design methods attach human-readable concepts only by retraining or altering the classifier. We propose Language-Anchored Decomposition (LAD), a post-hoc framework that delivers concepts which are simultaneously named, faithful, and obtained without modifying the model. For each class, a large language model proposes a concept vocabulary that CLIP-based similarity maps localize across image regions. Inverting standard non-negative matrix factorization, LAD fixes these language-grounded maps as the coefficient matrix and learns only a concept basis that reconstructs the frozen encoder's activations, so naming becomes a structural constraint and the model's own feature geometry determines which concepts are retained. Removing this anchor preserves accuracy but collapses attribution faithfulness. Across natural-image, scene, and medical-imaging benchmarks, LAD produces spatially precise explanations that are decision-relevant under both concept insertion and deletion, while uniquely providing stable, human-interpretable concept names.
The rapid advancement of large-scale generative models has accelerated the spread of highly deceptive AI-generated images, making generalized synthetic image detection a critical imperative. Existing forensic networks often struggle with cross-model generalization and realworld degradations due to their reliance on single-domain representations and conventional binary classification optimization. To overcome these limitations, we propose RNSIDNet, a novel forensic framework that achieves robust detection through enhanced RGB-Noise representation learning. Specifically, our method employs a dual-branch architecture where global RGB semantics, extracted by an attention-refined CLIP backbone, dynamically modulate highfrequency noise artifacts captured by Bayar convolutions via a Feature-wise Linear Modulation (FiLM) module. To further enhance the learned representations, we design a Hard Sample-aware Contrastive Learning (HSCL) strategy. By explicitly penalizing challenging training samples, HSCL reshapes the latent feature space to maximize the discriminative margin between pristine and synthetic domains. Extensive experiments across eight public benchmark datasets verify that our model achieves state-of-the-art performance, delivering superior generalization ability, robustness, and computational efficiency. Code and dataset will be publicly available on https://github.com/multimediaFor/RNSIDNet.
Traditional semantic segmentation models operate under a closed-set assumption and struggle to recognize unknown or unexpected objects-an essential capability for autonomous driving. As a result, such models often misclassify or overlook out-of-distribution (OOD) road anomalies, posing safety risks in open-world environments. We present a lightweight, postprocessing, road-aware anomaly segmentation framework that requires no retraining, no OOD data, and no auxiliary supervision. Our approach builds on a mask transformer-based segmentation network by exploiting query-level mask confidence and deriving a polygonal road prior to detect gap regions that may correspond to anomalies. To further suppress false positives, we introduce a CLIP-based zero-shot semantic filtering module using in-distribution prompts, with optional generalized OOD prompts. By jointly leveraging spatial priors and semantic verification, our framework produces robust and interpretable anomaly predictions. Evaluation on three public benchmarks-Fishyscapes, SMIYC, and RoadAnomaly-shows consistently strong performance. In particular, our method outperforms the training-free baseline Maskomaly on most metrics and achieves the highest AP on Fishyscapes LostAndFound. These results demonstrate the practicality and deployability of our approach for real-world autonomous driving systems.
Bohan Liu, Wenqian Ye, Guangzhi Xiong +3cs.CV cs.CL
Models trained via Contrastive Language-Image Pretraining (CLIP) serve as the foundational vision encoders for most modern Large Vision Language Models (LVLMs). Despite their widespread adoption, CLIP models exhibit a critical yet underexplored failure mode: irrelevant text appearing within images confounds visual representations, biasing them toward lexical meaning rather than true visual semantics. This robustness issue, commonly described as a Typographic Attack (TA), exposes a vulnerability that poses a significant risk to safety-critical applications such as autonomous driving. To achieve interpretable and effective robustness against TA, we propose a novel, training-free mechanistic interpretability method. Our method provides sampling-based interpretations of hidden state representations and quantitatively attributes semantic versus lexical focus to individual attention heads. Through probabilistic analysis and circuit mining, we isolate specific Vision Transformer (ViT) components that disproportionately encode lexical information, thereby identifying the mechanistic source of TA. We further show that simple interventions applied directly to the identified circuits, without any additional training, can substantially improve robustness against Typographic Attacks in object classification. These interventions, such as selective adjustment of attention weights, also outperform both supervised and training-free defense methods. Our experiments demonstrate that applying the proposed intervention to the vision encoders of several state-of-the-art LVLMs yields substantial gains in Visual Question Answering accuracy under Typographic Attack interference on RIO-Bench. These results confirm both the efficacy and the generalizability of our mechanistic approach. Code is released at https://github.com/Liu-524/SamplingTAR.
Text-to-image diffusion models achieve impressive visual quality, yet demographic bias remains a challenge, as neutral prompts consistently produce stereotypical representations across gender and race. Existing approaches remain limited by costly retraining or by inference-time interventions that often degrade image quality and semantic alignment. We propose Text Embedding Steering (TES), a training-free framework that mitigates demographic bias by directly optimizing conditional text embeddings during the diffusion process. We show that a two-stage strategy - early-stage global alignment followed by iterative denoising-time refinement with CLIP-based feedback - enables stable and controllable attribute steering without modifying model parameters. Extensive experiments on Stable Diffusion demonstrate that TES outperforms existing training-free baselines in fairness while maintaining competitive image quality. These results highlight that inference-time text embedding optimization is a practical and scalable solution for fairness-aware generation in diffusion models.
Blind image deblurring demands the recovery of high-fidelity details and coherent structures from complex, unknown degradations. Current blind image deblurring methods struggle with real-world, spatially varying degradations, and lack the semantic awareness necessary to reliably differentiate valid textures from artifacts. To bridge this gap, we propose CogSENet, a dynamic, semantic-aligned reconstruction framework inspired by the eagle's visual system. By mimicking the eagle's active saccadic scanning, we devise a Semantic-Driven State Space Module (SDSSM) with semantic-aware token regrouping via differentiable routing, enabling prompt-conditioned long-range dependency modeling. To ensure physically interpretable recovery of textures and structures, a BiFreqFusionBlock (BFFB) mirrors functional differentiation of the eagle's retina by decomposing features into high and low frequencies using wavelet transforms. Finally, we estimate a continuous Blur Field (CBF) from blur image and fuse it with CLIP semantic priors to modulate the deepest latent features, emulating focal adaptation and enabling adaptive restoration under spatially non-uniform blur. Extensive experiments demonstrate that CogSENetoutperforms state-of-the-art deblurring methods in both visual quality and structural fidelity with fewer parameters, while also performing favorably on dehazing, deraining, and denoising tasks.
To address data overload and inefficient shape-level annotation in robotic visual inspection, this paper proposes a hardware-software integrated optoelectronic architecture. A non-imaging, low-data paradigm is established to minimize annotation dependency. First, a sensor-in-the-loop strategy reconfigures a Digital Micromirror Device (DMD) as a physical optical convolutional layer, enabling photonic-domain feature extraction that unifies sensing hardware and processing software. To suppress data volume at the source, a block-based compressed sensing strategy encodes spatial information into low-dimensional temporal signals, drastically reducing redundancy. Subsequently, to bypass laborious manual defect shape annotation, natural language descriptions guide the network to align with highly generalizable features from Contrastive Language-Image Pre-training (CLIP), steering the attention maps of the optoelectronic neural network toward defect shapes. Furthermore, a Localization Accuracy for Attention (LAA) metric is proposed to quantify shape-level defect localization performance. Experiments on transparent material defect detection validate the system's effectiveness. Parametric analysis reveals how measurement matrices, compression ratios, and block sizes affect accuracy. Results show that, compared to traditional imaging, the proposed architecture maintains equivalent accuracy while reducing data volume by 90% for Vision Transformers and computational workload by 60% for Convolutional Neural Networks. This low-data paradigm offers an efficient solution for industrial automation scenarios involving massive data streams, high acquisition costs, or constrained edge resources.
Livia Betti, Sebastian Ricke, Ivica Obadic +2cs.LG
Geographic implicit neural representations (INRs) learn to map any coordinate on Earth to a location embedding, implicitly encoding geospatial data into the weights of a neural network. Location embeddings are widely used off the shelf as general-purpose geospatial representations, yet users lack principled tools to audit what geographic or semantic information these embeddings capture. In this work, we analyze the information content of geographic INRs through their location embeddings. We decompose these embeddings into human-interpretable features$\unicode{x2014}$namely, (i) sparse latent concepts, (ii) natural language concepts, and (iii) visual features. The latent concept embeddings are learned using sparse autoencoders. To recover natural language concepts, we apply sparse linear concept embeddings (SpLiCE) over a predefined geospatial dictionary. Finally, visual features are extracted using saliency maps derived from CLIP Surgery. We show that location embeddings can be decomposed into human-interpretable representations while retaining high reconstruction capability, revealing interpretable geographic structures such as forests, deserts, and urban features. Across methods, sparse decompositions expose systematic differences in encoded information, ranging from urban structures to broader biome and climate signals, and pretraining-space saliency maps further highlight complementary features such as roads and landmarks. We hope this work provides a first step toward interpretable geospatial representations.