Overview

Vision foundation models learn reusable visual representations from large image or video collections, often through self-supervised methods such as contrastive learning, masked prediction, and feature prediction, or through image-text and generative training. Researchers adapt them using frozen-feature probes, prompts, lightweight fine-tuning, distillation, and feature fusion for tasks ranging from classification and segmentation to depth, 3D geometry, and video understanding. These methods can transfer across tasks and domains, but specialized pretraining or adaptation can still be important in settings such as medical imaging, remote sensing, and scientific vision. A central open tension is whether broad pretrained representations can provide reliable spatial, temporal, and domain-robust understanding while remaining efficient and consistently evaluated.

Latest papers

152 papers16 papers in the last four weeks, up 7% on the four weeks before. 0.1% of all new papers.

Sep 28, 2026

Hard Vision, Easy Vision: What GPT-6 Astra Reveals Across Computer Vision

Frontier general-purpose systems are rapidly expanding beyond visual understanding into capabilities traditionally handled by dedicated computer-vision models. As these capabilities expand, a central question for the computer-vision community is how far this reach extends, and what remains hard. We evaluate GPT-6 Astra alongside five frontier general-purpose AI systems across 34 capabilities and 55 benchmarks spanning nine areas of computer vision. We compare their performance with dedicated models and humans where suitable references are available. Astra demonstrates broad visual capability, with substantial gains over other frontier systems in visual and spatial reasoning and several forms of structured prediction. Across the state-of-the-art systems, a consistent pattern emerges. Capabilities involving semantic interpretation, reasoning, and object-centric prediction increasingly approach or reach available reference levels. In contrast, larger gaps remain when tasks require metric geometric accuracy, faithful reconstruction, temporally consistent dense prediction, or specialized fine-grained visual knowledge. Additional reasoning and specialist tools close selected gaps, but their benefits vary across capabilities. These results map a changing landscape of computer vision in which increasingly sophisticated visual tasks are accessible through a general-purpose interface, while precise and fidelity-sensitive perception remains an important frontier.
Sep 28, 2026

Representation Risk in Pretrained Image Encoders

Applied researchers increasingly convert images into features with pretrained encoders, then use those features in a downstream prediction model. The encoder is often treated as an implementation detail. We show that it can instead be a consequential source of model uncertainty. We call this uncertainty representation risk: plausible pretrained encoders map the same images into different feature spaces and can yield sharply different out-of-sample conclusions from predictive performance. We compare ten modern and legacy frozen encoders across applications involving house prices, racehorse performance, breast-cancer histology, chest radiographs, continuous facial age, and rice disease. With common dimension control, heads, and group-safe splits, validation selects SigLIP 2 for houses, raising test R2R^2 from 0.396 for ResNet50 to 0.629, and DINOv2 for horses, raising R2R^2 from 0.029 to 0.105. No encoder is best in every task. Candidate procedures are constructed using training data and compared on a separate validation partition. The selected procedure reaches 0.658 for houses and 0.979 accuracy for pneumonia. Fixed-split gains are small for horses and rice, while repeated partitions reveal instability in horse feature union. Continuous age selects SigLIP 2 at 4.786 years MAE. The principal representation gaps persist with neural heads, similarly sized DINOv2 and ViT models, and limited adaptation. These results support a simple workflow: benchmark plausible representations, select on locked validation data, combine only when separate validation evidence justifies the additional cost, and report paired and split-level uncertainty. We implement this workflow in LOOKAGAIN-ML, the software package used to conduct the analyses in this paper.
Sep 26, 2026

Scalable In-Domain Self-Supervised Foundation Model for Dense Representation Transfer in High-Resolution Plant Imaging

High-resolution plant imaging enables detailed characterization of plant morphology, but dense scientific analysis remains limited by costly pixel-level annotations, large image pixel dimensions, and substantial variation in imaging conditions. This work proposes a scalable in-domain self-supervised pretrained foundation model for high-resolution, high-pixel-dimension multi-species plant imagery. A masked autoencoder with a ViT backbone is pretrained on more than 10 million multi-view plant image tiles using distributed training. Following scalable pretraining, the learned foundation-model representations are comprehensively benchmarked across fine-grained dense prediction and coarse global feature recognition, with particular emphasis on limited supervision and realistic downstream imaging conditions. This work focuses on the domain gap of existing foundation models in dense feature representation and transfer. Through extensive experiments involving limited annotations, cross-view variation, and resolution degradation, the in-domain FM achieves a Mean Dice of 0.8686 and a Pooled Dice of 0.8959, outperforming an MAE counterpart pretrained on large-scale natural-image data by 0.0694 and 0.0613, respectively. The results further indicate that increasing pretraining scale produces consistent improvements in dense feature transfer. Overall, these findings suggest that scaling in-domain self-supervised pretraining can reduce the domain gap and improve transferable dense representations for high-pixel-dimension scientific imaging.
Sep 23, 2026

UltraBench 2: Towards Robust Evaluation of Vision Foundation Models on Ultrasound

Benchmarking is an increasingly critical part of research in machine learning and the domains where it is applied, including healthcare. Yet, despite the steady development of new ultrasound foundation models in recent years, the development of well-designed benchmarks to evaluate them has lagged behind. This deficiency has led to fragmented and inconsistent evaluations of competing models, making it difficult to measure progress. To address this issue, we introduce UltraBench 2, a comprehensive benchmark with wide anatomical and task coverage, and a focus on standardization, reproducibility, and ease-of-use. Using this benchmark, we compare existing vision foundation models for ultrasound image analysis. Our analyses demonstrate that ultrasound-specific pretraining still leads on classification, but that state-of-the-art general-purpose models have drawn level on segmentation.
Sep 20, 2026

Towards robust multimodal 3D object detection via visual foundation models

Multimodal 3D object detection is fundamental to robust perception in autonomous driving because it integrates complementary information from LiDAR and camera sensors. However, existing methods often fail to maintain robustness under out-of-distribution (OOD) corruptions caused by sensor noise, adverse weather, and environmental changes. To address this problem, we propose RoboDistill, a robust and generalizable multimodal 3D object detection framework that leverages visual foundation models (VFMs), such as the Segment Anything Model (SAM). First, we introduce SAM-AD, a domain-specific pretraining strategy that fine-tunes SAM on autonomous-driving imagery to extract feature representations with rich semantic information. Second, we design the AD Feature Pyramid Network (AD-FPN) to refine and upsample SAM features at multiple scales for seamless fusion with LiDAR features. Third, we develop the Depth-Guided Wavelet Attention (DGWA) module, which suppresses high-frequency sensor noise while preserving critical contextual information. Finally, we introduce KD Fusion, in which the pretrained SAM-AD serves as a teacher that distills high-quality visual knowledge into a lightweight point-cloud network, thereby improving robustness under noisy conditions. Extensive experiments across 27 challenging OOD corruption settings show that RoboDistill generally delivers stronger or competitive detection performance and robustness relative to representative state-of-the-art methods. This work bridges the gap between VFMs and 3D object detection and advances robust multimodal perception for real-world autonomous-driving applications.
Sep 19, 2026

Human-Level Accuracy, Non-Human Strategies: Revealing Model-Human Divergence in Video Physical Reasoning

Video foundation models now reach human-level accuracy on physical-reasoning benchmarks, yet such tasks require predicting unobserved physical outcomes. Do these models perform human-like forward simulation, or do they exploit statistical regularities in visible scenes? Accuracy alone cannot distinguish these strategies. We introduce a distributional evaluation framework that treats model seeds and human raters as populations, enabling comparison of consensus, uncertainty, and strategy. On the Physion benchmark, we evaluate three ViT-L architectures (V-JEPA2, VideoMAEv2, DINOv2). V-JEPA2 narrows the accuracy gap to ~1 percentage point (73.2% vs. 74.2%), yet model-human disagreement reaches 26.4%, far exceeding human-human disagreement (4.8%), with substantially lower agreement (kappa ~ 0.48 vs. 0.91). The divergence follows forward-simulation demands: models outperform humans on geometric reasoning (linking, +11.8 pp) but underperform on gravitational dynamics (rolling, -11.8 pp) and causal chains (dominoes, -10.5 pp). Strategy fingerprinting confirms all three architectures share non-human strategies while none aligns with humans. Attribution analysis suggests that unobservable outcome features, rather than visible scene properties, predict this divergence, consistent with models relying more on scene-level statistical regularities than on explicit forward simulation, a systematic divergence that accuracy alone cannot reveal. Code is available at https://github.com/fanhong-li/model-human-divergence.
Sep 15, 2026

Exploring 2D backbone effects for indoor semantic occupancy prediction

Semantic occupancy prediction gives an embodied agent a voxel-level account of where space is free, occupied, and semantically meaningful. In RGB-D pipelines such as EmbodiedScan, the image encoder is often left as a default module, even though its features are the visual evidence later sampled into the 3D grid. We study this design choice directly. A central finding is that changing the 2D backbone improves occupancy accuracy more than several carefully designed occupancy architectures or modules. We keep the main RGB-D projection, depth branch, and occupancy head fixed, and replace only the image backbone. The compared encoders are CLIP-ResNet, CLIP-ViT, BLIP2, and DINOv2. Under the controlled setting, the measured mIoU changes substantially: DINOv2 obtains 30.55%, BLIP2 obtains 29.49%, CLIP-ViT obtains 24.33%, and CLIP-ResNet obtains 17.41%. The stronger encoders also exceed the original EmbodiedScan ResNet-50 baseline without modifying the downstream 3D fusion pipeline. Class-level results give a more detailed picture: DINOv2 is stronger on many layout and structural categories, whereas BLIP2 remains close on several object-centered classes. CLIP-ViT improves clearly over CLIP-ResNet, showing that the way CLIP features are exposed as dense tokens matters for voxel lifting. These results indicate that the image backbone is not a secondary engineering detail in embodied semantic occupancy, but a major source of variation in the final 3D prediction.
Sep 8, 2026

"World Knowledge" in the Weights: Reading Concept Circuits of Vision Transformers

Vision transformers (ViTs) have achieved remarkable generalization across visual domains, yet little is known about how they internally represent the structure of the world. To address this gap, we use Cross-Layer Transcoders (CLTs) to read concept circuits from ViTs: directed graphs whose nodes correspond to sparse, interpretable concepts and edges capture concept interactions across layers. Our method yields two complementary views of model behavior. The global concept circuit is input-invariant and can be recovered directly from learned cross-layer weights, exposing the reusable "world knowledge" encoded in the model. The instance concept circuit is input-dependent and identifies the concepts and pathways actually used for a specific prediction, enabling faithful example-level explanations. We demonstrate the utility of concept circuits in three ways: (1) Automatic spurious correlation discovery: leveraging the statistics of our global concept circuits to identify shortcut dependencies within the model. (2) Spurious correlation removal: intervening on the instance concept circuit to steer the model towards correct predictions. Empirical results show that our method outperforms existing counterparts by 11.0% on the Waterbird dataset. (3) Model comparison: contrasting the global concept circuits of different foundation models (e.g., CLIP vs. DINO) to reveal how supervision paradigms shape representational structure. Our code is available at https://github.com/deep-real/VisionCLT
Sep 7, 2026

CrACK: Adversarial Attacks on Cross-Model Consistency in Collaborative Vision Foundation Models

Training-free collaborative pipelines that integrate Vision Foundation Models such as CLIP, SAM, and DINO achieve strong open-vocabulary dense prediction and are increasingly deployed in safety-critical applications. The security of these systems is commonly assumed to follow from the robustness of their individual models. We challenge this assumption. We identify a vulnerability shared by every collaborative pipeline: each model consumes the intermediate output of another without verifying semantic consistency, an unverified premise that we term the semantic-spatial alignment dependency. Existing adversarial attacks target a single model and overlook this premise, leaving the inter-model interface entirely unguarded. We propose CrACK (Cross-model Adversarial Consistency attack), an inference-time attack that exploits this interface without modifying any input pixel, model weight, or training data. CrACK operates in two stages: Adversarial Affinity Contradiction Injection corrupts the cross-modal affinity matrix by inverting SAM encoder features under the guidance of CLIP patch-level semantics, and Semantic Interface Poisoning steers the prediction through a max-distance label permutation derived from CLIP text embeddings. Experiments on four collaborative pipelines across eight benchmarks show that CrACK causes catastrophic degradation while every individual model continues to produce its unchanged standalone output, rendering per-model defenses structurally blind. The corruption further cascades into large vision-language model reasoning, driving models such as LLaVA to produce erroneous responses from visually intact inputs. Our results show that the security of a collaborative AI system cannot be reduced to the robustness of its components, and that inter-model feature interfaces must be treated as first-class security boundaries.
Sep 6, 2026

VidaForge: Open Research Infrastructure for Video Pretraining Data Recipes

Video foundation models increasingly rely on large-scale pretraining data, yet the end-to-end data pipelines behind them remain largely closed and difficult to inspect or reuse. Researchers seeking to understand how video data recipes affect model pretraining often need to build substantial infrastructure before testing even a focused hypothesis. We present VIDAFORGE, an open research infrastructure that represents a video data recipe as an executable five-stage workflow from raw videos to training datasets. A decision in this workflow can be varied to construct alternative datasets while preserving how every sample was produced. To demon strate this research workflow, we compare data recipes with different coverage and quality in early from-scratch pretraining of Wan 2.1 and V-JEPA 2.1. Across both learning objectives, the broader-coverage recipe achieves the highest downstream benchmark scores, while loss-based evaluation favors different recipes. This study demonstrates how VidaForge connects data-recipe choices to downstream model performance. We further release VIDAFORGE-3M, containing 3.14 million scene level clips totaling 6,475 hours, with fine-grained annotations and curation signals for video data-recipe research.
Sep 3, 2026

DART: Depth-as-Target Pretraining for Surgical Vision Foundation Models

Vision foundation models (VFMs) are valuable in data-scarce domains such as surgery, where a single pretrained backbone can provide rich representations for many downstream tasks. Yet the dominant self-supervised pretraining paradigm uses only RGB images, leaving readily available complementary signals, such as depth maps, unused. This is a particular missed opportunity in surgery, where natural-image VFMs transfer poorly while the scene geometry is rich and informative. With strong off-the-shelf models now able to produce pseudo-labeled dense depth for any image corpus, we hypothesize that such signals can be folded into pretraining to learn better representations. We present DART, an RGB-D pretraining recipe that builds on DINOv2 with a simple modification: a pixel-space depth reconstruction objective applied to masked iBOT patches, supervised by pseudo-labeled depth. Depth is used only during pretraining, so fine-tuning and inference remain RGB-only. We find that this pixel-level reconstruction head improves representation quality rather than disrupting it. We further show that depth, which encodes scene geometry, is more effective as a target than alternative dense signals such as Canny edges, confirming that the gains stem from depth rather than added supervision alone. Across eight surgical benchmarks spanning segmentation, depth estimation, and image-level recognition, DART outperforms both natural-image and in-domain baselines, including a vanilla DINOv2 trained on identical data, improving dense prediction while also strengthening image-level understanding. More broadly, DART shows that freely available geometric pseudo-labels can strengthen foundation model pretraining without extra labels or added inference cost, pointing toward stronger backbones for surgery.
Sep 3, 2026

What Do Scan-Derived Class Prototypes Add? Disentangling Supervision, Prototype Content and Query Protocol in Recognition over Frozen Foundation Features

A scan supplies labeled images and a geometric reference. We separate their contributions in a recognizer whose scan-derived prototype matrix acts as a supervised head's fixed output layer. On T-LESS, HOPE and 18 self-collected industrial parts, we test real, random and exactly permuted prototypes, matched geometry-free classifiers, stronger appearance rules and paired background protocols. Across DINOv2-giant and MetaCLIP-H with real-background queries, the largest fused-accuracy advantage of the real prototypes over either control is one percentage point; larger differences favor controls, by up to 2.8 points in arm means. On HOPE with DINOv2-giant the head alone is 2.8 points above exact permutations (95% interval: 0.8-4.7); this advantage does not reach fusion and is not observed on MetaCLIP-H. On DINOv2-giant, matched logistic regression comes within 0.5 points of fusion on T-LESS and exceeds it on HOPE and the self-collected parts. Against white cutouts, real HOPE query backgrounds lower image-prototype accuracy by 43 points on DINOv2-giant and 13 on MetaCLIP-H. The audit separates prototype content, label supervision and query protocol.
Sep 1, 2026

What, Where, and How: Probing Spatiotemporal Representations in Video Foundation Models

Self-supervised video foundation models learn rich spatiotemporal representations, yet it remains unclear what visual concepts these representations encode, where they emerge across transformer layers, and how they are geometrically organized. In this work, we tackle these three questions through a systematic layer-wise analysis of V-JEPA 2 and VideoMAE-v2. We leverage lightweight probes trained to discover three temporally grounded properties: (i) camera motion understanding, (ii) intuitive physics, and (iii) anomaly detection. Both models encode camera motion, with best results (>90>90 ROC AUC) emerging at 60-70% of network depth, and achieve moderate anomaly detection performance (>60>60 ROC AUC), but remain near chance on intuitive-physics tasks, suggesting a limited encoding of deeper physical reasoning. Beyond classification, we find that temporal features from individual videos form smooth low-dimensional trajectories in representation space, suggesting that camera motion is not only linearly decodable but also geometrically organized. Based on these results, we apply geometry-aware spline-based steering in the model's latent representations to interpolate camera motion, yielding steered videos with smoother trajectories and more coherent temporal progression than linear interpolation.
Sep 1, 2026

Do Satellites See Commuters? A Critical Benchmark of Vision Foundation Models

Satellite foundation models offer a globally available alternative to census data for commuting origin-destination (OD) generation, yet no study has systematically compared encoder paradigms within a single downstream pipeline. We ablate four satellite vision encoders: language-supervised (RemoteCLIP), self-supervised (DINOv3), and geographically grounded (SatCLIP, AlphaEarth) within an identical WeDAN graph diffusion framework across 1,925 US counties, 325 UK districts, and 14 global cities under five random seeds. Three main findings emerge. First, language-supervised features achieve the strongest in-distribution performance (RemoteCLIP CPC 0.602), while geographically grounded encoders transfer more reliably zero-shot: AlphaEarth improves CPC by 33% over RemoteCLIP on UK districts. Second, pretraining corpus scale alone is insufficient: DINOv3, trained on a substantially larger satellite corpus, underperforms RemoteCLIP by 0.091 CPC in-distribution and collapses to CPC 0.022 globally. Third, no encoder transfers usefully to global cities (best CPC 0.122 for RemoteCLIP, 0.022 for DINOv3), confirming cross-continental OD generation remains an open problem. We additionally clarify the semantics of the census noise parameter ηη, whose ordering reverses under cross-continental evaluation, a distinction critical to correctly interpreting prior results. Our code is available at: https://github.com/cruiseresearchgroup/satellites-see-commuters.
Sep 1, 2026

Restrict, Don't Retrain: Inference-Time VLM Guidance for Zero-Shot Aerial Segmentation

Global welfare often depends on the correct interpretation of aerial and satellite imagery. Acting on such imagery (mapping flooded ground, crop extent, or damaged infrastructure) demands pixel-level segmentation to ensure perfect class localization. Pretrained general foundation models, when applied directly, often miss important features and cannot always find all the classes belonging to a given scene, overlooking smaller objects that matter most. We use a single consumer-grade GPU running a vision-language model (VLM) to supply this missing guidance, improving segmentation while producing structured, auditable evidence that drives the result and can be inspected on its own. We fuse three approaches: the frozen foundation model that labels every pixel, and two queries to a VLM, one to choose the classes that matter, and one to locate the small objects the base model misses. Evaluating across four aerial datasets, we see consistent gains at each stage where the base model is competent.
Aug 30, 2026

Evaluating 2D and 3D-Aware Vision Foundation Models for Vehicle Attribute Recognition

Vehicle attribute recognition is an important task in intelligent transportation systems, particularly when Automatic License Plate Recognition (ALPR) is unavailable or unreliable. Although vision foundation models have shown strong transferability across domains, their effectiveness for fine-grained vehicle classification remains underexplored. Moreover, given the inherently three-dimensional structure of vehicles, it is unclear whether emerging 3D-aware foundation models offer advantages over standard 2D architectures. This paper presents an empirical benchmark of 14 state-of-the-art 2D and 3D-aware vision foundation models. Using the challenging real-world UFPR-VeSV dataset, we evaluate these models as frozen feature extractors via linear probing for vehicle type, make, and model recognition. We further stress-test the best-performing models under few-shot learning and Out-of-Distribution (OOD) domain shifts. Our results show that standard 2D self-supervised models, particularly DINOv3, substantially outperform 3D-aware models in fine-grained tasks, achieving over 93% Macro-Accuracy for make and model recognition. However, the 3D-aware Depth Anything v2 exhibits stronger invariance to viewing angles in vehicle type classification. These findings motivate hybrid approaches that combine 2D and 3D priors for robust vehicle recognition. Our code is publicly available at https://github.com/UFPR-IPASPPR/3D-Vision-Benchmark/.
Aug 26, 2026

Visual General Intelligence: A White Paper

This paper reconsiders intelligence from a vision-centered perspective and examines whether intelligence emerging from visual experience and learning may provide a pathway toward AGI. In the language domain, beginning with the introduction of the Transformer architecture, the GPT series has demonstrated transfer to unseen tasks through autoregressive language modeling on web-scale text combined with aggressive scaling. This raises a natural question, namely, what capabilities and forms of intelligence can emerge from visual modalities such as images, videos, and geometry? In this paper, we discuss whether visual intelligence can serve as a pathway toward AGI, referred to in this paper as visual general intelligence (VGI), by bringing together contributors from diverse standpoints and affiliations. Our aim is not to offer a single definition of visual intelligence, but to clarify the principles that computer vision should pursue in the AGI era, the visual input modalities, the benchmarks, the learning paradigms, and the relationship between vision, when taken as the core, and other modalities such as language.
Aug 21, 2026

WildFin: An In-the-Wild Dataset for Fish Behavioral Recognition

Recent advances in field technology have led to a massive influx of in-the-wild video data for ecological science. The primary bottleneck in leveraging this data is the high cost of expert annotation. While computer vision offers a potential solution, current models frequently fail when deployed in complex marine environments. To characterize these failures, we introduce WildFin, a novel benchmark for fish behavior recognition collected and annotated by ecologists. WildFin spans two critical real-world paradigms: stationary cameras monitoring groups of fish and dynamic divers following individual subjects. The dataset represents a massive curation effort, involving 1,350 hours of fieldwork and 600 hours of expert annotation to produce 9 hours of behavioral data with over 2 million frame-by-frame labels. We benchmark modern vision foundation models and quantify tradeoffs between static and spatiotemporal architectures, revealing the substantial gap that remains between current model capabilities and the demands of real-world underwater behavioral analysis. Project website: https://team-wildfin.github.io/.
Aug 21, 2026

Semantically Compatible Knowledge Distillation for Cross-Domain Object Detection with Vision Foundation Models

Vision foundation models (VFMs) offer strong generalization capabilities for domain-adaptive object detection (DAOD). However, existing VFM-based methods overlook the spatial-scale discrepancy between teacher and student feature maps, resulting in semantic incompatibility that weakens both feature alignment and pseudo-label learning. Moreover, domain shift can cause source-trained VFM teachers to miss target-domain objects, limiting the quality of their pseudo-labels. To address these issues, we propose the Semantic Localization-Enhanced Teacher (SLE-T), a semantically compatible knowledge-distillation framework built around a lightweight SLE Adapter for DINOv2. SLE Adapter injects pretrained local-texture priors into DINOv2 to improve cross-domain recognition and reformulates its features into dense representations that are spatially and semantically compatible with the student detector. SLE-T transfers the resulting teacher knowledge through either pseudo-label learning or feature alignment. We instantiate SLE-T with DINOv2-B and DINOv2-L (the ViT-B and ViT-L variants) and compare them with the larger DINOv2-G teacher. Extensive experiments on three DAOD benchmarks demonstrate that our method achieves state-of-the-art performance, and ablation studies confirm the importance of teacher-student semantic compatibility. Notably, SLE-T with DINOv2-B produces competitive or superior pseudo-labels using approximately one-quarter of the training time of DINOv2-G and substantially less GPU memory, demonstrating efficient VFM knowledge transfer under limited computational resources.
Aug 20, 2026

Robust Cross-Modal Foundation Model Perception for Underwater Robots under Degraded Visual Conditions

Reliable underwater robotic perception remains difficult because optical imagery degrades under turbidity, wavelength-dependent attenuation, low illumination, scattering, and blur. Although sonar provides complementary information that is less affected by optical visibility, prior visual-sonar research has largely focused on feature alignment and nominal detection performance. We investigate cross-modal robustness as visual reliability deteriorates and assess whether pretrained visual foundation-model representations can be complemented by sonar under severe degradation. We use frozen DINOv2 as the visual encoder and construct a controlled five-level benchmark ranging from clean to extreme visual conditions. We compare conventional visual detection, frozen foundation-model representations, sonar context, fixed multimodal fusion, clean-trained adaptive gating, and degradation-aware gated fusion. Our method trains the fusion mechanism across the full range of degradation while keeping the visual and sonar encoders frozen, allowing modality contributions to adapt without fine-tuning the pretrained backbone. Under extreme combined degradation, the DINOv2 baseline achieves 0.4610 balanced accuracy, while degradation-aware visual-sonar fusion reaches 0.6152, a 33.5% relative improvement. The learned sonar contribution increases from 14.2% under clean conditions to 41.3% under extreme degradation, demonstrating adaptive redistribution of cross-modal reliance. Fusion provides the largest gains under severe turbidity and blur, whereas color attenuation alone yields little additional benefit. These results show that foundation-model representations remain valuable but insufficient under severe information loss, and that explicitly adapting fusion to modality reliability can improve robust underwater multimodal perception.
Aug 13, 2026

A Controlled Study of Self-Supervised Image and Video Pretraining under Limited Resources

Visual foundation models are a cornerstone of image and video understanding but typically require large amounts of data and computation. The current scale required for pretraining visual foundation models may be unsustainable or unnecessary, and significant benefits arise when effective models can be obtained with fewer resources. To better understand how self-supervised learning (SSL) objectives behave under resource constraints, we conduct a controlled study of image and video SSL objectives under matched data, architecture, and compute budgets. We compare contrastive, reconstruction, feature-prediction, and diffusion objectives and evaluate both standalone and jointly trained image-video SSL formulations across a diverse set of image and video understanding tasks. Our results show that DINOv2-style pretraining consistently provides the strongest overall performance under limited resources. Furthermore, combining DINOv2 with video SSL objectives such as VideoMAE substantially improves image classification and segmentation performance, but degrades video tracking and camera-pose estimation performance, revealing an important tradeoff between semantic and geometric representation learning. These findings suggest that combining image and video SSL objectives can be beneficial in resource-limited settings, while highlighting the need for improved methods that better balance semantic, temporal, and geometric supervision.
Aug 12, 2026

Understanding Why Foundation Models Work for Diffusion-Generated Image Detection

Vision foundation models have recently emerged as powerful feature extractors for detecting AI-generated images, achieving strong generalization across generators and robustness to common image degradations. However, the reason behind their effectiveness is poorly understood. In this work, we investigate what cues are exploited by foundation-model-based detectors to distinguish real images from diffusion-generated ones. To this end, we design an ad hoc analysis protocol based on DDIM inversion. Given a real image we generate a sequence of synthetic copies by changing the depth of DDIM inversion. Even though most copies are semantically identical to the real reference, the detector score varies significantly across them due to subtle traces introduced by the diffusion synthesis, showing that its decision is not primarily driven by semantic failures. Through a frequency-swapping analysis, we further reveal that the discriminative cues exploited by the detectors are mainly localized in the low-to-mid frequency range, rather than only in the high-frequency range, as is the case for artifacts commonly associated with generative models. Finally, a latent-space analysis shows that regenerated images exhibit reduced variance and effective dimensionality, indicating that diffusion models do not fully reproduce the variability of real data. Overall, our results suggest that foundation-model-based detectors succeed by capturing non-semantic low-to-mid frequency distributional discrepancies between real and diffusion-generated images. These findings provide new insight into the robustness and generalization of such detectors and suggest directions for more interpretable forensic methods.
Aug 10, 2026

P3CA: Encoder-Agnostic Interpretation of Vision Foundation Model Embeddings via Spatial Probing

Vision foundation models are increasingly used as reusable encoders in medical image computing, yet their high-dimensional spatial embeddings are difficult to inspect beyond downstream task performance or global dimensionality reduction. We propose position-prompted PCA (P3CA), an encoder-agnostic method for local probing of channel-rich spatial tensors. Given a user-selected spatial prompt, P3CA estimates the feature normalization and dominant covariance directions within that region, then applies the resulting projection to the full tensor to visualize where locally informative directions are expressed. This produces a region-conditioned representation lens without modifying the encoder, retraining, or requiring task-specific labels. We implement P3CA in EmbedVision, an interactive 3D Slicer-based workflow, and evaluate it across natural images, colorectal pathology foundation-model embeddings, and spatial transcriptomic tensors. Across these settings, prompted projections reveal local structure suppressed by global PCA, improve prompt-matched pathology discrimination from frozen three-dimensional projections, and support comparison between learned and measured spatial representations.
Aug 10, 2026

LoRA-based Adaptation Alone Is Not Enough: Understanding the Limits of Foundation Models for Face Presentation Attack Detection

Face presentation attack detection (PAD) aims to reliably detect a wide range of presentation attacks. While PAD methods achieve strong performance within individual datasets, their performance degrades under cross-dataset evaluation. Variations in sensors or lighting conditions can reduce the effectiveness of detectors from near-perfect to nearly random. Foundation models (FMs) have emerged as a promising alternative because typical PAD datasets, such as the MCIO benchmarks (MSU-MFSD, CASIA-FASD, Replay-Attack, and OULU-NPU), are small relative to the scale used for web-based pretraining. However, existing PAD systems primarily focus on CLIP-based foundation models, while overlooking other FMs with different architectures and training procedures. This study addresses this question by systematically evaluating 32 FMs. Zero-shot prompting achieves performance near chance across model families and scales. The vision encoders, when low-rankadapted (LoRA) with fewer than 1% trainable weights, achieve below 2% intra-dataset ACER in most cases, while cross-dataset ACER is substantially higher. LoRA primarily refines the decision boundary within a dataset, suggesting that pretrained representations and the adaptation dataset play a larger role in cross-dataset generalization than the evaluated lightweight adaptation strategy.
Aug 10, 2026

Foundation Models are Implicit Deepfake Detectors

Pretrained self-supervised representations have emerged as a core component of current deepfake detection methods, yet it remains unclear which of their properties make real and fake media distinguishable. In this work, we uncover a surprisingly consistent phenomenon: across multiple pretrained models, datasets, and both image and video domains, fake samples systematically produce lower-magnitude representations than their real counterparts. Motivated by this finding, we formulate deepfake detection as an anomaly detection problem and show that simple statistics of feature magnitude achieve competitive performance with far more sophisticated deepfake detection methods. We further investigate the origin of this effect and demonstrate that reduced feature magnitude is primarily associated with semantic shifts introduced by fake content, while low-level generative fingerprints play a comparatively smaller role. Finally, we show that this discriminative signal strengthens as the size of the underlying foundation model grows, suggesting that advances in representation learning naturally translate into stronger zero-shot deepfake detectors.
Aug 9, 2026

Damage Classification for 3D Point Cloud Data via 3D Data Analysis and Vision Foundation Model-based 2D Projections

Fine-grained damage classification of 3D point cloud data (PCD) remains a persistent challenge, constrained by high computational demands and limited labeled data. This study examines two methods: 3D PCD-based damage assessment (3PDA) algorithm and 2D projection damage assessment (2PDA) In our 3PDA analysis algorithm, TDA is used to derive compact representations of 3D PCD segmented by pointNet, which are then integrated with anomaly detection algorithms to quantify structural degradation. We show that TDA effectively compresses geometric structure from VFM-segmented components into discriminative feature vectors and that anomaly detection models can reliably distinguish components with varying damage severity using only 3D PCD inputs. In the 2D projection analysis algorithm, we leverage large VFMs for granular damage detection by projecting 3D PCD into 2D views. These projections allow VFM based models to achieve competitive classification performance while requiring only a fraction of the computational cost associated with full 3D data processing. Our results demonstrate that 2D VFM pipelines in 2PDA can perform strongly on fine-grained damage classification tasks, highlighting their viability as lightweight, resource-efficient alternatives to traditional 3PDA architectures. Comparative evaluation shows that the 3PDA attains higher accuracy but only for a narrow subset of object geometries and at substantially higher computational cost due to its reliance on TDA and the scarcity of high-fidelity 3D datasets. In contrast, the 2PDA algorithm yields slightly lower accuracy but offers an order of magnitude reduction in time complexity and generalizes across a far broader range of object categories.
Aug 5, 2026

Invisible Shortcuts: Why Vision Encoders Know Your Camera

Deep vision models exploit shortcuts, relying on cues that correlate with supervision signals. Prior work has focused on visible biases, such as object-background or texture correlations. We identify a different source of shortcut learning: invisible metadata traces embedded at the pixel level, for metadata such as image processing and photo acquisition. We hypothesize that large-scale semantic supervision, whether through categorical labels (ImageNet) or billion-scale captions (LAION), naturally induces metadata-semantics correlations during pretraining, leading models to convert low-level signals into predictive features. By introducing controlled metadata-semantics correlations, we show that stronger ones produce systematically higher sensitivity to metadata traces and larger performance degradation under metadata distribution shifts. We further explore mitigation strategies applied during and after pretraining that reduce sensitivity not only to targeted metadata but also to unseen ones, without sacrificing performance on downstream tasks. Metadata sensitivity also has a positive side: it partly explains the strong generated-image detection ability of some encoders, while its mitigation can improve out-of-distribution generalization. Code: https://github.com/ryan-caesar-ramos/visual-encoder-traces
Aug 1, 2026

Foveated Probes Recover Localized Binding Information in Vision Foundation Models

Frozen vision foundation models are commonly evaluated through a single global image embedding, but this interface can conflate missing information with information lost at readout time. We study this distinction by keeping a pretrained vision encoder frozen and varying only the readout applied to its final patch tokens. We compare standard global readouts against a lightweight foveated readout, which attention-pools patch tokens using a learned or question-conditioned query, and against an oracle readout with access to the annotated target region. We evaluate these interfaces on three localized binding problems: a controlled synthetic color--shape binding task under clutter, a color-free crowded shape-detection variant, and a GQA-derived natural-image task where paired questions ask for the colors of different same-category objects in the same image. Global readouts perform near perfectly when the synthetic target appears alone, but collapse under clutter and counterfactual target edits, whereas the foveated readout recovers most of the oracle-accessible signal. On the GQA-derived task, question-independent global image vectors improve only modestly over question-only priors, while question-conditioned foveation substantially improves paired localized color accuracy. A counterfactual nuisance-to-signal ratio explains the synthetic failures: global pooling dilutes localized label-changing evidence while exposing the probe to nuisance variation from irrelevant objects. These results indicate that apparent spatial blindness in frozen vision models can arise from the global embedding interface rather than from an absence of spatial information in the frozen patch tokens.
Aug 1, 2026

Generated Images Are Easier to Forget: A Machine Unlearning Perspective for Synthetic Image Detection

Robust detection of generated images is critical to counter the misuse of generative models. Existing methods primarily depend on learning from human-annotated training datasets, limiting their generalization to unseen distributions. In contrast, large-scale vision models (LVMs) pre-trained on web-scale datasets exhibit exceptional generalization power through exposure to diverse distributions, offering a transformative paradigm for this task. However, our experimental results reveal that LVMs pre-trained on natural-image-dominated data can effectively capture the features of both natural and generated images, yielding comparably low losses and thus limited discriminative capacity between them. This prompts a key question: When and how do LVMs exhibit different behaviors when capturing features of natural and generated images? This investigation reveals an insight: during unlearning, LVMs exhibit disparate forgetting dynamics with feature degradation for generated images escalating faster than natural ones. Inspired by the disparate dynamics, we introduce two detection methods: 1) data-free detection, which prunes model parameters to induce unlearning without data access, and 2) data-driven detection, which optimizes LVMs to unlearn knowledge tied to generated images. Extensive experiments conducted on various benchmarks demonstrate that our unlearning-based approach outperforms conventional detection methods. By recasting the detection task as a problem of machine unlearning, our work establishes a new paradigm for generated image detection.
Jul 29, 2026

Foundation Models for Face Presentation Attack Detection: A Unified Linear-Probing Benchmark

Face presentation attack detection (PAD) remains challenging under cross-dataset evaluation, where domain shift degrades models trained on a single dataset. The scarcity of large-scale labeled data motivates adapting pretrained vision models rather than training task-specific architectures from scratch, raising a fundamental question: do general-purpose vision foundation models encode PAD-relevant information accessible with minimal task-specific training? To investigate, we systematically evaluate 24 frozen encoders, including self-supervised vision transformers, vision-language encoders, and supervised CNNs, using a unified linear-probing protocol on the MCIO benchmark (MSU-MFSD, CASIA-FASD, Replay-Attack, OULU-NPU). The backbone remains fixed, and only a lightweight linear head is trained to isolate the PAD information already present in the pretrained representation. Results show that frozen foundation-model representations can support strong intra-dataset PAD performance with only a linear classifier, but this performance does not reliably transfer across datasets. Model scale is beneficial within several families, although the effect is not monotonic and is strongly mediated by architecture and pretraining. InternViT-6B achieves the lowest mean intra-dataset error, whereas CLIP ViT-B/32 offers the most favorable cross-dataset transfer-compute trade-off among the evaluated probes. These findings suggest that while pretrained representations contain PAD-relevant information, explicit adaptation remains necessary to address domain shift.