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.

Nov 18, 2026

Unified Law for Visual Tasks (The Four Laws of Computer Vision)

In this paper, we propose a Unified Law for Visual Tasks, which could be further divided into Four Laws of Computer Vision, with the goal of enabling a single framework or method to uniformly train on and perform nearly all visual tasks. Our starting point is an inspiration and hypothesis: if the world information is encoded into image, which we call World Snapshots, then image processing is world processing, and image generation is World Snapshot generation. Furthermore, if all visual and visualization tasks can be reformulated as generation tasks, then a unified method, World Snapshot Models (WSMs), can perform the vast majority of visual tasks, much as large language models have largely unified natural language processing. This paper validates these laws across more than 1,000 trained and evaluated visual tasks (defined by distinct input-output mappings) and presents a unified training framework for future visual-task research and applications.
Nov 18, 2026

ALIVE: Bring Foundation Features to Life with Events

Visual foundation models provide powerful representations for diverse tasks, but their features become stale as visual content changes between image observations. Refreshing them requires new frames and repeated runs of a computationally expensive image encoder. We introduce ALIVE, an event-based method that incrementally updates native foundation model features between image observations. Given an anchor image, ALIVE uses subsequent events to transport and correct existing multi-layer features extracted by a frozen DINOv3 encoder. Its lightweight updater is trained end-to-end using only synthetic data, without downstream task supervision, and transfers directly to real-world event streams without any adaptation. Across three real-world datasets and five task-dataset settings spanning object detection, semantic segmentation, and depth estimation, ALIVE outperforms all evaluated event-based methods in every metric. On DSEC, when the anchor frame is 200,ms old, ALIVE improves over anchor-image predictions by 30.9,mAP in detection, 11.6,mIoU in segmentation, and 15.9% in depth AbsRel. Compared with image re-encoding, ALIVE reduces FLOPs by 82.4% and latency by a factor of 2.5.
Nov 17, 2026

GameRealm: Grounding Visual Learning in Interactive Games

Visual foundation models are central to physical intelligence, yet their development remains constrained by fragmented training datasets, inconsistent annotation quality, and limited multimodal supervision. To address these challenges, we introduce GameRealm, a large-scale, richly annotated dataset that grounds visual observations in synchronized, game-native supervision. GameRealm encompasses over 120 million RGB frames and 1,000 hours of gameplay footage across all views. Beyond its unprecedented scale, GameRealm provides rich multimodal supervision, including metric depth, camera parameters, synchronized audio, control inputs, and 3D point trajectories. Its diversity further extends to flexible view configurations, spanning single-view, stereo, multi-view, and panoramic recordings. To enable scalable acquisition across diverse game environments, we develop a unified data collection pipeline that combines human-operated interaction with agentic exploration. Extensive experiments across diverse visual learning tasks demonstrate that GameRealm consistently strengthens visual foundation models in both visual generation and geometric learning. These results establish GameRealm as a versatile and valuable data resource, unlocking new opportunities for advancing visual intelligence.
Nov 13, 2026

Beyond Endpoints: Bézier Paths for Vision Foundation Model Adaptation

Frozen vision foundation models can be adapted by learning a few prompt tokens shared across all images. When prompts are updated as data shift or tasks change, the distribution of image features moves from one endpoint to another. Prior theory asks which endpoints prompts can reach; we ask whether they can travel between them along the shortest path. Because shared prompts compete with each image's own content for attention, they often cannot. We quantify this restriction as the extra transport cost that shared attention imposes over the shortest path. In an exactly solvable attention model, the reachable features form a cone; unrolling it into a plane gives the exact minimum action for every pair of endpoints and explicit optimal prompts on short arcs. The analysis reveals a sharp threshold in content normalization, which measures how much attention each image's own tokens absorb: for same-sign endpoints, at most two distinct levels allow prompts to follow the shortest path, whereas three or more force a strictly positive extra cost unless the endpoints differ only by rescaling. This gap persists with more prompts, small content perturbations, and aligned multihead outputs. On ViT-B/16 with VTAB-1k, a quadratic Bézier prompt path that minimizes feature action reduces held-out transport action relative to linear and cross-entropy-trained paths between the same endpoints and improves intermediate and online accuracy under unlabeled distribution shift, while cross-entropy-trained paths retain higher minimum accuracy when labeled calibration data are available. Reachable endpoints do not guarantee a reachable shortest path: how prompts move matters, not only where they end. Project page: https://xixiaouab.github.io/Beyond-Endpoints/.
Nov 10, 2026

Adapting a Vision Foundation Model for Fracture Characterization in Acoustic Image Logs

Natural fractures (NFs) and drilling-induced tensile fractures (DITFs) in acoustic image logs provide insights into rock mass structure and in situ stress, but manual characterization is time-consuming. Deep learning can automate this task, although expert-labeled data are scarce. We investigate whether adapting a pretrained vision foundation model (VFM) outperforms task-specific models trained from scratch with limited annotations. Using 80 annotated acoustic image samples, we fine-tuned DINOv3, pretrained on satellite imagery, and evaluated it on four boreholes excluded from fine-tuning. The adapted model estimated NF and DITF parameters comparably to human interpretation. Relative to U-Net, YOLOv8, and Swin Transformer models trained from scratch, it yielded fewer false positives and negatives and more accurate fracture counts and orientations. However, its performance advantage generally narrowed as training sample size decreased. We also assessed decoder architecture, fine-tuning strategy, and image enhancement. A progressive upsampling decoder with full fine-tuning performed best among the DINOv3 configurations tested, while fracture-focused image enhancement did not consistently improve accuracy. These results demonstrate the potential of VFM adaptation for fracture characterization in acoustic image logs with limited labeled data and suggest a transferable workflow for other annotation-scarce geoscientific imaging tasks.
Nov 8, 2026

Probing Transformer Representations Layer by Layer, Module by Module

Which representation should be probed in a transformer-based foundation model? Recent work shows that the standard practice of probing the final layer is often suboptimal, as intermediate layers can provide better features. Yet, representations are still largely studied along depth alone, treating each transformer block as an atomic unit. In this work, we analyze transformer representations layer by layer, module by module. Through more than 144,000 linear probing evaluations of vision foundation models (VFMs) and large language models (LLMs), we show that layer and module choices should be considered jointly: the best layer-module pair yields substantially larger gains over the final-layer residual stream than the best layer with the residual stream fixed. Moreover, module performance can vary substantially across layers, showing that useful representations cannot be characterized by depth alone. Building on these findings, we evaluate simple fixed probing strategies: the final-layer feedforward activation (Act@100) for vision and the second normalization layer at 60% depth (LN2@60) for language. These strategies improve over the probing baseline in a majority of evaluations and recover, on average, 42% and 86% of the gains from probing the best layer-module pair for vision and language, respectively. For VFMs, these gains increase with the transfer difficulty between pretraining and downstream data.
Nov 7, 2026

PoseBench: Probing Pose Awareness in Visual Foundation Models

Pose awareness characterizes the extent to which a learned feature space aligns with the underlying object-pose manifold, such that observations with similar poses exhibit greater feature similarity than those with increasingly different poses. While such structure is desirable for pose-sensitive 3D vision tasks, visual foundation models pre-trained on individual 2D images typically lack explicit 3D or pose supervision. This raises a fundamental question: Do 2D visual foundation models exhibit pose awareness after large-scale pretraining? To answer this question, we introduce PoseBench, a systematic benchmark to probe pose awareness in pretrained visual features. Our evaluation follows a progressive design: held-out rendered views probe intrinsic pose-aware structure under controlled conditions, while real-image queries evaluate how robustly this structure remains accessible under real-world variations. We further introduce pose-graph-based refinement as a complementary probe of compatibility between feature neighborhoods and the underlying pose manifold. Across 31 publicly available checkpoints from 10 pretraining families, PoseBench reveals that pose awareness is an implicit but non-uniform property of 2D visual foundation models, and consistent gains from pose-graph refinement suggest that feature spaces retain structure compatible with the underlying pose manifold. PoseBench therefore provides a systematic basis for understanding and selecting pretrained features for pose-sensitive 3D vision tasks.
Nov 5, 2026

InnerLogic-Surg: An Open Surgical Foundation Model via Continual Pretraining

Vision foundation models are the backbone of modern surgical image analysis, yet most surgical systems still rely on encoders pretrained on natural images or on specialist models trained from scratch on comparatively small datasets. Pretraining encoders on large amounts of surgery-specific data has the potential to generally enhance the performance of surgical image analysis methods simply by providing stronger backbones. Here we present InnerLogic-Surg, an open family of surgical vision encoders (21M--300M parameters) obtained by continually pretraining a transformer-based image encoder on millions of surgical images curated from 11,965 internet-sourced surgical videos spanning ~2,900 hours and more than fifteen procedure types. Our training recipe inherits DINOv3's pretraining stage to preserve representation ability after surgical adaptation, and additionally applies a learned feature-upsampling module for high-resolution dense prediction at low token cost. Across a benchmark covering workflow, instrument and anatomy recognition, segmentation, detection, and monocular depth, under the same architecture, the largest InnerLogic-Surg model achieves the best linear-probe performance in all 20 surgical tasks. We demonstrate that surgical continual pretraining yields transferable representations that generalize across procedures, anatomical regions, and task types, showing our surgical image encoders as a promising and scalable alternative to the procedure-specific representation learning approaches that have traditionally dominated surgical computer vision.
Nov 1, 2026

SpecUltra: A Unified Foundation Model for Heterogeneous Spectral Images

Spectral imagery, including hyperspectral image (HSI) and multispectral image (MSI), exhibits substantial heterogeneity in band numbers, wavelengths, spectral resolutions, hindering unified modeling and cross-sensor transfer. From the perspective of imaging physics, these heterogeneous observations can be regarded as discrete measurements of an underlying continuous spectrum acquired through different sampling processes. Motivated by this perspective, we propose SpecUltra, a unified foundation model for heterogeneous spectral images. Specifically, SpecUltra employs 3D convolutional tokenization and wavelength-aware 3D RoPE to encode relative wavelength and spatial relationships, enabling fine-grained spatial-spectral modeling within a shared coordinate system. We further construct MultiHyper-200K, a large-scale multi-sensor HSI dataset, and pretrain SpecUltra using auxiliary-token-assisted mask image modeling (MIM), where a level token indicates the radiometric product type and register tokens absorb sensor- and acquisition-dependent variations. To bridge the spectral-granularity gap between HSI and MSI, we introduce Spectral Kernel Transformation (SKT), which transfers pretrained knowledge in parameter space by constructing MSI embedding kernels from fine-grained HSI-pretrained kernel slices. This design preserves pretrained spectral priors while maintaining compatibility with the pretrained backbone. Experiments on HSI and MSI classification and segmentation benchmarks demonstrate state-of-the-art or competitive performance across disparate sensors and downstream tasks.
Oct 30, 2026

Evaluating SAM-based and Convolutional Architectures in Thin Retinal Vessel Segmentation: Model Capacity, Computational Cost and Foundation Pre-training

In retinal vessel segmentation, plain Vision Transformers (ViTs) are commonly expected to perform worse than convolutional neural networks (CNNs) on thin vessels, for reasons such as their coarse patch tokenization, their weak locality bias and their need for large training sets. Although this expectation is widely shared in the literature, how plain ViTs actually compare with CNNs on thin vessels has not been systematically quantified. We address this question by comparing a model with a plain ViT encoder (MedSAM), a hybrid Transformer (LiteMedSAM) and three CNNs of different sizes on three retinal datasets (DRIVE, HRF, FIVES) and, as a second-domain check, a microscopy dataset (VessMap). We measured the centerline recall and precision of thin vessels and compared the models while also considering their computational cost and how many epochs they needed to train. We found that the pre-trained MedSAM had a centerline recall of thin vessels that was not significantly different from that of the best CNN on all four datasets (-0.6 to +2.5 percentage points). A compact plain ViT obtained through architecture search but trained from scratch did not retain this performance. For pre-trained plain ViTs, our results do not support the expectation that they miss thin vessels, although similar thin-vessel performance can be obtained with much smaller models.
Oct 29, 2026

Probing Spatial State Representations in Visual Foundation Models

An observer sees only part of the surrounding world, yet must keep track of what lies beyond the current view. Spatial perception therefore involves more than recognizing individual images: it calls for a persistent spatial state that is updated under self-motion and used to guide action. Do visual foundation models provide representations that support such a state? We study this question through four levels of spatial capability---spatial structure, object persistence, predictive spatial update, and actionability---using frozen-feature probes on ten representations under streaming egocentric observations of static indoor environments. The resulting capability profiles vary across tasks, network layers, and prefix lengths: strong readout of the visible scene does not ensure accurate object persistence or future-feature retrieval. Matched controls separate readout accuracy from reliance on visual history and motion. Native future prediction retains recoverable spatial structure, but a spatial reader trained on real features does not transfer to predicted features unchanged. Finally, goal-dependent candidate selection does not establish reliable continuous action utility: predicted actions respond to goal changes, yet supplying the correct goal does not reliably improve progress. Across the four levels, object persistence and actionability build on camera-pose readout, whereas feature prediction builds on visible-scene readout. Together, these findings motivate evaluating spatial representations not only by what can be decoded from them, but by whether that information is maintained beyond observation, updated predictively, and used for goal-directed action.
Oct 27, 2026

Venus-DefakerOne2: A Vision-Forensics Foundation Model with Broader Coverage

Real-world forgery is inherently heterogeneous: it manifests not only as digital manipulation but also as physical spoofing, and is carried not only by images but also by videos. Despite substantial progress, a unified anti-forgery foundation model covering such a broad threat spectrum remains absent. In this paper, we present Venus-DefakerOne2, which substantially extends the forensic capability boundary along both the forgery-form and media-type dimensions. Trained on tens of millions of open-source and in-house samples, it incorporates two key designs. Tide-Style-Tuning alternates LoRA and full-parameter fine-tuning to balance computational efficiency, generalization, and hard-sample fitting while mitigating catastrophic forgetting. Entropy-based Test Adaptation determines the optimal prediction modality by analyzing prediction entropy. Venus-DefakerOne2 achieves new state-of-the-art performance across a comprehensive suite of benchmarks, including DeepFake faces (96.7%), AI-generated images (93.3%), tampered documents (87.6%), manipulated natural images (89.6%), physical spoofs (93.1%), and AI-generated videos (96.0%). It significantly outperforms both specialized expert models and its predecessor, DefakerOne, establishing itself as a powerful and versatile foundation model for binary-decision forensic tasks.
Oct 27, 2026

A note on data-efficient finetuning of vision foundation models

Adapting foundation models to a new task within or beyond the training domain is the de facto standard for building specialized machine learning models. Recently, Visual Prompting (VP) has emerged as a promising technique for efficient knowledge transfer. As new foundation model families~(like Mamba) get introduced and VP pipelines such as AutoVP reach greater maturity, we find a growing need for a systematic evaluation of current approaches. In this work, we assess the performance of the latest models, comparing them to earlier architectures and alternative fine-tuning methods, to better understand the progress, challenges and opportunities in the field of efficient fine-tuning under resource limitations. Towards this goal, this paper provides a concise empirical overview of the interactions among foundation model families (Attention-, Convolution-, and Mamba-based) and transfer paradigms: VP, Linear Probing (LP), Full Fine-tuning (FFT) and Low Rank Adaptation (LoRA). Our work builds up on previous findings by broadening the selection of evaluated models and techniques while tuning hyperparameters. In the interest of delivering practical guidelines for the user, we also explore application of prevalent regularization techniques to boost performance in the context of VP.
Oct 26, 2026

GIFT: Representation Geometry Matters for Single-Domain Generalized Object Detection

Single-Domain Generalized Object Detection (S-DGOD) aims to generalize from a single source domain to unseen domains under distribution shifts. Existing efforts mainly focus on simulating unseen domains or suppressing domain-specific factors. However, they largely overlook the intrinsic generalization capability embedded in Vision Foundation Models (VFMs). In this work, we show that unconstrained fine-tuning disrupts the geometric structure of pre-trained representations, which encodes domain-invariant semantics. Motivated by this, we propose Geometric-Invariance Fine-Tuning (GIFT), a geometric-aware fine-tuning method for VFMs. Specifically, we introduce a series of learnable Householder reflections to the pre-trained weights to encourage geometry-consistent adaptation under domain shift. In addition, we propose a lightweight channel-wise scaling to facilitate effective task-specific adaptation. Extensive experiments demonstrate that GIFT consistently outperforms state-of-the-art (SOTA) approaches on two S-DGOD benchmarks. Notably, by introducing fewer than 1% additional trainable parameters into the frozen backbone, the proposed method achieves improvements of +7.9% and +2.9% in mPC over the previous SOTA on the Cityscapes-C and Diverse Weather Dataset, respectively. This finding highlights that preserving representation geometry is crucial for robust cross-domain generalization, establishing GIFT as a principled and efficient solution to domain shift.
Oct 26, 2026

Zero-Shot Monocular Depth Estimation for Visual Perception in Lower-Limb Prostheses

Monocular depth estimation has advanced rapidly with the development of foundation models that demonstrate strong zero-shot generalisation across diverse visual domains. However, the applicability of these models to visual perception for lower-limb prosthetics remains largely unexplored. Reliable depth perception is particularly important for prosthesis users, who often experience significant challenges when navigating stairs, uneven terrain, and other complex environments. To address this gap, we introduce the Lower-Limb Prosthesis (LP) dataset, a benchmark designed to evaluate depth perception from an emulated knee-level prosthetic viewpoint. Using this dataset, we evaluate four state-of-the-art monocular depth estimation foundation models. Our results provide the first systematic benchmark of monocular depth estimation performance from a prosthetic viewpoint, while highlighting both the challenges and the potential of deploying foundation models in prosthetic perception systems. The code and dataset are available at https://github.com/Ishivijay/lp_dataset.
Oct 24, 2026

UniQuery: Unifying Vision Models with a Joint Query Head for Cross-Domain Pedestrian Attribute Recognition

Pedestrian Attribute Recognition (PAR) predicts semantic attributes of individuals, such as gender, age, clothing and accessories, and its descriptions support attribute-based person retrieval and person re-identification. However, accurately recognizing fine-grained attributes across domains remains challenging due to occlusion, appearance variations, and shifts in image distributions. Ensembles of independently trained foundation encoders combine their logits with fixed fusion weights, so every attribute uses the same mix and the encoders are never trained together. We propose UniQuery, which instead jointly fine-tunes DINOv3 and CLIP for attribute-specific fusion. It combines two query heads, one for each encoder, with a joint query head whose attribute queries attend to tokens from both encoders, and the prediction is the average of the logits of the three query heads. We evaluate UniQuery on the benchmark of the UPAR Challenge 2027 at the WACV 2027 Real-World Surveillance workshop, where models are trained on PA100k, PETA and Market-1501 with a shared set of attributes. In leave-one-dataset-out experiments, UniQuery outperforms the independent ensemble in Challenge_Avg on both held-out datasets. Ablations show that learning fusion over independently trained encoders does not improve upon logit averaging in the mean over the two held-out datasets, while freezing the encoders lowers the Challenge_Avg of UniQuery. On the hidden test set of the benchmark, UniQuery achieves a Challenge_Avg of 0.8001.
Oct 23, 2026

Compression of Vision Foundation Models for Retinal OCT via Growing-Neural-Gas Codebooks and Graph-Guided Masked Image Modeling

Vision foundation models for retinal imaging are large and computationally demanding, limiting point-of-care deployment. We ask how much of such a model's downstream utility survives compression into a smaller network. We propose a topology-aware discretization framework that distills a continuous foundation model into a discrete codebook, then uses it as the reconstruction target for masked image modeling (MIM). It combines gradient-free codebook learning via Growing Neural Gas, which yields an adjacency graph over the codes, with a graph-distance soft cross-entropy loss that replaces one-hot MIM targets with topologically graded supervision. With student pretraining on 383 public Duke iAMD optical coherence tomography (OCT) volumes, the pipeline distills an 86M-parameter teacher into a 22M-parameter student with at least 3.6x fewer FLOPs. The student performs close to, and in some settings above, its teacher across three classification and two segmentation benchmarks, with the magnitude and direction of the difference depending on the teacher, task, metric, and student capacity. On the external OLIVES benchmark, it surpasses its MIRAGE teacher on every metric (98.06% vs. 95.22% balanced accuracy). It works with two Vision Transformer (ViT-Base) teachers, retinal (MIRAGE) and natural-image (DINOv3); with the out-of-domain teacher, the student outperforms it on retinal-layer segmentation, indicating that the added value of retina-specific pretraining is task-dependent: largest for segmentation, while a general-domain teacher is competitive for classification. These results yield compact retinal models potentially suitable for deployment, transferring across OCT datasets, devices, and clinical tasks, including three datasets not used in the student's pretraining.
Oct 23, 2026

The Locality Trade-off in Masked Patch Self-Distillation

Masked patch self-distillation is a key component of modern vision foundation models, including DINOv2 and DINOv3, yet its implications for local visual information remain unclear. We show that its semantic benefits can coexist with a scale-dependent locality cost. Using two annotation-free diagnostics, we distinguish recoverable local appearance from spatial correspondence. In smaller models, adding iBOT substantially weakens patch correspondence while largely preserving appearance decodability. In larger models, correspondence improves, but pixel and texture evidence becomes less linearly accessible in deep features. Thus, scaling changes the form of the locality cost rather than uniformly eliminating it. Our analysis of Gram anchoring further shows that preserving patch relations does not guarantee simultaneous improvements in both properties. These findings motivate evaluating and designing visual representations around the complementary demands of semantic abstraction, appearance preservation, and spatial correspondence. Our code is available at https://github.com/winci-ai/jea.
Oct 19, 2026

Foundation Model Driven Optical Flow

We investigate how frozen pretrained visual and depth representations support dense optical flow when coupled with a supervised matching network. Building on TTSFlow, this paper studies Dense Prediction Transformer (DPT) feature decoding and one coarse-to-fine refinement stage. Our method fuses features from frozen DINOv2 and Depth Anything V2 models and estimates optical flow through a trainable transformer-based matching pipeline. After optical-flow training on FlyingChairs and FlyingThings3D, it achieves 2.37 endpoint error (EPE) on the Sintel training split, Final pass, compared with 2.52 for the FlowSeek (T) configuration reported here. Our RoCo-Spring 2026 entry, RoCo-29.VSAI, obtains a robustness-balanced score of 1.1038, Spring EPE of 1.411, and RobustSpring proxy EPE of 4.739, with a lower balanced score than seven of the twelve listed reference baselines. These results characterize the complete pipeline, which reuses frozen foundation backbones and performs one refinement stage. Code is available at: https://github.com/Visual-and-Spatial-AI-Lab/FFlow
Oct 16, 2026

A100: A Hybrid Vision Foundation Model for Agricultural Plant Image Analysis

The use of accurate vision models for crop monitoring, disease diagnosis, and plant phenotyping in agriculture is hindered by the lack of expertly annotated data, especially in developing countries. We propose A100, an agricultural vision foundation model that is pre-trained using self-supervised learning on 3 million unlabeled images from fields covering more than 100 crop species. A100 uses a hybrid CNN-Transformer backbone that fuses MBConv stages for local feature extraction with Transformer stages for global contextual representation. A100 uses pre-training with the integration of DINO self-distillation, iBOT masked patch reconstruction, and KoLeo uniformity regularization, resulting in features that are better than ImageNet-12K supervised representations for agricultural applications without any fine-tuning. By utilizing a frozen backbone, the accuracy for crop classification of A100 is 95.37%, which is +2.56% better than ImageNet-12K, and reduces the leaf count error by 10.4% on the GroMo25 phenotyping benchmark, proving that the same backbone can be used for classification and regression tasks. After full fine-tuning, the performance of A100 is 99.90% on 32-class crop identification and 99.02% on 44-class disease detection. On the difficult PlantDoc dataset, A100 outperforms all baselines, such as CvT-13 (64.89% against 63.56%). Remarkably, A100 obtains 98.01% accuracy on plant classification using just 1% of the labelled data (about 284 images per category), showing great label efficiency for use cases with limited labeling. Under an identical protocol, A100 also outperforms the publicly released agricultural foundation models FoMo4Wheat and LeafVision by more than 14 points at 1% labels. We plan to publicly release our pretrained model to allow agricultural professionals to adapt A100 for various plant analysis applications using limited annotations.
Oct 12, 2026

Vision Models Capture Complementary Representations in Children That Converge in Adults

How the relationship between image-text aligned and vision-only models changes across development remains unknown. We applied time-resolved representational similarity analysis to EEG data from 8-year-olds, 12-year-olds, and adults. CLIP showed significant model-brain alignment already in children, whereas vision-only models aligned more strongly later in development. Residual analyses revealed complementary structure, with CLIP retaining alignment beyond an architecture-matched vision-only model from age 8. Much of the model-brain alignment reflected semantic structure, with EEG representations of category and animacy strengthening toward adulthood, while alignment with low-level image properties remained weak across age. Together, these findings suggest that image-text aligned and vision-only models capture complementary aspects of developing visual representations and may provide useful constraints for AI systems that aim to learn and adapt over time.
Oct 12, 2026

Carve VFM: Carving Specialists out of Generalists, Automatic Task-Specific Compression of Vision Foundation Models

Vision foundation models (VFMs) provide transferable visual representations, but their inference cost limits deployment in many settings. Fine-tuning specializes a VFM without shrinking it; distillation shrinks it, but through expensive teacher-guided training. Structured pruning offers a more direct route by extracting a compact subnetwork from the VFM. However, as compression gets aggressive, existing methods fail because they treat subnetwork selection and recovery as separate axes. Across layers, capacity must follow task sensitivity rather than spread uniformly. Within a layer, a removal cannot be judged by importance alone; its impact also depends on how well the lost contribution can be recovered from what remains. Selection and recovery must therefore be decided jointly. We introduce Carve VFM, which resolves this interdependence. Given a VFM, a calibration set, and a parameter budget, it (i) searches for task-specific capacity allocations across layers; (ii) guides the search via an adaptive recovery-aware fitness criterion; and (iii) reconstructs removed behavior through iterative closed-form low-rank corrections merged into the surviving weights. The result is a dense encoder produced in minutes on a single GPU without backbone retraining. Across twelve benchmarks and four models from DINOv3 and MetaCLIP-2, Carve VFM achieves the highest reported mean accuracy in seven of eight backbone-budget configurations among the evaluated dense structured-pruning methods. Further fine-tuning surpasses the strongest evaluated edge model (88.1% vs. 85.8%), and carved models cut compute and memory while delivering 1.3-2.0x higher GPU throughput than the uncompressed one. These results show that the compact model you need is inside the VFM; it only has to be carved out.
Oct 9, 2026

EV-JEPA: Elastic V-JEPA for Any Device and Cross-Device Inference

Foundation models like V-JEPA are increasingly deployed across a heterogeneous edge-cloud continuum spanning smart glasses, phones, robots, and servers. Yet, they are typically trained as rigid, fixed-capacity networks bound to one device, supporting neither reconfigurable compute budgets nor collaborative execution. Existing approaches address only one axis at a time: slimmable and NAS-based methods provide width elasticity but assume single-device inference, whereas split-computing methods enable cross-device collaboration only for fixed partitions and static capacities. We introduce Elastic V-JEPA (EV-JEPA), a single model trained once and deployable everywhere along the continuum. Its width is reconfigurable, and it is natively partitionable across heterogeneous devices. At its core is Transition-Aware Triplet Training (T3). At each step we pick a narrow and a wide width, and train three paths: each width alone, plus a hybrid path that runs the early layers narrow and the rest wide. The hybrid switch point is where one device hands off to another, so the same weights learn both single-device and cross-device execution. Across various benchmarks, EV-JEPA falls within 1 accuracy point of independently trained baselines while improving over previous elastic training approaches by up to 11 accuracy points. Narrow-to-wide cross-device hybrid execution recovers to within 1.1 points of the highest-capacity standalone model and supports many-to-one, any-to-any, and variable-split configurations. Finally, we show that the elastic principle extends beyond the backbone to an elastic token compression mechanism that exposes controllable trade-offs among on-device computation, latency, and accuracy.
Oct 8, 2026

One Block, Multiple Depths: Recurrent Vision Transformers with Depth-Programmed Experts

In this work, we show that a single Transformer block, applied recurrently, can match the accuracy of a full-depth vision encoder at comparable inference FLOPs without intermediate feature distillation. reViT restores depth-specific transformations by representing the FFN at each recurrent depth as a convex combination of a small shared expert bank. A continuous normalized-depth coordinate programs this mixture, defining a resampleable trajectory through FFN parameter space. We evaluate this design in two regimes: supervised ImageNet-1k training and distillation from a DINOv2 teacher. Across both regimes, controlled adaptations identify weight-space merging as the strongest tested MoE family at a matching one-FFN budget, ahead of the token-dispatch and output-mixture alternatives. Trained from scratch, reViT-B/16 attains DeiT III accuracy with about 70% fewer stored parameters. An 8-experts model distilled using only the teacher's output features retains nearly all of its DINOv2 teacher's linear-probe accuracy and transfers across classification, segmentation, and depth prediction. Elastic-depth training allows one checkpoint (trained model) to operate at multiple tested depths by resampling the same normalized coordinate interval. For fixed-depth deployment, the recurrent block can be materialized as a conventional dense graph, removing online routing and merging without changing the one-FFN-per-depth compute but expanding deployment storage.
Oct 7, 2026

DisParQ: Self-Supervised Part Concepts for Interpretable Vision Foundation Models

Concept-based vision models represent images through an intermediate layer of human-inspectable concepts, so what a model relies on can be traced to those concepts. However, those models are often limited to fixed categories or depend on language to define their concepts. We introduce DisParQ (Discrete Parts with Quantized attributes), a method that learns spatially grounded, discrete concept representations from a powerful frozen vision-only self-supervised backbone. It requires no class labels and no language supervision. Each image patch is assigned to exactly one concept from a learnable prototype dictionary, and only a sparse subset of concepts may activate per image. To capture how each concept varies across images (e.g., the type of a "wheel"), we learn continuous residuals alongside the concepts and then quantize them into discrete attributes. A spatial decoder reconstructs the backbone's representation from the concepts and attributes alone, so successful reconstruction means that the discrete representation preserves the backbone's information. We evaluate DisParQ across seven datasets, from general recognition (ImageNet, PartImageNet, Places) to fine-grained benchmarks (CUB, Cars, Dogs, Flowers). We show that DisParQ closely matches its frozen DINOv2 teacher on ImageNet linear probing (83.2% top-1), achieves higher concept consistency than language-aligned models, remains competitive on fine-grained recognition, and enables cross-category part-based retrieval.
Oct 4, 2026

SPACE-CLIPv2: Decoding Local Geometry from Frozen CLIP for Monocular Depth Estimation

Vision-language foundation models such as CLIP provide strong semantic representations, but their patch tokens are not directly optimized for dense metric geometry. SPACE-CLIP showed that frozen CLIP features can support monocular depth estimation through layer-group feature fusion, yet it leaves open how neighboring CLIP tokens should be combined to recover fine local structure. We present SPACE-CLIPv2, a frozen-backbone depth decoder that aggregates fixed local neighborhoods in CLIP token space. At selected decoder stages, the model samples a fixed token stencil, predicts aggregation weights, and injects the resulting response through a gated residual update. A token-space high-pass branch further preserves shallow local contrast. On NYU Depth V2, SPACE-CLIPv2 improves over a matched SPACE-CLIP baseline, while five-seed experiments consistently favor fixed over learned-offset sampling. Zero-shot iBims-1 evaluation further improves boundary and planar-geometry measures. These results support constrained local token aggregation as a practical mechanism for decoding geometry from frozen CLIP representations.
Oct 2, 2026

Geometry-Aligned Semantic Matching for Cross-Modal Planar Image Registration

Cross-modal image matching establishes stable and accurate geometric correspondences across modalities for planar registration. Existing semantic representations provide cross-modal consistency, but semantic similarity does not necessarily imply geometric correspondence. Meanwhile, fine-grained CNN features provide accurate local details but lack global cross-modal semantic guidance for stable refinement. To address these issues, we propose CDPM, which first establishes geometrically consistent semantic representations and then preserves their dominant role in correspondence estimation during fine-grained localization. Specifically, we progressively adapt DINOv3 using geometrically consistent cross-modal patch pairs, enabling feature similarity to better reflect true cross-modal spatial correspondences. We then construct a DINO-Centric Feature Pyramid, where multi-scale DINO representations maintain stable cross-modal correspondences, while a lightweight CNN branch provides auxiliary structural details for precise local refinement. Extensive experiments on three cross-modal datasets demonstrate the superior performance of CDPM. On VIS-IR, compared with the dense matcher RoMa, CDPM improves AUC@3/5/10/20 by 7.36, 13.40, 13.75, and 10.42 percentage points, respectively, and reduces mACE from 5.83 to 2.78 pixels. It also outperforms RoMa v2 across all metrics while requiring 45.6% fewer FLOPs. The online demo and dataset are available, and the code will be released on our project page at https://warren-wzw.github.io/CDPM/.
Oct 2, 2026

OmniAct3D: Leveraging Foundation Geometry and Evidence-Grounded Reasoning for Panoramic 3D Detection

Accurate 3D detection is essential for mobile embodied agents, while Vision Foundation Models (VFMs) offer transferable visual and geometric priors. Yet existing VFM-based 3D detectors rely on narrow-view monocular images or discrete perspective views, limiting coherent surround perception; equirectangular projection (ERP) instead encodes a continuous 360 scene in a single image. Direct transfer remains difficult because ERP organizes geometry and visual information differently, making object-relevant cues hard to model, localize, and preserve. We propose OmniAct3D, a framework that adapts perspective-trained VFM detectors to ERP while preserving transferable VFM priors. To resolve geometric mismatch, the ERP-Ray Geometry Adapter (ERGA-Ray) models spherical viewing rays and periodic spatial structure. To localize evidence in scene-wide context, the Visual-Action Reasoning Chain (VARC) grounds each hypothesis in relevant panoramic evidence and converts it into a structured geometric action. To recover local cues lost under fixed token budgets, the Appearance-Guided Heading Expert (AGHE) re-encodes object regions at higher resolution for heading estimation. Experiments show that OmniAct3D improves over the previous best 3D detector by 2.96 NDS points on Spheriverse and over the unadapted VFM baseline by 24.87 mAP points on PanoMMOcc. With target-specific geometry adaptation, VARC retains 95--98% of the same-configuration mAP, indicating reusable object-level 3D reasoning across sensing configurations. The source code will be made publicly available at https://github.com/FeiT-FeiTeng/OmniAct3D.
Oct 1, 2026

GRAFT: Growing Agglomerative Foundation Models via Continual Teacher Distillation

Vision foundation models such as DINOv2, SigLIP2, and MASt3R develop complementary capabilities from different pretraining objectives, yet their knowledge remains distributed across separate, specialized models. Multi-teacher knowledge distillation offers a path toward consolidating these capabilities into a single agglomerative backbone, but existing approaches assume a fixed set of teachers, and incorporating a new teacher requires repeating expensive joint distillation over the entire teacher set. We introduce GRAFT, a continual multi-teacher distillation framework that enables a unified backbone to progressively acquire capabilities from an open-ended sequence of foundation models. When a new teacher arrives, GRAFT treats the previously distilled model as a teacher for preserving learned capabilities, while the current student jointly learns from both the previous model and the incoming teacher. Furthermore, to reconcile the incompatible representation geometries of heterogeneous teachers, we introduce Teacher Specific Readout Tokens, which grant each teacher an independent read-out of the shared encoder, together with Geometry Agnostic Relational Loss that aligns a vision-language teacher by matching image-text similarity structures rather than raw feature values. We provide GRAFT model, which is a single, continually extensible backbone that unifies five domains, including image understanding, 2D dense prediction, 3D human pose estimation, 3D vision, and vision-language, delivering strong performance across all of them while acquiring each new capability at the cost of a single distillation rather than a full re-distillation.
Sep 29, 2026

PolyOCR-Venus: Unified OCR Foundation Models for Text-Centric Visual Intelligence

Optical Character Recognition (OCR) is evolving from plain-text transcription toward general visual intelligence, requiring models to recognize, localize, and reason over textual information in complex visual environments. However, existing OCR systems often excel at only some tasks and struggle to balance recognition, parsing, and reasoning across scenarios. In this report, we present PolyOCR, a family of unified OCR foundation models of varying scales. PolyOCR combines a shared instruction-following framework with a large-scale data engine that converts heterogeneous visual resources into quality-verified OCR supervision. We introduce Competence-Guided Policy Optimization, which combines verifier-based Group Relative Policy Optimization with on-policy distillation through sample-wise routing based on teacher reliability and the teacher--student competence gap. We also introduce OCRBench v2.1, our revision of OCRBench v2 with manually verified annotation corrections and task-aligned scoring metrics. Extensive experiments across OCRBench v2.1, CC-OCR, in-house KIE Benchmark, OmniDocBench v1.6 and MDPBench demonstrate that PolyOCR achieves state-of-the-art or highly competitive performance.