CertEx uses certified half-precision lower bounds to exactly answer kNN queries at up to 18x the throughput of exhaustive GPU search.
mega hub Mixed citations
Artificial intelligence for a rare disease
Mixed citation behavior. Most common role is background (46%).
1231
Pith papers citing it
6
external citations · Crossref
Background
46% of classified citations
abstract
Self-supervised learning holds the promise of eliminating the need for manual data annotation, enabling models to scale effortlessly to massive datasets and larger architectures. By not being tailored to specific tasks or domains, this training paradigm has the potential to learn visual representations from diverse sources, ranging from natural to aerial images -- using a single algorithm. This technical report introduces DINOv3, a major milestone toward realizing this vision by leveraging simple yet effective strategies. First, we leverage the benefit of scaling both dataset and model size by careful data preparation, design, and optimization. Second, we introduce a new method called Gram anchoring, which effectively addresses the known yet unsolved issue of dense feature maps degrading during long training schedules. Finally, we apply post-hoc strategies that further enhance our models' flexibility with respect to resolution, model size, and alignment with text. As a result, we present a versatile vision foundation model that outperforms the specialized state of the art across a broad range of settings, without fine-tuning. DINOv3 produces high-quality dense features that achieve outstanding performance on various vision tasks, significantly surpassing previous self- and weakly-supervised foundation models. We also share the DINOv3 suite of vision models, designed to advance the state of the art on a wide spectrum of tasks and data by providing scalable solutions for diverse resource constraints and deployment scenarios.
hub tools
citation-role summary
background 137
method 50
baseline 11
dataset 4
other 3
citation-polarity summary
claims ledger
- abstract Self-supervised learning holds the promise of eliminating the need for manual data annotation, enabling models to scale effortlessly to massive datasets and larger architectures. By not being tailored to specific tasks or domains, this training paradigm has the potential to learn visual representations from diverse sources, ranging from natural to aerial images -- using a single algorithm. This technical report introduces DINOv3, a major milestone toward realizing this vision by leveraging simple yet effective strategies. First, we leverage the benefit of scaling both dataset and model size by
authors
mega hub controls
export citing contexts JSON
export graph JSON
export full bundle JSON
open full Pith review
annotated reader queued
Recognition alignment
counterfactual ablation
co-cited works
representative citing papers
The paper presents a theoretical and practical framework for making language the low-rate part of an ordered visual code, enabling unified generation and understanding with bounded pixel cost.
Species-level label-free visual clustering of frozen embeddings is both clustering-limited and representation-limited; paired DNA audits quantify the remaining gap at about two bits.
Hyper-RED transfers semantic structure from images to events via hypergraph distillation, improving event representation learning across recognition, segmentation, depth.
A 53M-parameter model for open-vocabulary relation prediction that runs at 20 ms per frame and generalizes across region sources and predicate vocabularies.
A geometry-grounded benchmark reveals that pose-free visual generators systematically trade off camera-motion accuracy for scene-state preservation, with the pairwise teacher-forced objective itself identified as the binding constraint.
Exact solutions show abstraction is set by input/target geometry, rises with depth, peaks under small init, and is attenuated by nonlinearities—improving LLM probes via GELU ablation.
Every9D-21M supplies 21.8M real-world 9D pose annotations for 700 everyday categories by propagating manual canonical poses through cross-instance alignment in object-centric videos and verifying them multiview.
LeJEPA achieves linear identifiability of latent variables uniquely when the latents are Gaussian in worlds with stationary additive-noise transitions.
Dimension d = O(m^{-2} log n) nearly achieves the optimal margin m^rd(+∞, A) for retrieval embeddings, with matching lower bounds showing d = O(k log(n/k)) suffices and is necessary for m = Θ(k^{-1/2}) on k-sparse query matrices.
RICA replaces ICA's global generative model with local Riemannian geometry, introducing a disentanglement tensor based on the Hessian of the log-likelihood and Ricci curvature to measure pointwise disentanglement, which recovers sources across manifolds in controlled tests.
CanViT is the first task- and policy-agnostic AVFM pretrained via passive-to-active dense latent distillation on 13.2M scenes and 1B random glimpses, achieving 38.5% ADE20K mIoU in one glimpse and 84.5% ImageNet-1k top-1 after fine-tuning.
ConceptPose delivers state-of-the-art zero-shot relative pose estimation by matching open-vocabulary 3D concept vectors derived from VLM saliency maps, beating the strongest baseline by 62% in ADD(-S) without training.
A 22M-parameter encoder pre-trained on audio and music achieves state-of-the-art zero-shot KNN performance across 19 industrial signal datasets spanning four modalities by modeling spectrogram sub-bands rather than fixed-rate inputs.
REKI is a 2.4B vision-language-action model pretrained on public data that reaches near state-of-the-art success rates using a low-cost register-token memory.
SurgStreamer detects requested surgical events by separating video-to-state perception from reusable pointer-based reasoning, beating streaming vision-language models on sealed tests and transferring its reasoning weights to a new procedure type while perception requires retraining.
Synthetic multispectral scenes from a 3D radiative transfer model are used to learn concept vectors in geo-foundation model embeddings, revealing physical sensitivities that predict downstream performance.
TEAL performs pixel-level active learning for semantic segmentation by selecting diverse tokens from a frozen ViT and then refining with uncertainty on the decoder grid, consistently outperforming prior methods at budgets of 1-10 pixels per image.
WinARDet uses compressed window auto-regression to break the box repetition loop in MLLM-based dense object detection, improving F1 by 7.2 points on the extremely dense RSDense benchmark.
A benchmark that evaluates action-conditioned robot world models against paired real-world executions across four tasks, three embodiments, and seven models.
Normal-reference logical anomaly detectors lose up to 34 AUROC points under brightness and zoom-out shifts, while a classifier trained on synthetic anomalies from an LLM and image editor loses less than 1 point.
A representation's inherent 3D structure, measured by cross-view consistency for features and reconstruction accuracy for geometry, dictates its effectiveness for spatial reasoning, and the optimal transfer mechanism depends on representation form.
DIME pretrains on same-identity face pairs with a direction-of-difference loss, learning fine-grained facial representations that scale with data and outperform prior methods on seven facial analysis tasks.
Barycentric Guidance uses barycentric weights over a triangle mesh of discrete emotion anchors to combine noise predictions from multiple prompts, yielding continuous and accurate VA control in diffusion-based editors.
citing papers explorer
-
CertEx: Exact Visual Retrieval at Scale by Certified Exclusion
CertEx uses certified half-precision lower bounds to exactly answer kNN queries at up to 18x the throughput of exhaustive GPU search.
-
Making Language the Low-Rate Code of Vision: Successive Refinement from Words to Pixels
The paper presents a theoretical and practical framework for making language the low-rate part of an ordered visual code, enabling unified generation and understanding with bounded pixel cost.
-
InfoTaxa: Information-Calibrated Label-Free Clustering for Fine-Grained Visual Taxonomy
Species-level label-free visual clustering of frozen embeddings is both clustering-limited and representation-limited; paired DNA audits quantify the remaining gap at about two bits.
-
Hyper-RED: Scalable Event Pre-training via Semantic Hypergraph Distillation
Hyper-RED transfers semantic structure from images to events via hypergraph distillation, improving event representation learning across recognition, segmentation, depth.
-
RelateAnything: Real-Time Open-Vocabulary Relation Prediction From Any Inputs
A 53M-parameter model for open-vocabulary relation prediction that runs at 20 ms per frame and generalizes across region sources and predicate vocabularies.
-
Beyond Visual Quality: Evaluating Physical Consistency under Ego-Motion with EgoGenEval
A geometry-grounded benchmark reveals that pose-free visual generators systematically trade off camera-motion accuracy for scene-state preservation, with the pairwise teacher-forced objective itself identified as the binding constraint.
-
How are linear representations learned? Exact solutions to the dynamics of abstraction
Exact solutions show abstraction is set by input/target geometry, rises with depth, peaks under small init, and is attenuated by nonlinearities—improving LLM probes via GELU ablation.
-
Every9D-21M: Large-Scale Real-World 9D Canonicalization of Everyday Objects
Every9D-21M supplies 21.8M real-world 9D pose annotations for 700 everyday categories by propagating manual canonical poses through cross-instance alignment in object-centric videos and verifying them multiview.
-
When Does LeJEPA Learn a World Model?
LeJEPA achieves linear identifiability of latent variables uniquely when the latents are Gaussian in worlds with stationary additive-noise transitions.
-
Is Dimensionality a Barrier for Retrieval Models?
Dimension d = O(m^{-2} log n) nearly achieves the optimal margin m^rd(+∞, A) for retrieval embeddings, with matching lower bounds showing d = O(k log(n/k)) suffices and is necessary for m = Θ(k^{-1/2}) on k-sparse query matrices.
-
Disentanglement Beyond Generative Models with Riemannian ICA
RICA replaces ICA's global generative model with local Riemannian geometry, introducing a disentanglement tensor based on the Hessian of the log-likelihood and Ricci curvature to measure pointwise disentanglement, which recovers sources across manifolds in controlled tests.
-
CanViT: Toward Active-Vision Foundation Models
CanViT is the first task- and policy-agnostic AVFM pretrained via passive-to-active dense latent distillation on 13.2M scenes and 1B random glimpses, achieving 38.5% ADE20K mIoU in one glimpse and 84.5% ImageNet-1k top-1 after fine-tuning.
-
ConceptPose: Training-Free Zero-Shot Object Pose Estimation using Concept Vectors
ConceptPose delivers state-of-the-art zero-shot relative pose estimation by matching open-vocabulary 3D concept vectors derived from VLM saliency maps, beating the strongest baseline by 62% in ADD(-S) without training.
-
FISHER: A Foundation Model for Multi-Modal Industrial Signal Comprehensive Representation
A 22M-parameter encoder pre-trained on audio and music achieves state-of-the-art zero-shot KNN performance across 19 industrial signal datasets spanning four modalities by modeling spectrogram sub-bands rather than fixed-rate inputs.
-
REKI: Registering Knowledge in VLAs
REKI is a 2.4B vision-language-action model pretrained on public data that reaches near state-of-the-art success rates using a low-cost register-token memory.
-
SurgStreamer: proactive, request-driven monitoring of surgical video streams with structured state and a reusable reasoning layer
SurgStreamer detects requested surgical events by separating video-to-state perception from reusable pointer-based reasoning, beating streaming vision-language models on sealed tests and transferring its reasoning weights to a new procedure type while perception requires retraining.
-
PiCo: Physics-Informed Concept Probing for Geo-Foundation Models
Synthetic multispectral scenes from a 3D radiative transfer model are used to learn concept vectors in geo-foundation model embeddings, revealing physical sensitivities that predict downstream performance.
-
Pixels to Tokens: Token Space Efficient Active Learning for Low-Budget Semantic Segmentation
TEAL performs pixel-level active learning for semantic segmentation by selecting diverse tokens from a frozen ViT and then refining with uncertainty on the decoder grid, consistently outperforming prior methods at budgets of 1-10 pixels per image.
-
WinARDet: Compressed Window Auto-Regression for Long-Horizon Vision-Language Grounding
WinARDet uses compressed window auto-regression to break the box repetition loop in MLLM-based dense object detection, improving F1 by 7.2 points on the extremely dense RSDense benchmark.
-
WorldOracle: Benchmarking Action-Conditioned World Models against Real-World Executions
A benchmark that evaluates action-conditioned robot world models against paired real-world executions across four tasks, three embodiments, and seven models.
-
Robust Logical Anomaly Detection under Capture Shifts via Synthetic Anomalies
Normal-reference logical anomaly detectors lose up to 34 AUROC points under brightness and zoom-out shifts, while a classifier trained on synthetic anomalies from an LLM and image editor loses less than 1 point.
-
What Makes a Good Visual Representation for Spatial Understanding in MLLMs?
A representation's inherent 3D structure, measured by cross-view consistency for features and reconstruction accuracy for geometry, dictates its effectiveness for spatial reasoning, and the optimal transfer mechanism depends on representation form.
-
DIME: Scaling Facial Representation Learning via Differential Masked Autoencoding
DIME pretrains on same-identity face pairs with a direction-of-difference loss, learning fine-grained facial representations that scale with data and outperform prior methods on seven facial analysis tasks.
-
Barycentric Guidance: Turning Foundation Image Editors into Continuous Affective Controllers
Barycentric Guidance uses barycentric weights over a triangle mesh of discrete emotion anchors to combine noise predictions from multiple prompts, yielding continuous and accurate VA control in diffusion-based editors.
-
PathPilot: An Adaptive Question-Conditioned Multi-Magnification Navigation Agent for Whole-Slide Image Analysis
PathPilot achieves state-of-the-art open-ended question answering on whole-slide images by coupling pathologist-derived multi-magnification diagnostic localization with question-conditioned adaptive region exploration.
-
Variational Filtering for Bayesian Last Layers with an Independent Binary Probit Surrogate
A closed-form variational filter for Bayesian last layers, using an independent binary probit surrogate, achieves better accuracy and calibration than gradient-based baselines on large image classifiers.
-
RINO: Renormalization Group Distillation with No Labels
RINO uses multi-scale kT clustering views and DINO-style self-distillation to learn jet representations that, with 1% of labels, match fully supervised top-tagging performance.
-
ThinkDrag: Semantic Drag-Based Image Editing with Visual Reasoning
ThinkDrag combines visual reasoning with drag-based image editing to interpret sparse point constraints as semantic transformations, achieving state-of-the-art generation quality while maintaining competitive point-following precision.
-
NOFACE3D: NOise-scheduled Focused Anchoring and Controlled Enhancement for 3D Generation
A training-free method that sharpens fine geometry and texture in generated 3D assets by rescaling selected regions and resampling them with a boundary-distance noise schedule.
-
Unified Generative Video Compression with Autoregressive and Bidirectional Video Diffusion
A unified generative video codec that shares one bitstream between two diffusion decoders: a bidirectional mode for quality and an autoregressive mode for real-time streaming.
-
ROSETTA: Rotation Subspace Estimation for Training-free Transmodal Alignment
The relative transformation between independently trained encoders is a spatially varying rotation in whitened space; ROSETTA estimates it locally via a soft EM procedure without training.
-
Beyond Resolution: Object-to-Image Ratio Mismatch in Instance Retrieval
Object-to-image ratio mismatch, not resolution loss, is the dominant cause of cross-scale retrieval failures in visual instance retrieval.
-
SAGE: Sink-Aware Guided Emphasis for Visual Grounding in Vision-Language Decoders
SAGE reveals that attention sinks in VLM decoders are layer-dependent (static at early/late layers, query-driven at mid layers) and uses backbone-derived ROI masks to steer attention away from static sinks, improving grounding and reducing hallucinations.
-
CUSP: CUSUM-Governed Survival Hazard Alarms at the Perception Onset for Off-Road Navigation
A survival-model plus CUSUM alarm trained on human takeovers catches off-road hazards at the moment a human would first judge driving unsafe, detecting 85 of 142 events at five new sites versus 26 for the best baseline.
-
GS-Pool: Object-Level Change Detection in 3D Gaussian Splatting
GS-Pool detects changed objects between two independent 3D Gaussian Splatting reconstructions by pooling per-primitive evidence into objects and calibrating a change threshold per scene, outperforming prior methods on PASLCD and CL-Splats.
-
dIon: Fragmentation-Based Invariance for Self-Supervised Learning of Tandem Mass Spectra
A self-supervised objective using precursor-conditioned mixture views and precursor-withheld partial views produces spectrum representations that improve de novo peptide sequencing by 2.3 to 8.4 percentage points over matched training from scratch.
-
Estimation of the Label-Noise Transition Matrix with Performance Guarantees via Selective Classification
A posterior-free route to estimating the label-noise transition matrix via one-sided selective classification, with finite-sample guarantees at near-parametric rates and no curse of dimensionality.
-
SAGE: Salient Factor Discovery and Generation with Visual Foundation Representations
SAGE factorizes a frozen representation autoencoder's spatial latent into common and salient parts, enabling label-free subtype discovery and subtype-preserving generation from only dataset labels.
-
Convergence of Practical Muon with Finite Newton-Schulz Iterations and Nesterov Momentum
Muon with finite Newton-Schulz iterations and Nesterov momentum achieves an O(T^{-1/4}) bound on expected average Frobenius gradient norm under coupled learning-rate and momentum schedules, with no explicit rank or dimension factors.
-
Adversarial Training for Pixel Diffusion
Adversarial post-training on predicted clean images at non-high-noise timesteps jointly improves fidelity, coverage, alignment, and quality of pixel diffusion by restoring missing high-frequency spectral power.
-
Mutual Information Constrained Chernoff Bottleneck
The error-exponent-maximizing compression curve C(R) rises strictly up to the entropy of the minimal sufficient statistic H(V), saturates there at the uncompressed exponent, and needs only k+1 output symbols.
-
The Domain Is a Residue: Adapting Self-Supervised Features, Not Generators
A 2.9M-parameter adapter that moves only the 'domain residue' inside frozen DINO features matches or beats pixel-fed translators on several weather and sim-to-real benchmarks, with far fewer trainable parameters but measurably more scene-structure loss.
-
Reducing the Adaptation Gap Through Reachable Fisher Geometry
The reachable-Fisher trace is a useful subgroup-difficulty diagnostic and training heuristic for parameter-efficient fine-tuning, but the paper's own matrix-level audit shows it neither certifies nor reliably controls local subgroup-gap behavior.
-
JRDB-AVR: An Active Visual Reasoning Benchmark for Embodied Agents in Real-World Environments
JRDB-AVR is an active visual reasoning benchmark that evaluates both the final answer and the supporting visual evidence, and reveals a large gap between answer and evidence accuracy in current VLMs.
-
From Instrument-Mounted Demonstrations to In-Vivo Execution: Learning Bimanual Laparoscopic Appendectomy Without Robot-Collected Demonstrations
A sensor logger on hand-held laparoscopic instruments captured enough demonstration data to train a diffusion policy that completed live-rabbit appendectomies with electrosurgery in three of four animals, with the surgeon selecting the surgical phase.
-
TouchSight: Bare-Handed Tactile Prediction from Egocentric Video via Generative Visual Augmentation
From monocular egocentric RGB video, TouchSight predicts per-taxel contact force across the full hand using generative visual augmentation to bridge the gloved-to-bare-hand appearance gap.
-
Semantic-ITC: A Frame-wise Indoor Mobile Laser Scanning Dataset and Benchmark for Semantic Segmentation
Semantic-ITC is the first public dataset with per-frame point-level semantic labels for indoor mobile laser scanning, covering 52 sequences with 1.23 billion labeled points across 16 classes.
-
An Open-Source Hardware and Software Toolkit to Enable Agentic RHEED-Guided Thin-Film Synthesis
Open-source hardware and software for automated RHEED beam control, self-symmetry-based azimuthal alignment, and extensible quantitative analysis accessible to AI agents.
-
Same Encoder, Different Winner: A Paired-View Framework for Cell Painting Encoder Evaluation
Single-metric rankings of Cell Painting encoders are sensitive to the evaluation protocol, as background signal inflates within-study scores but does not transfer across batches.
-
DUET-DINO: Simultaneous Cross-View World Modeling for Latent Planning in Robot Manipulation
DUET-DINO combines side and wrist camera views with cross-attention to create a latent world model that supports zero-shot 7-DoF action planning for robot manipulation.