Self-Supervised Visual Representation Learning

What can a distribution-matching regularizer such as SIGReg in LeJEPA certify about contrastive learning? We study shared Gaussianization (SG), a characteristic-function Gaussianity test on the average of two normalized views, scaled by an independent $χ_d$ radius. Because disagreeing views shorten the average, one test detects both misalignment and non-uniformity. SG vanishes exactly at the aligned, uniform minimizers of population InfoNCE, and under equal marginals it bounds the InfoNCE excess by $4\cdot 3^{3/4}β$ times the square root of the SG loss, plus a term linear in the loss. The square-root rate and this dimension-free constant are sharp, and no squared mean-embedding distance on view pairs achieves a faster rate. With an explicit alignment term, a rotation-invariant uniformity test gives a linear bound if and only if its spectrum dominates that of InfoNCE's kernel $e^{βu^\top v}$; SG's own test does, Gaussian kernels $e^{-γ\|u-v\|^2}$ qualify exactly when $γ\ge β/2$, and moment matching never does. Away from the optimum, the objectives differ. Along an isotropic nuisance channel, pure SG lowers its loss by adding per-view nuisance whenever the shared code is non-uniform. An alignment weight above the channel's gain makes the nuisance-free solution a strict local minimizer; for LeJEPA, the same rule gives a critical SIGReg weight that decreases with the batch size. At finite batch size, an off-diagonal U-statistic removes a plug-in bias toward misalignment. In controlled latent-variable models, pure SG retains per-view style, an alignment weight above the measured gain removes it, and for LeJEPA at three batch sizes the measured gain separates the encoders that retain style from those that do not. InfoNCE training also reaches a lower SG$_{0.2}$ loss than SG$_{0.2}$ training from scratch, which points to an optimization gap.

Self-supervised learning (SSL) at scale produces powerful visual representations. However, most scalable SSL methods rely on ad hoc combinations of multiple objectives and stabilization mechanisms. Taking a step back, we ask if we can design a high-performing, yet principled SSL algorithm. Starting from the multi-view assumption, stipulating that task-relevant content is captured by the information common to different views, we construct an information-theoretic objective decomposing into interpretable terms. This derivation yields JEM, a student-teacher method that learns by aligning corresponding patch representations across views, explicitly regularized by information and structure preservation losses. JEM trains stably from 300M to 7B parameters, and, to our knowledge, is the first latent-space patch-level method demonstrated at 7B scale. Across all scales, JEM reaches strong performance on both global and dense probing tasks, on segmentation benchmarks consistently surpassing the DINOv2 algorithm, an influential foundation for today's strongest visual SSL methods. Notably, at 7B parameters, it exceeds the performance of DINOv3 on panoptic segmentation, despite being trained on $12\times$ less data without refinement stages. These results demonstrate that we can indeed design an SSL algorithm that learns strong representations, is principled and stable.

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.

Vision-Language Models (VLMs) such as CLIP excel in global semantic alignment but often lack fine-grained perceptual capabilities. This hinders dense prediction tasks and bottlenecks the visual potential of Multimodal Large Language Models (MLLMs). Existing research has attempted to enhance CLIP's visual representations by incorporating geometric priors from vision-centric models. However, these strategies often struggle to achieve deep alignment for both local spatial structures and global semantics, potentially even distorting the original image-text space. To address these limitations, we propose SALM, an unsupervised embedding alignment framework based on structurally-aware latent mask modeling. SALM effectively synergizes local and global alignment via a dual-path design combining explicit and implicit mechanisms, without requiring any image-text pairs. First, we introduce a dual-matrix alignment strategy that explicitly calibrates intra-sample spatial correlations and activation intensities, thereby effectively injecting local geometric priors. Based on this, we further design a latent mask modeling mechanism to guide CLIP to restore the missing semantic details of the target model, thereby implicitly aggregating fine-grained structures into the global semantic space. Furthermore, driven by the empirical observations that CLIP's shallow features inherently possess strong spatial observational capabilities, we naturally extend SALM to a highly efficient self-distillation paradigm, SALM-Self. This unlocks CLIP's intrinsic fine-grained potential without relying on any external models. Extensive experiments demonstrate that SALM not only significantly improves performance in dense prediction tasks but also boosts CLIP's zero-shot accuracy, effectively enhancing the fine-grained understanding capabilities of MLLMs. Project page at https://qzfm.github.io/salm_project_page/.

Strong image generation models are conditioned on class labels, aligned to frozen pretrained encoders, or built on separately trained autoencoders. While effective, generation then depends on supervision or pretraining: labels must be annotated, and encoders or autoencoders pretrained for the target domain. We study joint generative and self-supervised representation learning in a single model, enabling self-conditioned generation without labels or pretrained models. This is challenging because the objectives are mismatched: contrastive learning consumes clean augmented views and favors coarse, invariant semantics, while flow matching consumes noisy images and must preserve the fine detail and spatial layout that contrastive learning discards. We propose SCION (Self-conditioned Generation on Self-supervised representation), whose core is a single pixel-space encoder conditioned on the flow timestep and an embedding. For representation learning, this conditioning embedding is a learned global vector shared across images, with the encoder's [CLS] token yielding the semantic representation trained by the contrastive loss. For generative training, the conditioning embedding is the image's own [CLS] representation, while patch tokens pass through a decoder to predict the image. To sample without a reference image at inference, we jointly learn a prior over the embedding. Gradient-norm balancing and stop-gradient mechanisms enable joint optimization in one run. SCION is self-supervised and self-contained, with no labels or pretrained models. On ImageNet 256x256, with the JiT-B recipe and no representation guidance, SCION reaches 8.92 FID, surpassing class-unconditional iREPA, which aligns to pretrained DINOv2 (46.44), and RCG, which conditions on it (14.27). With JiT-L, SCION achieves 5.89 FID without guidance and 3.47 with representation guidance, outperforming RCG with the ADM recipe (6.24).

Self-supervised visual representation learning learns useful features without manual annotations during representation training. The image-based joint-embedding predictive architecture (I-JEPA) predicts latent representations of masked image regions, but its objective does not explicitly model responses to specified visual interventions. We introduce CI-JEPA, a counterfactual intervention-aware extension that learns to predict the representation change $ΔZ = Z_{\mathrm{CF}} - Z$ between an original image and a modified counterpart. We assess representation robustness through selective sensitivity: stronger responses to task-relevant semantic changes than to nuisance changes. Experiments on Flowers102 use flower-center occlusion as a candidate semantic intervention and background blur and tint as candidate nuisance interventions. With frozen-encoder linear probing, CI-JEPA achieves a best validation accuracy of 78.14%, compared with 77.55% for both the pretrained ViT-B/16 and the I-JEPA baseline, a gain of 0.59 percentage points. The reported mean $L_2$ representation changes are 4.42 for center occlusion, 3.48 for background tint, and 2.83 for background blur. This ordering is consistent with relative semantic selectivity for the evaluated interventions, rather than complete nuisance invariance. The accuracy comparison is complementary and does not establish improved robustness over the baselines. These controlled image modifications provide a framework for studying intervention-induced changes in JEPA representations; they do not establish causal feature discovery or robustness to all visual changes.

Why can masked prediction learn useful representations that unmasked reconstruction misses? We study this question in a high-dimensional model of a masked autoencoder (MAE) trained on data with shared latent structure and heterogeneous noise. We prove that masked linear reconstruction can recover the latent feature at linear sample complexity in regimes where unmasked linear reconstruction, equivalent to PCA, fails. The analysis also quantifies the statistical advantage of mask resampling, an established ingredient of masked pretraining. By introducing a fixed collection of $K$ masks per sample, we characterize its effect on feature recovery and downstream performance, identifying regimes where greater mask diversity lowers sample complexity. Guided by this prediction, we find that random cropping and flipping in standard image-training pipelines can obscure the advantage of mask resampling by renewing the prediction task even when the patch mask is fixed. Removing these transformations reveals a downstream advantage for dynamic over static masking in CNN autoencoders and vision transformers. A complementary BERT pilot finds benefits from greater mask diversity on downstream language tasks. Our results separate the benefit of the masked prediction objective from that of mask diversity, and show how a tractable theory can guide experiments that uncover advantages hidden by standard training practices.

End-to-end backpropagation has been the dominant mode of training in deep learning, allowing for the coordination of parameter updates across layers of a neural network. Prior studies have explored alternative -- and, in some cases, simpler -- training mechanisms, showing that they can sometimes achieve performance similar to backpropagation. However, the architectural conditions under which locally optimized networks, which avoid end-to-end backpropagation of error, can learn representations comparable to those learned through end-to-end training remain unclear. We aim to answer this question in the context of self-supervised learning, an important framework for large-scale pretraining in artificial intelligence. Here, we investigate how network width and depth affect the efficacy of greedy layer-wise and end-to-end self-supervised training in convolutional networks. We find that in wider networks, the benefits of end-to-end backpropagation over greedy layer-wise training shrink: in relatively shallow and very wide networks, we even observed higher performance in models trained with greedy layer-wise training. Subsequent analysis of the representations formed by these networks shows that very wide greedy-trained networks exhibit more favorable representational geometry than do networks trained end-to-end with backpropagation. This work shows that width can compensate for restricted credit assignment and identifies differences in representational geometry as a potential mechanism for their improved performance.

Self-supervised learning draws inspiration from infant visual development, yet standard training pipelines bear little resemblance to it: images are independently sampled and globally shuffled across epochs. We study self-supervised learning from continuous video streams, where frames are consumed in temporal order using strict sliding-window batches, without global reshuffling or multi-epoch replay. To this end, we construct WT++, a 95-hour urban walking-tour video dataset for streaming pretraining. Combined with a comprehensive evaluation suite we find that contrastive and distillation-based methods struggle in this setting, while MAE is more robust but still falls short of standard i.i.d. pretraining. We find that high inter-batch similarity, caused by sliding-window consumption across consecutive batches, does not explain this gap. The main challenge is high intra-batch similarity, where frames within each batch are near-duplicates. To mitigate this, we propose StreamMAE, which preserves the core MAE reconstruction objective while adapting the input pipeline with stream-aware regularization and motion-biased crop selection. StreamMAE outperforms streaming baselines, matches i.i.d. MAE trained on the same video data, remains competitive with ImageNet-pretrained MAE, and scales positively as the pretraining stream grows from 12 to 95 hours.

Over a century ago, Henri Poincaré argued that a motionless observer cannot acquire the notion of space. Yet, most visual representation learning methods operate on individual images, while those that leverage multiple views rely on RGB reconstruction, entangling geometry with appearance. We propose Poincar3, a self-supervised method that learns representations from multiple views through self-distillation instead of RGB reconstruction. We combine masked patch and image-level distillation with a teacher that observes additional views, enabling training from scratch without explicit 3D supervision. Poincar3 outperforms both previous single and multi-view self-supervised approaches such as DINOv3, MuM, and Muskie on correspondence estimation, camera pose estimation, and 3D reconstruction. Using a lightweight Poincaré adapter, we also find that our learned features encode camera motion more accurately than existing self-supervised representations.

Self-supervised learning (SSL) removes the need for annotations and makes models that are capable across more domains than supervised learning. The autoencoder SSL framework learns by reconstructing its own input after information loss through a bottleneck or noise injection. Masked autoencoders (MAE) are the most successful instantiation of this framework: they encode a random subset of patches, then decode the masked-out patches. In this work, we introduce key modifications to improve MAEs. Our method augments an image in two different ways, then masks and encodes each view separately. It then exchanges the global representations (CLS tokens) between views before decoding the masked patches. By design, our Masked Swingers encourages learning a view-agnostic summary of the image to facilitate efficient transfer. We perform extensive experiments, and find Masked Swingers outperforms MAE by +3-5% on ImageNet-1K kNN and provides large gains on fine-grained tasks, e.g., relative gains of +45% on instance retrieval, +22% on animal re-ID, and +76% on Omniglot character recognition. To boot, Swingers reduces error -64% relative to MAE on three new state-probing datasets, opening the door to world modeling. Welcome to our Swingers party.

RGB-T object tracking leverages the complementary characteristics of visible and thermal infrared modalities to improve robustness under adverse conditions. Existing supervised methods typically rely on costly modality-aligned bounding box annotations, while most self-supervised approaches follow a two-stage pseudo-labeling paradigm, making tracker training sensitive to pseudo-label quality and preventing joint end-to-end optimization. In this paper, we propose ESMTrack, a fully end-to-end self-supervised RGB-T tracking framework without offline pseudo-label generation or dense frame-level bounding box annotations. Given only the standard initial-frame annotation used in visual tracking, ESMTrack learns discriminative and temporally consistent representations through two complementary objectives: a grounding triplet loss on annotated initial frames and a cross-frame temporal triplet loss on unlabeled search frames, with reliable samples selected by forward-backward consistency. To address modality dominance bias, ESMTrack employs a three-branch architecture consisting of a fusion branch and two unimodal branches for RGB and thermal inputs. We quantify modality contributions using the Average Peak-to-Correlation Energy by measuring response discrepancies between the fusion and unimodal branches. The resulting reliability estimates guide a training-time modality decoupling mechanism that suppresses dominant-modality shortcuts and adaptively weights cross-modal contrastive learning for task-level alignment. Extensive experiments on five RGB-T tracking benchmarks show that ESMTrack achieves competitive state-of-the-art performance, strong cross-dataset generalization, and real-time inference speed. The source code is available at https://github.com/LiShenglana/ESMTrack.

Generative and representation learning remain asymmetrically connected: semantic representations are used to improve diffusion generation, whereas the models' own representations are often treated as a by-product of synthesis. We ask whether diffusion models can instead be trained to learn substantially stronger semantic representations without sacrificing generation quality. SelfFlow takes a step in this direction by introducing self-supervised patch alignment into flow matching, but its main gains remain in faster convergence and improved generation. Inspired by DINO and iBOT, we extend this framework with cross-view class-token alignment to further strengthen semantic representations. Specifically, we form two independently noised, dual-timestep observations of each image and align each student class-token representation with the stop-gradient EMA-teacher target from the other observation. This objective is optimized jointly with the inherited flow-matching and local patch objectives. Notably, although the additional objective acts only on the class token, it strengthens both class-token and patch representations. Compared with a matched two-view baseline, ImageNet linear-probing accuracy improves by 9.4% using the class token and 10.1% using mean-pooled patch tokens, while frozen-backbone VOC2012 segmentation improves by 3.6 mIoU. These representation gains are achieved while maintaining comparable ImageNet generation FID. In text-to-image training, the same objective also improves generation FID, reducing it from 2.52 to 2.37 at matched checkpoints. Our results show that representation need not remain a by-product of generation or merely a tool for improving it: it can be directly optimized as a first-class capability of diffusion pretraining alongside generation.

A deeper understanding of brain function requires a precise, structured characterization of behavior. Yet, extracting behavioral representations from video in a form suitable for scientific analysis remains a fundamental challenge. Many prior studies represent behavior via pose estimation or nonlinear video embeddings. However, pose tracking discards rich information beyond predefined keypoints, while nonlinear video embeddings lack interpretability. We address this limitation with SABLE (Sparse-view Animal Behavior Latent Embeddings), a self-supervised framework that leverages a geometric inductive bias to learn behavior representations. By augmenting a multi-view transformer with priors from monocular depth and pose estimation, SABLE reconstructs 3D animal behavior from extremely sparse views while learning explicit 3D latent structure. Without ground-truth 3D labels, it reliably recovers 3D behavior from two-view videos, whereas state-of-the-art (SOTA) methods fail or yield degenerate solutions. Across the International Brain Lab and Cheese3D datasets, we demonstrate that SABLE learns 3D representations that match or exceed prior SOTA performance in neural encoding and decoding. Once pretrained across animals, SABLE serves as an off-the-shelf model that generalizes zero-shot to unseen animals without animal-specific calibration or retraining. Our method establishes 3D-aware video embeddings that capture complex behavior, opening new avenues for studying brain-behavior relationships.

Images capture the world at one moment, whereas video reveals how it changes. Image self-supervision learns spatial structure from a single moment, while video methods commonly learn temporal relationships inside a representation computed jointly from several frames. We ask whether watching a scene change can instead improve features available from one image without sacrificing the ability to represent video. We introduce W2Rep, a masked feature-prediction framework in which an independently encoded source image participates in prediction at the same or another moment. The predictor is conditioned on visible video context, the queried location, and the signed time interval between source and target. This gives the cross-frame objective two complementary roles: the image path learns features that remain useful across time, while the video path must gather evidence that is missing from the source image. Across model scales and downstream tasks, W2Rep improves frozen and fine-tuned recognition under our comparison protocol, while joint video encoding provides further gains over frame-wise aggregation. Controlled experiments show that these gains depend on directly updating the source-image features and on using both video context and temporal displacement. Overall, change across a video can supervise a visual encoder whose representations remain useful at either image or video granularity. Code is available at~https://wenooi.github.io/W2Rep.

Evaluating semantic similarity between videos is a fundamental challenge in computer vision, essential for tasks ranging from out-of-distribution (OOD) detection to video retrieval. However, defining and labeling video similarity is notoriously difficult and expensive due to the complex spatio-temporal nature. In this paper, we propose a novel self-supervised approach that leverages generative uncertainty from text-to-video (T2V) diffusion models to learn semantic similarity without human annotations. Our method is based on the observation that T2V models produce consistent outputs for familiar concepts but exhibit high variance and uncertainty when prompted with specialized concepts. We utilize this behavior to identify stable semantic features within existing pretrained representations, such as VideoMAE and V-JEPA. Specifically, we learn a mask over these embeddings using purely generated data, encouraging the model to retain features that remain consistent across generations of general concepts while discarding those associated with generative noise or uncertainty. Experimental results across three key tasks demonstrate that our learned feature subspaces consistently outperform original pretrained features and baseline feature selection methods.

Joint-embedding self-supervised learning typically combines an invariance objective across augmented views with additional mechanisms to prevent representational collapse. These objectives are often applied after a projection head, while downstream tasks use the backbone representation before the projector. We find that this mismatch does not necessarily prevent dimensional collapse in the backbone, which can retain low effective rank and potentially limit downstream transfer. To address this, we introduce SACReg, a spectral anti-collapse regularizer motivated by an analysis of $λ$-balance, which captures the relative scale of weight matrices across layers. In a two-layer linear network, we show that (i) $λ$-balance prevents collapse, and (ii) our regularizer applied to the backbone induces $λ$-balance. In the nonlinear case, this regularizer leads to anti-collapse as well and, in realistic architectures on ImageNet100, it empirically increases the representations' ranks. We apply SACReg to JEPA and propose $λ$-JEPA, which improves over LeJEPA and VISReg on ImageNet-1k classification and in average linear-probe transfer performance across eight downstream image datasets. On video self-supervised learning, $λ$-JEPA improves over LeVJEPA and V-JEPA 2 on the Something-Something-v2 and Kinetics-400 benchmarks. Code is available at https://github.com/berkerdemirel/lambda-jepa.

Self-supervised pretraining learns representations that are broadly transferable across downstream tasks, yet direct fine-tuning can be suboptimal due to misalignment between self-supervised and downstream task objectives, potentially degrading pretrained features beneficial to the downstream task. The BiSSL framework addressed this by introducing a transitional training stage formulated as a bilevel optimization problem, in which the downstream task objective guides the self-supervised learning process in refining pretrained representations to better facilitate subsequent fine-tuning. However, BiSSL relies on conventional bilevel optimization solving techniques whose costly implicit hypergradient approximations render the method increasingly impractical for contemporary model architectures. To make it efficient and scalable, we introduce BiSSLight, which combines M-FAC-based implicit gradient approximation with parameter-efficient fine-tuning via LoRA, enabling efficient application at larger scales that were previously impractical. Evaluation across multiple downstream tasks and contemporary model architectures shows that BiSSLight consistently improves downstream performance, with gains becoming more pronounced as model size increases despite stronger baselines. The method is highly computationally efficient, reducing computation time by more than a factor of ten compared to its predecessor on a ViT-H backbone.

Geometric view synthesis can expand a single wine-label photograph into a training set, but its value with pretrained image encoders is unclear. We study this on a public WineSensed-derived benchmark of 1,000 classes, one enrollment photograph per class, and 4,295 real queries. With the earlier DINO vision transformer (ViT-S/16) recipe, geometric views raise top-1 accuracy from 34.1% to 62.6-63.7%, about three times the gain from two-dimensional (2D) augmentation. Frozen SigLIP 2-B already reaches 94.7%. A linear head over its frozen features gains 1.2-1.3 percentage points with the two geometric pipelines localized by the Segment Anything Model (SAM), while the other pipelines gain an inconclusive 0.3-0.6 points. Low-rank adaptation (LoRA) and validation-selected full fine-tuning show no clear gain within the reported confidence intervals; fixed-budget full fine-tuning loses 9-24 points. SAM localization supplies all six views for 99% of sources, compared with 43% for the edge-based front end. Recognition differences between the two cylinder constructions depend on the training recipe and are confounded by their crop and canvas conventions. Rendered-cylinder tests show different responses to source tilt, but an uncalibrated rim-ratio proxy establishes no corresponding trend in recognition on real photographs. An author-confirmed audit of 50 residual errors identifies 21 query-enrollment appearance mismatches, without establishing an irreducible error rate. These results support geometric synthesis for the tested self-supervised recipe and a smaller benefit through frozen-feature adaptation of the text-supervised encoder.

A useful visual representation is a statistic of the observed past that retains the latent factors shared with the future and discards patch-private noise. We formalize this requirement as a conditional predictive sufficient statistic (CPSS). Under a shared-factor model of image patches, the mutual information between the past and the next patch equals the information the past carries about the shared factor, up to a remainder that the next patch itself fails to reveal. Predicting the next patch embedding with a cosine loss is maximum likelihood for a von Mises-Fisher model of that embedding's direction, and is therefore a tractable surrogate for the predictive information. The same population loss is also minimized by a constant embedding, so stop-gradient does not by itself select the sufficient statistic; it only blocks the symmetric gradient that implements the constant solution in one step. The regression target is a shallow embedding, which forces the network output back into that shallow range and leaves the sufficient statistic in intermediate blocks. Small causal Transformers on MNIST and CIFAR-10 are used as diagnostics, not as a leaderboard. On MNIST the future shift and the stop-gradient move probe accuracy by tens of points, and the CPSS readout peaks before the output. On CIFAR-10, with the same short budget and no augmentation, every objective lands near a linear classifier on pixels. What still matches the derivation is the geometry: the CPSS output is a worse readout than its best intermediate block, next-pixel regression does not pay that penalty, and removing the stop-gradient collapses the effective rank of the embedding even when the pretext loss looks perfect.

Explicit geometric references offer a direct way to structure self-supervised representations. Existing adversarial distribution-matching formulations, however, require costly encoder-critic optimization. We introduce Flow-Based Distribution Matching (FBDM), a non-adversarial framework that learns this reference-directed geometry through spherical conditional velocity regression. An ETF-inspired reference allows its number of components K' to exceed the auxiliary flow dimension d* while retaining structured geometric separation. We assign both augmented views of each image to the same target, while limiting how many images each reference center can receive. An explicit alignment loss further pulls the two views' representations closer together. Experiments across benchmarks ranging from CIFAR to ImageNet show that FBDM achieves performance nearly on par with DM and remains competitive with existing SSL methods. Matched training-cost comparisons show a 1.48- to 1.83-fold speedup over DM with a negligible increase in GPU memory usage. We also provide a theoretical explanation for the usefulness of the learned representations: under stated conditions, we bound the downstream misclassification rate in terms of the FBDM pretraining loss.

Self-supervised vision transformers trained with DINO-style objectives exhibit striking emergent semantic representation quality across visual tasks, yet the mechanisms underlying this behavior remain unclear. We present a systematic empirical dissection of the DINO family and show that semantic representations arise primarily from enforcing consistency between geometrically distinct global views of the same image instance. This instance-specific global alignment acts as the semantic anchor of DINO-style learning. Across controlled retraining experiments evaluated on semantic correspondence and a diverse suite of 2D and 3D downstream tasks, we find that patch-level masking objectives enhance semantics only when trained jointly with this global alignment, indicating that the iBOT objective refines and densifies existing semantic structure rather than creating it independently. In contrast, local-to-global view alignment does not substantially improve semantic qualities at fixed compute beyond a purely global alignment. Beyond training design, we revisit how semantic representation quality should be evaluated: while classification accuracy is the standard validation score, semantic correspondence provides a complementary axis that more reliably predicts downstream task performance. Together, these findings provide a functional decomposition of DINO-style learning and represent an important step toward understanding how semantic representations emerge in self-supervised vision models.

Contrastive self-supervised learning has achieved strong performance by learning representations from multiple augmented views of the same image. However, most existing methods construct positive pairs using independently sampled stochastic augmentations, which may alter semantic content and ignore the intrinsic geometry of the data distribution. In this work, we propose OTCLR, an optimal transport-aware framework for contrastive learning representations that generates geometry-consistent positive samples. Instead of directly contrasting two randomly augmented views, we construct intermediate views between the original image and its augmented variants through entropic optimal-transport displacement interpolation. These transport-interpolated samples serve as positive views that better preserve image structure while explicitly modeling spatial distributional geometry. To further promote smooth representation learning, we evaluate auxiliary Sinkhorn regularization terms that encourage transport-interpolated views to remain consistent with their endpoint images. The proposed method can be incorporated into standard contrastive learning pipelines without modifying the encoder architecture. Experiments on multiple benchmark datasets show that our approach improves representation quality and transfer learning performance compared with conventional augmentation-based contrastive learning baselines.

New Guinea is the world's richest island flora (~2,856 orchid species), yet most species are represented by only a handful of photographs, far fewer than direct species-level classification requires. Methods for fine-grained identification in such species-rich, data-poor floras are needed, and it remains unclear which backbone architecture and pretraining strategy best support them. We built a two-stage system that first predicts the genus of a query photograph, then retrieves visually similar reference images of candidate species using FAISS. We compared four pretrained backbones -- two Vision Transformers (ViTs; DINOv2, BioCLIP 2) and two CNNs (ConvNeXt V2-L, EfficientNetV2-L) -- fine-tuned under an identical protocol on a fixed, species-stratified partition of 16,701 photographs spanning 120 genera and 1,350 species, assessing accuracy, calibration, error structure, species retrieval, and open-set detection of novel genera. DINOv2 attained the best genus performance (macro top-1 66.9%, 95% CI 63.7-70.6; global top-1 88.9%); both ViTs outranked both CNNs, and general-purpose self-supervised pretraining (DINOv2) outperformed domain-matched biological pretraining (BioCLIP 2) by 7.1 points of macro top-1. Errors concentrated on two abundant genera acting as error attractors. DINOv2 embeddings achieved species Recall@5 of 86.6% and genus Recall@5 of 98.7%; temperature scaling reduced every backbone's Expected Calibration Error to about 0.03; and a distance-based open-set gate flagged unseen genera (mean AUROC 0.958). A self-supervised Vision-Transformer backbone combined with embedding retrieval is an effective, deployable strategy for fine-grained identification in species-rich, data-poor floras. The system is released as an open web application (the New Guinea Orchid Identifier), offering a practical template for other hyperdiverse, under-documented taxa.

We present ParticleSplat, a self-supervised object-centric learning method that decomposes scenes into a set of latent ''particles'' representing semantic entities through feedforward 3D Gaussian Splatting. Building on the Deep Latent Particles (DLP) framework, which represents images as a set of particles with attributes such as position, scale, and visual appearance, we address a key limitation of DLP: its inherently 2D nature, which prevents explicit 3D spatial and geometric reasoning that are critical for downstream tasks such as robotic manipulation. Leveraging the structural similarity between latent particles and 3D Gaussian primitives, we introduce a 3D latent particle space trained with a novel view synthesis objective. Our model jointly encodes multiple views with camera poses into a shared 3D object-centric latent space, then transforms particles into particle-aligned 3D Gaussians whose composition reconstructs the full scene. On simulated and real-world datasets, we show that this formulation inherently learns object masks without supervision and supports controllable 3D scene editing, such as moving objects by modifying particles in the latent space. We further establish that the learned 3D representation improves downstream performance on robotic manipulation tasks.

Vision Transformers (ViT) excel in semantic understanding but fail to discriminate between object instances (e.g., identical embeddings for two dogs), limiting their use in instance-level tasks such as object detection and instance segmentation. We propose Contrastive Vision Transformer (CoViT), a self-supervised learning framework that injects instance-awareness into ViT through geometry-guided contrastive learning. CoViT uniquely coordinates ViT's attention maps and embeddings by constructing triplets: (1) Attention-guided masking: Refine multi-head attention via adaptive thresholding and morphological operations to generate instance masks, identifying foreground anchors; (2) Hardest contrastive mining: For each anchor, computing pairwise embedding similarities to select the intra-instance hardest positive (least similar patch within its mask) and inter-instance hardest negative (most similar patch from other instances), with intra-instance regions masked during negative search. These triplets drive a contrastive loss that simultaneously compresses intra-instance variance and expands inter-instance margins, forcing ViT to discern subtle geometric and appearance differences between instances. CoViT consistently achieves stable performance gains of over 2 AP points across multiple instance-level perception tasks by using ViT as backbone architecture. Notably, CoViT requires no extra decoders or labels, demonstrating that a pure ViT can learn instance-aware representations via inherent attention priors and targeted contrastive constraints. Code and models will be released.

Face forgery detectors often achieve strong results on controlled benchmarks, but their reliability under realistic image degradations remains limited. This paper presents a standardized benchmark for face forgery detection using the Multi-Dimensional Face Forgery Image (MFFI) dataset and evaluates performance on both clean and degraded test partitions. We compare six model families, including convolutional networks, transformer-based models, and a frozen self-supervised DINOv3 backbone, across spatial, spectral, and hybrid input representations. The results show that clean-set performance is not a reliable indicator of robustness under compression, resizing, and blurring. Xception with RGB obtains the best clean performance, reaching 0.884 mean ROC-AUC, but degrades substantially on the harder partition. In contrast, frozen DINOv3 achieves the strongest degraded-set result, with 0.726 mean ROC-AUC, while training only a linear classification head. The representation analysis indicates that Fourier-domain cues are most useful when combined with RGB information, whereas purely spectral inputs consistently underperform spatial representations. Qualitative attribution maps further suggest that convolutional detectors focus on localized artifacts, while DINOv3 relies on broader facial structure. These findings reinforce the need for degraded evaluation protocols and highlight self-supervised visual representations as a promising direction for robust face forgery detection. Our source code is publicly available at https://github.com/lucasdocunha/FaceForgery-Benchmark/.

Dense self-supervised learning (SSL) is a powerful paradigm for learning without annotations the local descriptors required to solve dense medical imaging tasks. We present Pix2Rep-v2, a framework for SSL of pixel- and voxel-level representations suitable for few-shot downstream applications. Pix2Rep-v2 addresses the main challenges of dense SSL by leveraging a redundancy reduction objective at the pixel-level with a principle of equivariance of dense representations, that scales efficiently to 3D or wide field-of-view applications. We evaluate our method on four datasets, across multiple tasks, multiple modalities and anatomical structures using multiple backbones in 2D and 3D, and under various data regimes. As an alternative to linear probing or full fine-tuning on the downstream task, we also propose an in-context variant, without downstream training, based on a dense prototype approach. Pix2Rep-v2 shows substantially higher data-efficiency in few-shot scenarios compared to fully supervised baselines, and is competitive with the state-of-the-art e.g., +9.3 Dice points in one-shot segmentation on the M&Ms-2 dataset. Our code and pre-trained models are publicly available at https://github.com/BioMedTP/pix2rep-v2.

Cardiac magnetic resonance (CMR) imaging provides complementary information on cardiac anatomy, function, and tissue characterization across multiple sequences and views. In this work, we investigate foundation model pretraining for 2D CMR and introduce CMRVision, a CMR-specific foundation model trained using DINOv3-style self-supervised learning on a multi-center, multi-sequence cohort of 36 million CMR images. We systematically evaluate architectural and training design choices for domain-specific pretraining. CMRVision is evaluated on two downstream tasks: multi-task segmentation across cine, late gadolinium enhancement (LGE), and mapping sequences, and cine view classification. Our experiments show that CMR-specific pretraining, smaller patch sizes, and patch-level objectives consistently improve downstream performance. Across a multi-task segmentation benchmark, CMRVision achieved the strongest overall performance, outperforming prior natural-image (NI), medical-image, supervised, and CMR foundation model baselines. Improvements were modest but consistent across structures and sequences, with Dice scores ranging from 0.940-0.967 for LV and 0.855-0.905 for myocardium, and reaching 0.929 for RV, 0.920 for LA, and 0.931 for RA. The largest gains were observed for myocardium segmentation in LGE and mapping images. In a zero-shot segmentation task on unseen LGE long-axis views, the model achieved an average Dice score of 0.692, demonstrating cross-view generalization. For cine view classification, CMRVision achieved the highest average accuracy (0.906), compared to prior methods reported in the literature. These results highlight the potential of CMRVision to support robust and generalizable cardiac MRI analysis across multiple sequences and views.

We introduce ViTAMINS, a method that integrates synthetic hard negatives into unsupervised vision transformer pretraining to improve representation quality. Our approach is thoroughly benchmarked on ImageNet and transfer learning, image retrieval, copy detection, and image, video segmentation tasks. Notably, our proposed negatives give rise to emergent properties, where learned representations contain explicit information about the semantic content of an image and serve as excellent classifiers (up to +11.3% over baselines). ViTAMINS achieves these benefits through simple modifications to existing contrastive frameworks and outperforms competing methods while being more resource efficient, e.g., our ViT-B surpasses V-JEPA with ViT-L. Our findings motivate reconsidering contrastive learning as a simpler yet powerful alternative to dominant generative and self-distillation approaches.