Vision Foundation Model Adaptation

Large vision models provide useful representations for remote-sensing segmentation but are often too expensive for deployment at the satellite or field edge. Existing feature-level distillation methods also tend to assume similar teacher and student architectures and often stop feature alignment when task training begins. We introduce JEDI (JEPA-to-Edge Distillation), a two-stage framework that transfers representations from a large I-JEPA Vision Transformer teacher to a compact SegFormer student. First, JEDI aligns the student's terminal representation with the teacher's token space using cross-architecture projection and spatial alignment. It then jointly optimizes supervised segmentation, temperature-scaled response distillation, and persistent feature alignment throughout task adaptation. On CalCROP21, JEDI-B0 achieves 68.0 mean Intersection-over-Union (mIoU) with 4.04M parameters, improving over the standalone student by 16.0 points and coming within 2.0 points of the 70.0 mIoU achieved by the 639M-parameter teacher. We evaluate SegFormer B0, B1, and B2 students with 4.04M, 14.33M, and 28M parameters, respectively. Across all three variants, JEDI consistently outperforms response-, structure-, channel-, and relational-distillation baselines under the same teacher-student setting. These results show that persistent representation alignment is especially valuable under aggressive compression, substantially reducing model size and computation while preserving segmentation performance.

Vision Foundation Models (VFMs) provide transferable patch representations for few-shot industrial anomaly detection, but their attention computation is typically inherited from pretraining objectives centered on semantic aggregation. This creates a potential mismatch: token relations that support semantic recognition may not adequately expose the localized texture and structural deviations required for anomaly localization. We therefore investigate the hypothesis that the attention computation of a frozen VFM can be reconfigured as a task-relevant component of anomaly detection. We instantiate this idea with Power-Law Self-Correlation Enhanced Attention (PL-SCEA), which retains the semantic context of pretrained query-key attention while constructing token-adaptive self-correlations over contextualized value features. Positive-correlation filtering and power-law reweighting then emphasize relations that are salient relative to each token's relational background, without introducing additional trainable attention projections. The resulting features are modeled by a lightweight variational autoencoder that provides a fixed-size reconstruction-based representation of category-specific normality. The two stages serve complementary roles: attention reconfiguration shapes how local relational deviations are represented, while reconstruction-based modeling converts deviations from learned normality into anomaly scores. Across MVTec AD and VisA, the complete framework achieves competitive image-level detection and consistently strong pixel-level localization across the evaluated few-shot settings. Ablations further show that PL-SCEA improves localization with either the VAE or a memory bank under the tested setting. These results support the view that task-aligned attention reconfiguration can improve the anomaly-localization capability of frozen pretrained representations.

Promptable segmentation foundation models such as SAM3 accept an open-vocabulary text concept and return every instance matching it, but adapting them to a specialized domain by full fine-tuning is computationally prohibitive for the organizations that would benefit most. This study applies Low-Rank Adaptation (LoRA) to SAM3 for multi-class structural defect segmentation and examines both how such a model can be supervised from conventional annotation and whether the resulting efficiency gain transfers across datasets. Two contributions are methodological. First, we describe a supervision procedure that trains a concept-promptable model directly from COCO-style class-labeled instance segmentation by using the category name itself as the prompt, requiring no prompt templates, no synonym expansion, and no learned class embeddings. Second, we identify and mitigate a failure mode specific to this setting: because a conventional annotation file yields positive prompts exclusively, the model's presence prediction decouples from the text condition and degenerates into responding to any prompt, a collapse that is invisible to every metric computed on positive prompts alone. Exhaustive hard-negative prompting, in which every dataset category absent from an image is issued as a zero-detection query, addresses this at no annotation cost. Two adapter placements were compared under an identical protocol, updating 0.121% and 1.341% of model parameters. On a purpose-built tunnel lining dataset, pixel intersection-over-union improved from 0.017 to 0.338 and instance-level recall from 0.375 to 0.672; on the independent public Structural Defects Dataset, from 0.017 to 0.855 and from 0.574 to 1.000. Improvements were directionally consistent across ten metrics on both datasets, and the largest per-category gains occurred precisely where zero-shot competence was absent.

Off-road navigation can fail when physical structures induce irrecoverable states such as high-centering or entrapment, requiring human interventions. Identifying these structures is crucial, yet challenging. Such failure events are rare and costly to collect, resulting in limited training data. Moreover, the collected data associate frames with outcomes, but do not indicate the visual cues responsible for the failure. Learning directly from these data can therefore exploit scenario-specific visual cues, leading to poor generalization. We propose \textbf{Geo-Anchored Fine-Tuning (GAFT)}, a parameter-efficient method that adapts a vision foundation model with a geometry-derived prior. It guides LoRA adaptation by aligning a spatial attention-rollout map with the geometry prior, while preserving pretrained representations. On an intervention-verified forest hazard benchmark, across ten independently trained adaptations, GAFT consistently outperforms frozen DINOv2 and supervised PEFT baselines, improving the repeated leave-one-scenario-out mean $F_2$ from 0.0607 to 0.3757 with statistical significance under paired analysis. Within these independently trained models, the best-performing GAFT model achieves a repeated-LOSO $F_2$ of 0.570. Code and benchmark: https://github.com/Xu-Yanran/geo_anchored_fine_tuning

Face recognition systems are increasingly deployed in security-critical applications, yet they remain vulnerable to presentation and morph attacks. Presentation attack detection (PAD) and morphing attack detection (MAD) are therefore essential components of trustworthy face biometrics. Despite advancements in PAD and MAD methods, existing detectors suffer from limited generalization and degrade in cross-dataset evaluation. In this paper, we systematically investigate whether general-purpose foundation models (FMs) and multimodal large language models (MLLMs) encode PAD-relevant and MAD-relevant information, and how such models can best be deployed for both tasks. We study five approaches with increasing access to the internal information of the model: (i) zero-shot prompting of off-the-shelf MLLMs; (ii) training a shallow model on the next-token logit probabilities at the output of the MLLM; (iii) parameter-efficient fine-tuning on task-specific question-answer data, yielding two specialized MLLMs, called PADLLM and MADLLM, which additionally provide textual reasoning for their decisions; (iv) linear probing of frozen vision encoders; and (v) fine-tuning of vision encoders of FMs and MLLMs. We benchmark 16 open-weight MLLMs and 30 vision encoder backbones on four PAD datasets (MSU-MFSD, CASIA-FASD, Replay-Attack, and OULU-NPU) and four MAD datasets (FFHQ, FRGC, FRLL, and FERET). Our experiments show that FMs and MLLMs can achieve significant performance for PAD and MAD. In addition, the fine-tuned models achieve state-of-the-art detection performance in cross-dataset evaluation, indicating that general-purpose pretrained representations carry substantial attack-relevant information. Source code of all our experiments will be publicly released.

We aim to improve frozen DINOv3 dense-prediction models under distribution shift by adding inference computation inside the visual backbone, without changing model weights, task adapters, or prediction heads. The challenge is that repeated transformer-block computation must refine dense features without disrupting the pairwise patch relations that DINOv3 uses to preserve spatial structure. We introduce GramLoop, a training-free framework that replays a short transformer window and controls each replay through final-layer cosine-Gram consistency. Each proposal is propagated through the frozen suffix, measured against the standard DINOv3 trajectory, and accepted through a patchwise gate at the replay-window endpoint. Across object detection and semantic segmentation under corruptions, perturbations, and natural shifts, GramLoop improves all five shifted benchmarks over the paired DINOv3 baseline. On COCO-O, it improves mAP by +0.252 and Effective Robustness by +0.250, while preserving clean ADE20K performance. Code will be released at https://github.com/cheyan9/GramLoop.

Foundation models are increasingly adapted for downstream medical imaging tasks, yet the influence of the chosen adaptation strategy on subgroup fairness remains poorly understood. We investigate how three parameter-efficient adaptation techniques, including linear heads on the raw CLS token, an MLP, and an attention-pooling module over multi-layer patch features, affect both pathology classification performance and subgroup disparities when applied to the frozen Rad-DINO chest X-ray encoder. Using MIMIC-CXR, we evaluate eight pathologies across race, sex, and imaging-view subgroups on a prevalence-preserving, demographically balanced test set, and additionally probe how strongly each adapter encodes protected attributes. We find that attention pooling achieves the strongest overall discriminative performance and encodes attributes, particularly race, most strongly, but that improved overall performance does not consistently reduce subgroup disparities. Notably, stronger attribute encoding did not correspond to larger disparities: early network layers encoded race most weakly yet produced the largest subgroup performance gaps. Exploring different attention-pooling layer combinations further revealed no consistent relationship between the layers pooled, attribute encoding strength, and subgroup fairness. Our results indicate that richer, more expressive representations can improve accuracy while leaving fairness implications task-dependent and unpredictable, which must be assessed directly and per-task rather than inferred from encoding strength or overall performance alone.

Spatio-temporal PV data are essential for understanding adoption processes in off-grid regions, yet such data remain largely unavailable. Automated segmentation of remote sensing (RS) imagery offers a promising solution; yet, residential PV systems remain challenging targets because of their small size and sparse distribution, resulting in severe target-background imbalance. Vision-language foundation models (FMs) provide a data-efficient paradigm through prompt-based semantic and spatial guidance, but the relative contribution of different prompt types remains unclear. We systematically evaluate SAM3 for small-scale PV segmentation in RS imagery by comparing textual, geometric, and hybrid prompting, under varying supervision levels, training strategies, spatial resolutions, and imaging conditions. Multi-temporal aerial imagery from a large off-grid rural region serves as a study site, with findings validated across three additional datasets. Prompting strategy emerged as the dominant factor governing model behavior. Textual prompting consistently produced the lowest performance and showed the greatest sensitivity to supervision and imaging conditions. In contrast, spatial guidance substantially improved both segmentation accuracy and robustness. Hybrid prompting achieved the highest accuracy and stability, indicating that semantic and spatial guidance provide complementary information. Most performance gains were achieved with only a few hundred annotated samples, demonstrating strong data efficiency. Transfer learning had limited overall impact, with only modest improvements observed for textual prompting under limited supervision. Overall, our findings establish prompting strategy as a key determinant of SAM3 adaptation, robustness, and generalization, highlighting the potential of promptable FMs for scalable PV mapping in data-constrained off-grid regions.

Recent Vision Foundation Models (VFMs) predict depth, camera pose, and pointmap in a single forward pass without per-scene optimization, achieving strong generalization. However, enforcing explicit multi-view geometric consistency, e.g., through bundle adjustment, is computationally costly and is thus not imposed during VFM pretraining, so such inconsistency can arise. To address this, implicit self-consistency derived from model outputs (e.g., pointmaps, features), though enforced at test-time in prior work, delivers inherently limited performance gain, especially on scenes where the pretrained VFM is highly inaccurate. In contrast to this implicit signal, we propose Self-Geometry, a plug-and-play test-time adaptation pipeline that directly imposes explicit multi-view geometric constraints using 2D pixel correspondences as pseudo ground-truth. Our proposed Self-Geometry consists of Geometric Disentanglement Optimization, which combines Multi-View Consistency and Epipolar Consistency losses with Gradient Disentanglement to prevent gradient conflict; Frame Angular-Neighbor, a view sampler based on SO(3) geodesic distances for lightly imposing these constraints; and Lightweight TTA, which adapts VFMs via LoRA. Our method achieves consistent improvements in both pose and geometry estimation across six VFMs (VGGT, $π^3$, DA3-Giant/Large/Base/Small) and four benchmarks (7Scenes, ETH3D, ScanNet++, HiRoom).

Leukemia cell image classification is challenged by real-world domain shifts from acquisition, staining, illumination, and site protocols, causing single-dataset models to generalize poorly in real clinical scenarios. This work presents a robust framework for leukemia classification across multiple heterogeneous datasets using a two-stage pipeline with a pretrained vision foundation model. Stage 1 performs binary classification (leukemia vs. non-leukemia) and is trained using 122,167 single-cell images. Stage 2 is conditionally applied to Stage 1 positives to perform subtype classification into Acute Lymphoblastic Leukemia (ALL) and Acute Myeloid Leukemia (AML), trained using 69,400 single-cell images. Labels are harmonized across five heterogeneous datasets to enable cross-dataset training, and performance is evaluated on a held-out dataset protocol to assess domain-shift generalization. Within this pipeline, three encoders are benchmarked (DinoBloom, pretrained on single-cell images; BiomedCLIP, pretrained on biomedical data; and CLIP as a general-purpose model) under linear probing, Low-Rank Adaptation (LoRA), and a Retrieval-Augmented Classification (RAC) module that retrieves the top-k most similar cell images to provide cytomorphological grounding. The objective is to quantify how much domain-specific pretraining contributes to performance under domain shift, and whether cost-effective adaptation and retrieval can be a viable alternative to expensive domain-specialized pretraining. The held-out protocol additionally serves as a diagnostic tool, revealing when classification performance is attributable to dataset-specific artifacts rather than to cytomorphological features.

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

Newly triggered landslides rarely carry immediate annotations, so cross-domain transferability determines the value of landslide mapping for emergency response and regional risk assessment. Vision foundation models have strengthened representational transfer, yet on unseen regions, events, and data sources they still generate high-confidence false alarms. Terrain, material, and rainfall triggering can constrain such errors, but their supports are local, regional, and event-scale, so that resampling onto a 10~m grid misaligns them with the segmentation decision unit and compounds the uncertain geographic context problem (UGCoP). We propose GeoPhysAdapter, which anchors on a frozen vision foundation model, restricts terrain, material, and triggering to dense spatial guidance, regional modulation, and event-timing forcing, and applies bounded adaptation at two decision units, the pixel and the candidate landslide body, reverting exactly to the visual prediction where support is insufficient. On an event-isolated PILD dataset of four public sources, 55 global landslide events, and 7,890 test samples, 70.3% of cross-domain false-positive mass lies in near-pure spurious bodies of median equivalent diameter 207m, matching coarse-prior support rather than the pixel. Pixel-level adaptation removes a net 507,817 erroneous pixels and reduces error by 7.76%, whereas raising the decision unit to the candidate body, under identical samples, anchor, and baseline, increases error reduction to 23.99%, approximately 3.1 times the pixel-level effect, improves IoU by 0.031 (14.2% relative), and corrects 9.92 pixels per pixel harmed. The data and code are publicly available at: https://github.com/Liu-Zhihang/geophysadapter.

Semi-supervised adaptation of vision foundation models (VFMs) commonly freezes the pretrained backbone and updates lightweight modules such as LoRA. However, pseudo-labels have mixed reliability, and a single LoRA adapter must absorb reliable, ambiguous, and noisy gradients in the same low-rank space. This can make VFM adaptation sensitive to pseudo-label noise. We propose \textbf{TriNoL}, a \textbf{Tri}ple-expert learning framework from \textbf{No}isy \textbf{L}abels for semi-supervised VFM adaptation. TriNoL routes unlabeled samples into three confidence regions and assigns them to three LoRA experts: a Positive Expert for high-confidence pseudo-labels, an Alignment Expert for medium-confidence ambiguous samples, and a Negative Expert for low-confidence noisy samples. The VFM backbone remains frozen, and only the LoRA experts and classifier head are updated. By separating different pseudo-label reliability regions into specialized adaptation paths, TriNoL improves robustness to noisy supervision while keeping the training cost low.

Parameter-Efficient Fine-Tuning (PEFT) has become the de facto standard for adapting Vision Transformers (ViTs) to downstream tasks. While parameter count has been the dominant efficiency metric in PEFT, it does not imply \textit{compute efficiency}: parameter-sparse methods can still incur full-model training cost per step, and typically need long schedules to reach peak accuracy. We introduce Circuit Fine-Tuning (CFT), a compute-efficient framework that uses circuit discovery---conventionally used to explain trained models---to select modules for fine-tuning before training. Whereas attribution is conventionally formulated against a trained task head, we formulate it against a near-zero-initialized probe head, which isolates the response of the backbone to the target distribution rather than the preferences of a particular classifier. CFT then fine-tunes only the recovered subgraph. CFT needs no learning-rate warmup and reaches peak accuracy in ${\sim}20$ epochs on average---versus $44$--$96$ for strong PEFT baselines---yielding $2.3$--$6.6\times$ fewer training FLOPs and up to $16\times$ less wall-clock time, while adding zero parameters and no inference operations. Experiments across a standard visual transfer benchmark (VTAB-1k), hierarchical backbones (Swin), domain-shifted medical imaging (CBIS-DDSM), and a vision-language model (Gemma-3 on CUB-200) demonstrate the effectiveness of CFT. Code is available at https://github.com/UriKialy/CFT

Vision foundation models (VFMs) such as DINO are pretrained for single-image representation, whereas remote sensing change detection requires reasoning over a bi-temporal pair. Existing VFM-based methods usually encode the two images independently and compare them only afterward, leaving the VFM backbone unaware of cross-temporal relations. To bridge this mismatch, we present AdaDINO, a pair-aware in-backbone adaptation framework that equips a frozen DINO encoder with bi-temporal interaction for efficient change detection. Its core component, Change-aware Gated Local Adaptation (CGLA), couples the two streams after selected frozen blocks and injects a shared temporal residual into them with opposite signs, enhancing genuine change responses while preserving the pair midpoint. Batch-Shared Chunk Selection (BSCS) further reduces feed-forward network (FFN) computation by retaining a batch-shared subset of channel chunks that can be executed as a compact dense FFN. A CGLA-Prior-Guided Refinement (CPGR) decoder reuses encoder-side change responses for coarse-to-fine prediction. Experiments on four remote sensing change detection benchmarks show that AdaDINO achieves competitive or superior performance against VFM-based baselines, with the largest gain on the category-agnostic SYSU-CD dataset. With 62.5% of the FFN hidden width removed, AdaDINO still achieves an F1 score of 85.29% on SYSU-CD while delivering a 1.41$\times$ throughput speedup. The code will be released.

Parameter-efficient fine-tuning enables the adaptation of vision foundation models to biomedical tasks under limited computational resources, but a single low-rank update can constrain all task-specific changes to one narrow parameter subspace. This restriction may prevent the model from simultaneously representing globally shared task structure and localized residual directions required for generalization to unseen imaging domains. We introduce LoRSA, a global--residual adaptation framework that jointly learns a dense low-rank component and a dynamically structured-sparse low-rank component. The dense component captures globally coordinated task adaptation, while the structured component provides complementary residual corrections whose support evolves during training. We characterize the representational capacity, approximation properties, rank structure, and singular-subspace complementarity of this decomposition. We evaluate LoRSA for four-class breast-density classification using DINOv3-Base, with VinDr-Mammo as the source domain and MammosighTR and RSNA as unseen external domains. LoRSA remains competitive on the internal validation set and achieves the best external macro-F1 on both target datasets, improving upon the strongest competing method by 2.15 percentage points on MammosighTR and 3.09 percentage points on RSNA. Weight-matrix analysis further shows that approximately $92\%$ of the energy of each adaptation component lies outside the bilateral singular subspace of the other, indicating that the two components learn largely complementary update directions. These results suggest that organizing adaptation capacity into distinct global and residual paths can improve the external-domain generalization of parameter-efficiently adapted biomedical vision models.

Generative traffic video forecasting aims to synthesize long-horizon, temporally coherent future videos of traffic scenes from a short observation history and textual descriptions. In this paper, we present CosmosAlign, a generative traffic video forecasting framework built upon the pretrained Cosmos3-Nano world foundation model. Our approach is motivated by the observation that successfully adapting large pretrained world models to downstream forecasting tasks depends primarily on distribution alignment rather than increased model capacity. To this end, we propose a two-stage LoRA adaptation strategy that first aligns the conditioning-mode distribution with the target forecasting task, and then aligns the training captions with the model's native structured prompting interface through an LLM-based re-captioning pipeline. During inference, we further improve prediction quality using a fully training-free procedure consisting of consensus-based medoid sample selection and motion-adaptive blending of static scene regions. CosmosAlign achieves a final score of 76.49 on the AI City Challenge 2026 Track 5 benchmark, ranking first on the final leaderboard. Our code is publicly available at https://quangminhdinh.github.io/CosmosAlign/.

Foundation models have shown strong transferability in cardiac MRI (CMR), but their effectiveness for heterogeneous multi-view and multi-sequence CMR analysis remains unclear. In this work, we explore the effectiveness of fine-tuning and combining different CMR foundation models for the Universal Multi-Sequence, Multi-Center and Multi-View CMR Segmentation (CMR-Multi) Challenge. CineMA was fine-tuned for cine and late gadolinium enhancement (LGE) segmentation across short-axis and long-axis views. For direct left-ventricular ejection fraction (LVEF) estimation, we used two recent frozen CMR foundation models to extract embedding vectors that were then combined using attention-based multiple-instance learning for LVEF regression. In the challenge validation set, cine segmentation achieved Dice scores of 0.862, 0.883, and 0.902 for short-axis, two-chamber and four-chamber cine MRI, respectively. LGE segmentation achieved Dice scores between 0.621 and 0.846 across views. The direct LVEF regression model achieved an MAE of 4.96 percentage points and a Pearson correlation of 0.91. These results indicate that foundation models can be effectively adapted and combined for multi-view CMR analysis, while accurate LGE scar segmentation remains a challenging task.

Prompt tuning, a leading parameter-efficient adaptation paradigm in NLP, has recently been extended to computer vision. Visual prompt tuning (VPT) adapts pre-trained vision transformers (ViTs) by updating a small set of additional prompt parameters. However, existing visual prompts are randomly initialized and do not exploit prior knowledge, such as instructions in NLP. We address this gap by injecting two complementary semantic priors into VPT. Fundamental image priors, including color, texture, and shape, are extracted with classical hand-crafted operators and injected into the input space, while self-attention maps provide instance-aware semantics in the feature space. We further propose a cascaded scheme that integrates both priors throughout ViT adaptation. Experiments on 34 challenging image classification datasets demonstrate superior downstream adaptation while tuning only 0.74% of ViT parameters. Project page: https://xixiaouab.github.io/Cascaded-Semantics/.

Fossil leaves are rarely preserved whole -- sedimentary rock hides, breaks, and erodes the lamina, yet paleobotany depends on the complete shape and outline of the leaf. We cast the recovery of the missing tissue as amodal reconstruction and present AmodalDINO, a multi-head dense-prediction model that predicts four masks from a single RGB image: visible leaf, amodal complete leaf, amodal main vein, and fine veins. Unlike essentially all prior amodal work, AmodalDINO is given no visible mask. It predicts the visible and amodal regions jointly, so it needs no upstream instance segmenter at runtime. Two simple but effective changes adapt the model to the amodal segmentation task: fully fine-tune a DINOv3 ViT-L/16 at a small learning rate instead of freezing it, and attach auxiliary venation heads alongside the leaf heads. These two changes enable the model to learn the structural shape prior of leaves. Trained only on synthetic leaf fossil images, AmodalDINO reaches 95.0% Dice / 90.5% IoU on the validation set and transfers well to real fossil specimens. Stripped to two heads, the same recipe can run on two benchmark datasets, reaching 85.05 full mIoU / 66.65 occluded mIoU on KINS and 80.90 / 38.15 on COCOA-cls. The model is also practical: by quantizing to 4-bit weights, it runs entirely offline in a browser, matching the original model with an IoU of 0.910. We also add ruler-based calibration to estimate surface area, and a generative visualization of living leaves on local devices.

Self-supervised Vision Foundation Models (VFMs) have become essential backbones for downstream tasks due to their strong and transferable visual representations. However, their patch-token-level features are often too coarse for dense prediction tasks such as semantic segmentation and depth estimation when accurate fine-grained predictions are required. Feature upsampling methods have been developed to recover pixel-level detail but still face limitations. Learnable upsamplers are often designed for a specific encoders and must be retrained for different encoders. Image-guided methods that use shallow pixel encoders often introduce textural artifacts and lack the semantic guidance needed for accurate downstream predictions. We introduce PixelUp, a zero-shot VFM-agnostic upsampler achieving semantic awareness through a coarse-to-fine chain of windowed cross-attention architecture guided by multi-scale semantic features. We demonstrate that PixelUp outperforms both VFM-specific and VFM-agnostic upsamplers, achieving state-of-the-art performance on dense prediction tasks with an average improvement of +1.2 mIoU on semantic segmentation and +0.25 $δ_1$, on NYUv2 depth estimation across VFMs. PixelUp further improves training-free open-vocabulary and unsupervised semantic segmentation by an average of +1.3 mIoU and +0.5 mIoU, respectively. Code available at https://pixelup-project.vercel.app/

Deploying vision-language models (VLMs) for safety-critical spatial reasoning on resource-constrained autonomous driving platforms requires both compact model size and reliable metric grounding. We present MoRAL (Multimodal Reasoning for Autonomous Language Models), a two-stage fine-tuning pipeline that teaches Cosmos-Reason2-2B to first read a physics-encoded Bird's Eye View (BEV) representation and then reason over it for driving decisions. The BEV image encodes LiDAR metric distance as color bands, object class as cluster morphology, and radar Doppler velocity as directional wedge overlays, externalizing spatial perception into the input image so that no learned 3D backbone is required at inference. Stage 1 fine-tunes the vision encoder on 60,000 grounding records; zero-shot baselines produce no parseable BEV outputs, confirming the vocabulary requires explicit training. Stage 2 fine-tunes the full model (52M parameters, 2.4% of total) on 57,696 chain-of-thought records generated by Cosmos-Reason2-8B as teacher, spanning eight driving question types. On 2,304 held-out nuScenes frames evaluated by Gemma 4 (31B) calibrated against human review, MoRAL wins seven of eight question types over a zero-shot 8B baseline despite using four times fewer parameters, with the largest margins on question types requiring structured multi-step physics reasoning. Emergency braking recall improves from 10.8% to 47.8%, output degeneration falls from 94.1% to 20.8%, and the full pipeline fits a consumer 8 GB GPU at 42 tok/s without quantization. These results establish a reproducible foundation for compact, physics-grounded VLM reasoning on mobile edge platforms.

Monocular depth foundation models, benefiting from large-scale synthetic training data, have demonstrated strong generalization. However, they often hallucinate depth on non-Lambertian surfaces, estimating reflected content in mirrors or transmitted content behind glass rather than the physical surface itself. Adapting these models with real-world data is challenging because conventional depth sensors are also unreliable in such regions. We observe that while the appearance of a non-Lambertian surface varies with its reflected or transmitted environment, its underlying geometry remains unchanged. Based on this observation, we propose GIFT (Geometry-Invariant Fine-Tuning), a parameter-efficient post-training framework that requires no measured depth labels. We collect groups of RGB images under controlled appearance changes while keeping the camera and target geometry fixed. GIFT exploits geometric invariance across these observations to suppress non-Lambertian depth hallucinations while retaining general depth estimation capability. We further construct a controlled benchmark that evaluates non-Lambertian depth recovery, robustness to appearance changes, and performance retention in other regions. Experiments on our benchmark and an independent real-world dataset demonstrate that GIFT improves depth prediction for mirrors and transparent objects while largely preserving the base model's performance, providing a practical and low-cost approach for adapting monocular depth foundation models to non-Lambertian scenes.

Latents from vision foundation models (VFMs) are semantically rich and well suited for visual understanding. Recent representation autoencoder methods such as RAE have shown that they can provide promising latent spaces for image generation. However, VFM latents remain difficult to model directly: DiT-generated latents exhibit spectral mismatch with encoder latents, especially in high-frequency components. Our channel-wise spectral analysis further reveals that these high-frequency components are diffusely distributed across latent channels and entangled with semantic information, making the latent space difficult for DiT to model. To address these challenges, we propose SPAE, latent adaptation framework for generation. Specifically, SPAE employs a compact bottleneck to distill stable semantic information while suppressing high-frequency components, thereby improving the alignment between DiT-generated latents and encoder latents. In addition, we apply a channel-wise masking strategy to promote the decoupling of semantic information and high-frequency details across bottleneck channels. Experiments show that SPAE achieves a favorable balance among visual understanding, generation quality, and reconstruction fidelity.

Modern machine learning pipelines increasingly rely on reusing pretrained and foundation models across downstream tasks. These pretrained models can differ not only in performance but also in how they can be used: some only provide black-box predictions, while others may permit white-box access to internal representations that can be probed or fine-tuned. When deployed to the target domain in the presence of distribution shift, no single strategy, including zero-shot application, fine-tuning, or directly training a target-specific model, is uniformly the best. In this work, we propose Frontier Learning, a framework that treats a library of candidate models spanning different training histories and access regimes as complementary sources of information rather than mutually exclusive alternatives. Frontier Learning constructs a unified target-domain feature by concatenating internal representations from white-box candidates as well as prediction outputs from black-box candidates, then fits a lightweight, regularized supervised learner on this concatenated representation using labeled target data. Because the resulting hypothesis class contains predictors obtained by zero-shot reuse, fine-tuning, and direct training as special cases, empirical risk minimization over the frontier learner is guaranteed to be no worse, on the training sample, than any individual baseline. We evaluate the framework in simulations spanning varying degrees of source-target compatibility and in two real-world distribution-shift settings: visual domain adaptation on DomainNet/VisDA and clinical mortality prediction across intensive care unit domains using MIMIC-IV-Notes. Across all settings, Frontier Learning matches or outperforms the strongest individual reuse strategy, with the largest gains arising precisely when no single baseline is reliable across the range of shift considered.

Despite the widespread adoption of foundation models as feature extractors for medical imaging, relatively little is understood about how different pretraining strategies influence the transferability of learned representations to weakly supervised ophthalmic imaging tasks. We investigate this question in ultra-widefield (UWF) retinal imaging by evaluating foundation model representations within a patch-based multiple instance learning (MIL) framework for disease classification on UWF images. We compare Vision Transformer encoders pretrained with supervised, Masked Autoencoder (MAE), and self-distillation objectives, while keeping the downstream aggregation architecture unchanged. Within a controlled comparison of ViT-B encoders pretrained on ImageNet-1k, the choice of pretraining objective substantially influenced frozen representation transfer, with supervised and self-distillation-based models outperforming MAE. A contemporary DINOv3 model pretrained at a larger scale achieved the strongest overall performance, with a quadratic weighted kappa of 0.863 for five-class diabetic retinopathy grading, comparable with DINOv1. Attention analysis further revealed distinct patch-aggregation behaviours associated with the different pretrained representations, while partial fine-tuning substantially reduced the performance gap for MAE. These findings suggest that pretraining strategy influences both representation transferability and the subsequent aggregation of patch-level evidence within MIL, resulting in differences in downstream classification performance.

Existing cross-domain few-shot segmentation approaches suffer from high training costs due to source-domain episodic training and pixel-wise dense prediction, while often producing fragmented and noisy predictions. To overcome these issues, we propose a training-free entity-level few-shot segmentation framework for remote sensing images with advection refinement. Specifically, we first leverage SAM3's generic geometric priors to generate category-agnostic entity primitives. By reformulating few-shot inference from pixel-level prediction to entity-level reasoning, foreground and background prototypes are constructed and combined with dense textual semantic responses from SAM3 to build a multi-modal semantic potential field. Furthermore, an advection equation-based semantic refinement mechanism is introduced to propagate category-aware information across both feature and similarity spaces, enhancing semantic continuity and suppressing local texture noise. Extensive experiments on multiple remote sensing datasets demonstrate that the proposed framework effectively mitigates domain shift and local noise, substantially improving SAM3's adaptation capability for remote sensing few-shot segmentation without additional training. Our code will be publicly available at https://github.com/yu-ni1989/ELFSS-AR.

Remote sensing scene classification is increasingly relying on foundation models pre-trained on large-scale Earth-observation data. Moreover, transductive inference, which exploits the collective statistical structure of the entire unlabeled query set, appears to naturally match remote sensing pipelines where large images are routinely split into patches and inferred as a batch. In this work, we introduce LC-TIM (Locally Consistent Transductive Information Maximization), which extends the state-of-the-art Transductive Information Maximization for Few-Shot CLIP (TIM++) objective with a local consistency regularizer that enforces prediction agreement between each query sample and its $κ$ nearest feature-space neighbors. The regularizer enters as a single multiplicative factor in the closed-form $q$-update, adding negligible computational overhead. We further propose a multi-source extension that fuses the affinity graph from multiple remote sensing foundation model, further boosting classification accuracy. To assess these methods, we establish the first comprehensive, open-source benchmark for transductive few-shot RS scene classification, evaluating LP++, TransCLIP, TIM++, and LC-TIM across ten diverse datasets, two remote sensing vision-language models, and across various few-shot settings. Our experiments show that transductive methods consistently outperform zero-shot baselines, and that LC-TIM achieves state-of-the-art accuracy, with the largest gains in the low-shot regime where neighborhood cues are most informative. Code is publicly available at: https://github.com/elkhouryk/LC-TIM

Existing methods for adapting 2D foundation models such as SAM to 3D volumes either process slices independently---ignoring inter-slice context---or require substantial architectural changes and retraining. In this paper, we present \textbf{SAM+D}, a parameter-efficient framework that lifts SAM-family models by one spatial dimension---enabling 3D volumetric segmentation from 2D SAM and, for the first time via parameter-efficient fine-tuning, end-to-end 4D (3D+T) spatiotemporal segmentation from video-based SAM2---while keeping the vast majority of pre-trained parameters frozen. SAM+D introduces two lightweight, model-agnostic modules into frozen transformer blocks: (1)\textbf{Depth-Routed LoRA (DRLoRA)} experts with learned routing for spatially adaptive low-rank updates, and (2)\textbf{Depth Shift Modules (DSM)} for cross-slice feature exchange at zero additional parameter cost. Together, they provide volume-level context while tuning only ${\sim}$2.8% of parameters for SAM and ${\sim}$3.7% for SAM2. We evaluate SAM+D in two distinct settings, each lifting the base model by one spatial dimension: 3D segmentation, where SAM(2D$\,\to\,$3D) is evaluated on four CT benchmarks (KiTS, Pancreas, LiTS, Colon), and 4D segmentation, where SAM2 (2D+T$\,\to\,$3D+T) is evaluated on a cell tracking challenge (CTC) dataset (Fluo-N3DH-SIM+). In both settings SAM+D achieves competitive or superior results under the single-point prompt setting while using fewer trainable parameters than existing methods, demonstrating that SAM+D generalizes across SAM-family architectures, target dimensionalities (3D, 4D), and domains spanning medical imaging and bio-scene understanding. Code is publicly available at https://github.com/JerrySongCST/SAM-Plus-D.

Anterior eye segment (AES) segmentation is a key component of both ocular biometrics and emerging clinical image analysis applications. However, heterogeneous acquisition conditions and limited annotations in medical settings hinder the robustness and generalization of existing methods. Foundation models (FMs) such as DINOv3 offer strong transfer capabilities, but efficiently adapting their representations to dense prediction tasks remains challenging. In this study, we investigate robust AES segmentation in clinical settings, and propose a lightweight architecture built upon a distilled DINOv3 ViT-Small backbone. We introduce a step-attention feature refinement module that progressively adapts multi-level transformer representations before convolutional decoding, enabling efficient exploitation of pretrained features with few parameters. We evaluate the proposed approach on a private dataset of 333 clinically acquired AES images spanning eight ophthalmic acquisition protocols and annotated for seven anatomical classes. Compared with convolutional and transformer-based baselines, including DINOv3-based methods, our approach achieves the best overall performance, reaching 85.55% mIoU when fully fine-tuned. It also demonstrates the strongest robustness to domain shift across four unseen public AES segmentation datasets. These results establish a strong baseline for robust AES segmentation in clinical settings and highlight the importance of decoder design for effectively adapting FMs representations to medical segmentation tasks.