Vision Transformer

Vision Transformers (ViTs) and their variants have become state-of-the-art in many computer vision tasks and are widely used as backbones in large-scale vision and vision-language foundation models. While substantial research has focused on architectural improvements, the impact of patch size, a crucial initial design choice in ViTs, remains underexplored, particularly in medical domains where both two-dimensional (2D) and three-dimensional (3D) imaging modalities exist. In this study, using 12 medical imaging datasets from various imaging modalities (including seven 2D and five 3D datasets), we conduct a thorough evaluation of how different patch sizes affect ViT classification performance. Using a single graphical processing unit (GPU) and a range of patch sizes (1, 2, 4, 7, 14, 28), we fine-tune ViT models and observe consistent improvements in classification performance with smaller patch sizes (1, 2, and 4), which achieve the best results across nearly all datasets. More specifically, our results indicate improvements in balanced accuracy of up to 12.78% for 2D datasets (patch size 2 vs. 28) and up to 23.78% for 3D datasets (patch size 1 vs. 14), at the cost of increased computational expense. Moreover, by applying a straightforward ensemble strategy that fuses the predictions of the models trained with patch sizes 1, 2, and 4, we demonstrate a further boost in performance in most cases, especially for the 2D datasets. Our implementation is publicly available on GitHub: https://github.com/HealMaDe/MedViT

Vision Transformers, ViTs, have emerged as a powerful alternative to convolutional neural networks, CNNs, in a variety of image-based tasks. While CNNs have previously been evaluated for their ability to perform graphical perception tasks, which are essential for interpreting visualizations, the perceptual capabilities of ViTs remain largely unexplored. In this work, we investigate the performance of ViTs in elementary visual judgment tasks inspired by the foundational studies of Cleveland and McGill, which quantified the accuracy of human perception across different visual encodings. Inspired by their study, we benchmark ViTs against CNNs and human participants in a series of controlled graphical perception tasks. Our results reveal that, although ViTs demonstrate strong performance in general vision tasks, their alignment with human-like graphical perception in the visualization domain is limited. This study highlights key perceptual gaps and points to important considerations for the application of ViTs in visualization systems and graphical perceptual modeling.

Vision Transformers rely on positional embeddings and class tokens encoding fixed spatial priors. While effective for natural images, these priors may be suboptimal when spatial layout is weakly informative, a frequent condition in medical imaging. We introduce ZACH-ViT (Zero-token Adaptive Compact Hierarchical Vision Transformer), a compact Vision Transformer that removes positional embeddings and the [CLS] token, achieving permutation-invariant patch processing via global average pooling. Zero-token denotes removal of the dedicated aggregation token and positional encodings. Patch tokens remain unchanged. Adaptive residual projections preserve training stability under strict parameter constraints. We evaluate ZACH-ViT across seven MedMNIST datasets under a strict few-shot protocol (50 samples/class, fixed hyperparameters, five seeds). Results reveal regime-dependent behavior: ZACH-ViT (0.25M parameters, trained from scratch) achieves strongest advantage on BloodMNIST and remains competitive on PathMNIST, while relative advantage decreases on datasets with stronger anatomical priors (OCTMNIST, OrganAMNIST), consistent with our hypothesis. Component and pooling ablations show positional support becomes mildly beneficial as spatial structure increases, whereas reintroducing a [CLS] token is consistently unfavorable. These findings support that architectural alignment with data structure can outweigh universal benchmark dominance. Despite minimal size and no pretraining, ZACH-ViT achieves competitive performance under data-scarce conditions, relevant for compact medical imaging and low-resource settings. Code: https://github.com/Bluesman79/ZACH-ViT

Existing online video segmentation models typically combine a per-frame segmenter with complex specialized tracking modules. While effective, these modules introduce significant architectural complexity and computational overhead. Recent studies suggest that plain Vision Transformer (ViT) encoders, when scaled with sufficient capacity and large-scale pre-training, can conduct accurate image segmentation without requiring specialized modules. Motivated by this observation, we propose the Video Encoder-only Mask Transformer (VidEoMT), a simple encoder-only video segmentation model that eliminates the need for dedicated tracking modules. To enable temporal modeling in an encoder-only ViT, VidEoMT introduces a lightweight query propagation mechanism that carries information across frames by reusing queries from the previous frame. To balance this with adaptability to new content, it employs a query fusion strategy that combines the propagated queries with a set of temporally-agnostic learned queries. As a result, VidEoMT attains the benefits of a tracker without added complexity, achieving competitive accuracy while being 5x-10x faster, running at up to 160 FPS with a ViT-L backbone. Code: https://www.tue-mps.org/videomt/

Recent advances in foundation video generators such as Sora2, Veo3, and other commercial systems have produced highly realistic synthetic videos, exposing the limitations of existing detection methods that rely on shallow embedding trajectories, image-based adaptation, or computationally heavy MLLMs. We propose EA-Swin, an Embedding-Agnostic Swin Transformer that models spatiotemporal dependencies directly on pretrained video embeddings via a factorized windowed attention design, making it compatible with generic ViT-style patch-based encoders. Moreover, we construct the EA-Video dataset, a benchmark dataset comprising 130K videos that integrates newly collected samples with curated existing datasets, covering diverse commercial and open-source generators and including unseen-generator splits for rigorous cross-distribution evaluation. Extensive experiments show that EA-Swin achieves 0.97-0.99 accuracy across major generators, outperforming prior SoTA methods (typically 0.8-0.9) by a margin of 5-20%, while maintaining strong generalization to unseen distributions, establishing a scalable and robust solution for modern AI-generated video detection.

Vision Transformers (ViTs) have achieved remarkable success across various vision tasks, yet their deployment is often hindered by prohibitive computational costs. While structured weight pruning and token compression have emerged as promising solutions, they suffer from prolonged retraining and inter-layer dependencies that complicate optimization, respectively. We propose ToaSt, a decoupled framework applying specialized strategies to distinct ViT components. We apply coupled head-wise structured pruning to Multi-Head Self-Attention modules, leveraging attention operation characteristics to enhance robustness. For Feed-Forward Networks (over 60% of FLOPs), we introduce Token Channel Selection (TCS), a training-free method that filters redundant noise channels at inference time. Extensive evaluations across nine diverse models, including DeiT, ViT-MAE, and Swin Transformer, demonstrate that ToaSt achieves superior trade-offs between accuracy and efficiency, consistently outperforming existing baselines. On ViT-MAE-Huge, ToaSt achieves 88.52% accuracy (+1.64%p) with 39.4% FLOPs reduction. ToaSt also transfers effectively to diverse downstream tasks (COCO detection, ADE20K segmentation, CIFAR-100 classification), achieving 52.2 versus 51.9 mAP on COCO. Code: github.com/SHANNonLab-HUFS/ToaSt

General-purpose super-resolution models, particularly Vision Transformers, have achieved remarkable success but exhibit fundamental inefficiencies in common infrared imaging scenarios like surveillance and autonomous driving, which operate from fixed or nearly-static viewpoints. These models fail to exploit the strong, persistent spatial priors inherent in such scenes, leading to redundant learning and suboptimal performance. To address this, we propose the Regional Prior attention Transformer for infrared image Super-Resolution (RPT-SR), a novel architecture that explicitly encodes scene layout information into the attention mechanism. Our core contribution is a dual-token framework that fuses (1) learnable, regional prior tokens, which act as a persistent memory for the scene's global structure, with (2) local tokens that capture the frame-specific content of the current input. By utilizing these tokens into an attention, our model allows the priors to dynamically modulate the local reconstruction process. Extensive experiments validate our approach. While most prior works focus on a single infrared band, we demonstrate the broad applicability and versatility of RPT-SR by establishing new state-of-the-art performance across diverse datasets covering both Long-Wave (LWIR) and Short-Wave (SWIR) spectra

Vision Transformers (ViTs) have emerged as the state-of-the-art architecture in representation learning, leveraging self-attention mechanisms to excel in various tasks. ViTs split images into fixed-size patches, constraining them to a predefined size and necessitating pre-processing steps like resizing, padding, or cropping. This poses challenges in medical imaging, particularly with irregularly shaped structures like tumors. A fixed bounding box crop size produces input images with highly variable foreground-to-background ratios. Resizing medical images can degrade information and introduce artefacts, impacting diagnosis. Hence, tailoring variable-sized crops to regions of interest can enhance feature representation capabilities. Moreover, large images are computationally expensive, and smaller sizes risk information loss, presenting a computation-accuracy tradeoff. We propose VariViT, an improved ViT model crafted to handle variable image sizes while maintaining a consistent patch size. VariViT employs a novel positional embedding resizing scheme for a variable number of patches. We also implement a new batching strategy within VariViT to reduce computational complexity, resulting in faster training and inference times. In our evaluations on two 3D brain MRI datasets, VariViT surpasses vanilla ViTs and ResNet in glioma genotype prediction and brain tumor classification. It achieves F1-scores of 75.5% and 76.3%, respectively, learning more discriminative features. Our proposed batching strategy reduces computation time by up to 30% compared to conventional architectures. These findings underscore the efficacy of VariViT in image representation learning. Our code can be found here: https://github.com/Aswathi-Varma/varivit

Rapid and accurate building damage assessment in the immediate aftermath of tornadoes is critical for coordinating life-saving search and rescue operations, optimizing emergency resource allocation, and accelerating community recovery. However, current automated methods struggle with the unique visual complexity of tornado-induced wreckage, primarily due to severe domain shift from standard pre-training datasets and extreme class imbalance in real-world disaster data. To address these challenges, we introduce a systematic experimental framework evaluating 79 open-source deep learning models, encompassing both Convolutional Neural Networks (CNNs) and Vision Transformers, across over 2,300 controlled experiments on our newly curated Quad-State Tornado Damage (QSTD) benchmark dataset. Our findings reveal that achieving operational-grade performance hinges on a complex interaction between architecture and optimization, rather than architectural selection alone. Most strikingly, we demonstrate that optimizer choice can be more consequential than architecture: switching from Adam to SGD provided dramatic F1 gains of +25 to +38 points for Vision Transformer and Swin Transformer families, fundamentally reversing their ranking from bottom-tier to competitive with top-performing CNNs. Furthermore, a low learning rate of 1x10^(-4) proved universally critical, boosting average F1 performance by +10.2 points across all architectures. Our champion model, ConvNeXt-Base trained with these optimized settings, demonstrated strong cross-event generalization on the held-out Tuscaloosa-Moore Tornado Damage (TMTD) dataset, achieving 46.4% Macro F1 (+34.6 points over baseline) and retaining 85.5% Ordinal Top-1 Accuracy despite temporal and sensor domain shifts.

Quark-gluon discrimination provides a useful test case for studying how different machine-learning architectures learn the spatial structure of QCD radiation. In this work, we compare convolutional neural network (CNN), Vision Transformers (ViT), and hierarchical Swin Transformers using the same three-channel jet-image representation, consisting of charged-particle momentum, neutral-particle momentum, and charged-particle multiplicity from PYTHIA 8 jets. We study their performance for different training-set sizes and fine-tuning configurations, with particular attention to the role of local and global information in the jet images. CNN and Swin models consistently perform better than ViT in the cases studied. Since both CNN and Swin retain a strong local component in their architectures, this suggests that local jet substructure plays an important role in quark-gluon discrimination. The performance of the hierarchical Swin model also suggests that combining local features over larger spatial scales is useful. Block-wise fine-tuning improves the performance of the Transformer models, although the improvement becomes smaller and the training less stable as more blocks are unfrozen. We also find that self-supervised Momentum Contrast (MoCo) pretraining improves the model initialization, particularly when the amount of labeled training data is limited. Based on these observations, we developed a smaller Swin model adopted to the jet-image representation used in this study. It achieves comparable performance with substantially fewer parameters. The results show that it is important to adapt the model architecture and training procedure to the specific input characteristics of High Energy Physics (HEP) data when applying vision models in HEP.

With the rise of large-scale foundation models, efficiently adapting them to downstream tasks remains a central challenge. Linear probing, which freezes the backbone and trains a lightweight head, is computationally efficient but often restricted to last-layer representations. We show that task-relevant information is distributed across the network hierarchy rather than encoded solely in the last layers. To leverage this distribution of information, we apply an attentive probing mechanism that dynamically fuses representations from all layers of a Vision Transformer. This attentive layer fusion (ALF) learns to identify the most relevant layers for a target task and combines low-level structural cues with high-level semantic abstractions. Across 20 diverse datasets and multiple pretrained foundation models, ALF achieves consistent, substantial gains over standard linear probes. Attention heatmaps further reveal that tasks different from the pre-training domain benefit most from intermediate representations. Overall, our findings underscore the value of intermediate layers and demonstrate a principled, task-aware approach for unlocking their potential for probing-based adaptation.

Vision Transformers (ViTs) lack the hierarchical inductive biases inherent to Convolutional Neural Networks (CNNs), theoretically allowing them to maintain high-dimensional representations throughout all layers. However, recent observations suggest ViTs often spontaneously manifest a "U-shaped" entropy profile-compressing information in middle layers before expanding it for the final classification. In this work, we demonstrate that this "Inductive Bottleneck" is not an architectural artifact, but a data-dependent adaptation. By analyzing the layer-wise Effective Encoding Dimension (EED) of DINO-trained ViTs across datasets of varying compositional complexity (UC Merced, Tiny ImageNet, and CIFAR-100), we show that the depth of the bottleneck correlates strongly with the semantic abstraction required by the task. We find that while texture-heavy datasets preserve high-rank representations throughout, object-centric datasets drive the network to dampen high-frequency information in middle layers, effectively "learning" a bottleneck to isolate semantic features.

Convolutional Neural Networks and Vision Transformers are the two dominant architectural families in computer vision, defined by spatially local convolution and global self-attention respectively. Despite their apparent differences, we show that both operations are special cases of a single $k$-nearest neighbor aggregation framework: convolution selects neighbors by spatial proximity while attention selects by feature similarity, placing them at two ends of a shared operational spectrum. We introduce Convolutional Nearest Neighbors (ConvNN), a unified framework that exactly recovers standard and depthwise convolution, self-attention, and sparse attention variants including KVT-attention as special cases, and exposes the design space of neighbor-selection strategies between them through configurable similarity functions, positional encodings, and aggregation kernels. We validate ConvNN on ImageNet-1K classification across two complementary architectures: a hybrid branching layer in ResNet-50 that combines local and global feature learning, improving top-1 accuracy by 3.0% over the ResNet-50 baseline, and ConvNN-attention in ViT-Base that achieves 81.64% top-1 accuracy, surpassing standard multi-head self-attention by 0.7%. Together, these results demonstrate that ConvNN provides a principled foundation for designing operations that bridge convolutional and attention-based computation.

Model inversion is a widely adopted technique in data-free learning that reconstructs synthetic inputs from a pretrained model through iterative optimization, without access to original training data. Unfortunately, its application to state-of-the-art Vision Transformers (ViTs) poses a major computational challenge, due to their expensive self-attention mechanisms. To address this, Sparse Model Inversion (SMI) was proposed to improve efficiency by pruning and discarding seemingly unimportant patches, which were even claimed to be obstacles to knowledge transfer. However, our empirical findings suggest the opposite: even randomly selected patches can eventually acquire transferable knowledge through continued inversion. This reveals that discarding any prematurely inverted patches is inefficient, as it suppresses the extraction of class-agnostic features essential for knowledge transfer, along with class-specific features. In this paper, we propose Patch Rebirth Inversion (PRI), a novel approach that incrementally detaches the most important patches during the inversion process to construct sparse synthetic images, while allowing the remaining patches to continue evolving for future selection. This progressive strategy not only improves efficiency, but also encourages initially less informative patches to gradually accumulate more class-relevant knowledge, a phenomenon we refer to as the Re-Birth effect, thereby effectively balancing class-agnostic and class-specific knowledge. Experimental results show that PRI achieves up to 10x faster inversion than standard Dense Model Inversion (DMI) and 2x faster than SMI, while consistently outperforming SMI in accuracy and matching the performance of DMI.

Spatial transformations such as rotation and scale obscure the morphological cues needed for accurate image classification. Careful consideration is required for reliable use in high stakes settings. A model should stay robust under such transformations, expose why a correction was applied, and signal when its input is ambiguous. While geometrically equivariant architectures provide a mathematically grounded solution, they often limit model flexibility through strict symmetry constraints and incur significant computational overhead. Spatial Transformer Networks (STNs) offer a data-driven, flexible alternative for learning pseudo-equivariances to affine transformations. However, STNs have historically been restricted to convolutional architectures and suffer from training instability. To address this, we introduce a novel STN framework. It leverages the global modeling capabilities of transformers to regress the affine transformation acting on the input. For this, we decompose affine transformations into interpretable primitives, regressed under adaptable geometric constraints, thereby preventing the training instability typically caused by degenerate transformations. By sharing weights between the localization network and the classification backbone, the framework requires minimal computational overhead. Extensive experiments on challenging insect biodiversity and medical imaging benchmarks demonstrate that our approach achieves superior predictive performance under diverse spatial transformations while maintaining high efficiency. Code is available at https://github.com/johSchm/TokenSTN.

We propose a vision transformer (ViT)-based deep learning framework to refine disaster-affected area segmentation from remote sensing imagery, aiming to support and enhance the Emergent Value Added Product (EVAP) developed by the Taiwan Space Agency (TASA). The process starts with a small set of manually annotated regions. We then apply principal component analysis (PCA)-based feature space analysis and construct a confidence index (CI) to expand these labels, producing a weakly supervised training set. These expanded labels are then used to train ViT-based encoder-decoder models with multi-band inputs from Sentinel-2 and Formosat-5 imagery. Our architecture supports multiple decoder variants and multi-stage loss strategies to improve performance under limited supervision. During the evaluation, model predictions are compared with higher-resolution EVAP output to assess spatial coherence and segmentation consistency. Case studies on the 2022 Poyang Lake drought and the 2023 Rhodes wildfire demonstrate that our framework improves the smoothness and reliability of segmentation results, offering a scalable approach for disaster mapping when accurate ground truth is unavailable.

The integration of Large Language Model (LLM) blocks with Vision Transformers (ViTs) holds significant promise for vision-only tasks by leveraging the rich semantic knowledge and reasoning capabilities of LLMs. However, a fundamental challenge lies in the inherent modality mismatch between the text-centric pre-training of LLMs and the vision-centric training of ViTs. Direct fusion often fails to fully exploit the LLM's potential and suffers from unstable finetuning. Consequently, prior works typically keep LLM blocks frozen while learning only the vision components. To address these challenges, we introduce Language-Adapted Vision Enhancer (LAVIE), a novel framework that bridges this modality gap through a synergistic pre-training strategy. LAVIE co-adapts a ViT backbone and an LLM fusion block by (1) employing Masked Auto-Encoding (MAE) to pre-train the ViT for richer visual representations, and (2) concurrently training Low-Rank Adaptation (LoRA) layers within the LLM block using the same MAE objective. This joint optimization guides the ViT to produce LLM-aligned features and the LLM to effectively interpret visual information. We demonstrate through extensive experiments that LAVIE significantly improves performance in various downstream vision tasks, offering an effective and efficient way to enhance visual understanding using frozen LLM knowledge. Code is available at https://github.com/selimkuzucu/LAVIE

Adapting vision models for time series analysis is compelling, yet all existing approaches are falling short of dedicated time series foundation models (TSFMs) in classification. In this work, we propose Time Vision Transformer (TiViT), the first framework that successfully unlocks the representational power of frozen Vision Transformers (ViTs) pretrained on large-scale image datasets for time series classification. TiViT achieves state-of-the-art performance without any finetuning by utilizing the hidden representations of OpenCLIP models. We explore the structure of TiViT representations and find that intermediate ViT layers with high intrinsic dimension are the most effective for time series classification. Furthermore, we assess the alignment between TiViT and TSFM representation spaces and identify a strong complementarity, with additional performance gains achieved through feature concatenation. Finally, we unfreeze the ViT backbone of TiViT for continual pretraining and contrastive alignment with TSFMs on time series, enhancing the performance of lightweight TiViT variants. Our findings reveal a new direction for the domain and task adaptation of vision foundation models. Code is available at https://github.com/ExplainableML/TiViT.

Attention is the critical component of a transformer. Yet the quadratic computational complexity of vanilla full attention in the input size and the inability of its linear attention variant to focus have been challenges for computer vision tasks. We provide a mathematical definition of generalized attention and formulate both vanilla softmax attention and linear attention within the general framework. We prove that generalized attention disperses, that is, as the number of keys tends to infinity, the query assigns equal weights to all keys. Motivated by the dispersion property and recent development of Mamba form of attention, we design Scalable and Efficient Mamba like Attention (SEMA) which utilizes token localization to avoid dispersion and maintain focusing, complemented by theoretically consistent arithmetic averaging to capture global aspect of attention. We support our approach on Imagenet-1k where classification results show that SEMA is a scalable and effective alternative beyond linear attention, outperforming recent vision Mamba models on increasingly larger scales of images at similar model parameter sizes.

Natural images exhibit strong geometric regularities: local structures, such as edges, corners, and textures, appear in many orientations and mirror configurations. Since Vision Transformers (ViTs) operate on square image patches, these transformations naturally correspond to the dihedral symmetry group $\mathrm{D}_8$, also known as the octic group. Recent work has shown that ViTs can be made reflection equivariant and more efficient than standard ViTs simultaneously by implementing the linear layers in the Fourier domain of the reflection group. In this work, we extend the equivariance to reflections and rotations and analyze the scalability of the resulting networks. Our Quick ViTs, based on octic equivariant linear layers, achieve 5.33x reductions in FLOPs and up to 8x reductions in memory compared to ordinary linear layers. By analyzing the arithmetic intensity of these layers, we identify theoretical limits on how much the FLOP savings translate into throughput improvements on modern GPUs. However, these limitations disappear as the embedding dimensions increase. Enabled by their computational efficiency, we conduct a broader empirical evaluation of equivariant ViTs than in previous work. Upon training supervised (DeiT-III) and self-supervised (DINOv2) on ImageNet-1K, we find that our Quick ViTs match or exceed baseline accuracy while at the same time providing substantial efficiency gains.

Understanding the internal mechanisms of deep neural networks remains a central challenge in machine learning. In computer vision, one promising yet only preliminarily explored approach is feature inversion via inverse networks, which reconstructs images from intermediate representations using trained inverse networks. In this study, we revisit feature inversion via inverse networks, introducing a novel, modular variant that enables a computationally more efficient application of the technique while, in some cases, producing semantically more coherent image reconstructions. We apply our method to large-scale transformer-based vision models, specifically Detection Transformer, Vision Transformer, Swin Transformer, and Data-Efficient Image Transformer, analyzing the resulting reconstructions across network depth. Our main analysis compares Detection Transformer and Vision Transformer, which exhibit the most informative differences among the evaluated architectures. At the same time, results for Swin Transformer and Data-Efficient Image Transformer support the broader applicability of our framework. Our findings reveal gradual representational changes across transformer layers as a shared characteristic of Detection Transformer and Vision Transformer, as well as systematic differences in how they preserve contextual shape and fine-grained image details, and in their robustness to color perturbations. These findings deepen understanding of transformer-based vision models and demonstrate the utility of modular feature inversion as an interpretability tool.