Vision Transformer (ViT) design has become increasingly diverse, with backbones combining convolutional stems, windowed, linear, or multi-axis attention, patch merging, and spatial reduction in various configurations. This diversity poses challenges for existing attribution methods, whose assumptions often do not hold across ViT variants: Grad-CAM requires a terminal spatial feature map, attention rollout assumes global softmax attention, and layer-wise relevance propagation (LRP) requires module-specific rules. To the best of our knowledge, no existing method provides a unified attribution framework across this architectural space. We show that this architectural diversity can be captured by a simpler underlying structure. The attention and resolution-reduction operators in current ViTs can be decomposed into four operation types: linear maps, bilinear mixing, normalization or gating, and reindexing. Each operation admits a relevance rule that satisfies conservation. Based on these rules, HiLRP supports new backbones by construction rather than by architecture-specific derivation, and its attribution maps decompose the prediction rather than relying on heuristic assumptions. We prove conservation and conditional equivariance and verify both to machine precision. Across 14 attribution methods and 10 architectures, we find that no prior method remains reliable across ViT families, while Faithfulness Correlation becomes uninformative for backbones robust to spatial masking. HiLRP alone preserves conservation across windowed, spatial-reduction, multi-axis, and linear-attention models, where naive extensions can produce zero or inflated relevance. It also localizes attribution failures in class activation mapping, achieving 0.97 Pointing compared with 0.55 for competing methods on EfficientViT.
AI-generated image (AIGI) detection has become increasingly challenging due to the rapid advancement of generative models and the diminishing gap between synthetic and authentic content. Existing vision transformer-based detectors commonly rely on weighted-sum strategies to aggregate intermediate representations across transformer layers, often overlooking the inherently ordered semantic progression of hierarchical features from shallow texture cues to deep semantic representations. In this work, we propose \textbf{PE-Mamba}, a novel framework built upon a pre-trained PE-Core vision transformer with lightweight LoRA adaptation that introduces three complementary components for cross-layer feature aggregation and fusion. First, a bidirectional selective aggregator (BSA) processes layer-wise classification tokens through forward and backward selective scans, where the forward scan progressively accumulates shallow-to-deep forensic evidence, and the backward scan performs deep-to-shallow contextual refinement to reinterpret low-level cues in light of high-level semantic context. Second, a softmax-weighted aggregator (SWA) computes a learned global summary of all layer tokens as a complementary aggregation path. Third, a sigmoid-gated blend (SGA) adaptively fuses the BSA and SWA outputs via a learnable scalar gate, allowing the model to dynamically balance directional sequential evidence and global layer-wise aggregation. Extensive experiments on UniversalFakeDetect (96.6\% mACC, 99.5\% mAP) and AIGCDetect (95.3\% mACC, 98.1\% mAP) demonstrate that \methodname{} outperforms 18 detectors with superior generalization across diverse generative models, while training only 1.3\% of total parameters (0.13\% for LoRA alone).
Vision transformers face significant computational overheads in high-resolution dense prediction due to the quadratic complexity of self-attention. Linear attention offers efficiency but sacrifices local context modeling. We propose \textbf{HSMLA (Hierarchical Softmax Multi-scale Linear Attention)}, which combines ReLU-based linear attention for global context, selective softmax refinement for critical local features, and multi-scale token representations via depthwise convolutions. HSMLA achieves superior accuracy-efficiency trade-offs: up to $4.2\times$ inference-time speedup across dense prediction tasks, $87.3%$ Dice with $3.2\times$ speedup on CT organ segmentation, and $94.2%$ AUC with $4.1\times$ speedup on pathology WSI.
Self-supervised monocular depth estimation typically relies on photometric reconstruction losses that couple depth, pose, and appearance assumptions. In this paper, we propose JEPADepth, a self-supervised monocular depth framework that incorporates a complementary training objective inspired by Image Joint-Embedding Predictive Architectures (I-JEPA) for self-supervised depth learning. Our method augments a standard photometric pipeline with a masked prediction loss computed in the representation space of a pretrained DINOv3 Vision Transformer encoder. A predictor infers target-region embeddings from visible context-region embeddings under structured masking, and is discarded along with the target encoder at inference time, adding no deployment cost. On KITTI, adding the JEPA objective consistently improves performance over the same DINOv3-based photometric baseline, without changing the inference-time architecture. Compared to prior monocular self-supervised methods, JEPADepth is competitive with state-of-the-art transformer-based approaches and outperforms strong CNN-based baselines on the standard benchmark. In zero-shot transfer (trained on KITTI and evaluated without fine-tuning), JEPADepth achieves the best or near-best performance among the compared methods on both Make3D and Cityscapes across multiple metrics.
Thanh Cong Le, Michal Szczepanski, Martyna Porebacs.CV
Vision Transformer detectors now approach the accuracy of CNNs but remain difficult to deploy on NPUs and microcontrollers because key components, including deformable attention, feature fusion, and nonlinear activation functions, are not natively compatible with integer arithmetic. Existing quantized detectors either retain operators such as Softmax, GELU, and LayerNorm or focus on heavyweight backbones, leaving lightweight detection transformers without an end-to-end integer implementation. We address this gap with I-LW-DETR, the first fully integer-only lightweight DETR, in which every operation in the forward pass, including transformer nonlinearities, is executed in integer arithmetic. I-LW-DETR is built upon three key components: a scale-preserving split convolution that assigns independent activation scale to each branch of the multi-scale projector; SD-ShiftGELU, a sign-dependent GELU approximation that preserves element-wise behavior while avoiding the accuracy degradation; and a constrained Shiftmax that maintains stable Softmax normalization. Experimental results demonstrate that the proposed quantization pipeline consistently produces efficient fully integer-only models across different model scales. Across all model scales, the proposed pipeline incurs only a moderate accuracy degradation while reducing the model size by approximately $3.6\times$ and the computational cost by more than one order of magnitude.
Lukas Knobel, Andrew Zisserman, Yuki M. Asanocs.CV
Understanding motion in video is a fundamental challenge for visual learning, as frame-to-frame change entangles two sources of dynamics: camera motion and object motion. This decomposition has remained underexplored in representation learning, partly because these factors are tightly coupled in natural videos and difficult to supervise separately. Yet recovering it is important for learning robust motion representations that separate meaningful object dynamics from camera-induced variation. We study whether such structured motion representations can be recovered from frozen features of a pretrained image vision transformer. We propose the Structured Dynamics Model (SDM), which explicitly separates the dominant source of temporal change from residual dynamics through future-feature prediction, rather than representing video change with a single entangled latent or with unstructured, spatially dense transition tokens. Training combines self-supervised learning on real video with weak supervision of scene dynamics on synthetic Kubric data. We evaluate SDM on ProbeMotion, a new evaluation suite spanning synthetic and real videos with camera motion, object motion, and combined dynamics. SDM outperforms backbone baselines using global CLS or average-pooled features, and compares favorably to strongly supervised representations such as VGGT on several probes, despite using substantially weaker supervision. These results suggest that pretrained image models can be readily repurposed into structured video-dynamics representations, providing a useful inductive bias for learning and analyzing latent video dynamics.
Andrei-Timotei Ardelean, Michael Fischer, Tim Weyrich +1cs.GR cs.LG
Inverse rendering is traditionally solved via differentiable renderers and gradient descent, which requires substantial problem-specific engineering and is prone to getting stuck in local minima due to ambiguities. Derivative-free approaches alleviate engineering requirements, but often heavily depend on a good problem initialization. In this work, we propose Feature-Informed Diffusion Evolution (FIDE), a fully black-box framework that requires no gradients or specific initialization: the renderer is treated as an opaque function whose only requirement is to produce images. Our key insight is feature guiding: rather than reducing each candidate rendering to a scalar loss value, we use a Vision Transformer (ViT) to extract dense visual features from it. We subsequently use these features to train a diffusion-based candidate proposal model, allowing the network to use visual cues to predict parameters that would match the target image. The candidate solutions proposed by this diffusion model are then refined in a closed loop with a CMA evolution strategy, continuously narrowing the proposal region as optimization progresses. We validate across diverse inverse problems from path tracing, vector splines, Voronoi shaders, and robotics, and demonstrate that feature-guiding substantially improves convergence over scalar-loss baselines and reliably escapes local minima where gradient-based methods stall.
Vision Transformer (ViT) has been widely used as a powerful framework for modeling global dependencies among image patches. However, its core component, self-attention assigns softmax-normalized relative weights to all patches, making it difficult to evaluate the relevance between patches independently. In visual recognition, images often contain many background or redundant patches, yet self-attention cannot explicitly reject such irrelevant patches, which may introduce unnecessary information into feature aggregation. To address this limitation, Screening has been proposed in the field of language modeling, where the relevance of each token is independently evaluated based on query-key similarity and low-relevance tokens are explicitly excluded through thresholding. In this work, we propose VisionScreen, a new vision model that extends Screening mechanism to visual recognition. VisionScreen treats image patches as tokens arranged on a two-dimensional grid and extends absolute relevance estimation based on query-key similarity to the two-dimensional spatial domain. This allows each patch to selectively aggregate only content-wise and spatially relevant patches without relying on competition among patches. Experiments on image classification benchmarks demonstrate that the proposed method outperforms conventional ViT. These results suggest that Screening can be effective for visual recognition, offering an alternative to relative feature aggregation based on softmax attention.
Recently, Vision Transformer (ViT)-based models have exhibited remarkable performance in image super-resolution. However, the quadratic computational complexity of ViTs with respect to spatial resolution severely constrains their efficiency, leading to high latency and massive memory consumption. To alleviate this, various window-based attention mechanisms have been proposed; yet, they inherently compromise the long-range dependency modeling that is the primary advantage of ViTs. To overcome these limitations, we propose the Clustered Unit-level Similarity Transformer (CUST), a novel architecture that efficiently integrates global and local information. Specifically, CUST enables each patch to aggregate and attend to similar patches within a broadened regional scope outside its local window, thereby capturing extensive contextual understanding. Furthermore, it employs overlapping attention windows to capture local dependencies, while explicitly extracting high-frequency details by computing the residual difference between the original features and their downsampled-upsampled counterparts. Comprehensive experiments demonstrate that our proposed model achieves a practical balance between computational efficiency and restoration performance. It achieves a lower memory footprint and faster inference speed compared to recent global context or lightweight models under realistic constraints. Code is available at [https://github.com/jwgdmkj/CUST].
Dynamic expansion methods for class-incremental learning (CIL) protect task-specific knowledge by growing dedicated tokens or subnetworks, yet our analyses suggest that classification supervision alone does not sufficiently preserve task-agnostic shared backbone representations over long incremental sequences. We identify two intertwined challenges: cross-task confusion from sequential training on predominantly current-task data, which biases decision boundaries toward recent tasks; and under-optimized shared representations in the backbone that cap long-term discriminability as tasks accumulate. We propose the Decoupled Representation Dynamic Network (DRDN), which addresses these challenges via two orthogonal mechanisms. For shared backbone representations, DRDN continuously applies masked image modeling (MIM) at every incremental step, with reconstruction gradients routed exclusively through the backbone, encouraging it to retain general visual structure beyond class-discriminative cues. For task-specific discrimination, DRDN employs hierarchical task token expansion across all transformer layers, with a modified per-task attention rule that reduces inter-task interference. We support this design with accuracy degradation analysis and cross-task confusion rate measurements. In the from-scratch ViT CIL setting (no external pretraining), DRDN consistently improves over strong token-expansion baselines with comparable backbone scale. On CIFAR100-B0 (10 steps), DRDN achieves 77.19% average accuracy, outperforming DKT by 1.36 points and DyTox by 3.53 points, with an advantage that grows at longer incremental sequences. Multi-seed validation confirms stability (+/-0.31%). The MIM decoder is active only during training, adding no inference-time parameters or computation.
Recent recursive Transformer studies have primarily reused shared parameters across computation steps to construct compact, parameter-efficient models. In this work, we leverage recursion to build effectively deeper Transformers with stronger representational capacity. However, in Vision Transformers, simply increasing recursion depth does not reliably improve performance, as existing recursive approaches do not fully utilize the intermediate representations produced throughout recursive computation. We propose Soft Mixture-of-Recursions (SoftMoR) and its Vision Transformer instantiation, Soft Recursive Vision Transformer (SR-ViT). SoftMoR learns token-wise mixture weights to softly combine outputs from all recursion steps, allowing intermediate representations to be utilized in a learnable and flexible way. Across diverse vision tasks, SR-ViT consistently improves as recursion depth increases with minimal parameter overhead. On ImageNet-1K, increasing recursion depth from 1 to 4 improves SR-ViT-S top-1 accuracy from 79.83% to 82.48% with only 1.7M additional parameters, outperforming the substantially larger DeiT-B while using approximately 27% of its parameters. These results demonstrate that SoftMoR provides a parameter-efficient path to deeper and stronger Vision Transformers through recursion.
Most token reduction methods for Vision Transformers seek favorable tradeoffs between accuracy and efficiency by pruning, merging, or pooling patch tokens. REDI (Relevance for DINOv3 Token Reduction) studies this question through a controlled supervised reference: how should a fixed token budget be allocated across patches for image classification? REDI quantizes final block DINOv3 patch representations into a visual vocabulary and derives class conditioned corpus scores using supervised TF-IDF over visual words. For each validation image, the ground truth class selects a row of the TF-IDF table, and four transformed views produce a TF-IDF map aligned to a reference center crop. A separate dense pass on the same crop provides an attention map. After independent min max normalization, their elementwise product defines the REDI score. A fixed keep, merge, and compress operator then uses score rank to assign patch roles and score magnitude to weight merging and compression. With precomputed REDI scores, a frozen DINOv3 ViT-B/16 backbone, and the same linear classifier used for dense evaluation, the operator reduces the sequence length from 201 to 107 tokens, a 46.8% sequence reduction. The REDI variant based on incoming attention mass achieves 84.706% Top-1 accuracy on ImageNet-1K, compared with 83.514% for the dense baseline, 82.634% for incoming attention mass alone, and 81.796% for supervised TF-IDF alone. The same corpus term also improves reduced classification for three alternative attention formulations relative to their attention only counterparts. Together, these controlled comparisons indicate that class specific corpus statistics and image specific attention provide complementary signals for patch ranking in this setting.
Diffusion models enable probabilistic super-resolution and conditional generation, but pixel-space methods are computationally expensive and learned latent spaces often lack interpretable uncertainty quantification. We introduce Patch-PODiff-ViT, a structured latent diffusion framework in which the latent space is defined by patchwise Proper Orthogonal Decomposition (POD), a fixed linear orthonormal basis over local patches, rather than learned by a nonlinear autoencoder. This yields low-dimensional, variance-ordered tokens that preserve spatial structure and enable efficient diffusion in a structured low-dimensional latent space with a Vision Transformer. Because the decoder is fixed, linear, and orthonormal, latent coefficient uncertainty can be propagated directly to physical-space predictive variance, enabling analytic propagation of predictive variance through the linear decoder without Monte Carlo estimation in pixel space. Across sea surface temperature, medical imaging, and natural images, the method achieves strong reconstruction with fewer parameters and lower memory, while producing well-calibrated spatial uncertainty that closely matches empirical ensembles.
Md Irtiza Hossain, Humaira Ayesha, Junaid Ahmed Sifatcs.CV cs.DC
We present CLEAR-MoE, a four-phase post-training pipeline that converts a frozen pretrained Vision Transformer (ViT) into a sparse Mixture-of-Experts (MoE) model without updating backbone weights. The pipeline (i) scores feed-forward network (FFN) layers by sparsity, clusterability, and output sensitivity; (ii) decomposes selected layers into a shared low-rank SVD basis and per-cluster residual experts using k-means clustering; (iii) trains lightweight routers supervised by cluster labels; and (iv) dispatches tokens through pluggable CUDA backends. On Imagenette with DeiT-Small, CLEAR-MoE retains 99.9% of the dense model's accuracy (86.70 +/- 0.02% versus 86.73%). Extensive ablation studies reveal a consistent empirical finding: the shared SVD basis is the primary factor responsible for preserving accuracy. Random routing, learned routing, and three different router architectures produce nearly identical performance, with accuracy varying by at most 0.06 percentage points (86.62%-86.68%). Accuracy also remains stable across different SVD ranks, expert counts (2-8), calibration set sizes (50-500), and random seeds. This behavior generalizes across five ViT backbones (DeiT-Tiny, DeiT-Small, DeiT-Base, ViT-Small, and ViT-Base), covering models from 5.7M to 86.6M parameters, with accuracy differences <= 0.10 percentage points from their dense counterparts. On a GTX 960 GPU, routing and scatter-gather overhead make the CLEAR-MoE FFN 1.3-1.7x slower than the dense implementation. A dispatch microbenchmark further shows that routing is an order of magnitude more memory-bound than expert matrix multiplications, identifying fused dispatch kernels as a promising direction for future optimization.
Modern VLMs and VLA systems commonly adopt off-the-shelf ViTs such as SigLIP2 as visual encoders, but diverse downstream requirements in latency, temporal modeling, and VLM integration often call for customized SOTA-level ViTs. Training such encoders remains beyond the reach of much of the community, as it requires massive image-text data, while standard softmax attention makes high-resolution or dynamic-resolution pretraining prohibitively costly and often forces low-resolution pretraining followed by post-hoc adaptation. TuringViT addresses these challenges with three key designs: Turing Linear Attention (TLA) for efficient sequence modeling, VISTA-Curation to construct supervision-rich image-video training data, and native dynamic-resolution pretraining that supports flexible inputs from the start and transfers seamlessly to downstream VLMs. As a result, TuringViT outperforms leading open-source ViT baselines with only 10% of the data, achieves stronger downstream VLM performance, and delivers substantially better latency scaling on high-resolution inputs. Our scaling-law analysis further shows that TuringViT continues to improve predictably with curated data scale, far from saturation. Its fast adaptation, hardware-friendly design, and efficient deployment have made it a unified visual foundation across XPeng's AI systems. More broadly, TuringViT provides a reproducible pipeline that dramatically lowers the cost for the community to train, customize, and deploy SOTA-level ViTs, moving toward making such Vision Transformers accessible to all.
Tang Li, Yanlin Chen, Mengmeng Ma +1cs.CV cs.AI cs.LG
Despite high accuracy, Vision Transformer (ViT) predictions can be driven by spurious cues, raising the need to understand their inner workings before safe deployment. Sparse autoencoders (SAEs) provide a promising lens for decomposing model representations into human-interpretable concepts, yet adapting SAE-based interpretation to ViTs remains challenging due to limited control over concept coverage and subjective, non-scalable feature interpretation. To fill the gaps, motivated by neuroscience-inspired principles, we propose ViSAE, a mechanistic interpretability toolbox for understanding ViT inner workings through concept circuits. ViSAE consists of three components: (1) A probing suite with 64K images and a 16K visually grounded concept vocabulary, improving concept coverage efficiency by 20x over ImageNet and interpretation accuracy by 28.7% over existing concept sets. (2) Top-down concept reading and Bottom-up circuit tracing algorithms that automatically recover ViT inner workings via concept circuits. (3) Applications for auditing and steering ViT behavior. Through concept editing, ViSAE improves the worst-group accuracy on WaterBirds by 48.2%, outperforming existing methods by 23.8%. Our data and code: https://github.com/deep-real/ViSAE.
Catastrophic forgetting is often framed as a representational problem: after sequential training, a model appears to lose the features that supported performance on earlier tasks. We challenge the stronger form of this view. Across controlled continual-learning settings, we find that a significant portion of apparent forgetting can be attributed to interface drift between internal stages rather than permanent erasure of task-relevant computation. We study this phenomenon through a stitched evaluation protocol that combines early computation from a post-update network with late computation from its predecessor, optionally mediated by a compact, task-specific transport key. We describe transport keys at a systems level as compact interface-alignment operators estimated from a small set of paired anchor activations and evaluated through model stitching. On split CIFAR-100 with a ResNet-style network, transport keys recover most of the original Task A performance after sequential training on Task B. On a compact vision transformer, we observe a similar recovery pattern. These results suggest that continual learning may require better mechanisms for indexing and re-accessing latent computations, not only methods that prevent weight change.
While linear-complexity attention mechanisms offer a promising alternative to Softmax attention for overcoming the quadratic bottleneck, training such models from scratch remains prohibitively expensive. Inheriting weights from pretrained Transformers provides an appealing shortcut, yet the fundamental representational gap between Softmax and linear attention prevents effective weight transfer. In this work, we address this conversion challenge from two perspectives: architectural alignment and representational alignment. We identify Test-Time Training (TTT) as a linear-complexity architecture whose two-layer dynamic formulation is structurally aligned with Softmax attention, enabling direct inheritance of pretrained attention weights. To further align representational properties, including key shift-invariance and locality, we introduce key instance normalization and a lightweight locality enhancement module. We validate our approach by linearizing Stable Diffusion 3.5 and introduce SD3.5-T$^5$ (Transformer To Test Time Training). With only 1 hour of fine-tuning on 4$\times$H20 GPUs, SD3.5-T$^5$ achieves comparable text-to-image quality to the fine-tuned Softmax model, while accelerating inference by 1.32$\times$ and 1.47$\times$ at 1K and 2K resolutions.
Sparse Autoencoders (SAEs) have demonstrated significant success in interpreting Large Language Models (LLMs) by decomposing dense representations into sparse, semantic components. However, their potential for analyzing Vision Transformers (ViTs) remains largely under-explored. In this work, we present the first application of SAEs to the ViT [CLS] token for out-of-distribution (OOD) detection, addressing the limitation of existing methods that rely on entangled feature representations. We propose a novel framework utilizing a Top-k SAE to disentangle the dense [CLS] features into a structured latent space. Through this analysis, we reveal that in-distribution (ID) data exhibits consistent, class-specific activation patterns, which we formalize as Class Activation Profiles (CAPs). Our study uncovers a key structural invariant: while ID samples preserve a stable pattern within CAPs, OOD samples systematically disrupt this structure. Leveraging this insight, we introduce a scoring function based on the divergence of core energy profiles to quantify the deviation from ideal activation profiles. Our method achieves strong results on the FPR95 metric, critical for safety-sensitive applications across multiple benchmarks, while also achieving competitive AUROC. Overall, our findings demonstrate that the sparse, disentangled features revealed by SAEs can serve as a powerful, interpretable tool for robust OOD detection in vision models.
Laurenz Reichardt, Nikolas Ebert, Oliver Wasenmüllercs.CV
3D point cloud perception remains tightly coupled to custom CUDA operators for spatial operations, limiting portability and efficiency on non-NVIDIA, AMD, and embedded hardware. We introduce PointTransformerX (PTX), a fully PyTorch-native vision transformer backbone for 3D point clouds, removing all custom CUDA operators and external libraries while retaining competitive accuracy. PTX introduces 3D-GS-RoPE, a rotary positional embedding that encodes 3D spatial relationships directly in self-attention without neighborhood construction, and further replaces sparse convolutional patch embedding with a linear projection. PTX explores inference-time scaling of attention windows to improve accuracy without retraining. With a redesigned feed-forward network, PTX achieves 98.7\% of PointTransformer V3's accuracy on ScanNet with 79.2\% fewer parameters and executing 1.6\times faster while requiring just 253 MB memory. PTX runs natively on NVIDIA GPUs, AMD GPUs (ROCm), and CPUs, providing an efficient and portable foundation for point cloud perception.