Local Scale Equivariance with Latent Deep Equilibrium Canonicalizer

Scale variation is a fundamental challenge in computer vision. Objects of the same class can have different sizes, and their perceived size is further affected by the distance from the camera. These variations are local to the objects, i.e., different object sizes may change differently within the same image. To effectively handle scale variations, we present a deep equilibrium canonicalizer (DEC) to improve the local scale equivariance of a model. DEC can be easily incorporated into existing network architectures and can be adapted to a pre-trained model. Notably, we show that on the competitive ImageNet benchmark, DEC improves both model performance and local scale consistency across four popular pre-trained deep-nets, e.g., ViT, DeiT, Swin, and BEiT. Our code is available at https://github.com/ashiq24/local-scale-equivariance.

Calibration Attention: Instance-wise Temperature Scaling for Vision Transformers

Probability calibration is critical when Vision Transformers are deployed in risk-sensitive applications. The standard fix, post-hoc temperature scaling, uses a single global scalar and requires a held-out validation set. We introduce Calibration Attention (CalAttn), a drop-in module that learns an adaptive, per-instance temperature directly from the ViT's CLS token. Across CIFAR-10/100, MNIST, Tiny-ImageNet, and ImageNet-1K, CalAttn reduces calibration error by up to 4x on ViT-224, DeiT, and Swin, while adding under 0.1 percent additional parameters. The learned temperatures cluster tightly around 1.0, in contrast to the large global values used by standard temperature scaling. CalAttn is simple, efficient, and architecture-agnostic, and yields more trustworthy probabilities without sacrificing accuracy. Code: [https://github.com/EagleAdelaide/CalibrationAttention-CalAttn-](https://github.com/EagleAdelaide/CalibrationAttention-CalAttn-)

Variance-Based Pruning for Accelerating and Compressing Trained Networks

Increasingly expensive training of ever larger models such as Vision Transfomers motivate reusing the vast library of already trained state-of-the-art networks. However, their latency, high computational costs and memory demands pose significant challenges for deployment, especially on resource-constrained hardware. While structured pruning methods can reduce these factors, they often require costly retraining, sometimes for up to hundreds of epochs, or even training from scratch to recover the lost accuracy resulting from the structural modifications. Maintaining the provided performance of trained models after structured pruning and thereby avoiding extensive retraining remains a challenge. To solve this, we introduce Variance-Based Pruning, a simple and structured one-shot pruning technique for efficiently compressing networks, with minimal finetuning. Our approach first gathers activation statistics, which are used to select neurons for pruning. Simultaneously the mean activations are integrated back into the model to preserve a high degree of performance. On ImageNet-1k recognition tasks, we demonstrate that directly after pruning DeiT-Base retains over 70% of its original performance and requires only 10 epochs of fine-tuning to regain 99% of the original accuracy while simultaneously reducing MACs by 35% and model size by 36%, thus speeding up the model by 1.44x.

Frequency-Dynamic Attention Modulation for Dense Prediction

Vision Transformers (ViTs) have significantly advanced computer vision, demonstrating strong performance across various tasks. However, the attention mechanism in ViTs makes each layer function as a low-pass filter, and the stacked-layer architecture in existing transformers suffers from frequency vanishing. This leads to the loss of critical details and textures. We propose a novel, circuit-theory-inspired strategy called Frequency-Dynamic Attention Modulation (FDAM), which can be easily plugged into ViTs. FDAM directly modulates the overall frequency response of ViTs and consists of two techniques: Attention Inversion (AttInv) and Frequency Dynamic Scaling (FreqScale). Since circuit theory uses low-pass filters as fundamental elements, we introduce AttInv, a method that generates complementary high-pass filtering by inverting the low-pass filter in the attention matrix, and dynamically combining the two. We further design FreqScale to weight different frequency components for fine-grained adjustments to the target response function. Through feature similarity analysis and effective rank evaluation, we demonstrate that our approach avoids representation collapse, leading to consistent performance improvements across various models, including SegFormer, DeiT, and MaskDINO. These improvements are evident in tasks such as semantic segmentation, object detection, and instance segmentation. Additionally, we apply our method to remote sensing detection, achieving state-of-the-art results in single-scale settings. The code is available at \href{https://github.com/Linwei-Chen/FDAM}{https://github.com/Linwei-Chen/FDAM}.

Block-based Symmetric Pruning and Fusion for Efficient Vision Transformers

Vision Transformer (ViT) has achieved impressive results across various vision tasks, yet its high computational cost limits practical applications. Recent methods have aimed to reduce ViT's $O(n^2)$ complexity by pruning unimportant tokens. However, these techniques often sacrifice accuracy by independently pruning query (Q) and key (K) tokens, leading to performance degradation due to overlooked token interactions. To address this limitation, we introduce a novel {\bf Block-based Symmetric Pruning and Fusion} for efficient ViT (BSPF-ViT) that optimizes the pruning of Q/K tokens jointly. Unlike previous methods that consider only a single direction, our approach evaluates each token and its neighbors to decide which tokens to retain by taking token interaction into account. The retained tokens are compressed through a similarity fusion step, preserving key information while reducing computational costs. The shared weights of Q/K tokens create a symmetric attention matrix, allowing pruning only the upper triangular part for speed up. BSPF-ViT consistently outperforms state-of-the-art ViT methods at all pruning levels, increasing ImageNet classification accuracy by 1.3% on DeiT-T and 2.0% on DeiT-S, while reducing computational overhead by 50%. It achieves 40% speedup with improved accuracy across various ViTs.

DART: Differentiable Dynamic Adaptive Region Tokenizer for Vision Transformer and Mamba

Recently, non-convolutional models such as the Vision Transformer (ViT) and Vision Mamba (Vim) have achieved remarkable performance in computer vision tasks. However, their reliance on fixed-size patches often results in excessive encoding of background regions and omission of critical local details, especially when informative objects are sparsely distributed. To address this, we introduce a fully differentiable Dynamic Adaptive Region Tokenizer (DART), which adaptively partitions images into content-dependent patches of varying sizes. DART combines learnable region scores with piecewise differentiable quantile operations to allocate denser tokens to information-rich areas. Despite introducing only approximately 1 million (1M) additional parameters, DART improves accuracy by 2.1% on DeiT (ImageNet-1K). Unlike methods that uniformly increase token density to capture fine-grained details, DART offers a more efficient alternative, achieving 45% FLOPs reduction with superior performance. Extensive experiments on DeiT, Vim, and VideoMamba confirm that DART consistently enhances accuracy while incurring minimal or even reduced computational overhead. Code is available at https://github.com/HCPLab-SYSU/DART.

Token Transforming: A Unified and Training-Free Token Compression Framework for Vision Transformer Acceleration

Vision transformers have been widely explored in various vision tasks. Due to heavy computational cost, much interest has aroused for compressing vision transformer dynamically in the aspect of tokens. Current methods mainly pay attention to token pruning or merging to reduce token numbers, in which tokens are compressed exclusively, causing great information loss and therefore post-training is inevitably required to recover the performance. In this paper, we rethink token reduction and unify the process as an explicit form of token matrix transformation, in which all existing methods are constructing special forms of matrices within the framework. Furthermore, we propose a many-to-many Token Transforming framework that serves as a generalization of all existing methods and reserves the most information, even enabling training-free acceleration. We conduct extensive experiments to validate our framework. Specifically, we reduce 40% FLOPs and accelerate DeiT-S by $\times$1.5 with marginal 0.1% accuracy drop. Furthermore, we extend the method to dense prediction tasks including segmentation, object detection, depth estimation, and language model generation. Results demonstrate that the proposed method consistently achieves substantial improvements, offering a better computation-performance trade-off, impressive budget reduction and inference acceleration.

Is Attention Required for Transformer Inference? Explore Function-preserving Attention Replacement

While transformers excel across vision and language pretraining tasks, their reliance on attention mechanisms poses challenges for inference efficiency, especially on edge and embedded accelerators with limited parallelism and memory bandwidth. Hinted by the observed redundancy of attention at inference time, we hypothesize that though the model learns complicated token dependency through pretraining, the inference-time sequence-to-sequence mapping in each attention layer is actually ''simple'' enough to be represented with a much cheaper function. In this work, we explore FAR, a Function-preserving Attention Replacement framework that replaces all attention blocks in pretrained transformers with learnable sequence-to-sequence modules, exemplified by an LSTM. FAR optimize a multi-head LSTM architecture with a block-wise distillation objective and a global structural pruning framework to achieve a family of efficient LSTM-based models from pretrained transformers. We validate FAR on the DeiT vision transformer family and demonstrate that it matches the accuracy of the original models on ImageNet and multiple downstream tasks with reduced parameters and latency. Further analysis shows that FAR preserves the semantic token relationships and the token-to-token correlation learned in the transformer's attention module.

Stronger ViTs With Octic Equivariance

Recent efforts at scaling computer vision models have established Vision Transformers (ViTs) as the leading architecture. ViTs incorporate weight sharing over image patches as an important inductive bias. In this work, we show that ViTs benefit from incorporating equivariance under the octic group, i.e., reflections and 90-degree rotations, as a further inductive bias. We develop new architectures, octic ViTs, that use octic-equivariant layers and put them to the test on both supervised and self-supervised learning. Through extensive experiments on DeiT-III and DINOv2 training on ImageNet-1K, we show that octic ViTs yield more computationally efficient networks while also improving performance. In particular, we achieve approximately 40% reduction in FLOPs for ViT-H while simultaneously improving both classification and segmentation results.

Lossless Token Merging Even Without Fine-Tuning in Vision Transformers

Although Vision Transformers (ViTs) have become the standard architecture in computer vision, their massive sizes lead to significant computational overhead. Token compression techniques have attracted considerable attention to address this issue, but they often suffer from severe information loss, requiring extensive additional training to achieve practical performance. In this paper, we propose Adaptive Token Merging (ATM), a novel method that ensures lossless token merging, eliminating the need for fine-tuning while maintaining competitive performance. ATM adaptively reduces tokens across layers and batches by carefully adjusting layer-specific similarity thresholds, thereby preventing the undesirable merging of less similar tokens with respect to each layer. Furthermore, ATM introduces a novel token matching technique that considers not only similarity but also merging sizes, particularly for the final layers, to minimize the information loss incurred from each merging operation. We empirically validate our method across a wide range of pretrained models, demonstrating that ATM not only outperforms all existing training-free methods but also surpasses most training-intensive approaches, even without additional training. Remarkably, training-free ATM achieves over a 30% reduction in FLOPs for the DeiT-T and DeiT-S models without any drop in their original accuracy.