Daily Papers

Accurate reconstruction of both the geometric and topological details of a 3D object from a single 2D image embodies a fundamental challenge in computer vision. Existing explicit/implicit solutions to this problem struggle to recover self-occluded geometry and/or faithfully reconstruct topological shape structures. To resolve this dilemma, we introduce LIST, a novel neural architecture that leverages local and global image features to accurately reconstruct the geometric and topological structure of a 3D object from a single image. We utilize global 2D features to predict a coarse shape of the target object and then use it as a base for higher-resolution reconstruction. By leveraging both local 2D features from the image and 3D features from the coarse prediction, we can predict the signed distance between an arbitrary point and the target surface via an implicit predictor with great accuracy. Furthermore, our model does not require camera estimation or pixel alignment. It provides an uninfluenced reconstruction from the input-view direction. Through qualitative and quantitative analysis, we show the superiority of our model in reconstructing 3D objects from both synthetic and real-world images against the state of the art.

Convolutional Neural Networks define an exceptionally powerful class of models, but are still limited by the lack of ability to be spatially invariant to the input data in a computationally and parameter efficient manner. In this work we introduce a new learnable module, the Spatial Transformer, which explicitly allows the spatial manipulation of data within the network. This differentiable module can be inserted into existing convolutional architectures, giving neural networks the ability to actively spatially transform feature maps, conditional on the feature map itself, without any extra training supervision or modification to the optimisation process. We show that the use of spatial transformers results in models which learn invariance to translation, scale, rotation and more generic warping, resulting in state-of-the-art performance on several benchmarks, and for a number of classes of transformations.

Image deraining is an essential vision technique that removes rain streaks and water droplets, enhancing clarity for critical vision tasks like autonomous driving. However, current single-scale models struggle with fine-grained recovery and global consistency. To address this challenge, we propose Progressive Rain removal with Integrated State-space Modeling (PRISM), a progressive three-stage framework: Coarse Extraction Network (CENet), Frequency Fusion Network (SFNet), and Refine Network (RNet). Specifically, CENet and SFNet utilize a novel Hybrid Attention UNet (HA-UNet) for multi-scale feature aggregation by combining channel attention with windowed spatial transformers. Moreover, we propose Hybrid Domain Mamba (HDMamba) for SFNet to jointly model spatial semantics and wavelet domain characteristics. Finally, RNet recovers the fine-grained structures via an original-resolution subnetwork. Our model learns high-frequency rain characteristics while preserving structural details and maintaining global context, leading to improved image quality. Our method achieves competitive results on multiple datasets against recent deraining methods.

Spatial relationships between objects represent key scene information for humans to understand and interact with the world. To study the capability of current computer vision systems to recognize physically grounded spatial relations, we start by proposing precise relation definitions that permit consistently annotating a benchmark dataset. Despite the apparent simplicity of this task relative to others in the recognition literature, we observe that existing approaches perform poorly on this benchmark. We propose new approaches exploiting the long-range attention capabilities of transformers for this task, and evaluating key design principles. We identify a simple "RelatiViT" architecture and demonstrate that it outperforms all current approaches. To our knowledge, this is the first method to convincingly outperform naive baselines on spatial relation prediction in in-the-wild settings. The code and datasets are available in https://sites.google.com/view/spatial-relation.

Diffusion Transformers (DiTs) dominate video generation but their high computational cost severely limits real-world applicability, usually requiring tens of minutes to generate a few seconds of video even on high-performance GPUs. This inefficiency primarily arises from the quadratic computational complexity of 3D Full Attention with respect to the context length. In this paper, we propose a training-free framework termed Sparse VideoGen (SVG) that leverages the inherent sparsity in 3D Full Attention to boost inference efficiency. We reveal that the attention heads can be dynamically classified into two groups depending on distinct sparse patterns: (1) Spatial Head, where only spatially-related tokens within each frame dominate the attention output, and (2) Temporal Head, where only temporally-related tokens across different frames dominate. Based on this insight, SVG proposes an online profiling strategy to capture the dynamic sparse patterns and predicts the type of attention head. Combined with a novel hardware-efficient tensor layout transformation and customized kernel implementations, SVG achieves up to 2.28x and 2.33x end-to-end speedup on CogVideoX-v1.5 and HunyuanVideo, respectively, while preserving generation quality.

Recent Vision-Language-Action (VLA) models demonstrate remarkable generalization in robotics but are restricted by their substantial size and computational cost, limiting real-world deployment. However, conventional lightweighting methods often sacrifice critical capabilities, particularly spatial reasoning. This creates a trade-off between efficiency and performance. To address this challenge, our work reuses Register Tokens, which were introduced for artifact removal in Vision Transformers but subsequently discarded. We suppose that these tokens contain essential spatial information and propose RetoVLA, a novel architecture that reuses them directly by injecting them into the Action Expert. RetoVLA maintains a lightweight structure while leveraging this repurposed spatial context to enhance reasoning. We demonstrate RetoVLA's effectiveness through a series of comprehensive experiments. On our custom-built 7-DOF robot arm, the model achieves a 17.1%p absolute improvement in success rates for complex manipulation tasks. Our results confirm that reusing Register Tokens directly enhances spatial reasoning, demonstrating that what was previously discarded as an artifact is in fact a valuable, unexplored resource for robotic intelligence. A video demonstration is available at: https://youtu.be/2CseBR-snZg

This work introduces Video Diffusion Transformer (VDT), which pioneers the use of transformers in diffusion-based video generation. It features transformer blocks with modularized temporal and spatial attention modules to leverage the rich spatial-temporal representation inherited in transformers. We also propose a unified spatial-temporal mask modeling mechanism, seamlessly integrated with the model, to cater to diverse video generation scenarios. VDT offers several appealing benefits. 1) It excels at capturing temporal dependencies to produce temporally consistent video frames and even simulate the physics and dynamics of 3D objects over time. 2) It facilitates flexible conditioning information, \eg, simple concatenation in the token space, effectively unifying different token lengths and modalities. 3) Pairing with our proposed spatial-temporal mask modeling mechanism, it becomes a general-purpose video diffuser for harnessing a range of tasks, including unconditional generation, video prediction, interpolation, animation, and completion, etc. Extensive experiments on these tasks spanning various scenarios, including autonomous driving, natural weather, human action, and physics-based simulation, demonstrate the effectiveness of VDT. Additionally, we present comprehensive studies on how \model handles conditioning information with the mask modeling mechanism, which we believe will benefit future research and advance the field. Project page: https:VDT-2023.github.io

Convolutional networks and vision transformers have different forms of pairwise interactions, pooling across layers and pooling at the end of the network. Does the latter really need to be different? As a by-product of pooling, vision transformers provide spatial attention for free, but this is most often of low quality unless self-supervised, which is not well studied. Is supervision really the problem? In this work, we develop a generic pooling framework and then we formulate a number of existing methods as instantiations. By discussing the properties of each group of methods, we derive SimPool, a simple attention-based pooling mechanism as a replacement of the default one for both convolutional and transformer encoders. We find that, whether supervised or self-supervised, this improves performance on pre-training and downstream tasks and provides attention maps delineating object boundaries in all cases. One could thus call SimPool universal. To our knowledge, we are the first to obtain attention maps in supervised transformers of at least as good quality as self-supervised, without explicit losses or modifying the architecture. Code at: https://github.com/billpsomas/simpool.

Accurate depth estimation from monocular videos remains challenging due to ambiguities inherent in single-view geometry, as crucial depth cues like stereopsis are absent. However, humans often perceive relative depth intuitively by observing variations in the size and spacing of objects as they move. Inspired by this, we propose a novel method that infers relative depth by examining the spatial relationships and temporal evolution of a set of tracked 2D trajectories. Specifically, we use off-the-shelf point tracking models to capture 2D trajectories. Then, our approach employs spatial and temporal transformers to process these trajectories and directly infer depth changes over time. Evaluated on the TAPVid-3D benchmark, our method demonstrates robust zero-shot performance, generalizing effectively from synthetic to real-world datasets. Results indicate that our approach achieves temporally smooth, high-accuracy depth predictions across diverse domains.

We propose DOME, a diffusion-based world model that predicts future occupancy frames based on past occupancy observations. The ability of this world model to capture the evolution of the environment is crucial for planning in autonomous driving. Compared to 2D video-based world models, the occupancy world model utilizes a native 3D representation, which features easily obtainable annotations and is modality-agnostic. This flexibility has the potential to facilitate the development of more advanced world models. Existing occupancy world models either suffer from detail loss due to discrete tokenization or rely on simplistic diffusion architectures, leading to inefficiencies and difficulties in predicting future occupancy with controllability. Our DOME exhibits two key features:(1) High-Fidelity and Long-Duration Generation. We adopt a spatial-temporal diffusion transformer to predict future occupancy frames based on historical context. This architecture efficiently captures spatial-temporal information, enabling high-fidelity details and the ability to generate predictions over long durations. (2)Fine-grained Controllability. We address the challenge of controllability in predictions by introducing a trajectory resampling method, which significantly enhances the model's ability to generate controlled predictions. Extensive experiments on the widely used nuScenes dataset demonstrate that our method surpasses existing baselines in both qualitative and quantitative evaluations, establishing a new state-of-the-art performance on nuScenes. Specifically, our approach surpasses the baseline by 10.5% in mIoU and 21.2% in IoU for occupancy reconstruction and by 36.0% in mIoU and 24.6% in IoU for 4D occupancy forecasting.

We propose a flow-guided transformer, which innovatively leverage the motion discrepancy exposed by optical flows to instruct the attention retrieval in transformer for high fidelity video inpainting. More specially, we design a novel flow completion network to complete the corrupted flows by exploiting the relevant flow features in a local temporal window. With the completed flows, we propagate the content across video frames, and adopt the flow-guided transformer to synthesize the rest corrupted regions. We decouple transformers along temporal and spatial dimension, so that we can easily integrate the locally relevant completed flows to instruct spatial attention only. Furthermore, we design a flow-reweight module to precisely control the impact of completed flows on each spatial transformer. For the sake of efficiency, we introduce window partition strategy to both spatial and temporal transformers. Especially in spatial transformer, we design a dual perspective spatial MHSA, which integrates the global tokens to the window-based attention. Extensive experiments demonstrate the effectiveness of the proposed method qualitatively and quantitatively. Codes are available at https://github.com/hitachinsk/FGT.

Spatiotemporal predictive learning methods generally fall into two categories: recurrent-based approaches, which face challenges in parallelization and performance, and recurrent-free methods, which employ convolutional neural networks (CNNs) as encoder-decoder architectures. These methods benefit from strong inductive biases but often at the expense of scalability and generalization. This paper proposes PredFormer, a pure transformer-based framework for spatiotemporal predictive learning. Motivated by the Vision Transformers (ViT) design, PredFormer leverages carefully designed Gated Transformer blocks, following a comprehensive analysis of 3D attention mechanisms, including full-, factorized-, and interleaved-spatial-temporal attention. With its recurrent-free, transformer-based design, PredFormer is both simple and efficient, significantly outperforming previous methods by large margins. Extensive experiments on synthetic and real-world datasets demonstrate that PredFormer achieves state-of-the-art performance. On Moving MNIST, PredFormer achieves a 51.3% reduction in MSE relative to SimVP. For TaxiBJ, the model decreases MSE by 33.1% and boosts FPS from 533 to 2364. Additionally, on WeatherBench, it reduces MSE by 11.1% while enhancing FPS from 196 to 404. These performance gains in both accuracy and efficiency demonstrate PredFormer's potential for real-world applications. The source code will be released at https://github.com/yyyujintang/PredFormer .

Diffusion Transformers have established a new state-of-the-art in image synthesis, but the high computational cost of iterative sampling severely hampers their practical deployment. While existing acceleration methods often focus on the temporal domain, they overlook the substantial spatial redundancy inherent in the generative process, where global structures emerge long before fine-grained details are formed. The uniform computational treatment of all spatial regions represents a critical inefficiency. In this paper, we introduce Just-in-Time (JiT), a novel training-free framework that addresses this challenge by acceleration in the spatial domain. JiT formulates a spatially approximated generative ordinary differential equation (ODE) that drives the full latent state evolution based on computations from a dynamically selected, sparse subset of anchor tokens. To ensure seamless transitions as new tokens are incorporated to expand the dimensions of the latent state, we propose a deterministic micro-flow, a simple and effective finite-time ODE that maintains both structural coherence and statistical correctness. Extensive experiments on the state-of-the-art FLUX.1-dev model demonstrate that JiT achieves up to a 7x speedup with nearly lossless performance, significantly outperforming existing acceleration methods and establishing a new and superior trade-off between inference speed and generation fidelity.

Vision Transformer (ViT) extends the application range of transformers from language processing to computer vision tasks as being an alternative architecture against the existing convolutional neural networks (CNN). Since the transformer-based architecture has been innovative for computer vision modeling, the design convention towards an effective architecture has been less studied yet. From the successful design principles of CNN, we investigate the role of spatial dimension conversion and its effectiveness on transformer-based architecture. We particularly attend to the dimension reduction principle of CNNs; as the depth increases, a conventional CNN increases channel dimension and decreases spatial dimensions. We empirically show that such a spatial dimension reduction is beneficial to a transformer architecture as well, and propose a novel Pooling-based Vision Transformer (PiT) upon the original ViT model. We show that PiT achieves the improved model capability and generalization performance against ViT. Throughout the extensive experiments, we further show PiT outperforms the baseline on several tasks such as image classification, object detection, and robustness evaluation. Source codes and ImageNet models are available at https://github.com/naver-ai/pit

Distilling knowledge from convolutional neural networks (CNNs) is a double-edged sword for vision transformers (ViTs). It boosts the performance since the image-friendly local-inductive bias of CNN helps ViT learn faster and better, but leading to two problems: (1) Network designs of CNN and ViT are completely different, which leads to different semantic levels of intermediate features, making spatial-wise knowledge transfer methods (e.g., feature mimicking) inefficient. (2) Distilling knowledge from CNN limits the network convergence in the later training period since ViT's capability of integrating global information is suppressed by CNN's local-inductive-bias supervision. To this end, we present Cumulative Spatial Knowledge Distillation (CSKD). CSKD distills spatial-wise knowledge to all patch tokens of ViT from the corresponding spatial responses of CNN, without introducing intermediate features. Furthermore, CSKD exploits a Cumulative Knowledge Fusion (CKF) module, which introduces the global response of CNN and increasingly emphasizes its importance during the training. Applying CKF leverages CNN's local inductive bias in the early training period and gives full play to ViT's global capability in the later one. Extensive experiments and analysis on ImageNet-1k and downstream datasets demonstrate the superiority of our CSKD. Code will be publicly available.

Very recently, a variety of vision transformer architectures for dense prediction tasks have been proposed and they show that the design of spatial attention is critical to their success in these tasks. In this work, we revisit the design of the spatial attention and demonstrate that a carefully-devised yet simple spatial attention mechanism performs favourably against the state-of-the-art schemes. As a result, we propose two vision transformer architectures, namely, Twins-PCPVT and Twins-SVT. Our proposed architectures are highly-efficient and easy to implement, only involving matrix multiplications that are highly optimized in modern deep learning frameworks. More importantly, the proposed architectures achieve excellent performance on a wide range of visual tasks, including image level classification as well as dense detection and segmentation. The simplicity and strong performance suggest that our proposed architectures may serve as stronger backbones for many vision tasks. Our code is released at https://github.com/Meituan-AutoML/Twins .

A central objective in computer vision is to design models with appropriate 2-D inductive bias. Desiderata for 2D inductive bias include two-dimensional position awareness, dynamic spatial locality, and translation and permutation invariance. To address these goals, we leverage an expressive variation of the multidimensional State Space Model (SSM). Our approach introduces efficient parameterization, accelerated computation, and a suitable normalization scheme. Empirically, we observe that incorporating our layer at the beginning of each transformer block of Vision Transformers (ViT) significantly enhances performance for multiple ViT backbones and across datasets. The new layer is effective even with a negligible amount of additional parameters and inference time. Ablation studies and visualizations demonstrate that the layer has a strong 2-D inductive bias. For example, vision transformers equipped with our layer exhibit effective performance even without positional encoding

We introduce a novel training-free spatial grounding technique for text-to-image generation using Diffusion Transformers (DiT). Spatial grounding with bounding boxes has gained attention for its simplicity and versatility, allowing for enhanced user control in image generation. However, prior training-free approaches often rely on updating the noisy image during the reverse diffusion process via backpropagation from custom loss functions, which frequently struggle to provide precise control over individual bounding boxes. In this work, we leverage the flexibility of the Transformer architecture, demonstrating that DiT can generate noisy patches corresponding to each bounding box, fully encoding the target object and allowing for fine-grained control over each region. Our approach builds on an intriguing property of DiT, which we refer to as semantic sharing. Due to semantic sharing, when a smaller patch is jointly denoised alongside a generatable-size image, the two become "semantic clones". Each patch is denoised in its own branch of the generation process and then transplanted into the corresponding region of the original noisy image at each timestep, resulting in robust spatial grounding for each bounding box. In our experiments on the HRS and DrawBench benchmarks, we achieve state-of-the-art performance compared to previous training-free spatial grounding approaches.

We present BoTNet, a conceptually simple yet powerful backbone architecture that incorporates self-attention for multiple computer vision tasks including image classification, object detection and instance segmentation. By just replacing the spatial convolutions with global self-attention in the final three bottleneck blocks of a ResNet and no other changes, our approach improves upon the baselines significantly on instance segmentation and object detection while also reducing the parameters, with minimal overhead in latency. Through the design of BoTNet, we also point out how ResNet bottleneck blocks with self-attention can be viewed as Transformer blocks. Without any bells and whistles, BoTNet achieves 44.4% Mask AP and 49.7% Box AP on the COCO Instance Segmentation benchmark using the Mask R-CNN framework; surpassing the previous best published single model and single scale results of ResNeSt evaluated on the COCO validation set. Finally, we present a simple adaptation of the BoTNet design for image classification, resulting in models that achieve a strong performance of 84.7% top-1 accuracy on the ImageNet benchmark while being up to 1.64x faster in compute time than the popular EfficientNet models on TPU-v3 hardware. We hope our simple and effective approach will serve as a strong baseline for future research in self-attention models for vision

There still remains an extreme performance gap between Vision Transformers (ViTs) and Convolutional Neural Networks (CNNs) when training from scratch on small datasets, which is concluded to the lack of inductive bias. In this paper, we further consider this problem and point out two weaknesses of ViTs in inductive biases, that is, the spatial relevance and diverse channel representation. First, on spatial aspect, objects are locally compact and relevant, thus fine-grained feature needs to be extracted from a token and its neighbors. While the lack of data hinders ViTs to attend the spatial relevance. Second, on channel aspect, representation exhibits diversity on different channels. But the scarce data can not enable ViTs to learn strong enough representation for accurate recognition. To this end, we propose Dynamic Hybrid Vision Transformer (DHVT) as the solution to enhance the two inductive biases. On spatial aspect, we adopt a hybrid structure, in which convolution is integrated into patch embedding and multi-layer perceptron module, forcing the model to capture the token features as well as their neighboring features. On channel aspect, we introduce a dynamic feature aggregation module in MLP and a brand new "head token" design in multi-head self-attention module to help re-calibrate channel representation and make different channel group representation interacts with each other. The fusion of weak channel representation forms a strong enough representation for classification. With this design, we successfully eliminate the performance gap between CNNs and ViTs, and our DHVT achieves a series of state-of-the-art performance with a lightweight model, 85.68% on CIFAR-100 with 22.8M parameters, 82.3% on ImageNet-1K with 24.0M parameters. Code is available at https://github.com/ArieSeirack/DHVT.

In this work, we introduce Dual Attention Vision Transformers (DaViT), a simple yet effective vision transformer architecture that is able to capture global context while maintaining computational efficiency. We propose approaching the problem from an orthogonal angle: exploiting self-attention mechanisms with both "spatial tokens" and "channel tokens". With spatial tokens, the spatial dimension defines the token scope, and the channel dimension defines the token feature dimension. With channel tokens, we have the inverse: the channel dimension defines the token scope, and the spatial dimension defines the token feature dimension. We further group tokens along the sequence direction for both spatial and channel tokens to maintain the linear complexity of the entire model. We show that these two self-attentions complement each other: (i) since each channel token contains an abstract representation of the entire image, the channel attention naturally captures global interactions and representations by taking all spatial positions into account when computing attention scores between channels; (ii) the spatial attention refines the local representations by performing fine-grained interactions across spatial locations, which in turn helps the global information modeling in channel attention. Extensive experiments show our DaViT achieves state-of-the-art performance on four different tasks with efficient computations. Without extra data, DaViT-Tiny, DaViT-Small, and DaViT-Base achieve 82.8%, 84.2%, and 84.6% top-1 accuracy on ImageNet-1K with 28.3M, 49.7M, and 87.9M parameters, respectively. When we further scale up DaViT with 1.5B weakly supervised image and text pairs, DaViT-Gaint reaches 90.4% top-1 accuracy on ImageNet-1K. Code is available at https://github.com/dingmyu/davit.

Transformer first appears in the field of natural language processing and is later migrated to the computer vision domain, where it demonstrates excellent performance in vision tasks. However, recently, Retentive Network (RetNet) has emerged as an architecture with the potential to replace Transformer, attracting widespread attention in the NLP community. Therefore, we raise the question of whether transferring RetNet's idea to vision can also bring outstanding performance to vision tasks. To address this, we combine RetNet and Transformer to propose RMT. Inspired by RetNet, RMT introduces explicit decay into the vision backbone, bringing prior knowledge related to spatial distances to the vision model. This distance-related spatial prior allows for explicit control of the range of tokens that each token can attend to. Additionally, to reduce the computational cost of global modeling, we decompose this modeling process along the two coordinate axes of the image. Abundant experiments have demonstrated that our RMT exhibits exceptional performance across various computer vision tasks. For example, RMT achieves 84.1% Top1-acc on ImageNet-1k using merely 4.5G FLOPs. To the best of our knowledge, among all models, RMT achieves the highest Top1-acc when models are of similar size and trained with the same strategy. Moreover, RMT significantly outperforms existing vision backbones in downstream tasks such as object detection, instance segmentation, and semantic segmentation. Our work is still in progress.

Sensor fusion is critical to perception systems for task domains such as autonomous driving and robotics. Recently, the Transformer integrated with CNN has demonstrated high performance in sensor fusion for various perception tasks. In this work, we introduce a method for fusing data from camera and LiDAR. By employing Transformer modules at multiple resolutions, proposed method effectively combines local and global contextual relationships. The performance of the proposed method is validated by extensive experiments with two adversarial benchmarks with lengthy routes and high-density traffics. The proposed method outperforms previous approaches with the most challenging benchmarks, achieving significantly higher driving and infraction scores. Compared with TransFuser, it achieves 8% and 19% improvement in driving scores for the Longest6 and Town05 Long benchmarks, respectively.

Convolutional neural networks (CNNs) have so far been the de-facto model for visual data. Recent work has shown that (Vision) Transformer models (ViT) can achieve comparable or even superior performance on image classification tasks. This raises a central question: how are Vision Transformers solving these tasks? Are they acting like convolutional networks, or learning entirely different visual representations? Analyzing the internal representation structure of ViTs and CNNs on image classification benchmarks, we find striking differences between the two architectures, such as ViT having more uniform representations across all layers. We explore how these differences arise, finding crucial roles played by self-attention, which enables early aggregation of global information, and ViT residual connections, which strongly propagate features from lower to higher layers. We study the ramifications for spatial localization, demonstrating ViTs successfully preserve input spatial information, with noticeable effects from different classification methods. Finally, we study the effect of (pretraining) dataset scale on intermediate features and transfer learning, and conclude with a discussion on connections to new architectures such as the MLP-Mixer.

Recently, token-based generation have demonstrated their effectiveness in image synthesis. As a representative example, non-autoregressive Transformers (NATs) can generate decent-quality images in a few steps. NATs perform generation in a progressive manner, where the latent tokens of a resulting image are incrementally revealed. At each step, the unrevealed image regions are padded with mask tokens and inferred by NAT. In this paper, we delve into the mechanisms behind the effectiveness of NATs and uncover two important patterns that naturally emerge from NATs: Spatially (within a step), although mask and visible tokens are processed uniformly by NATs, the interactions between them are highly asymmetric. In specific, mask tokens mainly gather information for decoding, while visible tokens tend to primarily provide information, and their deep representations can be built only upon themselves. Temporally (across steps), the interactions between adjacent generation steps mostly concentrate on updating the representations of a few critical tokens, while the computation for the majority of tokens is generally repetitive. Driven by these findings, we propose EfficientNAT (ENAT), a NAT model that explicitly encourages these critical interactions inherent in NATs. At the spatial level, we disentangle the computations of visible and mask tokens by encoding visible tokens independently, while decoding mask tokens conditioned on the fully encoded visible tokens. At the temporal level, we prioritize the computation of the critical tokens at each step, while maximally reusing previously computed token representations to supplement necessary information. ENAT improves the performance of NATs notably with significantly reduced computational cost. Experiments on ImageNet-256, ImageNet-512 and MS-COCO validate the effectiveness of ENAT. Code is available at https://github.com/LeapLabTHU/ENAT.

Spike-based Transformer presents a compelling and energy-efficient alternative to traditional Artificial Neural Network (ANN)-based Transformers, achieving impressive results through sparse binary computations. However, existing spike-based transformers predominantly focus on spatial attention while neglecting crucial temporal dependencies inherent in spike-based processing, leading to suboptimal feature representation and limited performance. To address this limitation, we propose Spiking Transformer with Spatial-Temporal Attention (STAtten), a simple and straightforward architecture that efficiently integrates both spatial and temporal information in the self-attention mechanism. STAtten introduces a block-wise computation strategy that processes information in spatial-temporal chunks, enabling comprehensive feature capture while maintaining the same computational complexity as previous spatial-only approaches. Our method can be seamlessly integrated into existing spike-based transformers without architectural overhaul. Extensive experiments demonstrate that STAtten significantly improves the performance of existing spike-based transformers across both static and neuromorphic datasets, including CIFAR10/100, ImageNet, CIFAR10-DVS, and N-Caltech101. The code is available at https://github.com/Intelligent-Computing-Lab-Yale/STAtten

In this paper we introduce the Temporo-Spatial Vision Transformer (TSViT), a fully-attentional model for general Satellite Image Time Series (SITS) processing based on the Vision Transformer (ViT). TSViT splits a SITS record into non-overlapping patches in space and time which are tokenized and subsequently processed by a factorized temporo-spatial encoder. We argue, that in contrast to natural images, a temporal-then-spatial factorization is more intuitive for SITS processing and present experimental evidence for this claim. Additionally, we enhance the model's discriminative power by introducing two novel mechanisms for acquisition-time-specific temporal positional encodings and multiple learnable class tokens. The effect of all novel design choices is evaluated through an extensive ablation study. Our proposed architecture achieves state-of-the-art performance, surpassing previous approaches by a significant margin in three publicly available SITS semantic segmentation and classification datasets. All model, training and evaluation codes are made publicly available to facilitate further research.

Recent progress in vision Transformers exhibits great success in various tasks driven by the new spatial modeling mechanism based on dot-product self-attention. In this paper, we show that the key ingredients behind the vision Transformers, namely input-adaptive, long-range and high-order spatial interactions, can also be efficiently implemented with a convolution-based framework. We present the Recursive Gated Convolution (g^nConv) that performs high-order spatial interactions with gated convolutions and recursive designs. The new operation is highly flexible and customizable, which is compatible with various variants of convolution and extends the two-order interactions in self-attention to arbitrary orders without introducing significant extra computation. g^nConv can serve as a plug-and-play module to improve various vision Transformers and convolution-based models. Based on the operation, we construct a new family of generic vision backbones named HorNet. Extensive experiments on ImageNet classification, COCO object detection and ADE20K semantic segmentation show HorNet outperform Swin Transformers and ConvNeXt by a significant margin with similar overall architecture and training configurations. HorNet also shows favorable scalability to more training data and larger model sizes. Apart from the effectiveness in visual encoders, we also show g^nConv can be applied to task-specific decoders and consistently improve dense prediction performance with less computation. Our results demonstrate that g^nConv can be a new basic module for visual modeling that effectively combines the merits of both vision Transformers and CNNs. Code is available at https://github.com/raoyongming/HorNet

Recent progress in multimodal generation has increasingly combined autoregressive (AR) and diffusion-based approaches, leveraging their complementary strengths: AR models capture long-range dependencies and produce fluent, context-aware outputs, while diffusion models operate in continuous latent spaces to refine high-fidelity visual details. However, existing hybrids often lack systematic guidance on how and why to allocate model capacity between these paradigms. In this work, we introduce MADFormer, a Mixed Autoregressive and Diffusion Transformer that serves as a testbed for analyzing AR-diffusion trade-offs. MADFormer partitions image generation into spatial blocks, using AR layers for one-pass global conditioning across blocks and diffusion layers for iterative local refinement within each block. Through controlled experiments on FFHQ-1024 and ImageNet, we identify two key insights: (1) block-wise partitioning significantly improves performance on high-resolution images, and (2) vertically mixing AR and diffusion layers yields better quality-efficiency balances--improving FID by up to 75% under constrained inference compute. Our findings offer practical design principles for future hybrid generative models.

As it is empirically observed that Vision Transformers (ViTs) are quite insensitive to the order of input tokens, the need for an appropriate self-supervised pretext task that enhances the location awareness of ViTs is becoming evident. To address this, we present DropPos, a novel pretext task designed to reconstruct Dropped Positions. The formulation of DropPos is simple: we first drop a large random subset of positional embeddings and then the model classifies the actual position for each non-overlapping patch among all possible positions solely based on their visual appearance. To avoid trivial solutions, we increase the difficulty of this task by keeping only a subset of patches visible. Additionally, considering there may be different patches with similar visual appearances, we propose position smoothing and attentive reconstruction strategies to relax this classification problem, since it is not necessary to reconstruct their exact positions in these cases. Empirical evaluations of DropPos show strong capabilities. DropPos outperforms supervised pre-training and achieves competitive results compared with state-of-the-art self-supervised alternatives on a wide range of downstream benchmarks. This suggests that explicitly encouraging spatial reasoning abilities, as DropPos does, indeed contributes to the improved location awareness of ViTs. The code is publicly available at https://github.com/Haochen-Wang409/DropPos.

In recent years, Vision Transformers have attracted increasing interest from computer vision researchers. However, the advantage of these transformers over CNNs is only fully manifested when trained over a large dataset, mainly due to the reduced inductive bias towards spatial locality within the transformer's self-attention mechanism. In this work, we present a data-efficient vision transformer that does not rely on self-attention. Instead, it employs a novel generalization to multiple axes of the very recent Hyena layer. We propose several alternative approaches for obtaining this generalization and delve into their unique distinctions and considerations from both empirical and theoretical perspectives. Our empirical findings indicate that the proposed Hyena N-D layer boosts the performance of various Vision Transformer architectures, such as ViT, Swin, and DeiT across multiple datasets. Furthermore, in the small dataset regime, our Hyena-based ViT is favorable to ViT variants from the recent literature that are specifically designed for solving the same challenge, i.e., working with small datasets or incorporating image-specific inductive bias into the self-attention mechanism. Finally, we show that a hybrid approach that is based on Hyena N-D for the first layers in ViT, followed by layers that incorporate conventional attention, consistently boosts the performance of various vision transformer architectures.

Vision Transformers rely on positional embeddings and class tokens that encode fixed spatial priors. While effective for natural images, these priors may hinder generalization when spatial layout is weakly informative or inconsistent, a frequent condition in medical imaging and edge-deployed clinical systems. We introduce ZACH-ViT (Zero-token Adaptive Compact Hierarchical Vision Transformer), a compact Vision Transformer that removes both positional embeddings and the [CLS] token, achieving permutation invariance through global average pooling over patch representations. The term "Zero-token" specifically refers to removing the dedicated [CLS] aggregation token and positional embeddings; patch tokens remain unchanged and are processed normally. Adaptive residual projections preserve training stability in compact configurations while maintaining a strict parameter budget. Evaluation is performed across seven MedMNIST datasets spanning binary and multi-class tasks under a strict few-shot protocol (50 samples per class, fixed hyperparameters, five random seeds). The empirical analysis demonstrates regime-dependent behavior: ZACH-ViT (0.25M parameters, trained from scratch) achieves its strongest advantage on BloodMNIST and remains competitive with TransMIL on PathMNIST, while its relative advantage decreases on datasets with strong anatomical priors (OCTMNIST, OrganAMNIST), consistent with the architectural hypothesis. These findings support the view that aligning architectural inductive bias with data structure can be more important than pursuing universal benchmark dominance. Despite its minimal size and lack of pretraining, ZACH-ViT achieves competitive performance while maintaining sub-second inference times, supporting deployment in resource-constrained clinical environments. Code and models are available at https://github.com/Bluesman79/ZACH-ViT.

The resurgence of convolutional neural networks (CNNs) in visual recognition tasks, exemplified by ConvNeXt, has demonstrated their capability to rival transformer-based architectures through advanced training methodologies and ViT-inspired design principles. However, both CNNs and transformers exhibit a simplicity bias, favoring straightforward features over complex structural representations. Furthermore, modern CNNs often integrate MLP-like blocks akin to those in transformers, but these blocks suffer from significant information redundancies, necessitating high expansion ratios to sustain competitive performance. To address these limitations, we propose SpaRTAN, a lightweight architectural design that enhances spatial and channel-wise information processing. SpaRTAN employs kernels with varying receptive fields, controlled by kernel size and dilation factor, to capture discriminative multi-order spatial features effectively. A wave-based channel aggregation module further modulates and reinforces pixel interactions, mitigating channel-wise redundancies. Combining the two modules, the proposed network can efficiently gather and dynamically contextualize discriminative features. Experimental results in ImageNet and COCO demonstrate that SpaRTAN achieves remarkable parameter efficiency while maintaining competitive performance. In particular, on the ImageNet-1k benchmark, SpaRTAN achieves 77. 7% accuracy with only 3.8M parameters and approximately 1.0 GFLOPs, demonstrating its ability to deliver strong performance through an efficient design. On the COCO benchmark, it achieves 50.0% AP, surpassing the previous benchmark by 1.2% with only 21.5M parameters. The code is publicly available at [https://github.com/henry-pay/SpaRTAN].

Hyperspectral object tracking using snapshot mosaic cameras is emerging as it provides enhanced spectral information alongside spatial data, contributing to a more comprehensive understanding of material properties. Using transformers, which have consistently outperformed convolutional neural networks (CNNs) in learning better feature representations, would be expected to be effective for Hyperspectral object tracking. However, training large transformers necessitates extensive datasets and prolonged training periods. This is particularly critical for complex tasks like object tracking, and the scarcity of large datasets in the hyperspectral domain acts as a bottleneck in achieving the full potential of powerful transformer models. This paper proposes an effective methodology that adapts large pretrained transformer-based foundation models for hyperspectral object tracking. We propose an adaptive, learnable spatial-spectral token fusion module that can be extended to any transformer-based backbone for learning inherent spatial-spectral features in hyperspectral data. Furthermore, our model incorporates a cross-modality training pipeline that facilitates effective learning across hyperspectral datasets collected with different sensor modalities. This enables the extraction of complementary knowledge from additional modalities, whether or not they are present during testing. Our proposed model also achieves superior performance with minimal training iterations.

In recent years, the emergence of Transformers with self-attention mechanism has revolutionized the hyperspectral image (HSI) classification. However, these models face major challenges in computational efficiency, as their complexity increases quadratically with the sequence length. The Mamba architecture, leveraging a state space model (SSM), offers a more efficient alternative to Transformers. This paper introduces the Spatial-Spectral Morphological Mamba (MorpMamba) model in which, a token generation module first converts the HSI patch into spatial-spectral tokens. These tokens are then processed by morphological operations, which compute structural and shape information using depthwise separable convolutional operations. The extracted information is enhanced in a feature enhancement module that adjusts the spatial and spectral tokens based on the center region of the HSI sample, allowing for effective information fusion within each block. Subsequently, the tokens are refined through a multi-head self-attention which further improves the feature space. Finally, the combined information is fed into the state space block for classification and the creation of the ground truth map. Experiments on widely used HSI datasets demonstrate that the MorpMamba model outperforms (parametric efficiency) both CNN and Transformer models. The source code will be made publicly available at https://github.com/MHassaanButt/MorpMamba.

Pixel-level labels are particularly expensive to acquire. Hence, pretraining is a critical step to improve models on a task like semantic segmentation. However, prominent algorithms for pretraining neural networks use image-level objectives, e.g. image classification, image-text alignment a la CLIP, or self-supervised contrastive learning. These objectives do not model spatial information, which might be sub-optimal when finetuning on downstream tasks with spatial reasoning. In this work, we pretrain network with a location-aware (LOCA) self-supervised method which fosters the emergence of strong dense features. Specifically, we use both a patch-level clustering scheme to mine dense pseudo-labels and a relative location prediction task to encourage learning about object parts and their spatial arrangements. Our experiments show that LOCA pretraining leads to representations that transfer competitively to challenging and diverse semantic segmentation datasets.

Tables are widely used with various structures to organize and present data. Recent attempts on table understanding mainly focus on relational tables, yet overlook to other common table structures. In this paper, we propose TUTA, a unified pre-training architecture for understanding generally structured tables. Noticing that understanding a table requires spatial, hierarchical, and semantic information, we enhance transformers with three novel structure-aware mechanisms. First, we devise a unified tree-based structure, called a bi-dimensional coordinate tree, to describe both the spatial and hierarchical information of generally structured tables. Upon this, we propose tree-based attention and position embedding to better capture the spatial and hierarchical information. Moreover, we devise three progressive pre-training objectives to enable representations at the token, cell, and table levels. We pre-train TUTA on a wide range of unlabeled web and spreadsheet tables and fine-tune it on two critical tasks in the field of table structure understanding: cell type classification and table type classification. Experiments show that TUTA is highly effective, achieving state-of-the-art on five widely-studied datasets.

In this paper, we present DiT-MoE, a sparse version of the diffusion Transformer, that is scalable and competitive with dense networks while exhibiting highly optimized inference. The DiT-MoE includes two simple designs: shared expert routing and expert-level balance loss, thereby capturing common knowledge and reducing redundancy among the different routed experts. When applied to conditional image generation, a deep analysis of experts specialization gains some interesting observations: (i) Expert selection shows preference with spatial position and denoising time step, while insensitive with different class-conditional information; (ii) As the MoE layers go deeper, the selection of experts gradually shifts from specific spacial position to dispersion and balance. (iii) Expert specialization tends to be more concentrated at the early time step and then gradually uniform after half. We attribute it to the diffusion process that first models the low-frequency spatial information and then high-frequency complex information. Based on the above guidance, a series of DiT-MoE experimentally achieves performance on par with dense networks yet requires much less computational load during inference. More encouragingly, we demonstrate the potential of DiT-MoE with synthesized image data, scaling diffusion model at a 16.5B parameter that attains a new SoTA FID-50K score of 1.80 in 512times512 resolution settings. The project page: https://github.com/feizc/DiT-MoE.

As the rapid progression of practical applications based on Large Language Models continues, the importance of extrapolating performance has grown exponentially in the research domain. In our study, we identified an anomalous behavior in Transformer models that had been previously overlooked, leading to a chaos around closest tokens which carried the most important information. We've coined this discovery the "headache of Transformers". To address this at its core, we introduced a novel self-attention structure named Collinear Constrained Attention (CoCA). This structure can be seamlessly integrated with existing extrapolation, interpolation methods, and other optimization strategies designed for traditional Transformer models. We have achieved excellent extrapolating performance even for 16 times to 24 times of sequence lengths during inference without any fine-tuning on our model. We have also enhanced CoCA's computational and spatial efficiency to ensure its practicality. We plan to open-source CoCA shortly. In the meantime, we've made our code available in the appendix for reappearing experiments.

DETR has been recently proposed to eliminate the need for many hand-designed components in object detection while demonstrating good performance. However, it suffers from slow convergence and limited feature spatial resolution, due to the limitation of Transformer attention modules in processing image feature maps. To mitigate these issues, we proposed Deformable DETR, whose attention modules only attend to a small set of key sampling points around a reference. Deformable DETR can achieve better performance than DETR (especially on small objects) with 10 times less training epochs. Extensive experiments on the COCO benchmark demonstrate the effectiveness of our approach. Code is released at https://github.com/fundamentalvision/Deformable-DETR.

Vision Transformers (ViTs) have triggered the most recent and significant breakthroughs in computer vision. Their efficient designs are mostly guided by the indirect metric of computational complexity, i.e., FLOPs, which however has a clear gap with the direct metric such as throughput. Thus, we propose to use the direct speed evaluation on the target platform as the design principle for efficient ViTs. Particularly, we introduce LITv2, a simple and effective ViT which performs favourably against the existing state-of-the-art methods across a spectrum of different model sizes with faster speed. At the core of LITv2 is a novel self-attention mechanism, which we dub HiLo. HiLo is inspired by the insight that high frequencies in an image capture local fine details and low frequencies focus on global structures, whereas a multi-head self-attention layer neglects the characteristic of different frequencies. Therefore, we propose to disentangle the high/low frequency patterns in an attention layer by separating the heads into two groups, where one group encodes high frequencies via self-attention within each local window, and another group encodes low frequencies by performing global attention between the average-pooled low-frequency keys and values from each window and each query position in the input feature map. Benefiting from the efficient design for both groups, we show that HiLo is superior to the existing attention mechanisms by comprehensively benchmarking FLOPs, speed and memory consumption on GPUs and CPUs. For example, HiLo is 1.4x faster than spatial reduction attention and 1.6x faster than local window attention on CPUs. Powered by HiLo, LITv2 serves as a strong backbone for mainstream vision tasks including image classification, dense detection and segmentation. Code is available at https://github.com/ziplab/LITv2.

Leveraging pre-trained Diffusion Transformers (DiTs) for high-resolution (HR) image synthesis often leads to spatial layout collapse and degraded texture fidelity. Prior work mitigates these issues with complex pipelines that first perform a base-resolution (i.e., training-resolution) denoising process to guide HR generation. We instead explore the intrinsic generative mechanisms of DiTs and propose ResDiT, a training-free method that scales resolution efficiently. We identify the core factor governing spatial layout, position embeddings (PEs), and show that the original PEs encode incorrect positional information when extrapolated to HR, which triggers layout collapse. To address this, we introduce a PE scaling technique that rectifies positional encoding under resolution changes. To further remedy low-fidelity details, we develop a local-enhancement mechanism grounded in base-resolution local attention. We design a patch-level fusion module that aggregates global and local cues, together with a Gaussian-weighted splicing strategy that eliminates grid artifacts. Comprehensive evaluations demonstrate that ResDiT consistently delivers high-fidelity, high-resolution image synthesis and integrates seamlessly with downstream tasks, including spatially controlled generation.

Vision-Language-Action (VLA) models have recently shown impressive generalization and language-guided manipulation capabilities. However, their performance degrades on tasks requiring precise spatial reasoning due to limited spatial reasoning inherited from Vision-Language Models (VLMs). Existing VLAs rely on extensive action-data pretraining to ground VLMs in 3D space, which reduces training efficiency and is still insufficient for accurate spatial understanding. In this work, we present DepthVLA, a simple yet effective VLA architecture that explicitly incorporates spatial awareness through a pretrained depth prediction module. DepthVLA adopts a mixture-of-transformers design that unifies a VLM, a depth transformer, and an action expert with fully shared attentions, forming an end-to-end model with enhanced spatial reasoning. Extensive evaluations in both real-world and simulated environments show that DepthVLA outperforms state-of-the-art approaches, achieving 78.5% vs. 65.0% progress in real-world tasks, 94.9% vs. 93.6% in the LIBERO simulator, and 74.8% vs. 58.8% in the Simpler simulator. Our code will be made publicly available.

Omnidirectional videos (ODVs) are redefining viewer experiences in virtual reality (VR) by offering an unprecedented full field-of-view (FOV). This study extends the domain of saliency prediction to 360-degree environments, addressing the complexities of spherical distortion and the integration of spatial audio. Contextually, ODVs have transformed user experience by adding a spatial audio dimension that aligns sound direction with the viewer's perspective in spherical scenes. Motivated by the lack of comprehensive datasets for 360-degree audio-visual saliency prediction, our study curates YT360-EyeTracking, a new dataset of 81 ODVs, each observed under varying audio-visual conditions. Our goal is to explore how to utilize audio-visual cues to effectively predict visual saliency in 360-degree videos. Towards this aim, we propose two novel saliency prediction models: SalViT360, a vision-transformer-based framework for ODVs equipped with spherical geometry-aware spatio-temporal attention layers, and SalViT360-AV, which further incorporates transformer adapters conditioned on audio input. Our results on a number of benchmark datasets, including our YT360-EyeTracking, demonstrate that SalViT360 and SalViT360-AV significantly outperform existing methods in predicting viewer attention in 360-degree scenes. Interpreting these results, we suggest that integrating spatial audio cues in the model architecture is crucial for accurate saliency prediction in omnidirectional videos. Code and dataset will be available at https://cyberiada.github.io/SalViT360.

Monocular 3D human pose estimation remains a challenging task due to inherent depth ambiguities and occlusions. Compared to traditional methods based on Transformers or Convolutional Neural Networks (CNNs), recent diffusion-based approaches have shown superior performance, leveraging their probabilistic nature and high-fidelity generation capabilities. However, these methods often fail to account for the spatial and temporal correlations across predicted frames, resulting in limited temporal consistency and inferior accuracy in predicted 3D pose sequences. To address these shortcomings, this paper proposes StarPose, an autoregressive diffusion framework that effectively incorporates historical 3D pose predictions and spatial-temporal physical guidance to significantly enhance both the accuracy and temporal coherence of pose predictions. Unlike existing approaches, StarPose models the 2D-to-3D pose mapping as an autoregressive diffusion process. By synergically integrating previously predicted 3D poses with 2D pose inputs via a Historical Pose Integration Module (HPIM), the framework generates rich and informative historical pose embeddings that guide subsequent denoising steps, ensuring temporally consistent predictions. In addition, a fully plug-and-play Spatial-Temporal Physical Guidance (STPG) mechanism is tailored to refine the denoising process in an iterative manner, which further enforces spatial anatomical plausibility and temporal motion dynamics, rendering robust and realistic pose estimates. Extensive experiments on benchmark datasets demonstrate that StarPose outperforms state-of-the-art methods, achieving superior accuracy and temporal consistency in 3D human pose estimation. Code is available at https://github.com/wileychan/StarPose.

We present GLOD, a transformer-first architecture for object detection in high-resolution satellite imagery. GLOD replaces CNN backbones with a Swin Transformer for end-to-end feature extraction, combined with novel UpConvMixer blocks for robust upsampling and Fusion Blocks for multi-scale feature integration. Our approach achieves 32.95\% on xView, outperforming SOTA methods by 11.46\%. Key innovations include asymmetric fusion with CBAM attention and a multi-path head design capturing objects across scales. The architecture is optimized for satellite imagery challenges, leveraging spatial priors while maintaining computational efficiency.

Vision transformers have gained popularity recently, leading to the development of new vision backbones with improved features and consistent performance gains. However, these advancements are not solely attributable to novel feature transformation designs; certain benefits also arise from advanced network-level and block-level architectures. This paper aims to identify the real gains of popular convolution and attention operators through a detailed study. We find that the key difference among these feature transformation modules, such as attention or convolution, lies in their spatial feature aggregation approach, known as the "spatial token mixer" (STM). To facilitate an impartial comparison, we introduce a unified architecture to neutralize the impact of divergent network-level and block-level designs. Subsequently, various STMs are integrated into this unified framework for comprehensive comparative analysis. Our experiments on various tasks and an analysis of inductive bias show a significant performance boost due to advanced network-level and block-level designs, but performance differences persist among different STMs. Our detailed analysis also reveals various findings about different STMs, such as effective receptive fields and invariance tests. All models and codes used in this study are publicly available at https://github.com/OpenGVLab/STM-Evaluation.

Efficiently quantifying renal structures can provide distinct spatial context and facilitate biomarker discovery for kidney morphology. However, the development and evaluation of the transformer model to segment the renal cortex, medulla, and collecting system remains challenging due to data inefficiency. Inspired by the hierarchical structures in vision transformer, we propose a novel method using a 3D block aggregation transformer for segmenting kidney components on contrast-enhanced CT scans. We construct the first cohort of renal substructures segmentation dataset with 116 subjects under institutional review board (IRB) approval. Our method yields the state-of-the-art performance (Dice of 0.8467) against the baseline approach of 0.8308 with the data-efficient design. The Pearson R achieves 0.9891 between the proposed method and manual standards and indicates the strong correlation and reproducibility for volumetric analysis. We extend the proposed method to the public KiTS dataset, the method leads to improved accuracy compared to transformer-based approaches. We show that the 3D block aggregation transformer can achieve local communication between sequence representations without modifying self-attention, and it can serve as an accurate and efficient quantification tool for characterizing renal structures.

We present Multiscale Vision Transformers (MViT) for video and image recognition, by connecting the seminal idea of multiscale feature hierarchies with transformer models. Multiscale Transformers have several channel-resolution scale stages. Starting from the input resolution and a small channel dimension, the stages hierarchically expand the channel capacity while reducing the spatial resolution. This creates a multiscale pyramid of features with early layers operating at high spatial resolution to model simple low-level visual information, and deeper layers at spatially coarse, but complex, high-dimensional features. We evaluate this fundamental architectural prior for modeling the dense nature of visual signals for a variety of video recognition tasks where it outperforms concurrent vision transformers that rely on large scale external pre-training and are 5-10x more costly in computation and parameters. We further remove the temporal dimension and apply our model for image classification where it outperforms prior work on vision transformers. Code is available at: https://github.com/facebookresearch/SlowFast

Designed to learn long-range interactions on sequential data, transformers continue to show state-of-the-art results on a wide variety of tasks. In contrast to CNNs, they contain no inductive bias that prioritizes local interactions. This makes them expressive, but also computationally infeasible for long sequences, such as high-resolution images. We demonstrate how combining the effectiveness of the inductive bias of CNNs with the expressivity of transformers enables them to model and thereby synthesize high-resolution images. We show how to (i) use CNNs to learn a context-rich vocabulary of image constituents, and in turn (ii) utilize transformers to efficiently model their composition within high-resolution images. Our approach is readily applied to conditional synthesis tasks, where both non-spatial information, such as object classes, and spatial information, such as segmentations, can control the generated image. In particular, we present the first results on semantically-guided synthesis of megapixel images with transformers and obtain the state of the art among autoregressive models on class-conditional ImageNet. Code and pretrained models can be found at https://github.com/CompVis/taming-transformers .

High spatial frequency information, including fine details like textures, significantly contributes to the accuracy of semantic segmentation. However, according to the Nyquist-Shannon Sampling Theorem, high-frequency components are vulnerable to aliasing or distortion when propagating through downsampling layers such as strided-convolution. Here, we propose a novel Spatial Frequency Modulation (SFM) that modulates high-frequency features to a lower frequency before downsampling and then demodulates them back during upsampling. Specifically, we implement modulation through adaptive resampling (ARS) and design a lightweight add-on that can densely sample the high-frequency areas to scale up the signal, thereby lowering its frequency in accordance with the Frequency Scaling Property. We also propose Multi-Scale Adaptive Upsampling (MSAU) to demodulate the modulated feature and recover high-frequency information through non-uniform upsampling This module further improves segmentation by explicitly exploiting information interaction between densely and sparsely resampled areas at multiple scales. Both modules can seamlessly integrate with various architectures, extending from convolutional neural networks to transformers. Feature visualization and analysis confirm that our method effectively alleviates aliasing while successfully retaining details after demodulation. Finally, we validate the broad applicability and effectiveness of SFM by extending it to image classification, adversarial robustness, instance segmentation, and panoptic segmentation tasks. The code is available at https://github.com/Linwei-Chen/SFM.

Diffusion transformers (DiTs) achieve high generative quality but lock FLOPs to image resolution, limiting principled latency-quality trade-offs, and allocate computation uniformly across input spatial tokens, wasting resource allocation to unimportant regions. We introduce Elastic Latent Interface Transformer (ELIT), a drop-in, DiT-compatible mechanism that decouples input image size from compute. Our approach inserts a latent interface, a learnable variable-length token sequence on which standard transformer blocks can operate. Lightweight Read and Write cross-attention layers move information between spatial tokens and latents and prioritize important input regions. By training with random dropping of tail latents, ELIT learns to produce importance-ordered representations with earlier latents capturing global structure while later ones contain information to refine details. At inference, the number of latents can be dynamically adjusted to match compute constraints. ELIT is deliberately minimal, adding two cross-attention layers while leaving the rectified flow objective and the DiT stack unchanged. Across datasets and architectures (DiT, U-ViT, HDiT, MM-DiT), ELIT delivers consistent gains. On ImageNet-1K 512px, ELIT delivers an average gain of 35.3% and 39.6% in FID and FDD scores. Project page: https://snap-research.github.io/elit/

Most recent semantic segmentation methods adopt a fully-convolutional network (FCN) with an encoder-decoder architecture. The encoder progressively reduces the spatial resolution and learns more abstract/semantic visual concepts with larger receptive fields. Since context modeling is critical for segmentation, the latest efforts have been focused on increasing the receptive field, through either dilated/atrous convolutions or inserting attention modules. However, the encoder-decoder based FCN architecture remains unchanged. In this paper, we aim to provide an alternative perspective by treating semantic segmentation as a sequence-to-sequence prediction task. Specifically, we deploy a pure transformer (ie, without convolution and resolution reduction) to encode an image as a sequence of patches. With the global context modeled in every layer of the transformer, this encoder can be combined with a simple decoder to provide a powerful segmentation model, termed SEgmentation TRansformer (SETR). Extensive experiments show that SETR achieves new state of the art on ADE20K (50.28% mIoU), Pascal Context (55.83% mIoU) and competitive results on Cityscapes. Particularly, we achieve the first position in the highly competitive ADE20K test server leaderboard on the day of submission.

Image dehazing is a representative low-level vision task that estimates latent haze-free images from hazy images. In recent years, convolutional neural network-based methods have dominated image dehazing. However, vision Transformers, which has recently made a breakthrough in high-level vision tasks, has not brought new dimensions to image dehazing. We start with the popular Swin Transformer and find that several of its key designs are unsuitable for image dehazing. To this end, we propose DehazeFormer, which consists of various improvements, such as the modified normalization layer, activation function, and spatial information aggregation scheme. We train multiple variants of DehazeFormer on various datasets to demonstrate its effectiveness. Specifically, on the most frequently used SOTS indoor set, our small model outperforms FFA-Net with only 25% #Param and 5% computational cost. To the best of our knowledge, our large model is the first method with the PSNR over 40 dB on the SOTS indoor set, dramatically outperforming the previous state-of-the-art methods. We also collect a large-scale realistic remote sensing dehazing dataset for evaluating the method's capability to remove highly non-homogeneous haze.

Diffusion Transformer (DiT), an emerging diffusion model for visual generation, has demonstrated superior performance but suffers from substantial computational costs. Our investigations reveal that these costs primarily stem from the static inference paradigm, which inevitably introduces redundant computation in certain diffusion timesteps and spatial regions. To overcome this inefficiency, we propose Dynamic Diffusion Transformer (DyDiT), an architecture that dynamically adjusts its computation along both timestep and spatial dimensions. Specifically, we introduce a Timestep-wise Dynamic Width (TDW) approach that adapts model width conditioned on the generation timesteps. In addition, we design a Spatial-wise Dynamic Token (SDT) strategy to avoid redundant computation at unnecessary spatial locations. TDW and SDT can be seamlessly integrated into DiT and significantly accelerates the generation process. Building on these designs, we further enhance DyDiT in three key aspects. First, DyDiT is integrated seamlessly with flow matching-based generation, enhancing its versatility. Furthermore, we enhance DyDiT to tackle more complex visual generation tasks, including video generation and text-to-image generation, thereby broadening its real-world applications. Finally, to address the high cost of full fine-tuning and democratize technology access, we investigate the feasibility of training DyDiT in a parameter-efficient manner and introduce timestep-based dynamic LoRA (TD-LoRA). Extensive experiments on diverse visual generation models, including DiT, SiT, Latte, and FLUX, demonstrate the effectiveness of DyDiT.

The motion transfer task involves transferring motion from a source video to newly generated videos, requiring the model to decouple motion from appearance. Previous diffusion-based methods primarily rely on separate spatial and temporal attention mechanisms within 3D U-Net. In contrast, state-of-the-art video Diffusion Transformers (DiT) models use 3D full attention, which does not explicitly separate temporal and spatial information. Thus, the interaction between spatial and temporal dimensions makes decoupling motion and appearance more challenging for DiT models. In this paper, we propose DeT, a method that adapts DiT models to improve motion transfer ability. Our approach introduces a simple yet effective temporal kernel to smooth DiT features along the temporal dimension, facilitating the decoupling of foreground motion from background appearance. Meanwhile, the temporal kernel effectively captures temporal variations in DiT features, which are closely related to motion. Moreover, we introduce explicit supervision along dense trajectories in the latent feature space to further enhance motion consistency. Additionally, we present MTBench, a general and challenging benchmark for motion transfer. We also introduce a hybrid motion fidelity metric that considers both the global and local motion similarity. Therefore, our work provides a more comprehensive evaluation than previous works. Extensive experiments on MTBench demonstrate that DeT achieves the best trade-off between motion fidelity and edit fidelity.

Vision transformers (ViTs) are quickly becoming the de-facto architecture for computer vision, yet we understand very little about why they work and what they learn. While existing studies visually analyze the mechanisms of convolutional neural networks, an analogous exploration of ViTs remains challenging. In this paper, we first address the obstacles to performing visualizations on ViTs. Assisted by these solutions, we observe that neurons in ViTs trained with language model supervision (e.g., CLIP) are activated by semantic concepts rather than visual features. We also explore the underlying differences between ViTs and CNNs, and we find that transformers detect image background features, just like their convolutional counterparts, but their predictions depend far less on high-frequency information. On the other hand, both architecture types behave similarly in the way features progress from abstract patterns in early layers to concrete objects in late layers. In addition, we show that ViTs maintain spatial information in all layers except the final layer. In contrast to previous works, we show that the last layer most likely discards the spatial information and behaves as a learned global pooling operation. Finally, we conduct large-scale visualizations on a wide range of ViT variants, including DeiT, CoaT, ConViT, PiT, Swin, and Twin, to validate the effectiveness of our method.

Medical image segmentation is an essential prerequisite for developing healthcare systems, especially for disease diagnosis and treatment planning. On various medical image segmentation tasks, the u-shaped architecture, also known as U-Net, has become the de-facto standard and achieved tremendous success. However, due to the intrinsic locality of convolution operations, U-Net generally demonstrates limitations in explicitly modeling long-range dependency. Transformers, designed for sequence-to-sequence prediction, have emerged as alternative architectures with innate global self-attention mechanisms, but can result in limited localization abilities due to insufficient low-level details. In this paper, we propose TransUNet, which merits both Transformers and U-Net, as a strong alternative for medical image segmentation. On one hand, the Transformer encodes tokenized image patches from a convolution neural network (CNN) feature map as the input sequence for extracting global contexts. On the other hand, the decoder upsamples the encoded features which are then combined with the high-resolution CNN feature maps to enable precise localization. We argue that Transformers can serve as strong encoders for medical image segmentation tasks, with the combination of U-Net to enhance finer details by recovering localized spatial information. TransUNet achieves superior performances to various competing methods on different medical applications including multi-organ segmentation and cardiac segmentation. Code and models are available at https://github.com/Beckschen/TransUNet.

Large vision-language models (VLMs) still struggle with reliable 3D spatial reasoning, a core capability for embodied and physical AI systems. This limitation arises from their inability to capture fine-grained 3D geometry and spatial relationships. While recent efforts have introduced multi-view geometry transformers into VLMs, they typically fuse only the deep-layer features from vision and geometry encoders, discarding rich hierarchical signals and creating a fundamental bottleneck for spatial understanding. To overcome this, we propose SpatialStack, a general hierarchical fusion framework that progressively aligns vision, geometry, and language representations across the model hierarchy. Moving beyond conventional late-stage vision-geometry fusion, SpatialStack stacks and synchronizes multi-level geometric features with the language backbone, enabling the model to capture both local geometric precision and global contextual semantics. Building upon this framework, we develop VLM-SpatialStack, a model that achieves state-of-the-art performance on multiple 3D spatial reasoning benchmarks. Extensive experiments and ablations demonstrate that our multi-level fusion strategy consistently enhances 3D understanding and generalizes robustly across diverse spatial reasoning tasks, establishing SpatialStack as an effective and extensible design paradigm for vision-language-geometry integration in next-generation multimodal physical AI systems.

Text-to-video models have recently undergone rapid and substantial advancements. Nevertheless, due to limitations in data and computational resources, achieving efficient generation of long videos with rich motion dynamics remains a significant challenge. To generate high-quality, dynamic, and temporally consistent long videos, this paper presents ARLON, a novel framework that boosts diffusion Transformers with autoregressive models for long video generation, by integrating the coarse spatial and long-range temporal information provided by the AR model to guide the DiT model. Specifically, ARLON incorporates several key innovations: 1) A latent Vector Quantized Variational Autoencoder (VQ-VAE) compresses the input latent space of the DiT model into compact visual tokens, bridging the AR and DiT models and balancing the learning complexity and information density; 2) An adaptive norm-based semantic injection module integrates the coarse discrete visual units from the AR model into the DiT model, ensuring effective guidance during video generation; 3) To enhance the tolerance capability of noise introduced from the AR inference, the DiT model is trained with coarser visual latent tokens incorporated with an uncertainty sampling module. Experimental results demonstrate that ARLON significantly outperforms the baseline OpenSora-V1.2 on eight out of eleven metrics selected from VBench, with notable improvements in dynamic degree and aesthetic quality, while delivering competitive results on the remaining three and simultaneously accelerating the generation process. In addition, ARLON achieves state-of-the-art performance in long video generation. Detailed analyses of the improvements in inference efficiency are presented, alongside a practical application that demonstrates the generation of long videos using progressive text prompts. See demos of ARLON at http://aka.ms/arlon.

Change detection in remote sensing images is essential for tracking environmental changes on the Earth's surface. Despite the success of vision transformers (ViTs) as backbones in numerous computer vision applications, they remain underutilized in change detection, where convolutional neural networks (CNNs) continue to dominate due to their powerful feature extraction capabilities. In this paper, our study uncovers ViTs' unique advantage in discerning large-scale changes, a capability where CNNs fall short. Capitalizing on this insight, we introduce ChangeViT, a framework that adopts a plain ViT backbone to enhance the performance of large-scale changes. This framework is supplemented by a detail-capture module that generates detailed spatial features and a feature injector that efficiently integrates fine-grained spatial information into high-level semantic learning. The feature integration ensures that ChangeViT excels in both detecting large-scale changes and capturing fine-grained details, providing comprehensive change detection across diverse scales. Without bells and whistles, ChangeViT achieves state-of-the-art performance on three popular high-resolution datasets (i.e., LEVIR-CD, WHU-CD, and CLCD) and one low-resolution dataset (i.e., OSCD), which underscores the unleashed potential of plain ViTs for change detection. Furthermore, thorough quantitative and qualitative analyses validate the efficacy of the introduced modules, solidifying the effectiveness of our approach. The source code is available at https://github.com/zhuduowang/ChangeViT.

Transformer models underpin many recent advances in practical machine learning applications, yet understanding their internal behavior continues to elude researchers. Given the size and complexity of these models, forming a comprehensive picture of their inner workings remains a significant challenge. To this end, we set out to understand small transformer models in a more tractable setting: that of solving mazes. In this work, we focus on the abstractions formed by these models and find evidence for the consistent emergence of structured internal representations of maze topology and valid paths. We demonstrate this by showing that the residual stream of only a single token can be linearly decoded to faithfully reconstruct the entire maze. We also find that the learned embeddings of individual tokens have spatial structure. Furthermore, we take steps towards deciphering the circuity of path-following by identifying attention heads (dubbed adjacency heads), which are implicated in finding valid subsequent tokens.

Vision transformer (ViT) expands the success of transformer models from sequential data to images. The model decomposes an image into many smaller patches and arranges them into a sequence. Multi-head self-attentions are then applied to the sequence to learn the attention between patches. Despite many successful interpretations of transformers on sequential data, little effort has been devoted to the interpretation of ViTs, and many questions remain unanswered. For example, among the numerous attention heads, which one is more important? How strong are individual patches attending to their spatial neighbors in different heads? What attention patterns have individual heads learned? In this work, we answer these questions through a visual analytics approach. Specifically, we first identify what heads are more important in ViTs by introducing multiple pruning-based metrics. Then, we profile the spatial distribution of attention strengths between patches inside individual heads, as well as the trend of attention strengths across attention layers. Third, using an autoencoder-based learning solution, we summarize all possible attention patterns that individual heads could learn. Examining the attention strengths and patterns of the important heads, we answer why they are important. Through concrete case studies with experienced deep learning experts on multiple ViTs, we validate the effectiveness of our solution that deepens the understanding of ViTs from head importance, head attention strength, and head attention pattern.

The quadratic computational complexity to the number of tokens limits the practical applications of Vision Transformers (ViTs). Several works propose to prune redundant tokens to achieve efficient ViTs. However, these methods generally suffer from (i) dramatic accuracy drops, (ii) application difficulty in the local vision transformer, and (iii) non-general-purpose networks for downstream tasks. In this work, we propose a novel Semantic Token ViT (STViT), for efficient global and local vision transformers, which can also be revised to serve as backbone for downstream tasks. The semantic tokens represent cluster centers, and they are initialized by pooling image tokens in space and recovered by attention, which can adaptively represent global or local semantic information. Due to the cluster properties, a few semantic tokens can attain the same effect as vast image tokens, for both global and local vision transformers. For instance, only 16 semantic tokens on DeiT-(Tiny,Small,Base) can achieve the same accuracy with more than 100% inference speed improvement and nearly 60% FLOPs reduction; on Swin-(Tiny,Small,Base), we can employ 16 semantic tokens in each window to further speed it up by around 20% with slight accuracy increase. Besides great success in image classification, we also extend our method to video recognition. In addition, we design a STViT-R(ecover) network to restore the detailed spatial information based on the STViT, making it work for downstream tasks, which is powerless for previous token sparsification methods. Experiments demonstrate that our method can achieve competitive results compared to the original networks in object detection and instance segmentation, with over 30% FLOPs reduction for backbone. Code is available at http://github.com/changsn/STViT-R

To generate accurate videos, algorithms have to understand the spatial and temporal dependencies in the world. Current algorithms enable accurate predictions over short horizons but tend to suffer from temporal inconsistencies. When generated content goes out of view and is later revisited, the model invents different content instead. Despite this severe limitation, no established benchmarks on complex data exist for rigorously evaluating video generation with long temporal dependencies. In this paper, we curate 3 challenging video datasets with long-range dependencies by rendering walks through 3D scenes of procedural mazes, Minecraft worlds, and indoor scans. We perform a comprehensive evaluation of current models and observe their limitations in temporal consistency. Moreover, we introduce the Temporally Consistent Transformer (TECO), a generative model that substantially improves long-term consistency while also reducing sampling time. By compressing its input sequence into fewer embeddings, applying a temporal transformer, and expanding back using a spatial MaskGit, TECO outperforms existing models across many metrics. Videos are available on the website: https://wilson1yan.github.io/teco

Understanding temporal dynamics of video is an essential aspect of learning better video representations. Recently, transformer-based architectural designs have been extensively explored for video tasks due to their capability to capture long-term dependency of input sequences. However, we found that these Video Transformers are still biased to learn spatial dynamics rather than temporal ones, and debiasing the spurious correlation is critical for their performance. Based on the observations, we design simple yet effective self-supervised tasks for video models to learn temporal dynamics better. Specifically, for debiasing the spatial bias, our method learns the temporal order of video frames as extra self-supervision and enforces the randomly shuffled frames to have low-confidence outputs. Also, our method learns the temporal flow direction of video tokens among consecutive frames for enhancing the correlation toward temporal dynamics. Under various video action recognition tasks, we demonstrate the effectiveness of our method and its compatibility with state-of-the-art Video Transformers.

3D visual perception tasks, including 3D detection and map segmentation based on multi-camera images, are essential for autonomous driving systems. In this work, we present a new framework termed BEVFormer, which learns unified BEV representations with spatiotemporal transformers to support multiple autonomous driving perception tasks. In a nutshell, BEVFormer exploits both spatial and temporal information by interacting with spatial and temporal space through predefined grid-shaped BEV queries. To aggregate spatial information, we design spatial cross-attention that each BEV query extracts the spatial features from the regions of interest across camera views. For temporal information, we propose temporal self-attention to recurrently fuse the history BEV information. Our approach achieves the new state-of-the-art 56.9\% in terms of NDS metric on the nuScenes test set, which is 9.0 points higher than previous best arts and on par with the performance of LiDAR-based baselines. We further show that BEVFormer remarkably improves the accuracy of velocity estimation and recall of objects under low visibility conditions. The code is available at https://github.com/zhiqi-li/BEVFormer.

Vision Transformer (ViT) demonstrates that Transformer for natural language processing can be applied to computer vision tasks and result in comparable performance to convolutional neural networks (CNN), which have been studied and adopted in computer vision for years. This naturally raises the question of how the performance of ViT can be advanced with design techniques of CNN. To this end, we propose to incorporate two techniques and present ViT-ResNAS, an efficient multi-stage ViT architecture designed with neural architecture search (NAS). First, we propose residual spatial reduction to decrease sequence lengths for deeper layers and utilize a multi-stage architecture. When reducing lengths, we add skip connections to improve performance and stabilize training deeper networks. Second, we propose weight-sharing NAS with multi-architectural sampling. We enlarge a network and utilize its sub-networks to define a search space. A super-network covering all sub-networks is then trained for fast evaluation of their performance. To efficiently train the super-network, we propose to sample and train multiple sub-networks with one forward-backward pass. After that, evolutionary search is performed to discover high-performance network architectures. Experiments on ImageNet demonstrate that ViT-ResNAS achieves better accuracy-MACs and accuracy-throughput trade-offs than the original DeiT and other strong baselines of ViT. Code is available at https://github.com/yilunliao/vit-search.

Metalens is an emerging optical system with an irreplaceable merit in that it can be manufactured in ultra-thin and compact sizes, which shows great promise of various applications such as medical imaging and augmented/virtual reality (AR/VR). Despite its advantage in miniaturization, its practicality is constrained by severe aberrations and distortions, which significantly degrade the image quality. Several previous arts have attempted to address different types of aberrations, yet most of them are mainly designed for the traditional bulky lens and not convincing enough to remedy harsh aberrations of the metalens. While there have existed aberration correction methods specifically for metalens, they still fall short of restoration quality. In this work, we propose MetaFormer, an aberration correction framework for metalens-captured images, harnessing Vision Transformers (ViT) that has shown remarkable restoration performance in diverse image restoration tasks. Specifically, we devise a Multiple Adaptive Filters Guidance (MAFG), where multiple Wiener filters enrich the degraded input images with various noise-detail balances, enhancing output restoration quality. In addition, we introduce a Spatial and Transposed self-Attention Fusion (STAF) module, which aggregates features from spatial self-attention and transposed self-attention modules to further ameliorate aberration correction. We conduct extensive experiments, including correcting aberrated images and videos, and clean 3D reconstruction from the degraded images. The proposed method outperforms the previous arts by a significant margin. We further fabricate a metalens and verify the practicality of MetaFormer by restoring the images captured with the manufactured metalens in the wild. Code and pre-trained models are available at https://benhenryl.github.io/MetaFormer

We present a novel approach for the detection of deepfake videos using a pair of vision transformers pre-trained by a self-supervised masked autoencoding setup. Our method consists of two distinct components, one of which focuses on learning spatial information from individual RGB frames of the video, while the other learns temporal consistency information from optical flow fields generated from consecutive frames. Unlike most approaches where pre-training is performed on a generic large corpus of images, we show that by pre-training on smaller face-related datasets, namely Celeb-A (for the spatial learning component) and YouTube Faces (for the temporal learning component), strong results can be obtained. We perform various experiments to evaluate the performance of our method on commonly used datasets namely FaceForensics++ (Low Quality and High Quality, along with a new highly compressed version named Very Low Quality) and Celeb-DFv2 datasets. Our experiments show that our method sets a new state-of-the-art on FaceForensics++ (LQ, HQ, and VLQ), and obtains competitive results on Celeb-DFv2. Moreover, our method outperforms other methods in the area in a cross-dataset setup where we fine-tune our model on FaceForensics++ and test on CelebDFv2, pointing to its strong cross-dataset generalization ability.

Region-instructed layout control in text-to-image generation is highly practical, yet existing methods suffer from limitations: (i) training-based approaches inherit data bias and often degrade image quality, and (ii) current techniques struggle with occlusion order, limiting real-world usability. To address these issues, we propose LayerBind. By modeling regional generation as distinct layers and binding them during the generation, our method enables precise regional and occlusion controllability. Our motivation stems from the observation that spatial layout and occlusion are established at a very early denoising stage, suggesting that rearranging the early latent structure is sufficient to modify the final output. Building on this, we structure the scheme into two phases: instance initialization and subsequent semantic nursing. (1) First, leveraging the contextual sharing mechanism in multimodal joint attention, Layer-wise Instance Initialization creates per-instance branches that attend to their own regions while anchoring to the shared background. At a designated early step, these branches are fused according to the layer order to form a unified latent with a pre-established layout. (2) Then, Layer-wise Semantic Nursing reinforces regional details and maintains the occlusion order via a layer-wise attention enhancement. Specifically, a sequential layered attention path operates alongside the standard global path, with updates composited under a layer-transparency scheduler. LayerBind is training-free and plug-and-play, serving as a regional and occlusion controller across Diffusion Transformers. Beyond generation, it natively supports editable workflows, allowing for flexible modifications like changing instances or rearranging visible orders. Both qualitative and quantitative results demonstrate LayerBind's effectiveness, highlighting its strong potential for creative applications.

This work presents a systematic investigation into modernizing Vision Transformer backbones by leveraging architectural advancements from the past five years. While preserving the canonical Attention-FFN structure, we conduct a component-wise refinement involving normalization, activation functions, positional encoding, gating mechanisms, and learnable tokens. These updates form a new generation of Vision Transformers, which we call ViT-5. Extensive experiments demonstrate that ViT-5 consistently outperforms state-of-the-art plain Vision Transformers across both understanding and generation benchmarks. On ImageNet-1k classification, ViT-5-Base reaches 84.2\% top-1 accuracy under comparable compute, exceeding DeiT-III-Base at 83.8\%. ViT-5 also serves as a stronger backbone for generative modeling: when plugged into an SiT diffusion framework, it achieves 1.84 FID versus 2.06 with a vanilla ViT backbone. Beyond headline metrics, ViT-5 exhibits improved representation learning and favorable spatial reasoning behavior, and transfers reliably across tasks. With a design aligned with contemporary foundation-model practices, ViT-5 offers a simple drop-in upgrade over vanilla ViT for mid-2020s vision backbones.

Large-scale text-to-image (T2I) diffusion models excel at open-domain synthesis but still struggle with precise text rendering, especially for multi-line layouts, dense typography, and long-tailed scripts such as Chinese. Prior solutions typically require costly retraining or rigid external layout constraints, which can degrade aesthetics and limit flexibility. We propose FreeText, a training-free, plug-and-play framework that improves text rendering by exploiting intrinsic mechanisms of Diffusion Transformer (DiT) models. FreeText decomposes the problem into where to write and what to write. For where to write, we localize writing regions by reading token-wise spatial attribution from endogenous image-to-text attention, using sink-like tokens as stable spatial anchors and topology-aware refinement to produce high-confidence masks. For what to write, we introduce Spectral-Modulated Glyph Injection (SGMI), which injects a noise-aligned glyph prior with frequency-domain band-pass modulation to strengthen glyph structure and suppress semantic leakage (rendering the concept instead of the word). Extensive experiments on Qwen-Image, FLUX.1-dev, and SD3 variants across longText-Benchmark, CVTG, and our CLT-Bench show consistent gains in text readability while largely preserving semantic alignment and aesthetic quality, with modest inference overhead.

The Transformer architecture has achieved tremendous success in natural language processing, computer vision, and scientific computing through its self-attention mechanism. However, its core components-positional encoding and attention mechanisms-have lacked a unified physical or mathematical interpretation. This paper proposes a structural theoretical framework that integrates positional encoding, kernel integral operators, and attention mechanisms for in-depth theoretical investigation. We map discrete positions (such as text token indices and image pixel coordinates) to spatial functions on continuous manifolds, enabling a field-theoretic interpretation of Transformer layers as kernel-modulated operators acting over embedded manifolds.

Diffusion transformers have gained significant attention in recent years for their ability to generate high-quality images and videos, yet still suffer from a huge computational cost due to their iterative denoising process. Recently, feature caching has been introduced to accelerate diffusion transformers by caching the feature computation in previous timesteps and reusing it in the following timesteps, which leverage the temporal similarity of diffusion models while ignoring the similarity in the spatial dimension. In this paper, we introduce Cluster-Driven Feature Caching (ClusCa) as an orthogonal and complementary perspective for previous feature caching. Specifically, ClusCa performs spatial clustering on tokens in each timestep, computes only one token in each cluster and propagates their information to all the other tokens, which is able to reduce the number of tokens by over 90%. Extensive experiments on DiT, FLUX and HunyuanVideo demonstrate its effectiveness in both text-to-image and text-to-video generation. Besides, it can be directly applied to any diffusion transformer without requirements for training. For instance, ClusCa achieves 4.96x acceleration on FLUX with an ImageReward of 99.49%, surpassing the original model by 0.51%. The code is available at https://github.com/Shenyi-Z/Cache4Diffusion.

Transformer-based methods have swept the benchmarks on 2D and 3D detection on images. Because tokenization before the attention mechanism drops the spatial information, positional encoding becomes critical for those methods. Recent works found that encodings based on samples of the 3D viewing rays can significantly improve the quality of multi-camera 3D object detection. We hypothesize that 3D point locations can provide more information than rays. Therefore, we introduce 3D point positional encoding, 3DPPE, to the 3D detection Transformer decoder. Although 3D measurements are not available at the inference time of monocular 3D object detection, 3DPPE uses predicted depth to approximate the real point positions. Our hybriddepth module combines direct and categorical depth to estimate the refined depth of each pixel. Despite the approximation, 3DPPE achieves 46.0 mAP and 51.4 NDS on the competitive nuScenes dataset, significantly outperforming encodings based on ray samples. We make the codes available at https://github.com/drilistbox/3DPPE.

We consider the problem of localizing a spatio-temporal tube in a video corresponding to a given text query. This is a challenging task that requires the joint and efficient modeling of temporal, spatial and multi-modal interactions. To address this task, we propose TubeDETR, a transformer-based architecture inspired by the recent success of such models for text-conditioned object detection. Our model notably includes: (i) an efficient video and text encoder that models spatial multi-modal interactions over sparsely sampled frames and (ii) a space-time decoder that jointly performs spatio-temporal localization. We demonstrate the advantage of our proposed components through an extensive ablation study. We also evaluate our full approach on the spatio-temporal video grounding task and demonstrate improvements over the state of the art on the challenging VidSTG and HC-STVG benchmarks. Code and trained models are publicly available at https://antoyang.github.io/tubedetr.html.

The canonical approach to video captioning dictates a caption generation model to learn from offline-extracted dense video features. These feature extractors usually operate on video frames sampled at a fixed frame rate and are often trained on image/video understanding tasks, without adaption to video captioning data. In this work, we present SwinBERT, an end-to-end transformer-based model for video captioning, which takes video frame patches directly as inputs, and outputs a natural language description. Instead of leveraging multiple 2D/3D feature extractors, our method adopts a video transformer to encode spatial-temporal representations that can adapt to variable lengths of video input without dedicated design for different frame rates. Based on this model architecture, we show that video captioning can benefit significantly from more densely sampled video frames as opposed to previous successes with sparsely sampled video frames for video-and-language understanding tasks (e.g., video question answering). Moreover, to avoid the inherent redundancy in consecutive video frames, we propose adaptively learning a sparse attention mask and optimizing it for task-specific performance improvement through better long-range video sequence modeling. Through extensive experiments on 5 video captioning datasets, we show that SwinBERT achieves across-the-board performance improvements over previous methods, often by a large margin. The learned sparse attention masks in addition push the limit to new state of the arts, and can be transferred between different video lengths and between different datasets. Code is available at https://github.com/microsoft/SwinBERT

Humans can perform complex tasks with long-term objectives by planning, reasoning, and forecasting outcomes of actions. For embodied agents to achieve similar capabilities, they must gain knowledge of the environment transferable to novel scenarios with a limited budget of additional trial and error. Learning-based approaches, such as deep RL, can discover and take advantage of inherent regularities and characteristics of the application domain from data, and continuously improve their performances, however at a cost of large amounts of training data. This thesis explores the development of data-driven techniques for spatial reasoning and planning tasks, focusing on enhancing learning efficiency, interpretability, and transferability across novel scenarios. Four key contributions are made. 1) CALVIN, a differential planner that learns interpretable models of the world for long-term planning. It successfully navigated partially observable 3D environments, such as mazes and indoor rooms, by learning the rewards and state transitions from expert demonstrations. 2) SOAP, an RL algorithm that discovers options unsupervised for long-horizon tasks. Options segment a task into subtasks and enable consistent execution of the subtask. SOAP showed robust performances on history-conditional corridor tasks as well as classical benchmarks such as Atari. 3) LangProp, a code optimisation framework using LLMs to solve embodied agent problems that require reasoning by treating code as learnable policies. The framework successfully generated interpretable code with comparable or superior performance to human-written experts in the CARLA autonomous driving benchmark. 4) Voggite, an embodied agent with a vision-to-action transformer backend that solves complex tasks in Minecraft. It achieved third place in the MineRL BASALT Competition by identifying action triggers to segment tasks into multiple stages.

Diffusion models (DMs) have become the leading choice for generative tasks across diverse domains. However, their reliance on multiple sequential forward passes significantly limits real-time performance. Previous acceleration methods have primarily focused on reducing the number of sampling steps or reusing intermediate results, failing to leverage variations across spatial regions within the image due to the constraints of convolutional U-Net structures. By harnessing the flexibility of Diffusion Transformers (DiTs) in handling variable number of tokens, we introduce RAS, a novel, training-free sampling strategy that dynamically assigns different sampling ratios to regions within an image based on the focus of the DiT model. Our key observation is that during each sampling step, the model concentrates on semantically meaningful regions, and these areas of focus exhibit strong continuity across consecutive steps. Leveraging this insight, RAS updates only the regions currently in focus, while other regions are updated using cached noise from the previous step. The model's focus is determined based on the output from the preceding step, capitalizing on the temporal consistency we observed. We evaluate RAS on Stable Diffusion 3 and Lumina-Next-T2I, achieving speedups up to 2.36x and 2.51x, respectively, with minimal degradation in generation quality. Additionally, a user study reveals that RAS delivers comparable qualities under human evaluation while achieving a 1.6x speedup. Our approach makes a significant step towards more efficient diffusion transformers, enhancing their potential for real-time applications.

Contemporary models for generating images show remarkable quality and versatility. Swayed by these advantages, the research community repurposes them to generate videos. Since video content is highly redundant, we argue that naively bringing advances of image models to the video generation domain reduces motion fidelity, visual quality and impairs scalability. In this work, we build Snap Video, a video-first model that systematically addresses these challenges. To do that, we first extend the EDM framework to take into account spatially and temporally redundant pixels and naturally support video generation. Second, we show that a U-Net - a workhorse behind image generation - scales poorly when generating videos, requiring significant computational overhead. Hence, we propose a new transformer-based architecture that trains 3.31 times faster than U-Nets (and is ~4.5 faster at inference). This allows us to efficiently train a text-to-video model with billions of parameters for the first time, reach state-of-the-art results on a number of benchmarks, and generate videos with substantially higher quality, temporal consistency, and motion complexity. The user studies showed that our model was favored by a large margin over the most recent methods. See our website at https://snap-research.github.io/snapvideo/.

Mixture-of-Experts (MoE) has emerged as a powerful paradigm for scaling model capacity while preserving computational efficiency. Despite its notable success in large language models (LLMs), existing attempts to apply MoE to Diffusion Transformers (DiTs) have yielded limited gains. We attribute this gap to fundamental differences between language and visual tokens. Language tokens are semantically dense with pronounced inter-token variation, while visual tokens exhibit spatial redundancy and functional heterogeneity, hindering expert specialization in vision MoE. To this end, we present ProMoE, an MoE framework featuring a two-step router with explicit routing guidance that promotes expert specialization. Specifically, this guidance encourages the router to partition image tokens into conditional and unconditional sets via conditional routing according to their functional roles, and refine the assignments of conditional image tokens through prototypical routing with learnable prototypes based on semantic content. Moreover, the similarity-based expert allocation in latent space enabled by prototypical routing offers a natural mechanism for incorporating explicit semantic guidance, and we validate that such guidance is crucial for vision MoE. Building on this, we propose a routing contrastive loss that explicitly enhances the prototypical routing process, promoting intra-expert coherence and inter-expert diversity. Extensive experiments on ImageNet benchmark demonstrate that ProMoE surpasses state-of-the-art methods under both Rectified Flow and DDPM training objectives. Code and models will be made publicly available.

We present EmerNeRF, a simple yet powerful approach for learning spatial-temporal representations of dynamic driving scenes. Grounded in neural fields, EmerNeRF simultaneously captures scene geometry, appearance, motion, and semantics via self-bootstrapping. EmerNeRF hinges upon two core components: First, it stratifies scenes into static and dynamic fields. This decomposition emerges purely from self-supervision, enabling our model to learn from general, in-the-wild data sources. Second, EmerNeRF parameterizes an induced flow field from the dynamic field and uses this flow field to further aggregate multi-frame features, amplifying the rendering precision of dynamic objects. Coupling these three fields (static, dynamic, and flow) enables EmerNeRF to represent highly-dynamic scenes self-sufficiently, without relying on ground truth object annotations or pre-trained models for dynamic object segmentation or optical flow estimation. Our method achieves state-of-the-art performance in sensor simulation, significantly outperforming previous methods when reconstructing static (+2.93 PSNR) and dynamic (+3.70 PSNR) scenes. In addition, to bolster EmerNeRF's semantic generalization, we lift 2D visual foundation model features into 4D space-time and address a general positional bias in modern Transformers, significantly boosting 3D perception performance (e.g., 37.50% relative improvement in occupancy prediction accuracy on average). Finally, we construct a diverse and challenging 120-sequence dataset to benchmark neural fields under extreme and highly-dynamic settings.

Modern text-to-video synthesis models demonstrate coherent, photorealistic generation of complex videos from a text description. However, most existing models lack fine-grained control over camera movement, which is critical for downstream applications related to content creation, visual effects, and 3D vision. Recently, new methods demonstrate the ability to generate videos with controllable camera poses these techniques leverage pre-trained U-Net-based diffusion models that explicitly disentangle spatial and temporal generation. Still, no existing approach enables camera control for new, transformer-based video diffusion models that process spatial and temporal information jointly. Here, we propose to tame video transformers for 3D camera control using a ControlNet-like conditioning mechanism that incorporates spatiotemporal camera embeddings based on Plucker coordinates. The approach demonstrates state-of-the-art performance for controllable video generation after fine-tuning on the RealEstate10K dataset. To the best of our knowledge, our work is the first to enable camera control for transformer-based video diffusion models.

3D reconstruction from multi-view images is a core challenge in computer vision. Recently, feed-forward methods have emerged as efficient and robust alternatives to traditional per-scene optimization techniques. Among them, state-of-the-art models like the Visual Geometry Grounding Transformer (VGGT) leverage full self-attention over all image tokens to capture global relationships. However, this approach suffers from poor scalability due to the quadratic complexity of self-attention and the large number of tokens generated in long image sequences. In this work, we introduce FlashVGGT, an efficient alternative that addresses this bottleneck through a descriptor-based attention mechanism. Instead of applying dense global attention across all tokens, FlashVGGT compresses spatial information from each frame into a compact set of descriptor tokens. Global attention is then computed as cross-attention between the full set of image tokens and this smaller descriptor set, significantly reducing computational overhead. Moreover, the compactness of the descriptors enables online inference over long sequences via a chunk-recursive mechanism that reuses cached descriptors from previous chunks. Experimental results show that FlashVGGT achieves reconstruction accuracy competitive with VGGT while reducing inference time to just 9.3% of VGGT for 1,000 images, and scaling efficiently to sequences exceeding 3,000 images. Our project page is available at https://wzpscott.github.io/flashvggt_page/.

Sora unveils the potential of scaling Diffusion Transformer for generating photorealistic images and videos at arbitrary resolutions, aspect ratios, and durations, yet it still lacks sufficient implementation details. In this technical report, we introduce the Lumina-T2X family - a series of Flow-based Large Diffusion Transformers (Flag-DiT) equipped with zero-initialized attention, as a unified framework designed to transform noise into images, videos, multi-view 3D objects, and audio clips conditioned on text instructions. By tokenizing the latent spatial-temporal space and incorporating learnable placeholders such as [nextline] and [nextframe] tokens, Lumina-T2X seamlessly unifies the representations of different modalities across various spatial-temporal resolutions. This unified approach enables training within a single framework for different modalities and allows for flexible generation of multimodal data at any resolution, aspect ratio, and length during inference. Advanced techniques like RoPE, RMSNorm, and flow matching enhance the stability, flexibility, and scalability of Flag-DiT, enabling models of Lumina-T2X to scale up to 7 billion parameters and extend the context window to 128K tokens. This is particularly beneficial for creating ultra-high-definition images with our Lumina-T2I model and long 720p videos with our Lumina-T2V model. Remarkably, Lumina-T2I, powered by a 5-billion-parameter Flag-DiT, requires only 35% of the training computational costs of a 600-million-parameter naive DiT. Our further comprehensive analysis underscores Lumina-T2X's preliminary capability in resolution extrapolation, high-resolution editing, generating consistent 3D views, and synthesizing videos with seamless transitions. We expect that the open-sourcing of Lumina-T2X will further foster creativity, transparency, and diversity in the generative AI community.

Single-stream architectures using Vision Transformer (ViT) backbones show great potential for real-time UAV tracking recently. However, frequent occlusions from obstacles like buildings and trees expose a major drawback: these models often lack strategies to handle occlusions effectively. New methods are needed to enhance the occlusion resilience of single-stream ViT models in aerial tracking. In this work, we propose to learn Occlusion-Robust Representations (ORR) based on ViTs for UAV tracking by enforcing an invariance of the feature representation of a target with respect to random masking operations modeled by a spatial Cox process. Hopefully, this random masking approximately simulates target occlusions, thereby enabling us to learn ViTs that are robust to target occlusion for UAV tracking. This framework is termed ORTrack. Additionally, to facilitate real-time applications, we propose an Adaptive Feature-Based Knowledge Distillation (AFKD) method to create a more compact tracker, which adaptively mimics the behavior of the teacher model ORTrack according to the task's difficulty. This student model, dubbed ORTrack-D, retains much of ORTrack's performance while offering higher efficiency. Extensive experiments on multiple benchmarks validate the effectiveness of our method, demonstrating its state-of-the-art performance. Codes is available at https://github.com/wuyou3474/ORTrack.

Arabic handwritten text recognition (HTR) is challenging, especially for historical texts, due to diverse writing styles and the intrinsic features of Arabic script. Additionally, Arabic handwriting datasets are smaller compared to English ones, making it difficult to train generalizable Arabic HTR models. To address these challenges, we propose HATFormer, a transformer-based encoder-decoder architecture that builds on a state-of-the-art English HTR model. By leveraging the transformer's attention mechanism, HATFormer captures spatial contextual information to address the intrinsic challenges of Arabic script through differentiating cursive characters, decomposing visual representations, and identifying diacritics. Our customization to historical handwritten Arabic includes an image processor for effective ViT information preprocessing, a text tokenizer for compact Arabic text representation, and a training pipeline that accounts for a limited amount of historic Arabic handwriting data. HATFormer achieves a character error rate (CER) of 8.6% on the largest public historical handwritten Arabic dataset, with a 51% improvement over the best baseline in the literature. HATFormer also attains a comparable CER of 4.2% on the largest private non-historical dataset. Our work demonstrates the feasibility of adapting an English HTR method to a low-resource language with complex, language-specific challenges, contributing to advancements in document digitization, information retrieval, and cultural preservation.

Image captioning (IC) refers to the automatic generation of natural language descriptions for images, with applications ranging from social media content generation to assisting individuals with visual impairments. While most research has been focused on English-based models, low-resource languages such as Brazilian Portuguese face significant challenges due to the lack of specialized datasets and models. Several studies create datasets by automatically translating existing ones to mitigate resource scarcity. This work addresses this gap by proposing a cross-native-translated evaluation of Transformer-based vision and language models for Brazilian Portuguese IC. We use a version of Flickr30K comprised of captions manually created by native Brazilian Portuguese speakers and compare it to a version with captions automatically translated from English to Portuguese. The experiments include a cross-context approach, where models trained on one dataset are tested on the other to assess the translation impact. Additionally, we incorporate attention maps for model inference interpretation and use the CLIP-Score metric to evaluate the image-description alignment. Our findings show that Swin-DistilBERTimbau consistently outperforms other models, demonstrating strong generalization across datasets. ViTucano, a Brazilian Portuguese pre-trained VLM, surpasses larger multilingual models (GPT-4o, LLaMa 3.2 Vision) in traditional text-based evaluation metrics, while GPT-4 models achieve the highest CLIP-Score, highlighting improved image-text alignment. Attention analysis reveals systematic biases, including gender misclassification, object enumeration errors, and spatial inconsistencies. The datasets and the models generated and analyzed during the current study are available in: https://github.com/laicsiifes/transformer-caption-ptbr.

In this paper, we present LaVieID, a novel local autoregressive video diffusion framework designed to tackle the challenging identity-preserving text-to-video task. The key idea of LaVieID is to mitigate the loss of identity information inherent in the stochastic global generation process of diffusion transformers (DiTs) from both spatial and temporal perspectives. Specifically, unlike the global and unstructured modeling of facial latent states in existing DiTs, LaVieID introduces a local router to explicitly represent latent states by weighted combinations of fine-grained local facial structures. This alleviates undesirable feature interference and encourages DiTs to capture distinctive facial characteristics. Furthermore, a temporal autoregressive module is integrated into LaVieID to refine denoised latent tokens before video decoding. This module divides latent tokens temporally into chunks, exploiting their long-range temporal dependencies to predict biases for rectifying tokens, thereby significantly enhancing inter-frame identity consistency. Consequently, LaVieID can generate high-fidelity personalized videos and achieve state-of-the-art performance. Our code and models are available at https://github.com/ssugarwh/LaVieID.

Manufacturing wafers is an intricate task involving thousands of steps. Defect Pattern Recognition (DPR) of wafer maps is crucial to find the root cause of the issue and further improving the yield in the wafer foundry. Mixed-type DPR is much more complicated compared to single-type DPR due to varied spatial features, the uncertainty of defects, and the number of defects present. To accurately predict the number of defects as well as the types of defects, we propose a novel compact deformable convolutional transformer (DC Transformer). Specifically, DC Transformer focuses on the global features present in the wafer map by virtue of learnable deformable kernels and multi-head attention to the global features. The proposed method succinctly models the internal relationship between the wafer maps and the defects. DC Transformer is evaluated on a real dataset containing 38 defect patterns. Experimental results show that DC Transformer performs exceptionally well in recognizing both single and mixed-type defects. The proposed method outperforms the current state of the models by a considerable margin

The task of multimodal referring expression comprehension (REC), aiming at localizing an image region described by a natural language expression, has recently received increasing attention within the research comminity. In this paper, we specifically focus on referring expression comprehension with commonsense knowledge (KB-Ref), a task which typically requires reasoning beyond spatial, visual or semantic information. We propose a novel framework for Commonsense Knowledge Enhanced Transformers (CK-Transformer) which effectively integrates commonsense knowledge into the representations of objects in an image, facilitating identification of the target objects referred to by the expressions. We conduct extensive experiments on several benchmarks for the task of KB-Ref. Our results show that the proposed CK-Transformer achieves a new state of the art, with an absolute improvement of 3.14% accuracy over the existing state of the art.

Camera-controllable video generation aims to synthesize videos with flexible and physically plausible camera movements. However, existing methods either provide imprecise camera control from text prompts or rely on labor-intensive manual camera trajectory parameters, limiting their use in automated scenarios. To address these issues, we propose a novel Vision-Language-Camera model, termed CT-1 (Camera Transformer 1), a specialized model designed to transfer spatial reasoning knowledge to video generation by accurately estimating camera trajectories. Built upon vision-language modules and a Diffusion Transformer model, CT-1 employs a Wavelet-based Regularization Loss in the frequency domain to effectively learn complex camera trajectory distributions. These trajectories are integrated into a video diffusion model to enable spatially aware camera control that aligns with user intentions. To facilitate the training of CT-1, we design a dedicated data curation pipeline and construct CT-200K, a large-scale dataset containing over 47M frames. Experimental results demonstrate that our framework successfully bridges the gap between spatial reasoning and video synthesis, yielding faithful and high-quality camera-controllable videos and improving camera control accuracy by 25.7% over prior methods.

Despite strong single-turn performance, diffusion-based image compositing often struggles to preserve coherent spatial relations in pairwise or sequential edits, where subsequent insertions may overwrite previously generated content and disrupt physical consistency. We introduce PICS, a self-supervised composition-by-decomposition paradigm that composes objects in parallel while explicitly modeling the compositional interactions among (fully-/partially-)visible objects and background. At its core, an Interaction Transformer employs mask-guided Mixture-of-Experts to route background, exclusive, and overlap regions to dedicated experts, with an adaptive α-blending strategy that infers a compatibility-aware fusion of overlapping objects while preserving boundary fidelity. To further enhance robustness to geometric variations, we incorporate geometry-aware augmentations covering both out-of-plane and in-plane pose changes of objects. Our method delivers superior pairwise compositing quality and substantially improved stability, with extensive evaluations across virtual try-on, indoor, and street scene settings showing consistent gains over state-of-the-art baselines. Code and data are available at https://github.com/RyanHangZhou/PICS

With the development of diffusion models, enhancing spatial controllability in text-to-image generation has become a vital challenge. As a representative task for addressing this challenge, layout-to-image generation aims to generate images that are spatially consistent with the given layout condition. Existing layout-to-image methods typically introduce the layout condition by integrating adapter modules into the base generative model. However, the generated images often exhibit low visual quality and stylistic inconsistency with the base model, indicating a loss of pretrained knowledge. To alleviate this issue, we construct the Layout Synthesis (LaySyn) dataset, which leverages images synthesized by the base model itself to mitigate the distribution shift from the pretraining data. Moreover, we propose the Layout Control (Laytrol) Network, in which parameters are inherited from MM-DiT to preserve the pretrained knowledge of the base model. To effectively activate the copied parameters and avoid disturbance from unstable control conditions, we adopt a dedicated initialization scheme for Laytrol. In this scheme, the layout encoder is initialized as a pure text encoder to ensure that its output tokens remain within the data domain of MM-DiT. Meanwhile, the outputs of the layout control network are initialized to zero. In addition, we apply Object-level Rotary Position Embedding to the layout tokens to provide coarse positional information. Qualitative and quantitative experiments demonstrate the effectiveness of our method.

Positional embeddings (PE) play a crucial role in Vision Transformers (ViTs) by providing spatial information otherwise lost due to the permutation invariant nature of self attention. While absolute positional embeddings (APE) have shown theoretical advantages over relative positional embeddings (RPE), particularly due to the ability of sinusoidal functions to preserve spatial inductive biases like monotonicity and shift invariance, a fundamental challenge arises when mapping a 2D grid to a 1D sequence. Existing methods have mostly overlooked or never explored the impact of patch ordering in positional embeddings. To address this, we propose LOOPE, a learnable patch-ordering method that optimizes spatial representation for a given set of frequencies, providing a principled approach to patch order optimization. Empirical results show that our PE significantly improves classification accuracy across various ViT architectures. To rigorously evaluate the effectiveness of positional embeddings, we introduce the "Three Cell Experiment", a novel benchmarking framework that assesses the ability of PEs to retain relative and absolute positional information across different ViT architectures. Unlike standard evaluations, which typically report a performance gap of 4 to 6% between models with and without PE, our method reveals a striking 30 to 35% difference, offering a more sensitive diagnostic tool to measure the efficacy of PEs. Our experimental analysis confirms that the proposed LOOPE demonstrates enhanced effectiveness in retaining both relative and absolute positional information.

This work analyzes the training dynamics of Image Restoration (IR) Transformers and uncovers a critical yet overlooked issue: conventional LayerNorm (LN) drives feature magnitudes to diverge to a million scale and collapses channel-wise entropy. We analyze this in the perspective of networks attempting to bypass LN's constraints that conflict with IR tasks. Accordingly, we address two misalignments between LN and IR: 1) per-token normalization disrupts spatial correlations, and 2) input-independent scaling discards input-specific statistics. To address this, we propose Image Restoration Transformer Tailored Layer Normalization i-LN, a simple drop-in replacement that normalizes features holistically and adaptively rescales them per input. We provide theoretical insights and empirical evidence that this simple design effectively leads to both improved training dynamics and thereby improved performance, validated by extensive experiments.

Bone density prediction via CT scans to estimate T-scores is crucial, providing a more precise assessment of bone health compared to traditional methods like X-ray bone density tests, which lack spatial resolution and the ability to detect localized changes. However, CT-based prediction faces two major challenges: the high computational complexity of transformer-based architectures, which limits their deployment in portable and clinical settings, and the imbalanced, long-tailed distribution of real-world hospital data that skews predictions. To address these issues, we introduce MedConv, a convolutional model for bone density prediction that outperforms transformer models with lower computational demands. We also adapt Bal-CE loss and post-hoc logit adjustment to improve class balance. Extensive experiments on our AustinSpine dataset shows that our approach achieves up to 21% improvement in accuracy and 20% in ROC AUC over previous state-of-the-art methods.

Gaze following aims to interpret human-scene interactions by predicting the person's focal point of gaze. Prevailing approaches often adopt a two-stage framework, whereby multi-modality information is extracted in the initial stage for gaze target prediction. Consequently, the efficacy of these methods highly depends on the precision of the preceding modality extraction. Others use a single-modality approach with complex decoders, increasing network computational load. Inspired by the remarkable success of pre-trained plain vision transformers (ViTs), we introduce a novel single-modality gaze following framework called ViTGaze. In contrast to previous methods, it creates a novel gaze following framework based mainly on powerful encoders (relative decoder parameters less than 1%). Our principal insight is that the inter-token interactions within self-attention can be transferred to interactions between humans and scenes. Leveraging this presumption, we formulate a framework consisting of a 4D interaction encoder and a 2D spatial guidance module to extract human-scene interaction information from self-attention maps. Furthermore, our investigation reveals that ViT with self-supervised pre-training has an enhanced ability to extract correlation information. Many experiments have been conducted to demonstrate the performance of the proposed method. Our method achieves state-of-the-art (SOTA) performance among all single-modality methods (3.4% improvement in the area under curve (AUC) score, 5.1% improvement in the average precision (AP)) and very comparable performance against multi-modality methods with 59% number of parameters less.

We present COMEDIAN, a novel pipeline to initialize spatio-temporal transformers for action spotting, which involves self-supervised learning and knowledge distillation. Action spotting is a timestamp-level temporal action detection task. Our pipeline consists of three steps, with two initialization stages. First, we perform self-supervised initialization of a spatial transformer using short videos as input. Additionally, we initialize a temporal transformer that enhances the spatial transformer's outputs with global context through knowledge distillation from a pre-computed feature bank aligned with each short video segment. In the final step, we fine-tune the transformers to the action spotting task. The experiments, conducted on the SoccerNet-v2 dataset, demonstrate state-of-the-art performance and validate the effectiveness of COMEDIAN's pretraining paradigm. Our results highlight several advantages of our pretraining pipeline, including improved performance and faster convergence compared to non-pretrained models.

Recently, the tokens of images share the same static data flow in many dense networks. However, challenges arise from the variance among the objects in images, such as large variations in the spatial scale and difficulties of recognition for visual entities. In this paper, we propose a data-dependent token routing strategy to elaborate the routing paths of image tokens for Dynamic Vision Transformer, dubbed DiT. The proposed framework generates a data-dependent path per token, adapting to the object scales and visual discrimination of tokens. In feed-forward, the differentiable routing gates are designed to select the scaling paths and feature transformation paths for image tokens, leading to multi-path feature propagation. In this way, the impact of object scales and visual discrimination of image representation can be carefully tuned. Moreover, the computational cost can be further reduced by giving budget constraints to the routing gate and early-stopping of feature extraction. In experiments, our DiT achieves superior performance and favorable complexity/accuracy trade-offs than many SoTA methods on ImageNet classification, object detection, instance segmentation, and semantic segmentation. Particularly, the DiT-B5 obtains 84.8\% top-1 Acc on ImageNet with 10.3 GFLOPs, which is 1.0\% higher than that of the SoTA method with similar computational complexity. These extensive results demonstrate that DiT can serve as versatile backbones for various vision tasks.

In this study, we focus on the problem of 3D human mesh recovery from a single image under obscured conditions. Most state-of-the-art methods aim to improve 2D alignment technologies, such as spatial averaging and 2D joint sampling. However, they tend to neglect the crucial aspect of 3D alignment by improving 3D representations. Furthermore, recent methods struggle to separate the target human from occlusion or background in crowded scenes as they optimize the 3D space of target human with 3D joint coordinates as local supervision. To address these issues, a desirable method would involve a framework for fusing 2D and 3D features and a strategy for optimizing the 3D space globally. Therefore, this paper presents 3D JOint contrastive learning with TRansformers (JOTR) framework for handling occluded 3D human mesh recovery. Our method includes an encoder-decoder transformer architecture to fuse 2D and 3D representations for achieving 2D&3D aligned results in a coarse-to-fine manner and a novel 3D joint contrastive learning approach for adding explicitly global supervision for the 3D feature space. The contrastive learning approach includes two contrastive losses: joint-to-joint contrast for enhancing the similarity of semantically similar voxels (i.e., human joints), and joint-to-non-joint contrast for ensuring discrimination from others (e.g., occlusions and background). Qualitative and quantitative analyses demonstrate that our method outperforms state-of-the-art competitors on both occlusion-specific and standard benchmarks, significantly improving the reconstruction of occluded humans.

Convolutional neural networks (CNNs) have been the consensus for medical image segmentation tasks. However, they suffer from the limitation in modeling long-range dependencies and spatial correlations due to the nature of convolution operation. Although transformers were first developed to address this issue, they fail to capture low-level features. In contrast, it is demonstrated that both local and global features are crucial for dense prediction, such as segmenting in challenging contexts. In this paper, we propose HiFormer, a novel method that efficiently bridges a CNN and a transformer for medical image segmentation. Specifically, we design two multi-scale feature representations using the seminal Swin Transformer module and a CNN-based encoder. To secure a fine fusion of global and local features obtained from the two aforementioned representations, we propose a Double-Level Fusion (DLF) module in the skip connection of the encoder-decoder structure. Extensive experiments on various medical image segmentation datasets demonstrate the effectiveness of HiFormer over other CNN-based, transformer-based, and hybrid methods in terms of computational complexity, and quantitative and qualitative results. Our code is publicly available at: https://github.com/amirhossein-kz/HiFormer

Bird's eye view (BEV) semantic segmentation plays a crucial role in spatial sensing for autonomous driving. Although recent literature has made significant progress on BEV map understanding, they are all based on single-agent camera-based systems. These solutions sometimes have difficulty handling occlusions or detecting distant objects in complex traffic scenes. Vehicle-to-Vehicle (V2V) communication technologies have enabled autonomous vehicles to share sensing information, dramatically improving the perception performance and range compared to single-agent systems. In this paper, we propose CoBEVT, the first generic multi-agent multi-camera perception framework that can cooperatively generate BEV map predictions. To efficiently fuse camera features from multi-view and multi-agent data in an underlying Transformer architecture, we design a fused axial attention module (FAX), which captures sparsely local and global spatial interactions across views and agents. The extensive experiments on the V2V perception dataset, OPV2V, demonstrate that CoBEVT achieves state-of-the-art performance for cooperative BEV semantic segmentation. Moreover, CoBEVT is shown to be generalizable to other tasks, including 1) BEV segmentation with single-agent multi-camera and 2) 3D object detection with multi-agent LiDAR systems, achieving state-of-the-art performance with real-time inference speed. The code is available at https://github.com/DerrickXuNu/CoBEVT.

We propose global context vision transformer (GC ViT), a novel architecture that enhances parameter and compute utilization for computer vision tasks. The core of the novel model are global context self-attention modules, joint with standard local self-attention, to effectively yet efficiently model both long and short-range spatial interactions, as an alternative to complex operations such as an attention masks or local windows shifting. While the local self-attention modules are responsible for modeling short-range information, the global query tokens are shared across all global self-attention modules to interact with local key and values. In addition, we address the lack of inductive bias in ViTs and improve the modeling of inter-channel dependencies by proposing a novel downsampler which leverages a parameter-efficient fused inverted residual block. The proposed GC ViT achieves new state-of-the-art performance across image classification, object detection and semantic segmentation tasks. On ImageNet-1K dataset for classification, GC ViT models with 51M, 90M and 201M parameters achieve 84.3%, 84.9% and 85.6% Top-1 accuracy, respectively, surpassing comparably-sized prior art such as CNN-based ConvNeXt and ViT-based Swin Transformer. Pre-trained GC ViT backbones in downstream tasks of object detection, instance segmentation, and semantic segmentation on MS COCO and ADE20K datasets outperform prior work consistently, sometimes by large margins.

The widespread dissemination of Deepfakes demands effective approaches that can detect perceptually convincing forged images. In this paper, we aim to capture the subtle manipulation artifacts at different scales using transformer models. In particular, we introduce a Multi-modal Multi-scale TRansformer (M2TR), which operates on patches of different sizes to detect local inconsistencies in images at different spatial levels. M2TR further learns to detect forgery artifacts in the frequency domain to complement RGB information through a carefully designed cross modality fusion block. In addition, to stimulate Deepfake detection research, we introduce a high-quality Deepfake dataset, SR-DF, which consists of 4,000 DeepFake videos generated by state-of-the-art face swapping and facial reenactment methods. We conduct extensive experiments to verify the effectiveness of the proposed method, which outperforms state-of-the-art Deepfake detection methods by clear margins.

Diffusion models have demonstrated excellent capabilities in text-to-image generation. Their semantic understanding (i.e., prompt following) ability has also been greatly improved with large language models (e.g., T5, Llama). However, existing models cannot perfectly handle long and complex text prompts, especially when the text prompts contain various objects with numerous attributes and interrelated spatial relationships. While many regional prompting methods have been proposed for UNet-based models (SD1.5, SDXL), but there are still no implementations based on the recent Diffusion Transformer (DiT) architecture, such as SD3 and FLUX.1.In this report, we propose and implement regional prompting for FLUX.1 based on attention manipulation, which enables DiT with fined-grained compositional text-to-image generation capability in a training-free manner. Code is available at https://github.com/antonioo-c/Regional-Prompting-FLUX.

Multimodal Diffusion Transformers (MM-DiTs) have achieved remarkable progress in text-driven visual generation. However, even state-of-the-art MM-DiT models like FLUX struggle with achieving precise alignment between text prompts and generated content. We identify two key issues in the attention mechanism of MM-DiT, namely 1) the suppression of cross-modal attention due to token imbalance between visual and textual modalities and 2) the lack of timestep-aware attention weighting, which hinder the alignment. To address these issues, we propose Temperature-Adjusted Cross-modal Attention (TACA), a parameter-efficient method that dynamically rebalances multimodal interactions through temperature scaling and timestep-dependent adjustment. When combined with LoRA fine-tuning, TACA significantly enhances text-image alignment on the T2I-CompBench benchmark with minimal computational overhead. We tested TACA on state-of-the-art models like FLUX and SD3.5, demonstrating its ability to improve image-text alignment in terms of object appearance, attribute binding, and spatial relationships. Our findings highlight the importance of balancing cross-modal attention in improving semantic fidelity in text-to-image diffusion models. Our codes are publicly available at https://github.com/Vchitect/TACA

Vision Transformers (ViTs), with their ability to model long-range dependencies through self-attention mechanisms, have become a standard architecture in computer vision. However, the interpretability of these models remains a challenge. To address this, we propose LeGrad, an explainability method specifically designed for ViTs. LeGrad computes the gradient with respect to the attention maps of ViT layers, considering the gradient itself as the explainability signal. We aggregate the signal over all layers, combining the activations of the last as well as intermediate tokens to produce the merged explainability map. This makes LeGrad a conceptually simple and an easy-to-implement tool for enhancing the transparency of ViTs. We evaluate LeGrad in challenging segmentation, perturbation, and open-vocabulary settings, showcasing its versatility compared to other SotA explainability methods demonstrating its superior spatial fidelity and robustness to perturbations. A demo and the code is available at https://github.com/WalBouss/LeGrad.

Adapting image models to the video domain has emerged as an efficient paradigm for solving video recognition tasks. Due to the huge number of parameters and effective transferability of image models, performing full fine-tuning is less efficient and even unnecessary. Thus, recent research is shifting its focus toward parameter-efficient image-to-video adaptation. However, these adaptation strategies inevitably introduce extra computational costs to deal with the domain gap and temporal modeling in videos. In this paper, we present a new adaptation paradigm (ZeroI2V) to transfer the image transformers to video recognition tasks (i.e., introduce zero extra cost to the original models during inference). To achieve this goal, we present two core designs. First, to capture the dynamics in videos and reduce the difficulty of image-to-video adaptation, we exploit the flexibility of self-attention and introduce spatial-temporal dual-headed attention (STDHA). This approach efficiently endows the image transformers with temporal modeling capability at zero extra parameters and computation. Second, to handle the domain gap between images and videos, we propose a linear adaption strategy that utilizes lightweight densely placed linear adapters to fully transfer the frozen image models to video recognition. Thanks to the customized linear design, all newly added adapters could be easily merged with the original modules through structural reparameterization after training, enabling zero extra cost during inference. Extensive experiments on representative fully-supervised and few-shot video recognition benchmarks showcase that ZeroI2V can match or even outperform previous state-of-the-art methods while enjoying superior parameter and inference efficiency.

Image classification is a computer vision task where a model analyzes an image to categorize it into a specific label. Vision Transformers (ViT) improve this task by leveraging self-attention to capture complex patterns and long range relationships between image patches. However, a key challenge for ViTs is efficiently incorporating multiscale feature representations, which is inherent in CNNs through their hierarchical structure. In this paper, we introduce the Scale-Aware Graph Attention Vision Transformer (SAG-ViT), a novel framework that addresses this challenge by integrating multi-scale features. Using EfficientNet as a backbone, the model extracts multi-scale feature maps, which are divided into patches to preserve semantic information. These patches are organized into a graph based on spatial and feature similarities, with a Graph Attention Network (GAT) refining the node embeddings. Finally, a Transformer encoder captures long-range dependencies and complex interactions. The SAG-ViT is evaluated on benchmark datasets, demonstrating its effectiveness in enhancing image classification performance.

Insufficient overlap between the melt pools produced during Laser Powder Bed Fusion (L-PBF) can lead to lack-of-fusion defects and deteriorated mechanical and fatigue performance. In-situ monitoring of the melt pool subsurface morphology requires specialized equipment that may not be readily accessible or scalable. Therefore, we introduce a machine learning framework to correlate in-situ two-color thermal images observed via high-speed color imaging to the two-dimensional profile of the melt pool cross-section. Specifically, we employ a hybrid CNN-Transformer architecture to establish a correlation between single bead off-axis thermal image sequences and melt pool cross-section contours measured via optical microscopy. In this architecture, a ResNet model embeds the spatial information contained within the thermal images to a latent vector, while a Transformer model correlates the sequence of embedded vectors to extract temporal information. Our framework is able to model the curvature of the subsurface melt pool structure, with improved performance in high energy density regimes compared to analytical melt pool models. The performance of this model is evaluated through dimensional and geometric comparisons to the corresponding experimental melt pool observations.

Vision Transformers (ViT) have shown their competitive advantages performance-wise compared to convolutional neural networks (CNNs) though they often come with high computational costs. To this end, previous methods explore different attention patterns by limiting a fixed number of spatially nearby tokens to accelerate the ViT's multi-head self-attention (MHSA) operations. However, such structured attention patterns limit the token-to-token connections to their spatial relevance, which disregards learned semantic connections from a full attention mask. In this work, we propose a novel approach to learn instance-dependent attention patterns, by devising a lightweight connectivity predictor module to estimate the connectivity score of each pair of tokens. Intuitively, two tokens have high connectivity scores if the features are considered relevant either spatially or semantically. As each token only attends to a small number of other tokens, the binarized connectivity masks are often very sparse by nature and therefore provide the opportunity to accelerate the network via sparse computations. Equipped with the learned unstructured attention pattern, sparse attention ViT (Sparsifiner) produces a superior Pareto-optimal trade-off between FLOPs and top-1 accuracy on ImageNet compared to token sparsity. Our method reduces 48% to 69% FLOPs of MHSA while the accuracy drop is within 0.4%. We also show that combining attention and token sparsity reduces ViT FLOPs by over 60%.

We present UGround, a Unified visual Grounding paradigm that dynamically selects intermediate layers across Unrolled transformers as ``mask as prompt'', diverging from the prevailing pipeline that leverages the fixed last hidden layer as ``<SEG> as prompt''. UGround addresses two primary challenges posed by the prevailing paradigm: (1) its reliance on the fixed last hidden layer, which sequentially amplifies cumulative errors arising from layer-by-layer propagation without intermediate correction, and (2) its use of <SEG> as a prompt, which implicitly projects textual embeddings into visual space without explicit spatial cues (\eg, coordinates). Central to UGround is Policy-Prompted Masking, which comprises two key components: Stochastic Skip Connection (SSC) and Mask as Prompt (MasP). SSC is a reinforcement learning policy that, via stochastic sampling, allows each <SEG> token to slide across unrolled transformer layers, enabling dynamic layer selection at which it connects to the vision model (\eg, SAM) in a skip-connection fashion. Given the selected hidden layer, MasP uses the similarity map derived from the <SEG> token and image tokens as a soft logit mask to prompt SAM for mask generation, offering explicit spatial cues through its activation regions. To validate the effectiveness of UGround, we, for the first time, have unified visual grounding within a single framework from an attribute perspective, spanning from traditional refer expression segmentation to newly proposed reasoning segmentation, single-target to multi-target, positive query to false premise (empty target). All codes and models are publicly available at https://github.com/rui-qian/UGround{https://github.com/rui-qian/UGround}.

We present Lay-Your-Scene (shorthand LayouSyn), a novel text-to-layout generation pipeline for natural scenes. Prior scene layout generation methods are either closed-vocabulary or use proprietary large language models for open-vocabulary generation, limiting their modeling capabilities and broader applicability in controllable image generation. In this work, we propose to use lightweight open-source language models to obtain scene elements from text prompts and a novel aspect-aware diffusion Transformer architecture trained in an open-vocabulary manner for conditional layout generation. Extensive experiments demonstrate that LayouSyn outperforms existing methods and achieves state-of-the-art performance on challenging spatial and numerical reasoning benchmarks. Additionally, we present two applications of LayouSyn. First, we show that coarse initialization from large language models can be seamlessly combined with our method to achieve better results. Second, we present a pipeline for adding objects to images, demonstrating the potential of LayouSyn in image editing applications.

Vision Transformers (ViTs) mark a revolutionary advance in neural networks with their token mixer's powerful global context capability. However, the pairwise token affinity and complex matrix operations limit its deployment on resource-constrained scenarios and real-time applications, such as mobile devices, although considerable efforts have been made in previous works. In this paper, we introduce CAS-ViT: Convolutional Additive Self-attention Vision Transformers, to achieve a balance between efficiency and performance in mobile applications. Firstly, we argue that the capability of token mixers to obtain global contextual information hinges on multiple information interactions, such as spatial and channel domains. Subsequently, we construct a novel additive similarity function following this paradigm and present an efficient implementation named Convolutional Additive Token Mixer (CATM). This simplification leads to a significant reduction in computational overhead. We evaluate CAS-ViT across a variety of vision tasks, including image classification, object detection, instance segmentation, and semantic segmentation. Our experiments, conducted on GPUs, ONNX, and iPhones, demonstrate that CAS-ViT achieves a competitive performance when compared to other state-of-the-art backbones, establishing it as a viable option for efficient mobile vision applications. Our code and model are available at: https://github.com/Tianfang-Zhang/CAS-ViT

We present Muse, a text-to-image Transformer model that achieves state-of-the-art image generation performance while being significantly more efficient than diffusion or autoregressive models. Muse is trained on a masked modeling task in discrete token space: given the text embedding extracted from a pre-trained large language model (LLM), Muse is trained to predict randomly masked image tokens. Compared to pixel-space diffusion models, such as Imagen and DALL-E 2, Muse is significantly more efficient due to the use of discrete tokens and requiring fewer sampling iterations; compared to autoregressive models, such as Parti, Muse is more efficient due to the use of parallel decoding. The use of a pre-trained LLM enables fine-grained language understanding, translating to high-fidelity image generation and the understanding of visual concepts such as objects, their spatial relationships, pose, cardinality etc. Our 900M parameter model achieves a new SOTA on CC3M, with an FID score of 6.06. The Muse 3B parameter model achieves an FID of 7.88 on zero-shot COCO evaluation, along with a CLIP score of 0.32. Muse also directly enables a number of image editing applications without the need to fine-tune or invert the model: inpainting, outpainting, and mask-free editing. More results are available at https://muse-model.github.io

We explore the capability of plain Vision Transformers (ViTs) for semantic segmentation and propose the SegVit. Previous ViT-based segmentation networks usually learn a pixel-level representation from the output of the ViT. Differently, we make use of the fundamental component -- attention mechanism, to generate masks for semantic segmentation. Specifically, we propose the Attention-to-Mask (ATM) module, in which the similarity maps between a set of learnable class tokens and the spatial feature maps are transferred to the segmentation masks. Experiments show that our proposed SegVit using the ATM module outperforms its counterparts using the plain ViT backbone on the ADE20K dataset and achieves new state-of-the-art performance on COCO-Stuff-10K and PASCAL-Context datasets. Furthermore, to reduce the computational cost of the ViT backbone, we propose query-based down-sampling (QD) and query-based up-sampling (QU) to build a Shrunk structure. With the proposed Shrunk structure, the model can save up to 40% computations while maintaining competitive performance.

Learning to solve precision-based manipulation tasks from visual feedback using Reinforcement Learning (RL) could drastically reduce the engineering efforts required by traditional robot systems. However, performing fine-grained motor control from visual inputs alone is challenging, especially with a static third-person camera as often used in previous work. We propose a setting for robotic manipulation in which the agent receives visual feedback from both a third-person camera and an egocentric camera mounted on the robot's wrist. While the third-person camera is static, the egocentric camera enables the robot to actively control its vision to aid in precise manipulation. To fuse visual information from both cameras effectively, we additionally propose to use Transformers with a cross-view attention mechanism that models spatial attention from one view to another (and vice-versa), and use the learned features as input to an RL policy. Our method improves learning over strong single-view and multi-view baselines, and successfully transfers to a set of challenging manipulation tasks on a real robot with uncalibrated cameras, no access to state information, and a high degree of task variability. In a hammer manipulation task, our method succeeds in 75% of trials versus 38% and 13% for multi-view and single-view baselines, respectively.

Transformers do not scale very well to long sequence lengths largely because of quadratic self-attention complexity. In the recent months, a wide spectrum of efficient, fast Transformers have been proposed to tackle this problem, more often than not claiming superior or comparable model quality to vanilla Transformer models. To this date, there is no well-established consensus on how to evaluate this class of models. Moreover, inconsistent benchmarking on a wide spectrum of tasks and datasets makes it difficult to assess relative model quality amongst many models. This paper proposes a systematic and unified benchmark, LRA, specifically focused on evaluating model quality under long-context scenarios. Our benchmark is a suite of tasks consisting of sequences ranging from 1K to 16K tokens, encompassing a wide range of data types and modalities such as text, natural, synthetic images, and mathematical expressions requiring similarity, structural, and visual-spatial reasoning. We systematically evaluate ten well-established long-range Transformer models (Reformers, Linformers, Linear Transformers, Sinkhorn Transformers, Performers, Synthesizers, Sparse Transformers, and Longformers) on our newly proposed benchmark suite. LRA paves the way towards better understanding this class of efficient Transformer models, facilitates more research in this direction, and presents new challenging tasks to tackle. Our benchmark code will be released at https://github.com/google-research/long-range-arena.