Daily Papers

Transformers have become the primary backbone of the computer vision community due to their impressive performance. However, the unfriendly computation cost impedes their potential in the video recognition domain. To optimize the speed-accuracy trade-off, we propose Semantic-aware Temporal Accumulation score (STA) to prune spatio-temporal tokens integrally. STA score considers two critical factors: temporal redundancy and semantic importance. The former depicts a specific region based on whether it is a new occurrence or a seen entity by aggregating token-to-token similarity in consecutive frames while the latter evaluates each token based on its contribution to the overall prediction. As a result, tokens with higher scores of STA carry more temporal redundancy as well as lower semantics thus being pruned. Based on the STA score, we are able to progressively prune the tokens without introducing any additional parameters or requiring further re-training. We directly apply the STA module to off-the-shelf ViT and VideoSwin backbones, and the empirical results on Kinetics-400 and Something-Something V2 achieve over 30% computation reduction with a negligible ~0.2% accuracy drop. The code is released at https://github.com/Mark12Ding/STA.

Recently, Vision Large Language Models (VLLMs) integrated with vision encoders have shown promising performance in vision understanding. The key of VLLMs is to encode visual content into sequences of visual tokens, enabling VLLMs to simultaneously process both visual and textual content. However, understanding videos, especially long videos, remain a challenge to VLLMs as the number of visual tokens grows rapidly when encoding videos, resulting in the risk of exceeding the context window of VLLMs and introducing heavy computation burden. To restrict the number of visual tokens, existing VLLMs either: (1) uniformly downsample videos into a fixed number of frames or (2) reducing the number of visual tokens encoded from each frame. We argue the former solution neglects the rich temporal cue in videos and the later overlooks the spatial details in each frame. In this work, we present Balanced-VLLM (B-VLLM): a novel VLLM framework that aims to effectively leverage task relevant spatio-temporal cues while restricting the number of visual tokens under the VLLM context window length. At the core of our method, we devise a text-conditioned adaptive frame selection module to identify frames relevant to the visual understanding task. The selected frames are then de-duplicated using a temporal frame token merging technique. The visual tokens of the selected frames are processed through a spatial token sampling module and an optional spatial token merging strategy to achieve precise control over the token count. Experimental results show that B-VLLM is effective in balancing the number of frames and visual tokens in video understanding, yielding superior performance on various video understanding benchmarks. Our code is available at https://github.com/zhuqiangLu/B-VLLM.

Video large language models (LLMs) achieve strong video understanding by leveraging a large number of spatio-temporal tokens, but suffer from quadratic computational scaling with token count. To address this, we propose a training-free spatio-temporal token merging method, named STTM. Our key insight is to exploit local spatial and temporal redundancy in video data which has been overlooked in prior work. STTM first transforms each frame into multi-granular spatial tokens using a coarse-to-fine search over a quadtree structure, then performs directed pairwise merging across the temporal dimension. This decomposed merging approach outperforms existing token reduction methods across six video QA benchmarks. Notably, STTM achieves a 2times speed-up with only a 0.5% accuracy drop under a 50% token budget, and a 3times speed-up with just a 2% drop under a 30% budget. Moreover, STTM is query-agnostic, allowing KV cache reuse across different questions for the same video. The project page is available at https://www.jshyun.me/projects/sttm.

We present pure-transformer based models for video classification, drawing upon the recent success of such models in image classification. Our model extracts spatio-temporal tokens from the input video, which are then encoded by a series of transformer layers. In order to handle the long sequences of tokens encountered in video, we propose several, efficient variants of our model which factorise the spatial- and temporal-dimensions of the input. Although transformer-based models are known to only be effective when large training datasets are available, we show how we can effectively regularise the model during training and leverage pretrained image models to be able to train on comparatively small datasets. We conduct thorough ablation studies, and achieve state-of-the-art results on multiple video classification benchmarks including Kinetics 400 and 600, Epic Kitchens, Something-Something v2 and Moments in Time, outperforming prior methods based on deep 3D convolutional networks. To facilitate further research, we release code at https://github.com/google-research/scenic/tree/main/scenic/projects/vivit

We propose a novel Latent Diffusion Transformer, namely Latte, for video generation. Latte first extracts spatio-temporal tokens from input videos and then adopts a series of Transformer blocks to model video distribution in the latent space. In order to model a substantial number of tokens extracted from videos, four efficient variants are introduced from the perspective of decomposing the spatial and temporal dimensions of input videos. To improve the quality of generated videos, we determine the best practices of Latte through rigorous experimental analysis, including video clip patch embedding, model variants, timestep-class information injection, temporal positional embedding, and learning strategies. Our comprehensive evaluation demonstrates that Latte achieves state-of-the-art performance across four standard video generation datasets, i.e., FaceForensics, SkyTimelapse, UCF101, and Taichi-HD. In addition, we extend Latte to text-to-video generation (T2V) task, where Latte achieves comparable results compared to recent T2V models. We strongly believe that Latte provides valuable insights for future research on incorporating Transformers into diffusion models for video generation.

We introduce PDE-Transformer, an improved transformer-based architecture for surrogate modeling of physics simulations on regular grids. We combine recent architectural improvements of diffusion transformers with adjustments specific for large-scale simulations to yield a more scalable and versatile general-purpose transformer architecture, which can be used as the backbone for building large-scale foundation models in physical sciences. We demonstrate that our proposed architecture outperforms state-of-the-art transformer architectures for computer vision on a large dataset of 16 different types of PDEs. We propose to embed different physical channels individually as spatio-temporal tokens, which interact via channel-wise self-attention. This helps to maintain a consistent information density of tokens when learning multiple types of PDEs simultaneously. We demonstrate that our pre-trained models achieve improved performance on several challenging downstream tasks compared to training from scratch and also beat other foundation model architectures for physics simulations.

Human-centric motion control in video generation remains a critical challenge, particularly when jointly controlling camera movements and human poses in scenarios like the iconic Grammy Glambot moment. While recent video diffusion models have made significant progress, existing approaches struggle with limited motion representations and inadequate integration of camera and human motion controls. In this work, we present TokenMotion, the first DiT-based video diffusion framework that enables fine-grained control over camera motion, human motion, and their joint interaction. We represent camera trajectories and human poses as spatio-temporal tokens to enable local control granularity. Our approach introduces a unified modeling framework utilizing a decouple-and-fuse strategy, bridged by a human-aware dynamic mask that effectively handles the spatially-and-temporally varying nature of combined motion signals. Through extensive experiments, we demonstrate TokenMotion's effectiveness across both text-to-video and image-to-video paradigms, consistently outperforming current state-of-the-art methods in human-centric motion control tasks. Our work represents a significant advancement in controllable video generation, with particular relevance for creative production applications.

The Transformer architecture is shown to provide a powerful machine transduction framework for online handwritten gestures corresponding to glyph strokes of natural language sentences. The attention mechanism is successfully used to create latent representations of an end-to-end encoder-decoder model, solving multi-level segmentation while also learning some language features and syntax rules. The additional use of a large decoding space with some learned Byte-Pair-Encoding (BPE) is shown to provide robustness to ablated inputs and syntax rules. The encoder stack was directly fed with spatio-temporal data tokens potentially forming an infinitely large input vocabulary, an approach that finds applications beyond that of this work. Encoder transfer learning capabilities is also demonstrated on several languages resulting in faster optimisation and shared parameters. A new supervised dataset of online handwriting gestures suitable for generic handwriting recognition tasks was used to successfully train a small transformer model to an average normalised Levenshtein accuracy of 96% on English or German sentences and 94% in French.

The Transformer architecture is shown to provide a powerful framework as an end-to-end model for building expression trees from online handwritten gestures corresponding to glyph strokes. In particular, the attention mechanism was successfully used to encode, learn and enforce the underlying syntax of expressions creating latent representations that are correctly decoded to the exact mathematical expression tree, providing robustness to ablated inputs and unseen glyphs. For the first time, the encoder is fed with spatio-temporal data tokens potentially forming an infinitely large vocabulary, which finds applications beyond that of online gesture recognition. A new supervised dataset of online handwriting gestures is provided for training models on generic handwriting recognition tasks and a new metric is proposed for the evaluation of the syntactic correctness of the output expression trees. A small Transformer model suitable for edge inference was successfully trained to an average normalised Levenshtein accuracy of 94%, resulting in valid postfix RPN tree representation for 94% of predictions.

Diffusion Transformers have demonstrated remarkable capabilities in visual synthesis, yet they often struggle with high-level semantic reasoning and long-horizon planning. This limitation frequently leads to visual hallucinations and mis-alignments with user instructions, especially in scenarios involving complex scene understanding, human-object interactions, multi-stage actions, and in-context motion reasoning. To address these challenges, we propose Plan-X, a framework that explicitly enforces high-level semantic planning to instruct video generation process. At its core lies a Semantic Planner, a learnable multimodal language model that reasons over the user's intent from both text prompts and visual context, and autoregressively generates a sequence of text-grounded spatio-temporal semantic tokens. These semantic tokens, complementary to high-level text prompt guidance, serve as structured "semantic sketches" over time for the video diffusion model, which has its strength at synthesizing high-fidelity visual details. Plan-X effectively integrates the strength of language models in multimodal in-context reasoning and planning, together with the strength of diffusion models in photorealistic video synthesis. Extensive experiments demonstrate that our framework substantially reduces visual hallucinations and enables fine-grained, instruction-aligned video generation consistent with multimodal context.

Recently, large-scale pre-training methods like CLIP have made great progress in multi-modal research such as text-video retrieval. In CLIP, transformers are vital for modeling complex multi-modal relations. However, in the vision transformer of CLIP, the essential visual tokenization process, which produces discrete visual token sequences, generates many homogeneous tokens due to the redundancy nature of consecutive and similar frames in videos. This significantly increases computation costs and hinders the deployment of video retrieval models in web applications. In this paper, to reduce the number of redundant video tokens, we design a multi-segment token clustering algorithm to find the most representative tokens and drop the non-essential ones. As the frame redundancy occurs mostly in consecutive frames, we divide videos into multiple segments and conduct segment-level clustering. Center tokens from each segment are later concatenated into a new sequence, while their original spatial-temporal relations are well maintained. We instantiate two clustering algorithms to efficiently find deterministic medoids and iteratively partition groups in high dimensional space. Through this token clustering and center selection procedure, we successfully reduce computation costs by removing redundant visual tokens. This method further enhances segment-level semantic alignment between video and text representations, enforcing the spatio-temporal interactions of tokens from within-segment frames. Our method, coined as CenterCLIP, surpasses existing state-of-the-art by a large margin on typical text-video benchmarks, while reducing the training memory cost by 35\% and accelerating the inference speed by 14\% at the best case. The code is available at {https://github.com/mzhaoshuai/CenterCLIP}{{https://github.com/mzhaoshuai/CenterCLIP}}.

Spatio-temporal action detection encompasses the tasks of localizing and classifying individual actions within a video. Recent works aim to enhance this process by incorporating interaction modeling, which captures the relationship between people and their surrounding context. However, these approaches have primarily focused on fully-supervised learning, and the current limitation lies in the lack of generalization capability to recognize unseen action categories. In this paper, we aim to adapt the pretrained image-language models to detect unseen actions. To this end, we propose a method which can effectively leverage the rich knowledge of visual-language models to perform Person-Context Interaction. Meanwhile, our Context Prompting module will utilize contextual information to prompt labels, thereby enhancing the generation of more representative text features. Moreover, to address the challenge of recognizing distinct actions by multiple people at the same timestamp, we design the Interest Token Spotting mechanism which employs pretrained visual knowledge to find each person's interest context tokens, and then these tokens will be used for prompting to generate text features tailored to each individual. To evaluate the ability to detect unseen actions, we propose a comprehensive benchmark on J-HMDB, UCF101-24, and AVA datasets. The experiments show that our method achieves superior results compared to previous approaches and can be further extended to multi-action videos, bringing it closer to real-world applications. The code and data can be found in https://webber2933.github.io/ST-CLIP-project-page.

Multimodal large language models (MLLMs) have demonstrated strong general-purpose capabilities in open-world visual comprehension. However, most existing MLLMs primarily focus on holistic, scene-level understanding, often overlooking the need for fine-grained, object-centric reasoning. In this paper, we present PixelRefer, a unified region-level MLLM framework that enables advanced fine-grained understanding over user-specified regions across both images and videos. Motivated by the observation that LLM attention predominantly focuses on object-level tokens, we propose a Scale-Adaptive Object Tokenizer (SAOT) to generate compact and semantically rich object representations from free-form regions. Our analysis reveals that global visual tokens contribute mainly in early LLM layers, inspiring the design of PixelRefer-Lite, an efficient variant that employs an Object-Centric Infusion module to pre-fuse global context into object tokens. This yields a lightweight Object-Only Framework that substantially reduces computational cost while maintaining high semantic fidelity. To facilitate fine-grained instruction tuning, we curate PixelRefer-2.2M, a high-quality object-centric instruction dataset. Extensive experiments across a range of benchmarks validate that PixelRefer achieves leading performance with fewer training samples, while PixelRefer-Lite offers competitive accuracy with notable gains in efficiency.

Multimodal large language models (MLLMs) have made remarkable progress in either temporal or spatial localization. However, they struggle to perform spatio-temporal video grounding. This limitation stems from two major challenges. Firstly, it is difficult to extract accurate spatio-temporal information of each frame in the video. Secondly, the substantial number of visual tokens makes it challenging to precisely map visual tokens of each frame to their corresponding spatial coordinates. To address these issues, we introduce SpaceVLLM, a MLLM endowed with spatio-temporal video grounding capability. Specifically, we adopt a set of interleaved Spatio-Temporal Aware Queries to capture temporal perception and dynamic spatial information. Moreover, we propose a Query-Guided Space Decoder to establish a corresponding connection between the queries and spatial coordinates. Additionally, due to the lack of spatio-temporal datasets, we construct the Unified Spatio-Temporal Grounding (Uni-STG) dataset, comprising 480K instances across three tasks. This dataset fully exploits the potential of MLLM to simultaneously facilitate localization in both temporal and spatial dimensions. Extensive experiments demonstrate that SpaceVLLM achieves the state-of-the-art performance across 11 benchmarks covering temporal, spatial, spatio-temporal and video understanding tasks, highlighting the effectiveness of our approach. Our code, datasets and model will be released at https://github.com/Jayce1kk/SpaceVLLM.

Spatio-temporal video grounding (STVG) aims at localizing the spatio-temporal tube of a video, as specified by the input text query. In this paper, we utilize multimodal large language models (MLLMs) to explore a zero-shot solution in STVG. We reveal two key insights about MLLMs: (1) MLLMs tend to dynamically assign special tokens, referred to as grounding tokens, for grounding the text query; and (2) MLLMs often suffer from suboptimal grounding due to the inability to fully integrate the cues in the text query (e.g., attributes, actions) for inference. Based on these insights, we propose a MLLM-based zero-shot framework for STVG, which includes novel decomposed spatio-temporal highlighting (DSTH) and temporal-augmented assembling (TAS) strategies to unleash the reasoning ability of MLLMs. The DSTH strategy first decouples the original query into attribute and action sub-queries for inquiring the existence of the target both spatially and temporally. It then uses a novel logit-guided re-attention (LRA) module to learn latent variables as spatial and temporal prompts, by regularizing token predictions for each sub-query. These prompts highlight attribute and action cues, respectively, directing the model's attention to reliable spatial and temporal related visual regions. In addition, as the spatial grounding by the attribute sub-query should be temporally consistent, we introduce the TAS strategy to assemble the predictions using the original video frames and the temporal-augmented frames as inputs to help improve temporal consistency. We evaluate our method on various MLLMs, and show that it outperforms SOTA methods on three common STVG benchmarks. The code will be available at https://github.com/zaiquanyang/LLaVA_Next_STVG.

Graphical user interface (GUI) agents built on vision-language models have emerged as a promising approach to automate human-computer workflows. However, they also face the inefficiency challenge as they process long sequences of high-resolution screenshots and solving long-horizon tasks, making inference slow, costly and memory-bound. While key-value (KV) caching can mitigate this, storing the full cache is prohibitive for image-heavy contexts. Existing cache-compression methods are sub-optimal as they do not account for the spatial and temporal redundancy of GUIs. In this work, we first analyze attention patterns in GUI agent workloads and find that, unlike in natural images, attention sparsity is uniformly high across all transformer layers. This insight motivates a simple uniform budget allocation strategy, which we show empirically outperforms more complex layer-varying schemes. Building on this, we introduce GUI-KV, a plug-and-play KV cache compression method for GUI agents that requires no retraining. GUI-KV combines two novel techniques: (i) spatial saliency guidance, which augments attention scores with the L2 norm of hidden states to better preserve semantically important visual tokens, and (ii) temporal redundancy scoring, which projects previous frames' keys onto the current frame's key subspace to preferentially prune redundant history. Across standard GUI agent benchmarks and models, GUI-KV outperforms competitive KV compression baselines, closely matching full-cache accuracy at modest budgets. Notably, in a 5-screenshot setting on the AgentNetBench benchmark, GUI-KV reduces decoding FLOPs by 38.9% while increasing step accuracy by 4.1% over the full-cache baseline. These results demonstrate that exploiting GUI-specific redundancies enables efficient and reliable agent performance.

We present xGen-MM-Vid (BLIP-3-Video): a multimodal language model for videos, particularly designed to efficiently capture temporal information over multiple frames. BLIP-3-Video takes advantage of the 'temporal encoder' in addition to the conventional visual tokenizer, which maps a sequence of tokens over multiple frames into a compact set of visual tokens. This enables BLIP3-Video to use much fewer visual tokens than its competing models (e.g., 32 vs. 4608 tokens). We explore different types of temporal encoders, including learnable spatio-temporal pooling as well as sequential models like Token Turing Machines. We experimentally confirm that BLIP-3-Video obtains video question-answering accuracies comparable to much larger state-of-the-art models (e.g., 34B), while being much smaller (i.e., 4B) and more efficient by using fewer visual tokens. The project website is at https://www.salesforceairesearch.com/opensource/xGen-MM-Vid/index.html

Vision Transformers (ViT) have emerged as the de-facto choice for numerous industry grade vision solutions. But their inference cost can be prohibitive for many settings, as they compute self-attention in each layer which suffers from quadratic computational complexity in the number of tokens. On the other hand, spatial information in images and spatio-temporal information in videos is usually sparse and redundant. In this work, we introduce LookupViT, that aims to exploit this information sparsity to reduce ViT inference cost. LookupViT provides a novel general purpose vision transformer block that operates by compressing information from higher resolution tokens to a fixed number of tokens. These few compressed tokens undergo meticulous processing, while the higher-resolution tokens are passed through computationally cheaper layers. Information sharing between these two token sets is enabled through a bidirectional cross-attention mechanism. The approach offers multiple advantages - (a) easy to implement on standard ML accelerators (GPUs/TPUs) via standard high-level operators, (b) applicable to standard ViT and its variants, thus generalizes to various tasks, (c) can handle different tokenization and attention approaches. LookupViT also offers flexibility for the compressed tokens, enabling performance-computation trade-offs in a single trained model. We show LookupViT's effectiveness on multiple domains - (a) for image-classification (ImageNet-1K and ImageNet-21K), (b) video classification (Kinetics400 and Something-Something V2), (c) image captioning (COCO-Captions) with a frozen encoder. LookupViT provides 2times reduction in FLOPs while upholding or improving accuracy across these domains. In addition, LookupViT also demonstrates out-of-the-box robustness and generalization on image classification (ImageNet-C,R,A,O), improving by up to 4% over ViT.

Multimodal large language models (MLLMs) have made remarkable strides, largely driven by their ability to process increasingly long and complex contexts, such as high-resolution images, extended video sequences, and lengthy audio input. While this ability significantly enhances MLLM capabilities, it introduces substantial computational challenges, primarily due to the quadratic complexity of self-attention mechanisms with numerous input tokens. To mitigate these bottlenecks, token compression has emerged as an auspicious and critical approach, efficiently reducing the number of tokens during both training and inference. In this paper, we present the first systematic survey and synthesis of the burgeoning field of multimodal long context token compression. Recognizing that effective compression strategies are deeply tied to the unique characteristics and redundancies of each modality, we categorize existing approaches by their primary data focus, enabling researchers to quickly access and learn methods tailored to their specific area of interest: (1) image-centric compression, which addresses spatial redundancy in visual data; (2) video-centric compression, which tackles spatio-temporal redundancy in dynamic sequences; and (3) audio-centric compression, which handles temporal and spectral redundancy in acoustic signals. Beyond this modality-driven categorization, we further dissect methods based on their underlying mechanisms, including transformation-based, similarity-based, attention-based, and query-based approaches. By providing a comprehensive and structured overview, this survey aims to consolidate current progress, identify key challenges, and inspire future research directions in this rapidly evolving domain. We also maintain a public repository to continuously track and update the latest advances in this promising area.

We present Omni-RGPT, a multimodal large language model designed to facilitate region-level comprehension for both images and videos. To achieve consistent region representation across spatio-temporal dimensions, we introduce Token Mark, a set of tokens highlighting the target regions within the visual feature space. These tokens are directly embedded into spatial regions using region prompts (e.g., boxes or masks) and simultaneously incorporated into the text prompt to specify the target, establishing a direct connection between visual and text tokens. To further support robust video understanding without requiring tracklets, we introduce an auxiliary task that guides Token Mark by leveraging the consistency of the tokens, enabling stable region interpretation across the video. Additionally, we introduce a large-scale region-level video instruction dataset (RegVID-300k). Omni-RGPT achieves state-of-the-art results on image and video-based commonsense reasoning benchmarks while showing strong performance in captioning and referring expression comprehension tasks.

Modern video generation frameworks based on Latent Diffusion Models suffer from inefficiencies in tokenization due to the Frame-Proportional Information Assumption. Existing tokenizers provide fixed temporal compression rates, causing the computational cost of the diffusion model to scale linearly with the frame rate. The paper proposes the Duration-Proportional Information Assumption: the upper bound on the information capacity of a video is proportional to the duration rather than the number of frames. Based on this insight, the paper introduces VFRTok, a Transformer-based video tokenizer, that enables variable frame rate encoding and decoding through asymmetric frame rate training between the encoder and decoder. Furthermore, the paper proposes Partial Rotary Position Embeddings (RoPE) to decouple position and content modeling, which groups correlated patches into unified tokens. The Partial RoPE effectively improves content-awareness, enhancing the video generation capability. Benefiting from the compact and continuous spatio-temporal representation, VFRTok achieves competitive reconstruction quality and state-of-the-art generation fidelity while using only 1/8 tokens compared to existing tokenizers.

Current vision models typically maintain a fixed correspondence between their representation structure and image space. Each layer comprises a set of tokens arranged "on-the-grid," which biases patches or tokens to encode information at a specific spatio(-temporal) location. In this work we present Moving Off-the-Grid (MooG), a self-supervised video representation model that offers an alternative approach, allowing tokens to move "off-the-grid" to better enable them to represent scene elements consistently, even as they move across the image plane through time. By using a combination of cross-attention and positional embeddings we disentangle the representation structure and image structure. We find that a simple self-supervised objective--next frame prediction--trained on video data, results in a set of latent tokens which bind to specific scene structures and track them as they move. We demonstrate the usefulness of MooG's learned representation both qualitatively and quantitatively by training readouts on top of the learned representation on a variety of downstream tasks. We show that MooG can provide a strong foundation for different vision tasks when compared to "on-the-grid" baselines.

Generating both plausible and accurate full body avatar motion is the key to the quality of immersive experiences in mixed reality scenarios. Head-Mounted Devices (HMDs) typically only provide a few input signals, such as head and hands 6-DoF. Recently, different approaches achieved impressive performance in generating full body motion given only head and hands signal. However, to the best of our knowledge, all existing approaches rely on full hand visibility. While this is the case when, e.g., using motion controllers, a considerable proportion of mixed reality experiences do not involve motion controllers and instead rely on egocentric hand tracking. This introduces the challenge of partial hand visibility owing to the restricted field of view of the HMD. In this paper, we propose the first unified approach, HMD-NeMo, that addresses plausible and accurate full body motion generation even when the hands may be only partially visible. HMD-NeMo is a lightweight neural network that predicts the full body motion in an online and real-time fashion. At the heart of HMD-NeMo is the spatio-temporal encoder with novel temporally adaptable mask tokens that encourage plausible motion in the absence of hand observations. We perform extensive analysis of the impact of different components in HMD-NeMo and introduce a new state-of-the-art on AMASS dataset through our evaluation.

Video super-resolution (VSR) aims to restore a sequence of high-resolution (HR) frames from their low-resolution (LR) counterparts. Although some progress has been made, there are grand challenges to effectively utilize temporal dependency in entire video sequences. Existing approaches usually align and aggregate video frames from limited adjacent frames (e.g., 5 or 7 frames), which prevents these approaches from satisfactory results. In this paper, we take one step further to enable effective spatio-temporal learning in videos. We propose a novel Trajectory-aware Transformer for Video Super-Resolution (TTVSR). In particular, we formulate video frames into several pre-aligned trajectories which consist of continuous visual tokens. For a query token, self-attention is only learned on relevant visual tokens along spatio-temporal trajectories. Compared with vanilla vision Transformers, such a design significantly reduces the computational cost and enables Transformers to model long-range features. We further propose a cross-scale feature tokenization module to overcome scale-changing problems that often occur in long-range videos. Experimental results demonstrate the superiority of the proposed TTVSR over state-of-the-art models, by extensive quantitative and qualitative evaluations in four widely-used video super-resolution benchmarks. Both code and pre-trained models can be downloaded at https://github.com/researchmm/TTVSR.

Transformers have been successfully applied in the field of video-based 3D human pose estimation. However, the high computational costs of these video pose transformers (VPTs) make them impractical on resource-constrained devices. In this paper, we present a hierarchical plug-and-play pruning-and-recovering framework, called Hierarchical Hourglass Tokenizer (H_{2}OT), for efficient transformer-based 3D human pose estimation from videos. H_{2}OT begins with progressively pruning pose tokens of redundant frames and ends with recovering full-length sequences, resulting in a few pose tokens in the intermediate transformer blocks and thus improving the model efficiency. It works with two key modules, namely, a Token Pruning Module (TPM) and a Token Recovering Module (TRM). TPM dynamically selects a few representative tokens to eliminate the redundancy of video frames, while TRM restores the detailed spatio-temporal information based on the selected tokens, thereby expanding the network output to the original full-length temporal resolution for fast inference. Our method is general-purpose: it can be easily incorporated into common VPT models on both seq2seq and seq2frame pipelines while effectively accommodating different token pruning and recovery strategies. In addition, our H_{2}OT reveals that maintaining the full pose sequence is unnecessary, and a few pose tokens of representative frames can achieve both high efficiency and estimation accuracy. Extensive experiments on multiple benchmark datasets demonstrate both the effectiveness and efficiency of the proposed method. Code and models are available at https://github.com/NationalGAILab/HoT.

Omnimodal large language models (OmniLLMs) have attracted increasing research attention of late towards unified audio-video understanding, wherein processing audio-video token sequences creates a significant computational bottleneck, however. Existing token compression methods have yet to accommodate this emerging need of jointly compressing multimodal tokens. To bridge this gap, we present OmniZip, a training-free, audio-guided audio-visual token-compression framework that optimizes multimodal token representation and accelerates inference. Specifically, OmniZip first identifies salient audio tokens, then computes an audio retention score for each time group to capture information density, thereby dynamically guiding video token pruning and preserving cues from audio anchors enhanced by cross-modal similarity. For each time window, OmniZip compresses the video tokens using an interleaved spatio-temporal scheme. Extensive empirical results demonstrate the merits of OmniZip - it achieves 3.42X inference speedup and 1.4X memory reduction over other top-performing counterparts, while maintaining performance with no training.

Human image animation has gained increasing attention and developed rapidly due to its broad applications in digital humans. However, existing methods rely largely on 2D-rendered pose images for motion guidance, which limits generalization and discards essential 3D information for open-world animation. To tackle this problem, we propose MTVCrafter (Motion Tokenization Video Crafter), the first framework that directly models raw 3D motion sequences (i.e., 4D motion) for human image animation. Specifically, we introduce 4DMoT (4D motion tokenizer) to quantize 3D motion sequences into 4D motion tokens. Compared to 2D-rendered pose images, 4D motion tokens offer more robust spatio-temporal cues and avoid strict pixel-level alignment between pose image and character, enabling more flexible and disentangled control. Then, we introduce MV-DiT (Motion-aware Video DiT). By designing unique motion attention with 4D positional encodings, MV-DiT can effectively leverage motion tokens as 4D compact yet expressive context for human image animation in the complex 3D world. Hence, it marks a significant step forward in this field and opens a new direction for pose-guided human video generation. Experiments show that our MTVCrafter achieves state-of-the-art results with an FID-VID of 6.98, surpassing the second-best by 65%. Powered by robust motion tokens, MTVCrafter also generalizes well to diverse open-world characters (single/multiple, full/half-body) across various styles and scenarios. Our video demos and code are on: https://github.com/DINGYANB/MTVCrafter.

We propose MAMo, a novel memory and attention frame-work for monocular video depth estimation. MAMo can augment and improve any single-image depth estimation networks into video depth estimation models, enabling them to take advantage of the temporal information to predict more accurate depth. In MAMo, we augment model with memory which aids the depth prediction as the model streams through the video. Specifically, the memory stores learned visual and displacement tokens of the previous time instances. This allows the depth network to cross-reference relevant features from the past when predicting depth on the current frame. We introduce a novel scheme to continuously update the memory, optimizing it to keep tokens that correspond with both the past and the present visual information. We adopt attention-based approach to process memory features where we first learn the spatio-temporal relation among the resultant visual and displacement memory tokens using self-attention module. Further, the output features of self-attention are aggregated with the current visual features through cross-attention. The cross-attended features are finally given to a decoder to predict depth on the current frame. Through extensive experiments on several benchmarks, including KITTI, NYU-Depth V2, and DDAD, we show that MAMo consistently improves monocular depth estimation networks and sets new state-of-the-art (SOTA) accuracy. Notably, our MAMo video depth estimation provides higher accuracy with lower latency, when omparing to SOTA cost-volume-based video depth models.

We present Phenaki, a model capable of realistic video synthesis, given a sequence of textual prompts. Generating videos from text is particularly challenging due to the computational cost, limited quantities of high quality text-video data and variable length of videos. To address these issues, we introduce a new model for learning video representation which compresses the video to a small representation of discrete tokens. This tokenizer uses causal attention in time, which allows it to work with variable-length videos. To generate video tokens from text we are using a bidirectional masked transformer conditioned on pre-computed text tokens. The generated video tokens are subsequently de-tokenized to create the actual video. To address data issues, we demonstrate how joint training on a large corpus of image-text pairs as well as a smaller number of video-text examples can result in generalization beyond what is available in the video datasets. Compared to the previous video generation methods, Phenaki can generate arbitrary long videos conditioned on a sequence of prompts (i.e. time variable text or a story) in open domain. To the best of our knowledge, this is the first time a paper studies generating videos from time variable prompts. In addition, compared to the per-frame baselines, the proposed video encoder-decoder computes fewer tokens per video but results in better spatio-temporal consistency.

Referring Video Object Segmentation (RVOS) aims to segment target objects in video sequences based on natural language descriptions. While recent advances in Multi-modal Large Language Models (MLLMs) have improved RVOS performance through enhanced text-video understanding, several challenges remain, including insufficient exploitation of MLLMs' prior knowledge, prohibitive computational and memory costs for long-duration videos, and inadequate handling of complex temporal dynamics. In this work, we propose SVAC, a unified model that improves RVOS by scaling up input frames and segmentation tokens to enhance video-language interaction and segmentation precision. To address the resulting computational challenges, SVAC incorporates the Anchor-Based Spatio-Temporal Compression (ASTC) module to compress visual tokens while preserving essential spatio-temporal structure. Moreover, the Clip-Specific Allocation (CSA) strategy is introduced to better handle dynamic object behaviors across video clips. Experimental results demonstrate that SVAC achieves state-of-the-art performance on multiple RVOS benchmarks with competitive efficiency. Our code is available at https://github.com/lizhang1998/SVAC.

Transformers have been successfully applied in the field of video-based 3D human pose estimation. However, the high computational costs of these video pose transformers (VPTs) make them impractical on resource-constrained devices. In this paper, we present a plug-and-play pruning-and-recovering framework, called Hourglass Tokenizer (HoT), for efficient transformer-based 3D human pose estimation from videos. Our HoT begins with pruning pose tokens of redundant frames and ends with recovering full-length tokens, resulting in a few pose tokens in the intermediate transformer blocks and thus improving the model efficiency. To effectively achieve this, we propose a token pruning cluster (TPC) that dynamically selects a few representative tokens with high semantic diversity while eliminating the redundancy of video frames. In addition, we develop a token recovering attention (TRA) to restore the detailed spatio-temporal information based on the selected tokens, thereby expanding the network output to the original full-length temporal resolution for fast inference. Extensive experiments on two benchmark datasets (i.e., Human3.6M and MPI-INF-3DHP) demonstrate that our method can achieve both high efficiency and estimation accuracy compared to the original VPT models. For instance, applying to MotionBERT and MixSTE on Human3.6M, our HoT can save nearly 50% FLOPs without sacrificing accuracy and nearly 40% FLOPs with only 0.2% accuracy drop, respectively. Code and models are available at https://github.com/NationalGAILab/HoT.

Prompt tuning with large-scale pretrained vision-language models empowers open-vocabulary predictions trained on limited base categories, e.g., object classification and detection. In this paper, we propose compositional prompt tuning with motion cues: an extended prompt tuning paradigm for compositional predictions of video data. In particular, we present Relation Prompt (RePro) for Open-vocabulary Video Visual Relation Detection (Open-VidVRD), where conventional prompt tuning is easily biased to certain subject-object combinations and motion patterns. To this end, RePro addresses the two technical challenges of Open-VidVRD: 1) the prompt tokens should respect the two different semantic roles of subject and object, and 2) the tuning should account for the diverse spatio-temporal motion patterns of the subject-object compositions. Without bells and whistles, our RePro achieves a new state-of-the-art performance on two VidVRD benchmarks of not only the base training object and predicate categories, but also the unseen ones. Extensive ablations also demonstrate the effectiveness of the proposed compositional and multi-mode design of prompts. Code is available at https://github.com/Dawn-LX/OpenVoc-VidVRD.

Episodic memory -- the ability to recall specific events grounded in time and space -- is a cornerstone of human cognition, enabling not only coherent storytelling, but also planning and decision-making. Despite their remarkable capabilities, Large Language Models (LLMs) lack a robust mechanism for episodic memory: we argue that integrating episodic memory capabilities into LLM is essential for advancing AI towards human-like cognition, increasing their potential to reason consistently and ground their output in real-world episodic events, hence avoiding confabulations. To address this challenge, we introduce a comprehensive framework to model and evaluate LLM episodic memory capabilities. Drawing inspiration from cognitive science, we develop a structured approach to represent episodic events, encapsulating temporal and spatial contexts, involved entities, and detailed descriptions. We synthesize a unique episodic memory benchmark, free from contamination, and release open source code and datasets to assess LLM performance across various recall and episodic reasoning tasks. Our evaluation of state-of-the-art models, including GPT-4 and Claude variants, Llama 3.1, and o1-mini, reveals that even the most advanced LLMs struggle with episodic memory tasks, particularly when dealing with multiple related events or complex spatio-temporal relationships -- even in contexts as short as 10k-100k tokens.

Most text-video retrieval methods utilize the text-image pre-trained models like CLIP as a backbone. These methods process each sampled frame independently by the image encoder, resulting in high computational overhead and limiting practical deployment. Addressing this, we focus on efficient text-video retrieval by tackling two key challenges: 1. From the perspective of trainable parameters, current parameter-efficient fine-tuning methods incur high inference costs; 2. From the perspective of model complexity, current token compression methods are mainly designed for images to reduce spatial redundancy but overlook temporal redundancy in consecutive frames of a video. To tackle these challenges, we propose Temporal Token Merging (TempMe), a parameter-efficient and training-inference efficient text-video retrieval architecture that minimizes trainable parameters and model complexity. Specifically, we introduce a progressive multi-granularity framework. By gradually combining neighboring clips, we reduce spatio-temporal redundancy and enhance temporal modeling across different frames, leading to improved efficiency and performance. Extensive experiments validate the superiority of our TempMe. Compared to previous parameter-efficient text-video retrieval methods, TempMe achieves superior performance with just 0.50M trainable parameters. It significantly reduces output tokens by 95% and GFLOPs by 51%, while achieving a 1.8X speedup and a 4.4% R-Sum improvement. With full fine-tuning, TempMe achieves a significant 7.9% R-Sum improvement, trains 1.57X faster, and utilizes 75.2% GPU memory usage. The code is available at https://github.com/LunarShen/TempMe.

Predicting human mobility is inherently challenging due to complex long-range dependencies and multi-scale periodic behaviors. To address this, we introduce RHYTHM (Reasoning with Hierarchical Temporal Tokenization for Human Mobility), a unified framework that leverages large language models (LLMs) as general-purpose spatio-temporal predictors and trajectory reasoners. Methodologically, RHYTHM employs temporal tokenization to partition each trajectory into daily segments and encode them as discrete tokens with hierarchical attention that captures both daily and weekly dependencies, thereby significantly reducing the sequence length while preserving cyclical information. Additionally, we enrich token representations by adding pre-computed prompt embeddings for trajectory segments and prediction targets via a frozen LLM, and feeding these combined embeddings back into the LLM backbone to capture complex interdependencies. Computationally, RHYTHM freezes the pretrained LLM's backbone to reduce attention complexity and memory cost. We evaluate our model against state-of-the-art methods using three real-world datasets. Notably, RHYTHM achieves a 2.4% improvement in overall accuracy, a 5.0% increase on weekends, and a 24.6% reduction in training time. Code is publicly available at https://github.com/he-h/rhythm.

In this paper, we present LLaVA-Scissor, a training-free token compression strategy designed for video multimodal large language models. Previous methods mostly attempt to compress tokens based on attention scores, but fail to effectively capture all semantic regions and often lead to token redundancy. Differently, we propose to leverage the Semantic Connected Components (SCC) approach that assigns tokens to distinct semantic regions within the token set, ensuring comprehensive semantic coverage. The outcome is a two-step spatio-temporal token compression strategy that utilizes SCC in both spatial and temporal domains. This strategy can effectively compress tokens by representing the entire video with a set of non-overlapping semantic tokens. We conduct extensive evaluations of the token compression capabilities of LLaVA-Scissor across diverse video understanding benchmarks, including video question answering, long video understanding, and comprehensive multi-choices benchmarks. Experimental results show that the proposed LLaVA-Scissor outperforms other token compression methods, achieving superior performance in various video understanding benchmarks, particularly at low token retention ratios. Project page: https://github.com/HumanMLLM/LLaVA-Scissor.

Existing methods of privacy-preserving action recognition (PPAR) mainly focus on frame-level (spatial) privacy removal through 2D CNNs. Unfortunately, they have two major drawbacks. First, they may compromise temporal dynamics in input videos, which are critical for accurate action recognition. Second, they are vulnerable to practical attacking scenarios where attackers probe for privacy from an entire video rather than individual frames. To address these issues, we propose a novel framework STPrivacy to perform video-level PPAR. For the first time, we introduce vision Transformers into PPAR by treating a video as a tubelet sequence, and accordingly design two complementary mechanisms, i.e., sparsification and anonymization, to remove privacy from a spatio-temporal perspective. In specific, our privacy sparsification mechanism applies adaptive token selection to abandon action-irrelevant tubelets. Then, our anonymization mechanism implicitly manipulates the remaining action-tubelets to erase privacy in the embedding space through adversarial learning. These mechanisms provide significant advantages in terms of privacy preservation for human eyes and action-privacy trade-off adjustment during deployment. We additionally contribute the first two large-scale PPAR benchmarks, VP-HMDB51 and VP-UCF101, to the community. Extensive evaluations on them, as well as two other tasks, validate the effectiveness and generalization capability of our framework.

We present TUMTraffic-VideoQA, a novel dataset and benchmark designed for spatio-temporal video understanding in complex roadside traffic scenarios. The dataset comprises 1,000 videos, featuring 85,000 multiple-choice QA pairs, 2,300 object captioning, and 5,700 object grounding annotations, encompassing diverse real-world conditions such as adverse weather and traffic anomalies. By incorporating tuple-based spatio-temporal object expressions, TUMTraffic-VideoQA unifies three essential tasks-multiple-choice video question answering, referred object captioning, and spatio-temporal object grounding-within a cohesive evaluation framework. We further introduce the TUMTraffic-Qwen baseline model, enhanced with visual token sampling strategies, providing valuable insights into the challenges of fine-grained spatio-temporal reasoning. Extensive experiments demonstrate the dataset's complexity, highlight the limitations of existing models, and position TUMTraffic-VideoQA as a robust foundation for advancing research in intelligent transportation systems. The dataset and benchmark are publicly available to facilitate further exploration.

Traffic forecasting represents a crucial problem within intelligent transportation systems. In recent research, Large Language Models (LLMs) have emerged as a promising method, but their intrinsic design, tailored primarily for sequential token processing, introduces notable challenges in effectively capturing spatial dependencies. Specifically, the inherent limitations of LLMs in modeling spatial relationships and their architectural incompatibility with graph-structured spatial data remain largely unaddressed. To overcome these limitations, we introduce ST-LINK, a novel framework that enhances the capability of Large Language Models to capture spatio-temporal dependencies. Its key components are Spatially-Enhanced Attention (SE-Attention) and the Memory Retrieval Feed-Forward Network (MRFFN). SE-Attention extends rotary position embeddings to integrate spatial correlations as direct rotational transformations within the attention mechanism. This approach maximizes spatial learning while preserving the LLM's inherent sequential processing structure. Meanwhile, MRFFN dynamically retrieves and utilizes key historical patterns to capture complex temporal dependencies and improve the stability of long-term forecasting. Comprehensive experiments on benchmark datasets demonstrate that ST-LINK surpasses conventional deep learning and LLM approaches, and effectively captures both regular traffic patterns and abrupt changes.

Existing video tokenizers typically use the traditional Variational Autoencoder (VAE) architecture for video compression and reconstruction. However, to achieve good performance, its training process often relies on complex multi-stage training tricks that go beyond basic reconstruction loss and KL regularization. Among these tricks, the most challenging is the precise tuning of adversarial training with additional Generative Adversarial Networks (GANs) in the final stage, which can hinder stable convergence. In contrast to GANs, diffusion models offer more stable training processes and can generate higher-quality results. Inspired by these advantages, we propose CDT, a novel Conditioned Diffusion-based video Tokenizer, that replaces the GAN-based decoder with a conditional causal diffusion model. The encoder compresses spatio-temporal information into compact latents, while the decoder reconstructs videos through a reverse diffusion process conditioned on these latents. During inference, we incorporate a feature cache mechanism to generate videos of arbitrary length while maintaining temporal continuity and adopt sampling acceleration technique to enhance efficiency. Trained using only a basic MSE diffusion loss for reconstruction, along with KL term and LPIPS perceptual loss from scratch, extensive experiments demonstrate that CDT achieves state-of-the-art performance in video reconstruction tasks with just a single-step sampling. Even a scaled-down version of CDT (3times inference speedup) still performs comparably with top baselines. Moreover, the latent video generation model trained with CDT also exhibits superior performance. The source code and pretrained weights will be released shortly, so please stay tuned for updates!

Video large language models (video LLMs) excel at video comprehension but face significant computational inefficiency due to redundant video tokens. Existing token pruning methods offer solutions. However, approaches operating within the LLM (inner-LLM pruning), such as FastV, incur intrinsic computational overhead in shallow layers. In contrast, methods performing token pruning before the LLM (outer-LLM pruning) primarily address spatial redundancy within individual frames or limited temporal windows, neglecting the crucial global temporal dynamics and correlations across longer video sequences. This leads to sub-optimal spatio-temporal reduction and does not leverage video compressibility fully. Crucially, the synergistic potential and mutual influence of combining these strategies remain unexplored. To further reduce redundancy, we introduce HoliTom, a novel training-free holistic token merging framework. HoliTom employs outer-LLM pruning through global redundancy-aware temporal segmentation, followed by spatial-temporal merging to reduce visual tokens by over 90%, significantly alleviating the LLM's computational burden. Complementing this, we introduce a robust inner-LLM token similarity-based merging approach, designed for superior performance and compatibility with outer-LLM pruning. Evaluations demonstrate our method's promising efficiency-performance trade-off on LLaVA-OneVision-7B, reducing computational costs to 6.9% of FLOPs while maintaining 99.1% of the original performance. Furthermore, we achieve a 2.28x reduction in Time-To-First-Token (TTFT) and a 1.32x acceleration in decoding throughput, highlighting the practical benefits of our integrated pruning approach for efficient video LLMs inference.

Recent advancements in text-to-video (T2V) diffusion models have enabled high-fidelity and realistic video synthesis. However, current T2V models often struggle to generate physically plausible content due to their limited inherent ability to accurately understand physics. We found that while the representations within T2V models possess some capacity for physics understanding, they lag significantly behind those from recent video self-supervised learning methods. To this end, we propose a novel framework called VideoREPA, which distills physics understanding capability from video understanding foundation models into T2V models by aligning token-level relations. This closes the physics understanding gap and enable more physics-plausible generation. Specifically, we introduce the Token Relation Distillation (TRD) loss, leveraging spatio-temporal alignment to provide soft guidance suitable for finetuning powerful pre-trained T2V models, a critical departure from prior representation alignment (REPA) methods. To our knowledge, VideoREPA is the first REPA method designed for finetuning T2V models and specifically for injecting physical knowledge. Empirical evaluations show that VideoREPA substantially enhances the physics commonsense of baseline method, CogVideoX, achieving significant improvement on relevant benchmarks and demonstrating a strong capacity for generating videos consistent with intuitive physics. More video results are available at https://videorepa.github.io/.

Masked Video Autoencoder (MVA) approaches have demonstrated their potential by significantly outperforming previous video representation learning methods. However, they waste an excessive amount of computations and memory in predicting uninformative tokens/frames due to random masking strategies. (e.g., over 16 nodes with 128 NVIDIA A100 GPUs). To resolve this issue, we exploit the unequal information density among the patches in videos and propose EVEREST, a surprisingly efficient MVA approach for video representation learning that finds tokens containing rich motion features and discards uninformative ones during both pre-training and fine-tuning. We further present an information-intensive frame selection strategy that allows the model to focus on informative and causal frames with minimal redundancy. Our method significantly reduces the computation and memory requirements of MVA, enabling the pre-training and fine-tuning on a single machine with 8 GPUs while achieving comparable performance to computation- and memory-heavy baselines on multiple benchmarks and the uncurated Ego4D dataset. We hope that our work contributes to reducing the barrier to further research on video understanding.

Recognizing interactive action plays an important role in human-robot interaction and collaboration. Previous methods use late fusion and co-attention mechanism to capture interactive relations, which have limited learning capability or inefficiency to adapt to more interacting entities. With assumption that priors of each entity are already known, they also lack evaluations on a more general setting addressing the diversity of subjects. To address these problems, we propose an Interactive Spatiotemporal Token Attention Network (ISTA-Net), which simultaneously model spatial, temporal, and interactive relations. Specifically, our network contains a tokenizer to partition Interactive Spatiotemporal Tokens (ISTs), which is a unified way to represent motions of multiple diverse entities. By extending the entity dimension, ISTs provide better interactive representations. To jointly learn along three dimensions in ISTs, multi-head self-attention blocks integrated with 3D convolutions are designed to capture inter-token correlations. When modeling correlations, a strict entity ordering is usually irrelevant for recognizing interactive actions. To this end, Entity Rearrangement is proposed to eliminate the orderliness in ISTs for interchangeable entities. Extensive experiments on four datasets verify the effectiveness of ISTA-Net by outperforming state-of-the-art methods. Our code is publicly available at https://github.com/Necolizer/ISTA-Net

To bridge the physical and virtual worlds for rapidly developed VR/AR applications, the ability to realistically drive 3D full-body avatars is of great significance. Although real-time body tracking with only the head-mounted displays (HMDs) and hand controllers is heavily under-constrained, a carefully designed end-to-end neural network is of great potential to solve the problem by learning from large-scale motion data. To this end, we propose a two-stage framework that can obtain accurate and smooth full-body motions with the three tracking signals of head and hands only. Our framework explicitly models the joint-level features in the first stage and utilizes them as spatiotemporal tokens for alternating spatial and temporal transformer blocks to capture joint-level correlations in the second stage. Furthermore, we design a set of loss terms to constrain the task of a high degree of freedom, such that we can exploit the potential of our joint-level modeling. With extensive experiments on the AMASS motion dataset and real-captured data, we validate the effectiveness of our designs and show our proposed method can achieve more accurate and smooth motion compared to existing approaches.

Continuous-space video generation has advanced rapidly, while discrete approaches lag behind due to error accumulation and long-context inconsistency. In this work, we revisit discrete generative modeling and present Uniform discRete diffuSion with metric pAth (URSA), a simple yet powerful framework that bridges the gap with continuous approaches for the scalable video generation. At its core, URSA formulates the video generation task as an iterative global refinement of discrete spatiotemporal tokens. It integrates two key designs: a Linearized Metric Path and a Resolution-dependent Timestep Shifting mechanism. These designs enable URSA to scale efficiently to high-resolution image synthesis and long-duration video generation, while requiring significantly fewer inference steps. Additionally, we introduce an asynchronous temporal fine-tuning strategy that unifies versatile tasks within a single model, including interpolation and image-to-video generation. Extensive experiments on challenging video and image generation benchmarks demonstrate that URSA consistently outperforms existing discrete methods and achieves performance comparable to state-of-the-art continuous diffusion methods. Code and models are available at https://github.com/baaivision/URSA

Video Question Answering (VQA) in long videos poses the key challenge of extracting relevant information and modeling long-range dependencies from many redundant frames. The self-attention mechanism provides a general solution for sequence modeling, but it has a prohibitive cost when applied to a massive number of spatiotemporal tokens in long videos. Most prior methods rely on compression strategies to lower the computational cost, such as reducing the input length via sparse frame sampling or compressing the output sequence passed to the large language model (LLM) via space-time pooling. However, these naive approaches over-represent redundant information and often miss salient events or fast-occurring space-time patterns. In this work, we introduce BIMBA, an efficient state-space model to handle long-form videos. Our model leverages the selective scan algorithm to learn to effectively select critical information from high-dimensional video and transform it into a reduced token sequence for efficient LLM processing. Extensive experiments demonstrate that BIMBA achieves state-of-the-art accuracy on multiple long-form VQA benchmarks, including PerceptionTest, NExT-QA, EgoSchema, VNBench, LongVideoBench, and Video-MME. Code, and models are publicly available at https://sites.google.com/view/bimba-mllm.

Masked Autoencoders (MAEs) learn generalizable representations for image, text, audio, video, etc., by reconstructing masked input data from tokens of the visible data. Current MAE approaches for videos rely on random patch, tube, or frame-based masking strategies to select these tokens. This paper proposes AdaMAE, an adaptive masking strategy for MAEs that is end-to-end trainable. Our adaptive masking strategy samples visible tokens based on the semantic context using an auxiliary sampling network. This network estimates a categorical distribution over spacetime-patch tokens. The tokens that increase the expected reconstruction error are rewarded and selected as visible tokens, motivated by the policy gradient algorithm in reinforcement learning. We show that AdaMAE samples more tokens from the high spatiotemporal information regions, thereby allowing us to mask 95% of tokens, resulting in lower memory requirements and faster pre-training. We conduct ablation studies on the Something-Something v2 (SSv2) dataset to demonstrate the efficacy of our adaptive sampling approach and report state-of-the-art results of 70.0% and 81.7% in top-1 accuracy on SSv2 and Kinetics-400 action classification datasets with a ViT-Base backbone and 800 pre-training epochs.

Vision-Language-Action (VLA) models built on pretrained Vision-Language Models (VLMs) show strong potential but are limited in practicality due to their large parameter counts. To mitigate this issue, using a lightweight VLM has been explored, but it compromises spatiotemporal reasoning. Although some methods suggest that incorporating additional 3D inputs can help, they usually rely on large VLMs to fuse 3D and 2D inputs and still lack temporal understanding. Therefore, we propose SwiftVLA, an architecture that enhances a compact model with 4D understanding while preserving design efficiency. Specifically, our approach features a pretrained 4D visual geometry transformer with a temporal cache that extracts 4D features from 2D images. Then, to enhance the VLM's ability to exploit both 2D images and 4D features, we introduce Fusion Tokens, a set of learnable tokens trained with a future prediction objective to generate unified representations for action generation. Finally, we introduce a mask-and-reconstruct strategy that masks 4D inputs to the VLM and trains the VLA to reconstruct them, enabling the VLM to learn effective 4D representations and allowing the 4D branch to be dropped at inference with minimal performance loss. Experiments in real and simulated environments show that SwiftVLA outperforms lightweight baselines and rivals VLAs up to 7 times larger, achieving comparable performance on edge devices while being 18 times faster and reducing memory footprint by 12 times.

Multimodal Large Language Models (MLLMs) have shown promising progress in understanding and analyzing video content. However, processing long videos remains a significant challenge constrained by LLM's context size. To address this limitation, we propose LongVU, a spatiotemporal adaptive compression mechanism thats reduces the number of video tokens while preserving visual details of long videos. Our idea is based on leveraging cross-modal query and inter-frame dependencies to adaptively reduce temporal and spatial redundancy in videos. Specifically, we leverage DINOv2 features to remove redundant frames that exhibit high similarity. Then we utilize text-guided cross-modal query for selective frame feature reduction. Further, we perform spatial token reduction across frames based on their temporal dependencies. Our adaptive compression strategy effectively processes a large number of frames with little visual information loss within given context length. Our LongVU consistently surpass existing methods across a variety of video understanding benchmarks, especially on hour-long video understanding tasks such as VideoMME and MLVU. Given a light-weight LLM, our LongVU also scales effectively into a smaller size with state-of-the-art video understanding performance.

Learning discriminative spatiotemporal representation is the key problem of video understanding. Recently, Vision Transformers (ViTs) have shown their power in learning long-term video dependency with self-attention. Unfortunately, they exhibit limitations in tackling local video redundancy, due to the blind global comparison among tokens. UniFormer has successfully alleviated this issue, by unifying convolution and self-attention as a relation aggregator in the transformer format. However, this model has to require a tiresome and complicated image-pretraining phrase, before being finetuned on videos. This blocks its wide usage in practice. On the contrary, open-sourced ViTs are readily available and well-pretrained with rich image supervision. Based on these observations, we propose a generic paradigm to build a powerful family of video networks, by arming the pretrained ViTs with efficient UniFormer designs. We call this family UniFormerV2, since it inherits the concise style of the UniFormer block. But it contains brand-new local and global relation aggregators, which allow for preferable accuracy-computation balance by seamlessly integrating advantages from both ViTs and UniFormer. Without any bells and whistles, our UniFormerV2 gets the state-of-the-art recognition performance on 8 popular video benchmarks, including scene-related Kinetics-400/600/700 and Moments in Time, temporal-related Something-Something V1/V2, untrimmed ActivityNet and HACS. In particular, it is the first model to achieve 90% top-1 accuracy on Kinetics-400, to our best knowledge. Code will be available at https://github.com/OpenGVLab/UniFormerV2.

It is a challenging task to learn rich and multi-scale spatiotemporal semantics from high-dimensional videos, due to large local redundancy and complex global dependency between video frames. The recent advances in this research have been mainly driven by 3D convolutional neural networks and vision transformers. Although 3D convolution can efficiently aggregate local context to suppress local redundancy from a small 3D neighborhood, it lacks the capability to capture global dependency because of the limited receptive field. Alternatively, vision transformers can effectively capture long-range dependency by self-attention mechanism, while having the limitation on reducing local redundancy with blind similarity comparison among all the tokens in each layer. Based on these observations, we propose a novel Unified transFormer (UniFormer) which seamlessly integrates merits of 3D convolution and spatiotemporal self-attention in a concise transformer format, and achieves a preferable balance between computation and accuracy. Different from traditional transformers, our relation aggregator can tackle both spatiotemporal redundancy and dependency, by learning local and global token affinity respectively in shallow and deep layers. We conduct extensive experiments on the popular video benchmarks, e.g., Kinetics-400, Kinetics-600, and Something-Something V1&V2. With only ImageNet-1K pretraining, our UniFormer achieves 82.9%/84.8% top-1 accuracy on Kinetics-400/Kinetics-600, while requiring 10x fewer GFLOPs than other state-of-the-art methods. For Something-Something V1 and V2, our UniFormer achieves new state-of-the-art performances of 60.9% and 71.2% top-1 accuracy respectively. Code is available at https://github.com/Sense-X/UniFormer.

Effectively modeling long spatiotemporal sequences is challenging due to the need to model complex spatial correlations and long-range temporal dependencies simultaneously. ConvLSTMs attempt to address this by updating tensor-valued states with recurrent neural networks, but their sequential computation makes them slow to train. In contrast, Transformers can process an entire spatiotemporal sequence, compressed into tokens, in parallel. However, the cost of attention scales quadratically in length, limiting their scalability to longer sequences. Here, we address the challenges of prior methods and introduce convolutional state space models (ConvSSM) that combine the tensor modeling ideas of ConvLSTM with the long sequence modeling approaches of state space methods such as S4 and S5. First, we demonstrate how parallel scans can be applied to convolutional recurrences to achieve subquadratic parallelization and fast autoregressive generation. We then establish an equivalence between the dynamics of ConvSSMs and SSMs, which motivates parameterization and initialization strategies for modeling long-range dependencies. The result is ConvS5, an efficient ConvSSM variant for long-range spatiotemporal modeling. ConvS5 significantly outperforms Transformers and ConvLSTM on a long horizon Moving-MNIST experiment while training 3X faster than ConvLSTM and generating samples 400X faster than Transformers. In addition, ConvS5 matches or exceeds the performance of state-of-the-art methods on challenging DMLab, Minecraft and Habitat prediction benchmarks and enables new directions for modeling long spatiotemporal sequences.

Weather nowcasting is an essential task that involves predicting future radar echo sequences based on current observations, offering significant benefits for disaster management, transportation, and urban planning. Current prediction methods are limited by training and storage efficiency, mainly focusing on 2D spatial predictions at specific altitudes. Meanwhile, 3D volumetric predictions at each timestamp remain largely unexplored. To address such a challenge, we introduce a comprehensive framework for 3D radar sequence prediction in weather nowcasting, using the newly proposed SpatioTemporal Coherent Gaussian Splatting (STC-GS) for dynamic radar representation and GauMamba for efficient and accurate forecasting. Specifically, rather than relying on a 4D Gaussian for dynamic scene reconstruction, STC-GS optimizes 3D scenes at each frame by employing a group of Gaussians while effectively capturing their movements across consecutive frames. It ensures consistent tracking of each Gaussian over time, making it particularly effective for prediction tasks. With the temporally correlated Gaussian groups established, we utilize them to train GauMamba, which integrates a memory mechanism into the Mamba framework. This allows the model to learn the temporal evolution of Gaussian groups while efficiently handling a large volume of Gaussian tokens. As a result, it achieves both efficiency and accuracy in forecasting a wide range of dynamic meteorological radar signals. The experimental results demonstrate that our STC-GS can efficiently represent 3D radar sequences with over 16times higher spatial resolution compared with the existing 3D representation methods, while GauMamba outperforms state-of-the-art methods in forecasting a broad spectrum of high-dynamic weather conditions.

Video restoration aims to reconstruct high quality video sequences from low quality inputs, addressing tasks such as super resolution, denoising, and deblurring. Traditional regression based methods often produce unrealistic details and require extensive paired datasets, while recent generative diffusion models face challenges in ensuring temporal consistency. We introduce DiTVR, a zero shot video restoration framework that couples a diffusion transformer with trajectory aware attention and a wavelet guided, flow consistent sampler. Unlike prior 3D convolutional or frame wise diffusion approaches, our attention mechanism aligns tokens along optical flow trajectories, with particular emphasis on vital layers that exhibit the highest sensitivity to temporal dynamics. A spatiotemporal neighbour cache dynamically selects relevant tokens based on motion correspondences across frames. The flow guided sampler injects data consistency only into low-frequency bands, preserving high frequency priors while accelerating convergence. DiTVR establishes a new zero shot state of the art on video restoration benchmarks, demonstrating superior temporal consistency and detail preservation while remaining robust to flow noise and occlusions.

Large-scale video-language pre-training has made remarkable strides in advancing video-language understanding tasks. However, the heavy computational burden of video encoding remains a formidable efficiency bottleneck, particularly for long-form videos. These videos contain massive visual tokens due to their inherent 3D properties and spatiotemporal redundancy, making it challenging to capture complex temporal and spatial relationships. To tackle this issue, we propose an efficient method called TEmporal-Spatial Token Aggregation (TESTA). TESTA condenses video semantics by adaptively aggregating similar frames, as well as similar patches within each frame. TESTA can reduce the number of visual tokens by 75% and thus accelerate video encoding. Building upon TESTA, we introduce a pre-trained video-language model equipped with a divided space-time token aggregation module in each video encoder block. We evaluate our model on five datasets for paragraph-to-video retrieval and long-form VideoQA tasks. Experimental results show that TESTA improves computing efficiency by 1.7 times, and achieves significant performance gains from its scalability in processing longer input frames, e.g., +13.7 R@1 on QuerYD and +6.5 R@1 on Condensed Movie.

Current video generation techniques excel at single-shot clips but struggle to produce narrative multi-shot videos, which require flexible shot arrangement, coherent narrative, and controllability beyond text prompts. To tackle these challenges, we propose MultiShotMaster, a framework for highly controllable multi-shot video generation. We extend a pretrained single-shot model by integrating two novel variants of RoPE. First, we introduce Multi-Shot Narrative RoPE, which applies explicit phase shift at shot transitions, enabling flexible shot arrangement while preserving the temporal narrative order. Second, we design Spatiotemporal Position-Aware RoPE to incorporate reference tokens and grounding signals, enabling spatiotemporal-grounded reference injection. In addition, to overcome data scarcity, we establish an automated data annotation pipeline to extract multi-shot videos, captions, cross-shot grounding signals and reference images. Our framework leverages the intrinsic architectural properties to support multi-shot video generation, featuring text-driven inter-shot consistency, customized subject with motion control, and background-driven customized scene. Both shot count and duration are flexibly configurable. Extensive experiments demonstrate the superior performance and outstanding controllability of our framework.

An important challenge in vision-based action recognition is the embedding of spatiotemporal features with two or more heterogeneous modalities into a single feature. In this study, we propose a new 3D deformable transformer for action recognition with adaptive spatiotemporal receptive fields and a cross-modal learning scheme. The 3D deformable transformer consists of three attention modules: 3D deformability, local joint stride, and temporal stride attention. The two cross-modal tokens are input into the 3D deformable attention module to create a cross-attention token with a reflected spatiotemporal correlation. Local joint stride attention is applied to spatially combine attention and pose tokens. Temporal stride attention temporally reduces the number of input tokens in the attention module and supports temporal expression learning without the simultaneous use of all tokens. The deformable transformer iterates L-times and combines the last cross-modal token for classification. The proposed 3D deformable transformer was tested on the NTU60, NTU120, FineGYM, and PennAction datasets, and showed results better than or similar to pre-trained state-of-the-art methods even without a pre-training process. In addition, by visualizing important joints and correlations during action recognition through spatial joint and temporal stride attention, the possibility of achieving an explainable potential for action recognition is presented.

With the development of large language models (LLMs), efficient inference through Key-Value (KV) cache compression has attracted considerable attention, especially for long-context generation. To compress the KV cache, recent methods identify critical KV tokens through heuristic ranking with attention scores. However, these methods often struggle to accurately determine critical tokens as they neglect the temporal patterns in attention scores, resulting in a noticeable degradation in LLM performance. To address this challenge, we propose AttentionPredictor, which is the first learning-based critical token identification approach. Specifically, AttentionPredictor learns a lightweight convolution model to capture spatiotemporal patterns and predict the next-token attention score. An appealing feature of AttentionPredictor is that it accurately predicts the attention score while consuming negligible memory. Moreover, we propose a cross-token critical cache prefetching framework that hides the token estimation time overhead to accelerate the decoding stage. By retaining most of the attention information, AttentionPredictor achieves 16times KV cache compression with comparable LLM performance, significantly outperforming the state-of-the-art.

Encoding videos into discrete tokens could align with text tokens to facilitate concise and unified multi-modal LLMs, yet introducing significant spatiotemporal compression compared to continuous video representation. Previous discrete video VAEs experienced unstable training, long training time, and degraded reconstruction quality. Given the easier training and superior performance of continuous VAEs, an intuitive idea is to enhance discrete video VAEs by leveraging continuous VAEs. After rethinking the intrinsic link between discrete and continuous representations, we found that FSQ could effectively preserve pre-trained continuous VAE priors compared to other quantization methods. By leveraging continuous VAE priors, it converges several times faster than training from scratch and achieves superior performance at convergence. Meanwhile, two structural improvements are proposed. First, inspired by how continuous VAEs enhance reconstruction via enlarged latent dimensions, we introduce a multi-token quantization mechanism, which achieves nearly a 1 dB improvement in PSNR without compromising the token compression ratio. Second, to tackle reconstruction challenges in high-compression video VAEs, we strengthen first-frame reconstruction, enabling the causal VAE to leverage this information in subsequent frames and markedly improving the performance of 4 x 16 x 16 discrete VAEs. Furthermore, we propose a joint discrete-continuous optimization scheme that unifies the two paradigms and, for the first time, achieves competitive performance on both continuous and discrete representations within a single network. We name our method OneVAE to reflect this connection.

Popular video training methods mainly operate on a fixed number of tokens sampled from a predetermined spatiotemporal grid, resulting in sub-optimal accuracy-computation trade-offs due to inherent video redundancy. They also lack adaptability to varying computational budgets for downstream tasks, hindering applications of the most competitive model in real-world scenes. We thus propose a new test setting, Token Optimization, for maximized input information across budgets, which optimizes the size-limited set of input tokens through token selection from more suitably sampled videos. To this end, we propose a novel augmentation tool termed Flux. By making the sampling grid flexible and leveraging token selection, it is easily adopted in most popular video training frameworks, boosting model robustness with nearly no additional cost. We integrate Flux in large-scale video pre-training, and the resulting FluxViT establishes new state-of-the-art results across extensive tasks at standard costs. Notably, with 1/4 tokens only, it can still match the performance of previous state-of-the-art models with Token Optimization, yielding nearly 90\% savings. All models and data are available at https://github.com/OpenGVLab/FluxViT.

Streaming video clips with large-scale video tokens impede vision transformers (ViTs) for efficient recognition, especially in video action detection where sufficient spatiotemporal representations are required for precise actor identification. In this work, we propose an end-to-end framework for efficient video action detection (EVAD) based on vanilla ViTs. Our EVAD consists of two specialized designs for video action detection. First, we propose a spatiotemporal token dropout from a keyframe-centric perspective. In a video clip, we maintain all tokens from its keyframe, preserve tokens relevant to actor motions from other frames, and drop out the remaining tokens in this clip. Second, we refine scene context by leveraging remaining tokens for better recognizing actor identities. The region of interest (RoI) in our action detector is expanded into temporal domain. The captured spatiotemporal actor identity representations are refined via scene context in a decoder with the attention mechanism. These two designs make our EVAD efficient while maintaining accuracy, which is validated on three benchmark datasets (i.e., AVA, UCF101-24, JHMDB). Compared to the vanilla ViT backbone, our EVAD reduces the overall GFLOPs by 43% and improves real-time inference speed by 40% with no performance degradation. Moreover, even at similar computational costs, our EVAD can improve the performance by 1.1 mAP with higher resolution inputs. Code is available at https://github.com/MCG-NJU/EVAD.

Recent advances in diffusion models have enabled high-quality video generation, but the additional temporal dimension significantly increases computational costs, making training and inference on long videos prohibitively expensive. In this paper, we identify a phenomenon we term Spatiotemporal Energy Decay in video diffusion models: post-softmax attention scores diminish as spatial and temporal distance between tokens increase, akin to the physical decay of signal or waves over space and time in nature. Motivated by this, we propose Radial Attention, a scalable sparse attention mechanism with O(n log n) complexity that translates energy decay into exponentially decaying compute density, which is significantly more efficient than standard O(n^2) dense attention and more expressive than linear attention. Specifically, Radial Attention employs a simple, static attention mask where each token attends to spatially nearby tokens, with the attention window size shrinking with temporal distance. Moreover, it allows pre-trained video diffusion models to extend their generation length with efficient LoRA-based fine-tuning. Extensive experiments show that Radial Attention maintains video quality across Wan2.1-14B, HunyuanVideo, and Mochi 1, achieving up to a 1.9times speedup over the original dense attention. With minimal tuning, it enables video generation up to 4times longer while reducing training costs by up to 4.4times compared to direct fine-tuning and accelerating inference by up to 3.7times compared to dense attention inference.

Effective modeling of complex spatiotemporal dependencies in long-form videos remains an open problem. The recently proposed Structured State-Space Sequence (S4) model with its linear complexity offers a promising direction in this space. However, we demonstrate that treating all image-tokens equally as done by S4 model can adversely affect its efficiency and accuracy. To address this limitation, we present a novel Selective S4 (i.e., S5) model that employs a lightweight mask generator to adaptively select informative image tokens resulting in more efficient and accurate modeling of long-term spatiotemporal dependencies in videos. Unlike previous mask-based token reduction methods used in transformers, our S5 model avoids the dense self-attention calculation by making use of the guidance of the momentum-updated S4 model. This enables our model to efficiently discard less informative tokens and adapt to various long-form video understanding tasks more effectively. However, as is the case for most token reduction methods, the informative image tokens could be dropped incorrectly. To improve the robustness and the temporal horizon of our model, we propose a novel long-short masked contrastive learning (LSMCL) approach that enables our model to predict longer temporal context using shorter input videos. We present extensive comparative results using three challenging long-form video understanding datasets (LVU, COIN and Breakfast), demonstrating that our approach consistently outperforms the previous state-of-the-art S4 model by up to 9.6% accuracy while reducing its memory footprint by 23%.

Existing methods for Video Reasoning Segmentation rely heavily on a single special token to represent the object in the keyframe or the entire video, inadequately capturing spatial complexity and inter-frame motion. To overcome these challenges, we propose VRS-HQ, an end-to-end video reasoning segmentation approach that leverages Multimodal Large Language Models (MLLMs) to inject rich spatiotemporal features into hierarchical tokens.Our key innovations include a Temporal Dynamic Aggregation (TDA) and a Token-driven Keyframe Selection (TKS). Specifically, we design frame-level <SEG> and temporal-level <TAK> tokens that utilize MLLM's autoregressive learning to effectively capture both local and global information. Subsequently, we apply a similarity-based weighted fusion and frame selection strategy, then utilize SAM2 to perform keyframe segmentation and propagation. To enhance keyframe localization accuracy, the TKS filters keyframes based on SAM2's occlusion scores during inference. VRS-HQ achieves state-of-the-art performance on ReVOS, surpassing VISA by 5.9%/12.5%/9.1% in J&F scores across the three subsets. These results highlight the strong temporal reasoning and segmentation capabilities of our method. Code and model weights will be released at VRS-HQ.