The efficiency of attention is critical because its time complexity grows quadratically with sequence length. SageAttention2 addresses this by utilizing quantization to accelerate matrix multiplications (Matmul) in attention. To further accelerate SageAttention2, we propose to utilize the faster instruction of FP8 Matmul accumulated in FP16. The instruction is 2x faster than the FP8 Matmul used in SageAttention2. Our experiments show that SageAttention2++ achieves a 3.9x speedup over FlashAttention while maintaining the same attention accuracy as SageAttention2. This means SageAttention2++ effectively accelerates various models, including those for language, image, and video generation, with negligible end-to-end metrics loss. The code will be available at https://github.com/thu-ml/SageAttention.
Animating images with interactive motion control has garnered popularity for image-to-video (I2V) generation. Modern approaches typically rely on large Gaussian kernels to extend motion trajectories as condition without explicitly defining movement region, leading to coarse motion control and failing to disentangle object and camera moving. To alleviate these, we present MotionPro, a precise motion controller that novelly leverages region-wise trajectory and motion mask to regulate fine-grained motion synthesis and identify target motion category (i.e., object or camera moving), respectively. Technically, MotionPro first estimates the flow maps on each training video via a tracking model, and then samples the region-wise trajectories to simulate inference scenario. Instead of extending flow through large Gaussian kernels, our region-wise trajectory approach enables more precise control by directly utilizing trajectories within local regions, thereby effectively characterizing fine-grained movements. A motion mask is simultaneously derived from the predicted flow maps to capture the holistic motion dynamics of the movement regions. To pursue natural motion control, MotionPro further strengthens video denoising by incorporating both region-wise trajectories and motion mask through feature modulation. More remarkably, we meticulously construct a benchmark, i.e., MC-Bench, with 1.1K user-annotated image-trajectory pairs, for the evaluation of both fine-grained and object-level I2V motion control. Extensive experiments conducted on WebVid-10M and MC-Bench demonstrate the effectiveness of MotionPro. Please refer to our project page for more results: https://zhw-zhang.github.io/MotionPro-page/.
Camera sensor simulation serves as a critical role for autonomous driving (AD), e.g. evaluating vision-based AD algorithms. While existing approaches have leveraged generative models for controllable image/video generation, they remain constrained to generating multi-view video sequences with fixed camera viewpoints and video frequency, significantly limiting their downstream applications. To address this, we present a generalizable camera simulation framework DriveCamSim, whose core innovation lies in the proposed Explicit Camera Modeling (ECM) mechanism. Instead of implicit interaction through vanilla attention, ECM establishes explicit pixel-wise correspondences across multi-view and multi-frame dimensions, decoupling the model from overfitting to the specific camera configurations (intrinsic/extrinsic parameters, number of views) and temporal sampling rates presented in the training data. For controllable generation, we identify the issue of information loss inherent in existing conditional encoding and injection pipelines, proposing an information-preserving control mechanism. This control mechanism not only improves conditional controllability, but also can be extended to be identity-aware to enhance temporal consistency in foreground object rendering. With above designs, our model demonstrates superior performance in both visual quality and controllability, as well as generalization capability across spatial-level (camera parameters variations) and temporal-level (video frame rate variations), enabling flexible user-customizable camera simulation tailored to diverse application scenarios. Code will be avaliable at https://github.com/swc-17/DriveCamSim for facilitating future research.
Video action understanding tasks in real-world scenarios always suffer data limitations. In this paper, we address the data-limited action understanding problem by bridging data scarcity. We propose a novel method that employs a text-to-video diffusion transformer to generate annotated data for model training. This paradigm enables the generation of realistic annotated data on an infinite scale without human intervention. We proposed the information enhancement strategy and the uncertainty-based label smoothing tailored to generate sample training. Through quantitative and qualitative analysis, we observed that real samples generally contain a richer level of information than generated samples. Based on this observation, the information enhancement strategy is proposed to enhance the informative content of the generated samples from two aspects: the environments and the characters. Furthermore, we observed that some low-quality generated samples might negatively affect model training. To address this, we devised the uncertainty-based label smoothing strategy to increase the smoothing of these samples, thus reducing their impact. We demonstrate the effectiveness of the proposed method on four datasets across five tasks and achieve state-of-the-art performance for zero-shot action recognition.
Recent advances in video generation models have sparked interest in world models capable of simulating realistic environments. While navigation has been well-explored, physically meaningful interactions that mimic real-world forces remain largely understudied. In this work, we investigate using physical forces as a control signal for video generation and propose force prompts which enable users to interact with images through both localized point forces, such as poking a plant, and global wind force fields, such as wind blowing on fabric. We demonstrate that these force prompts can enable videos to respond realistically to physical control signals by leveraging the visual and motion prior in the original pretrained model, without using any 3D asset or physics simulator at inference. The primary challenge of force prompting is the difficulty in obtaining high quality paired force-video training data, both in the real world due to the difficulty of obtaining force signals, and in synthetic data due to limitations in the visual quality and domain diversity of physics simulators. Our key finding is that video generation models can generalize remarkably well when adapted to follow physical force conditioning from videos synthesized by Blender, even with limited demonstrations of few objects. Our method can generate videos which simulate forces across diverse geometries, settings, and materials. We also try to understand the source of this generalization and perform ablations that reveal two key elements: visual diversity and the use of specific text keywords during training. Our approach is trained on only around 15k training examples for a single day on four A100 GPUs, and outperforms existing methods on force adherence and physics realism, bringing world models closer to real-world physics interactions. We release all datasets, code, weights, and interactive video demos at our project page.
Diffusion Transformers (DiTs) are essential for video generation but suffer from significant latency due to the quadratic complexity of attention. By computing only critical tokens, sparse attention reduces computational costs and offers a promising acceleration approach. However, we identify that existing methods fail to approach optimal generation quality under the same computation budget for two reasons: (1) Inaccurate critical token identification: current methods cluster tokens based on position rather than semantics, leading to imprecise aggregated representations. (2) Excessive computation waste: critical tokens are scattered among non-critical ones, leading to wasted computation on GPUs, which are optimized for processing contiguous tokens. In this paper, we propose SVG2, a training-free framework that maximizes identification accuracy and minimizes computation waste, achieving a Pareto frontier trade-off between generation quality and efficiency. The core of SVG2 is semantic-aware permutation, which clusters and reorders tokens based on semantic similarity using k-means. This approach ensures both a precise cluster representation, improving identification accuracy, and a densified layout of critical tokens, enabling efficient computation without padding. Additionally, SVG2 integrates top-p dynamic budget control and customized kernel implementations, achieving up to 2.30x and 1.89x speedup while maintaining a PSNR of up to 30 and 26 on HunyuanVideo and Wan 2.1, respectively.
Video Diffusion Transformers (VDiTs) have achieved remarkable progress in high-quality video generation, but remain computationally expensive due to the quadratic complexity of attention over high-dimensional video sequences. Recent attention acceleration methods leverage the sparsity of attention patterns to improve efficiency; however, they often overlook inefficiencies of redundant long-range interactions. To address this problem, we propose \textbf{VORTA}, an acceleration framework with two novel components: 1) a sparse attention mechanism that efficiently captures long-range dependencies, and 2) a routing strategy that adaptively replaces full 3D attention with specialized sparse attention variants throughout the sampling process. It achieves a $1.76\times$ end-to-end speedup without quality loss on VBench. Furthermore, VORTA can seamlessly integrate with various other acceleration methods, such as caching and step distillation, reaching up to $14.41\times$ speedup with negligible performance degradation. VORTA demonstrates its efficiency and enhances the practicality of VDiTs in real-world settings.
Diffusion Transformers (DiTs) have emerged as the state-of-the-art architecture for video generation, yet their computational and memory demands hinder practical deployment. While post-training quantization (PTQ) presents a promising approach to accelerate Video DiT models, existing methods suffer from two critical limitations: (1) dependence on lengthy, computation-heavy calibration procedures, and (2) considerable performance deterioration after quantization. To address these challenges, we propose DVD-Quant, a novel Data-free quantization framework for Video DiTs. Our approach integrates three key innovations: (1) Progressive Bounded Quantization (PBQ) and (2) Auto-scaling Rotated Quantization (ARQ) for calibration data-free quantization error reduction, as well as (3) $\delta$-Guided Bit Switching ($\delta$-GBS) for adaptive bit-width allocation. Extensive experiments across multiple video generation benchmarks demonstrate that DVD-Quant achieves an approximately 2$\times$ speedup over full-precision baselines on HunyuanVideo while maintaining visual fidelity. Notably, DVD-Quant is the first to enable W4A4 PTQ for Video DiTs without compromising video quality. Code and models will be available at https://github.com/lhxcs/DVD-Quant.
Real-world driving requires people to observe the current environment, anticipate the future, and make appropriate driving decisions. This requirement is aligned well with the capabilities of world models, which understand the environment and predict the future. However, recent world models in autonomous driving are built explicitly, where they could predict the future by controllable driving video generation. We argue that driving world models should have two additional abilities: action control and action prediction. Following this line, previous methods are limited because they predict the video requires given actions of the same length as the video and ignore the dynamical action laws. To address these issues, we propose ProphetDWM, a novel end-to-end driving world model that jointly predicts future videos and actions. Our world model has an action module to learn latent action from the present to the future period by giving the action sequence and observations. And a diffusion-model-based transition module to learn the state distribution. The model is jointly trained by learning latent actions given finite states and predicting action and video. The joint learning connects the action dynamics and states and enables long-term future prediction. We evaluate our method in video generation and action prediction tasks on the Nuscenes dataset. Compared to the state-of-the-art methods, our method achieves the best video consistency and best action prediction accuracy, while also enabling high-quality long-term video and action generation.
The field of robotics has made significant strides toward developing generalist robot manipulation policies. However, evaluating these policies in real-world scenarios remains time-consuming and challenging, particularly as the number of tasks scales and environmental conditions change. In this work, we demonstrate that world models can serve as a scalable, reproducible, and reliable proxy for real-world robot policy evaluation. A key challenge is generating accurate policy videos from world models that faithfully reflect the robot actions. We observe that directly inputting robot actions or using high-dimensional encoding methods often fails to generate action-following videos. To address this, we propose Policy2Vec, a simple yet effective approach to turn a video generation model into a world simulator that follows latent action to generate the robot video. We then introduce WorldEval, an automated pipeline designed to evaluate real-world robot policies entirely online. WorldEval effectively ranks various robot policies and individual checkpoints within a single policy, and functions as a safety detector to prevent dangerous actions by newly developed robot models. Through comprehensive paired evaluations of manipulation policies in real-world environments, we demonstrate a strong correlation between policy performance in WorldEval and real-world scenarios. Furthermore, our method significantly outperforms popular methods such as real-to-sim approach.
Autonomous robots typically need to construct representations of their surroundings and adapt their motions to the geometry of their environment. Here, we tackle the problem of constructing a policy model for collision-free motion generation, consistent with the environment, from a single input RGB image. Extracting 3D structures from a single image often involves monocular depth estimation. Developments in depth estimation have given rise to large pre-trained models such as DepthAnything. However, using outputs of these models for downstream motion generation is challenging due to frustum-shaped errors that arise. Instead, we propose a framework known as Video-Generation Environment Representation (VGER), which leverages the advances of large-scale video generation models to generate a moving camera video conditioned on the input image. Frames of this video, which form a multiview dataset, are then input into a pre-trained 3D foundation model to produce a dense point cloud. We then introduce a multi-scale noise approach to train an implicit representation of the environment structure and build a motion generation model that complies with the geometry of the representation. We extensively evaluate VGER over a diverse set of indoor and outdoor environments. We demonstrate its ability to produce smooth motions that account for the captured geometry of a scene, all from a single RGB input image.
Leveraging the diffusion transformer (DiT) architecture, models like Sora, CogVideoX and Wan have achieved remarkable progress in text-to-video, image-to-video, and video editing tasks. Despite these advances, diffusion-based video generation remains computationally intensive, especially for high-resolution, long-duration videos. Prior work accelerates its inference by skipping computation, usually at the cost of severe quality degradation. In this paper, we propose SRDiffusion, a novel framework that leverages collaboration between large and small models to reduce inference cost. The large model handles high-noise steps to ensure semantic and motion fidelity (Sketching), while the smaller model refines visual details in low-noise steps (Rendering). Experimental results demonstrate that our method outperforms existing approaches, over 3$\times$ speedup for Wan with nearly no quality loss for VBench, and 2$\times$ speedup for CogVideoX. Our method is introduced as a new direction orthogonal to existing acceleration strategies, offering a practical solution for scalable video generation.
WonderPlay is a novel framework integrating physics simulation with video generation for generating action-conditioned dynamic 3D scenes from a single image. While prior works are restricted to rigid body or simple elastic dynamics, WonderPlay features a hybrid generative simulator to synthesize a wide range of 3D dynamics. The hybrid generative simulator first uses a physics solver to simulate coarse 3D dynamics, which subsequently conditions a video generator to produce a video with finer, more realistic motion. The generated video is then used to update the simulated dynamic 3D scene, closing the loop between the physics solver and the video generator. This approach enables intuitive user control to be combined with the accurate dynamics of physics-based simulators and the expressivity of diffusion-based video generators. Experimental results demonstrate that WonderPlay enables users to interact with various scenes of diverse content, including cloth, sand, snow, liquid, smoke, elastic, and rigid bodies -- all using a single image input. Code will be made public. Project website: https://kyleleey.github.io/WonderPlay/
Subject-to-Video (S2V) generation aims to create videos that faithfully incorporate reference content, providing enhanced flexibility in the production of videos. To establish the infrastructure for S2V generation, we propose OpenS2V-Nexus, consisting of (i) OpenS2V-Eval, a fine-grained benchmark, and (ii) OpenS2V-5M, a million-scale dataset. In contrast to existing S2V benchmarks inherited from VBench that focus on global and coarse-grained assessment of generated videos, OpenS2V-Eval focuses on the model's ability to generate subject-consistent videos with natural subject appearance and identity fidelity. For these purposes, OpenS2V-Eval introduces 180 prompts from seven major categories of S2V, which incorporate both real and synthetic test data. Furthermore, to accurately align human preferences with S2V benchmarks, we propose three automatic metrics, NexusScore, NaturalScore and GmeScore, to separately quantify subject consistency, naturalness, and text relevance in generated videos. Building on this, we conduct a comprehensive evaluation of 16 representative S2V models, highlighting their strengths and weaknesses across different content. Moreover, we create the first open-source large-scale S2V generation dataset OpenS2V-5M, which consists of five million high-quality 720P subject-text-video triples. Specifically, we ensure subject-information diversity in our dataset by (1) segmenting subjects and building pairing information via cross-video associations and (2) prompting GPT-Image-1 on raw frames to synthesize multi-view representations. Through OpenS2V-Nexus, we deliver a robust infrastructure to accelerate future S2V generation research.
Recent advancements in image-to-video (I2V) generation have shown promising performance in conventional scenarios. However, these methods still encounter significant challenges when dealing with complex scenes that require a deep understanding of nuanced motion and intricate object-action relationships. To address these challenges, we present Dynamic-I2V, an innovative framework that integrates Multimodal Large Language Models (MLLMs) to jointly encode visual and textual conditions for a diffusion transformer (DiT) architecture. By leveraging the advanced multimodal understanding capabilities of MLLMs, our model significantly improves motion controllability and temporal coherence in synthesized videos. The inherent multimodality of Dynamic-I2V further enables flexible support for diverse conditional inputs, extending its applicability to various downstream generation tasks. Through systematic analysis, we identify a critical limitation in current I2V benchmarks: a significant bias towards favoring low-dynamic videos, stemming from an inadequate balance between motion complexity and visual quality metrics. To resolve this evaluation gap, we propose DIVE - a novel assessment benchmark specifically designed for comprehensive dynamic quality measurement in I2V generation. In conclusion, extensive quantitative and qualitative experiments confirm that Dynamic-I2V attains state-of-the-art performance in image-to-video generation, particularly revealing significant improvements of 42.5%, 7.9%, and 11.8% in dynamic range, controllability, and quality, respectively, as assessed by the DIVE metric in comparison to existing methods.
Despite the remarkable generation quality of video Diffusion Transformer (DiT) models, their practical deployment is severely hindered by extensive computational requirements. This inefficiency stems from two key challenges: the quadratic complexity of self-attention with respect to token length and the multi-step nature of diffusion models. To address these limitations, we present Jenga, a novel inference pipeline that combines dynamic attention carving with progressive resolution generation. Our approach leverages two key insights: (1) early denoising steps do not require high-resolution latents, and (2) later steps do not require dense attention. Jenga introduces a block-wise attention mechanism that dynamically selects relevant token interactions using 3D space-filling curves, alongside a progressive resolution strategy that gradually increases latent resolution during generation. Experimental results demonstrate that Jenga achieves substantial speedups across multiple state-of-the-art video diffusion models while maintaining comparable generation quality (8.83$\times$ speedup with 0.01\% performance drop on VBench). As a plug-and-play solution, Jenga enables practical, high-quality video generation on modern hardware by reducing inference time from minutes to seconds -- without requiring model retraining. Code: https://github.com/dvlab-research/Jenga
Recent advances in scene-based video generation have enabled systems to synthesize coherent visual narratives from structured prompts. However, a crucial dimension of storytelling -- character-driven dialogue and speech -- remains underexplored. In this paper, we present a modular pipeline that transforms action-level prompts into visually and auditorily grounded narrative dialogue, enriching visual storytelling with natural voice and character expression. Our method takes as input a pair of prompts per scene, where the first defines the setting and the second specifies a character's behavior. While a story generation model such as Text2Story generates the corresponding visual scene, we focus on generating expressive character utterances from these prompts and the scene image. We apply a pretrained vision-language encoder to extract a high-level semantic feature from the representative frame, capturing salient visual context. This feature is then combined with the structured prompts and used to guide a large language model in synthesizing natural, character-consistent dialogue. To ensure contextual consistency across scenes, we introduce a Recursive Narrative Bank that conditions each dialogue generation on the accumulated dialogue history from prior scenes. This approach enables characters to speak in ways that reflect their evolving goals and interactions throughout a story. Finally, we render each utterance as expressive, character-consistent speech, resulting in fully-voiced video narratives. Our framework requires no additional training and demonstrates applicability across a variety of story settings, from fantasy adventures to slice-of-life episodes.
Recent advances in static 3D generation have intensified the demand for physically consistent dynamic 3D content. However, existing video generation models, including diffusion-based methods, often prioritize visual realism while neglecting physical plausibility, resulting in implausible object dynamics. Prior approaches for physics-aware dynamic generation typically rely on large-scale annotated datasets or extensive model fine-tuning, which imposes significant computational and data collection burdens and limits scalability across scenarios. To address these challenges, we present MAGIC, a training-free framework for single-image physical property inference and dynamic generation, integrating pretrained image-to-video diffusion models with iterative LLM-based reasoning. Our framework generates motion-rich videos from a static image and closes the visual-to-physical gap through a confidence-driven LLM feedback loop that adaptively steers the diffusion model toward physics-relevant motion. To translate visual dynamics into controllable physical behavior, we further introduce a differentiable MPM simulator operating directly on 3D Gaussians reconstructed from the single image, enabling physically grounded, simulation-ready outputs without any supervision or model tuning. Experiments show that MAGIC outperforms existing physics-aware generative methods in inference accuracy and achieves greater temporal coherence than state-of-the-art video diffusion models.
Generating photorealistic driving videos has seen significant progress recently, but current methods largely focus on ordinary, non-adversarial scenarios. Meanwhile, efforts to generate adversarial driving scenarios often operate on abstract trajectory or BEV representations, falling short of delivering realistic sensor data that can truly stress-test autonomous driving (AD) systems. In this work, we introduce Challenger, a framework that produces physically plausible yet photorealistic adversarial driving videos. Generating such videos poses a fundamental challenge: it requires jointly optimizing over the space of traffic interactions and high-fidelity sensor observations. Challenger makes this affordable through two techniques: (1) a physics-aware multi-round trajectory refinement process that narrows down candidate adversarial maneuvers, and (2) a tailored trajectory scoring function that encourages realistic yet adversarial behavior while maintaining compatibility with downstream video synthesis. As tested on the nuScenes dataset, Challenger generates a diverse range of aggressive driving scenarios-including cut-ins, sudden lane changes, tailgating, and blind spot intrusions-and renders them into multiview photorealistic videos. Extensive evaluations show that these scenarios significantly increase the collision rate of state-of-the-art end-to-end AD models (UniAD, VAD, SparseDrive, and DiffusionDrive), and importantly, adversarial behaviors discovered for one model often transfer to others.
Generative AI (GenAI) is rapidly advancing the field of Autonomous Driving (AD), extending beyond traditional applications in text, image, and video generation. We explore how generative models can enhance automotive tasks, such as static map creation, dynamic scenario generation, trajectory forecasting, and vehicle motion planning. By examining multiple generative approaches ranging from Variational Autoencoder (VAEs) over Generative Adversarial Networks (GANs) and Invertible Neural Networks (INNs) to Generative Transformers (GTs) and Diffusion Models (DMs), we highlight and compare their capabilities and limitations for AD-specific applications. Additionally, we discuss hybrid methods integrating conventional techniques with generative approaches, and emphasize their improved adaptability and robustness. We also identify relevant datasets and outline open research questions to guide future developments in GenAI. Finally, we discuss three core challenges: safety, interpretability, and realtime capabilities, and present recommendations for image generation, dynamic scenario generation, and planning.