Video Diffusion Models

Video object removal aims to erase target objects while reconstructing visually plausible and temporally coherent content. However, target objects often induce shadows, reflections, illumination changes and other effects that extend beyond the provided mask, making conventional mask-conditioned completion prone to visible residuals. We therefore formulate side-effect-aware object removal as an understanding-guided process that integrates object--effect relations, affected-region localization, and context-aware reconstruction. Specifically, we introduce Object-Induced Relation Distillation to transfer token-level object--effect relations from a pretrained vision foundation model to the video diffusion model. We then design Object-aware Framewise Context Cross-Attention to combine target-object semantics with per-frame background context for removal and reconstruction, and propose Attention-guided Region Localization to derive a soft spatial prior over the target object and its affected regions. Extensive experiments across multiple benchmarks demonstrate that our method achieves more complete object-and-effect removal and outperforms existing approaches in removal quality, side-effect suppression, and temporal consistency.

Despite recent progress, video diffusion models still struggle to synthesize realistic videos involving highly dynamic motions or requiring fine-grained motion controllability. A central limitation lies in the scarcity of such examples in commonly used training datasets. To address this, we introduce DynaVid, a video synthesis framework that leverages synthetic motion data in training, which is represented as optical flow and rendered using computer graphics pipelines. This approach offers two key advantages. First, synthetic motion offers diverse motion patterns and precise control signals that are difficult to obtain from real data. Second, unlike rendered videos with artificial appearances, rendered optical flow encodes only motion and is decoupled from appearance, thereby preventing models from reproducing the unnatural look of synthetic videos. Building on this idea, DynaVid adopts a two-stage generation framework: a motion generator first synthesizes motion, and then a motion-guided video generator produces video frames conditioned on that motion. This decoupled formulation enables the model to learn dynamic motion patterns from synthetic data while preserving visual realism from real-world videos. We validate our framework on two challenging scenarios, vigorous human motion generation and extreme camera motion control, where existing datasets are particularly limited. Extensive experiments demonstrate that DynaVid improves the realism and controllability in dynamic motion generation and camera motion control.

Motion transfer enables controllable video generation by transferring temporal dynamics from a reference video to synthesize a new video conditioned on a target caption. However, existing Diffusion Transformer (DiT)-based methods are limited to single-object videos, restricting fine-grained control in real-world scenes with multiple objects. In this work, we introduce MotionGrounder, a DiT-based framework that firstly handles motion transfer with multi-object controllability. Our Flow-based Motion Signal (FMS) in MotionGrounder provides a stable motion prior for target video generation, while our Object-Caption Alignment Loss (OCAL) grounds object captions to their corresponding spatial regions. We further propose a new Object Grounding Score (OGS), which jointly evaluates (i) spatial alignment between source video objects and their generated counterparts and (ii) semantic consistency between each generated object and its target caption. Our experiments show that MotionGrounder consistently outperforms recent baselines across quantitative, qualitative, and human evaluations.

Video diffusion models exhibit emergent reasoning capabilities like solving mazes and puzzles, yet little is understood about how they reason during generation. We take a first step towards understanding this and study the internal planning dynamics of video models using 2D maze solving as a controlled testbed. Our investigations reveal two findings. Our first finding is early plan commitment: video diffusion models commit to a high-level motion plan within the first few denoising steps, after which further denoising alters visual details but not the underlying trajectory. Our second finding is that path length, not obstacle density, is the dominant predictor of maze difficulty, with a sharp failure threshold at 12 steps. This means video models can only reason over long mazes by chaining together multiple sequential generations. To demonstrate the practical benefits of our findings, we introduce Chaining with Early Planning, or ChEaP, which only spends compute on seeds with promising early plans and chains them together to tackle complex mazes. This improves accuracy from 7% to 67% on long-horizon mazes and by 2.5x overall on hard tasks in Frozen Lake and VR-Bench across Wan2.2-14B and HunyuanVideo-1.5. Our analysis reveals that current video models possess deeper reasoning capabilities than previously recognized, which can be elicited more reliably with better inference-time scaling.

Recent methods have made notable progress in the visual quality of hand-object interaction video synthesis. However, most approaches rely on 2D control signals that lack spatial expressiveness and limit the utilization of synthetic 3D conditional data. To address these limitations, we propose HVG-3D, a unified framework for 3D-aware hand-object interaction (HOI) video synthesis conditioned on explicit 3D representations. Specifically, we develop a diffusion-based architecture augmented with a 3D ControlNet, which encodes geometric and motion cues from 3D inputs to enable explicit 3D reasoning during video synthesis. To achieve high-quality synthesis, HVG-3D is designed with two core components: (i) a 3D-aware HOI video generation diffusion architecture that encodes geometric and motion cues from 3D inputs for explicit 3D reasoning; and (ii) a hybrid pipeline for constructing input and condition signals, enabling flexible and precise control during both training and inference. During inference, given a single real image and a 3D control signal from either simulation or real data, HVG-3D generates high-fidelity, temporally consistent videos with precise spatial and temporal control. Experiments on the TASTE-Rob dataset demonstrate that HVG-3D achieves state-of-the-art spatial fidelity, temporal coherence, and controllability, while enabling effective utilization of both real and simulated data.

Egocentric "walking tour" videos provide a rich source of image data to develop rich and diverse visual models of environments around the world. However, the significant presence of humans in frames of these videos due to crowds and eye-level camera perspectives mitigates their usefulness in environment modeling applications. We focus on addressing this challenge by developing a generative algorithm that can realistically remove (i.e., inpaint) humans and their associated shadow effects from walking tour videos. Key to our approach is the construction of a rich semi-synthetic dataset of video clip pairs to train this generative model. Each pair in the dataset consists of an environment-only background clip, and a composite clip of walking humans with simulated shadows overlaid on the background. We randomly sourced both foreground and background components from real egocentric walking tour videos around the world to maintain visual diversity. We then used this dataset to fine-tune the state-of-the-art Casper video diffusion model for object and effects inpainting, and demonstrate that the resulting model performs far better than Casper both qualitatively and quantitatively at removing humans from walking tour clips with significant human presence and complex backgrounds. Finally, we show that the resulting generated clips can be used to build successful 3D/4D models of urban locations.

Sign language plays a crucial role in bridging communication gaps between the deaf and hard-of-hearing communities. However, existing sign language video generation models often rely on complex intermediate representations, which limits their flexibility and efficiency. In this work, we propose a novel pose-free framework for real-time sign language video generation. Our method eliminates the need for intermediate pose representations by directly mapping natural language text to sign language videos using a diffusion-based approach. We introduce two key innovations: (1) a pose-free generative model based on the a state-of-the-art diffusion backbone, which learns implicit text-to-gesture alignments without pose estimation, and (2) a Trainable Sliding Tile Attention (T-STA) mechanism that accelerates inference by exploiting spatio-temporal locality patterns. Unlike previous training-free sparsity approaches, T-STA integrates trainable sparsity into both training and inference, ensuring consistency and eliminating the train-test gap. This approach significantly reduces computational overhead while maintaining high generation quality, making real-time deployment feasible. Our method increases video generation speed by 3.07x without compromising video quality. Our contributions open new avenues for real-time, high-quality, pose-free sign language synthesis, with potential applications in inclusive communication tools for diverse communities. Code: https://github.com/AIGeeksGroup/FlashSign.

In video diffusion transformers, visual patch tokens maintain explicit correspondence to space and time. We hypothesize that their channel dimension can serve as a semantic control space, which we call Visual Dial Space V+. In this space, additive directions can be broadcast to the token stream to control appearance or motion attributes, enabling slider-style edits such as making a generated person look older or run faster. To verify the hypothesis, we present TokenDial, a framework for learning attribute directions in the proposed V+ space. TokenDial keeps the pretrained video generator frozen and optimizes only additive directions. Rather than requiring paired edited videos, TokenDial supervises each direction through its induced effect on generated videos: the edited video should move along the desired attribute while the remaining content stays stable. The learned directions become reusable visual dials that support continuous appearance and motion control, explicit spatiotemporal localization, composition, and reuse across prompts, resolutions, and video lengths. Experiments and human studies show that TokenDial achieves stronger slider controllability and better content preservation than prior video editing and slider-based methods.

Video prediction is a useful function for autonomous driving, enabling intelligent vehicles to reliably anticipate how driving scenes will evolve and thereby supporting reasoning and safer planning. However, existing models are constrained by multi-stage training pipelines and remain insufficient in modeling the diverse motion patterns in real driving scenes, leading to degraded temporal consistency and visual quality. To address these challenges, this paper introduces the historical motion priors-informed diffusion model (HMPDM), a video prediction model that leverages historical motion priors to enhance motion understanding and temporal coherence. The proposed deep learning system introduces three key designs: (i) a Temporal-aware Latent Conditioning (TaLC) module for implicit historical motion injection; (ii) a Motion-aware Pyramid Encoder (MaPE) for multi-scale motion representation; (iii) a Self-Conditioning (SC) strategy for stable iterative denoising. Extensive experiments on the Cityscapes and KITTI benchmarks demonstrate that HMPDM outperforms state-of-the-art video prediction methods with efficiency, achieving a 28.2% improvement in FVD on Cityscapes under the same monocular RGB input configuration setting. The implementation codes are publicly available at https://github.com/KELISBU/HMPDM.

Scaling video diffusion transformers is fundamentally bottlenecked by two compounding costs: the expensive quadratic complexity of attention per step, and the iterative sampling steps. In this work, we propose EFlow, an efficient few-step training framework, that tackles these bottlenecks simultaneously. To reduce sampling steps, we build on a solution-flow objective that learns a function mapping a noised state at time t to time s. Making this formulation computationally feasible and high-quality at video scale, however, demands two complementary innovations. First, we propose Gated Local-Global Attention, a token-droppable hybrid block which is efficient, expressive, and remains highly stable under aggressive random token-dropping, substantially reducing per-step compute. Second, we develop an efficient few-step training recipe. We propose Path-Drop Guided training to replace the expensive guidance target with a computationally cheap, weak path. Furthermore, we augment this with a Mean-Velocity Additivity regularizer to ensure high fidelity at extremely low step counts. Together, our EFlow enables a practical from-scratch training pipeline, achieving up to 2.5x higher training throughput over standard solution-flow, and 45.3x lower inference latency than standard iterative models with competitive performance on Kinetics and large-scale text-to-video datasets.

Large-scale video diffusion models achieve impressive visual quality, yet often fail to preserve geometric consistency. Prior approaches improve consistency either by augmenting the generator with additional modules or applying geometry-aware alignment. However, architectural modifications can compromise the generalization of internet-scale pretrained models, while existing alignment methods are limited to static scenes and rely on RGB-space rewards that require repeated VAE decoding, incurring substantial compute overhead and failing to generalize to highly dynamic real-world scenes. To preserve the pretrained capacity while improving geometric consistency, we propose VGGRPO (Visual Geometry GRPO), a latent geometry-guided framework for geometry-aware video post-training. VGGRPO introduces a Latent Geometry Model (LGM) that stitches video diffusion latents to geometry foundation models, enabling direct decoding of scene geometry from the latent space. By constructing LGM from a geometry model with 4D reconstruction capability, VGGRPO naturally extends to dynamic scenes, overcoming the static-scene limitations of prior methods. Building on this, we perform latent-space Group Relative Policy Optimization with two complementary rewards: a camera motion smoothness reward that penalizes jittery trajectories, and a geometry reprojection consistency reward that enforces cross-view geometric coherence. Experiments on both static and dynamic benchmarks show that VGGRPO improves camera stability, geometry consistency, and overall quality while eliminating costly VAE decoding, making latent-space geometry-guided reinforcement an efficient and flexible approach to world-consistent video generation.

Flow-matching video generators produce temporally coherent, high-fidelity outputs yet routinely violate elementary physics because their reconstruction objectives penalize per-frame deviations without distinguishing physically consistent dynamics from impossible ones. Contrastive flow matching offers a principled remedy by pushing apart velocity-field trajectories of differing conditions, but we identify a fundamental obstacle in the text-conditioned video setting: semantic-physics entanglement. Because natural-language prompts couple scene content with physical behavior, naive negative sampling draws conditions whose velocity fields largely overlap with the positive sample's, causing the contrastive gradient to directly oppose the flow-matching objective. We formalize this gradient conflict, deriving a precise alignment condition that reveals when contrastive learning helps versus harms training. Guided by this analysis, we introduce DiReCT (Disentangled Regularization of Contrastive Trajectories), a lightweight post-training framework that decomposes the contrastive signal into two complementary scales: a macro-contrastive term that draws partition-exclusive negatives from semantically distant regions for interference-free global trajectory separation, and a micro-contrastive term that constructs hard negatives sharing full scene semantics with the positive sample but differing along a single, LLM-perturbed axis of physical behavior; spanning kinematics, forces, materials, interactions, and magnitudes. A velocity-space distributional regularizer helps to prevent catastrophic forgetting of pretrained visual quality. When applied to Wan 2.1-1.3B, our method improves the physical commonsense score on VideoPhy by 16.7% and 11.3% compared to the baseline and SFT, respectively, without increasing training time.

We present THFM, a unified video foundation model for human-centric perception that jointly addresses dense tasks (depth, normals, segmentation, dense pose) and sparse tasks (2d/3d keypoint estimation) within a single architecture. THFM is derived from a pretrained text-to-video diffusion model, repurposed as a single-forward-pass perception model and augmented with learnable tokens for sparse predictions. Modulated by the text prompt, our single unified model is capable of performing various perception tasks. Crucially, our model is on-par or surpassing state-of-the-art specialized models on a variety of benchmarks despite being trained exclusively on synthetic data (i.e.~without training on real-world or benchmark specific data). We further highlight intriguing emergent properties of our model, which we attribute to the underlying diffusion-based video representation. For example, our model trained on videos with a single human in the scene generalizes to multiple humans and other object classes such as anthropomorphic characters and animals -- a capability that hasn't been demonstrated in the past.

Reference-to-video (R2V) generation is a controllable video synthesis paradigm that constrains the generation process using both text prompts and reference images, enabling applications such as personalized advertising and virtual try-on. In practice, existing R2V methods typically introduce additional high-level semantic or cross-modal features alongside the VAE latent representation of the reference image and jointly feed them into the diffusion Transformer (DiT). These auxiliary representations provide semantic guidance and act as implicit alignment signals, which can partially alleviate pixel-level information leakage in the VAE latent space. However, they may still struggle to address copy--paste artifacts and multi-subject confusion caused by modality mismatch across heterogeneous encoder features. In this paper, we propose RefAlign, a representation alignment framework that explicitly aligns DiT reference-branch features to the semantic space of a visual foundation model (VFM). The core of RefAlign is a reference alignment loss that pulls the reference features and VFM features of the same subject closer to improve identity consistency, while pushing apart the corresponding features of different subjects to enhance semantic discriminability. This simple yet effective strategy is applied only during training, incurring no inference-time overhead, and achieves a better balance between text controllability and reference fidelity. Extensive experiments on the OpenS2V-Eval benchmark demonstrate that RefAlign outperforms current state-of-the-art methods in TotalScore, validating the effectiveness of explicit reference alignment for R2V tasks.

Autoregressive video diffusion models have demonstrated remarkable progress, yet they remain bottlenecked by intractable linear KV-cache growth, temporal repetition, and compounding errors during long-video generation. To address these challenges, we present PackForcing, a unified framework that efficiently manages the generation history through a novel three-partition KV-cache strategy. Specifically, we categorize the historical context into three distinct types: (1) Sink tokens, which preserve early anchor frames at full resolution to maintain global semantics; (2) Mid tokens, which achieve a massive spatiotemporal compression (32x token reduction) via a dual-branch network fusing progressive 3D convolutions with low-resolution VAE re-encoding; and (3) Recent tokens, kept at full resolution to ensure local temporal coherence. To strictly bound the memory footprint without sacrificing quality, we introduce a dynamic top-$k$ context selection mechanism for the mid tokens, coupled with a continuous Temporal RoPE Adjustment that seamlessly re-aligns position gaps caused by dropped tokens with negligible overhead. Empowered by this principled hierarchical context compression, PackForcing can generate coherent 2-minute, 832x480 videos at 16 FPS on a single H200 GPU. It achieves a bounded KV cache of just 4 GB and enables a remarkable 24x temporal extrapolation (5s to 120s), operating effectively either zero-shot or trained on merely 5-second clips. Extensive results on VBench demonstrate state-of-the-art temporal consistency (26.07) and dynamic degree (56.25), proving that short-video supervision is sufficient for high-quality, long-video synthesis. https://github.com/ShandaAI/PackForcing

Recent advances in video generation models have achieved impressive results. However, these models heavily rely on the use of high-quality data that combines both high visual quality and high motion quality. In this paper, we identify a key challenge in video data curation: the Motion-Vision Quality Dilemma. We discovered that visual quality and motion intensity inherently exhibit a negative correlation, making it hard to obtain golden data that excels in both aspects. To address this challenge, we first examine the hierarchical learning dynamics of video diffusion models and conduct gradient-based analysis on quality-degraded samples. We discover that quality-imbalanced data can produce gradients similar to golden data at appropriate timesteps. Based on this, we introduce the novel concept of Timestep selection in Training Process. We propose Timestep-aware Quality Decoupling (TQD), which modifies the data sampling distribution to better match the model's learning process. For certain types of data, the sampling distribution is skewed toward higher timesteps for motion-rich data, while high visual quality data is more likely to be sampled during lower timesteps. Through extensive experiments, we demonstrate that TQD enables training exclusively on separated imbalanced data to achieve performance surpassing conventional training with better data, challenging the necessity of perfect data in video generation. Moreover, our method also boosts model performance when trained on high-quality data, showcasing its effectiveness across different data scenarios.

Recent progress in video-to-video (V2V) translation has enabled realistic resimulation of embodied AI demonstrations, a capability that allows pretrained robot policies to be transferable to new environments without additional data collection. However, prior works can only operate on a single view at a time, while embodied AI tasks are commonly captured from multiple synchronized cameras to support policy learning. Naively applying single-view models independently to each camera leads to inconsistent appearance across views, and standard transformer architectures do not scale to multi-view settings due to the quadratic cost of cross-view attention. We present VideoWeaver, the first multimodal multi-view V2V translation framework. VideoWeaver is initially trained as a single-view flow-based V2V model. To achieve an extension to the multi-view regime, we propose to ground all views in a shared 4D latent space derived from a feed-forward spatial foundation model, namely, Pi3. This encourages view-consistent appearance even under wide baselines and dynamic camera motion. To scale beyond a fixed number of cameras, we train views at distinct diffusion timesteps, enabling the model to learn both joint and conditional view distributions. This in turn allows autoregressive synthesis of new viewpoints conditioned on existing ones. Experiments show superior or similar performance to the state-of-the-art on the single-view translation benchmarks and, for the first time, physically and stylistically consistent multi-view translations, including challenging egocentric and heterogeneous-camera setups central to world randomization for robot learning.

We propose InstanceAnimator, a novel Diffusion Transformer framework for multi-instance sketch video colorization. Existing methods suffer from three core limitations: inflexible user control due to heavy reliance on single reference frames, poor instance controllability leading to misalignment in multi-character scenarios, and degraded detail fidelity in fine-grained regions. To address these challenges, we introduce three corresponding innovations. First, a Canvas Guidance Condition eliminates workflow fragmentation by allowing free placement of reference elements and background, enabling unprecedented user flexibility. Second, an Instance Matching Mechanism resolves misalignment by integrating instance features with the sketches, ensuring precise control over multiple characters. Third, an Adaptive Decoupled Control Module enhances detail fidelity by injecting semantic features from characters, backgrounds, and text conditions into the diffusion process. Extensive experiments demonstrate that InstanceAnimator achieves superior multi-instance colorization with enhanced user control, high visual quality, and strong instance consistency.

Generating long videos using pre-trained video diffusion models, which are typically trained on short clips, presents a significant challenge. Directly applying these models for long-video inference often leads to a notable degradation in visual quality. This paper identifies that this issue primarily stems from two out-of-distribution (O.O.D) problems: frame-level relative position O.O.D and context-length O.O.D. To address these challenges, we propose FreeLOC, a novel training-free, layer-adaptive framework that introduces two core techniques: Video-based Relative Position Re-encoding (VRPR) for frame-level relative position O.O.D, a multi-granularity strategy that hierarchically re-encodes temporal relative positions to align with the model's pre-trained distribution, and Tiered Sparse Attention (TSA) for context-length O.O.D, which preserves both local detail and long-range dependencies by structuring attention density across different temporal scales. Crucially, we introduce a layer-adaptive probing mechanism that identifies the sensitivity of each transformer layer to these O.O.D issues, allowing for the selective and efficient application of our methods. Extensive experiments demonstrate that our approach significantly outperforms existing training-free methods, achieving state-of-the-art results in both temporal consistency and visual quality. Code is available at https://github.com/Westlake-AGI-Lab/FreeLOC.

Long-trajectory video generation is a crucial yet challenging task for world modeling primarily due to the limited scalability of existing video diffusion models (VDMs). Autoregressive models, while offering infinite rollout, suffer from visual drift and poor controllability. To address these issues, we propose DCARL, a novel divide-and-conquer, autoregressive framework that effectively combines the structural stability of the divide-and-conquer scheme with the high-fidelity generation of VDMs. Our approach first employs a dedicated Keyframe Generator trained without temporal compression to establish long-range, globally consistent structural anchors. Subsequently, an Interpolation Generator synthesizes the dense frames in an autoregressive manner with overlapping segments, utilizing the keyframes for global context and a single clean preceding frame for local coherence. Trained on a large-scale internet long trajectory video dataset, our method achieves superior performance in both visual quality (lower FID and FVD) and camera adherence (lower ATE and ARE) compared to state-of-the-art autoregressive and divide-and-conquer baselines, demonstrating stable and high-fidelity generation for long trajectory videos up to 32 seconds in length.

Advances in diffusion-based video generation models, while significantly improving human animation, poses threats of misuse through the creation of fake videos from a specific person's photo and text prompts. Recent efforts have focused on adversarial attacks that introduce crafted perturbations to protect images from diffusion-based models. However, most existing approaches target image generation, while relatively few explicitly address image-to-video diffusion models (VDMs), and most primarily focus on UNet-based architectures. Hence, their effectiveness against Diffusion Transformer (DiT) models remains largely under-explored, as these models demonstrate improved feature retention, and stronger temporal consistency due to larger capacity and advanced attention mechanisms. In this work, we introduce Anti-I2V, a novel defense against malicious human image-to-video generation, applicable across diverse diffusion backbones. Instead of restricting noise updates to the RGB space, Anti-I2V operates in both the $L$$a$$b$* and frequency domains, improving robustness and concentrating on salient pixels. We then identify the network layers that capture the most distinct semantic features during the denoising process to design appropriate training objectives that maximize degradation of temporal coherence and generation fidelity. Through extensive validation, Anti-I2V demonstrates state-of-the-art defense performance against diverse video diffusion models, offering an effective solution to the problem.

Generative world models offer a compelling foundation for augmented-reality (AR) applications: by predicting future image sequences that incorporate deliberate visual edits, they enable temporally coherent, augmented future frames that can be computed ahead of time and cached, avoiding per-frame rendering from scratch in real time. In this work, we present SEGAR, a preliminary framework that combines a diffusion-based world model with a selective correction stage to support this vision. The world model generates augmented future frames with region-specific edits while preserving others, and the correction stage subsequently aligns safety-critical regions with real-world observations while preserving intended augmentations elsewhere. We demonstrate this pipeline in driving scenarios as a representative setting where semantic region structure is well defined and real-world feedback is readily available. We view this as an early step toward generative world models as practical AR infrastructure, where future frames can be generated, cached, and selectively corrected on demand.

Diffusion-based video editing has emerged as an important paradigm for high-quality and flexible content generation. However, despite their generality and strong modeling capacity, Diffusion Transformers (DiT) remain computationally expensive due to the iterative denoising process, posing challenges for practical deployment. Existing video diffusion acceleration methods primarily exploit denoising timestep-level feature reuse, which mitigates the redundancy in denoising process, but overlooks the architectural redundancy within the DiT that many attention operations over spatio-temporal tokens are redundantly executed, offering little to no incremental contribution to the model output. This work introduces HetCache, a training-free diffusion acceleration framework designed to exploit the inherent heterogeneity in diffusion-based masked video-to-video (MV2V) generation and editing. Instead of uniformly reuse or randomly sampling tokens, HetCache assesses the contextual relevance and interaction strength among various types of tokens in designated computing steps. Guided by spatial priors, it divides the spatial-temporal tokens in DiT model into context and generative tokens, and selectively caches the context tokens that exhibit the strongest correlation and most representative semantics with generative ones. This strategy reduces redundant attention operations while maintaining editing consistency and fidelity. Experiments show that HetCache achieves a noticeable acceleration, including a 2.67$\times$ latency speedup and FLOPs reduction over commonly used foundation models, with negligible degradation in editing quality.

Diffusion and flow matching models have unlocked unprecedented capabilities for creative content creation, such as interactive image and streaming video generation. The growing demand for higher resolutions, frame rates, and context lengths, however, makes efficient generation increasingly challenging, as computational complexity grows quadratically with the number of generated tokens. Our work seeks to optimize the efficiency of the generation process in settings where the user's gaze location is known or can be estimated, for example, by using eye tracking. In these settings, we leverage the eccentricity-dependent acuity of human vision: while a user perceives very high-resolution visual information in a small region around their gaze location (the foveal region), the ability to resolve detail quickly degrades in the periphery of the visual field. Our approach starts with a mask modeling the foveated resolution to allocate tokens non-uniformly, assigning higher token density to foveal regions and lower density to peripheral regions. An image or video is generated in a mixed-resolution token setting, yielding results perceptually indistinguishable from full-resolution generation, while drastically reducing the token count and generation time. To this end, we develop a principled mechanism for constructing mixed-resolution tokens directly from high-resolution data, allowing a foveated diffusion model to be post-trained from an existing base model while maintaining content consistency across resolutions. We validate our approach through extensive analysis and a carefully designed user study, demonstrating the efficacy of foveation as a practical and scalable axis for efficient generation.

State-of-the-art video generation models produce remarkable photorealism, but they lack the precise control required to align generated content with specific scene requirements. Furthermore, without an underlying explicit geometry, these models cannot guarantee 3D consistency. Conversely, 3D engines offer granular control over every scene element and provide native 3D consistency by design, yet their output often remains trapped in the "uncanny valley". Bridging this sim-to-real gap requires both structural precision, where the output must exactly preserve the geometry and dynamics of the input, and global semantic transformation, where materials, lighting, and textures must be holistically transformed to achieve photorealism. We present RealMaster, a method that leverages video diffusion models to lift rendered video into photorealistic video while maintaining full alignment with the output of the 3D engine. To train this model, we generate a paired dataset via an anchor-based propagation strategy, where the first and last frames are enhanced for realism and propagated across the intermediate frames using geometric conditioning cues. We then train an IC-LoRA on these paired videos to distill the high-quality outputs of the pipeline into a model that generalizes beyond the pipeline's constraints, handling objects and characters that appear mid-sequence and enabling inference without requiring anchor frames. Evaluated on complex GTA-V sequences, RealMaster significantly outperforms existing video editing baselines, improving photorealism while preserving the geometry, dynamics, and identity specified by the original 3D control.

Video-based world models offer a powerful paradigm for embodied simulation and planning, yet state-of-the-art models often generate physically implausible manipulations - such as object penetration and anti-gravity motion - due to training on generic visual data and likelihood-based objectives that ignore physical laws. We present ABot-PhysWorld, a 14B Diffusion Transformer model that generates visually realistic, physically plausible, and action-controllable videos. Built on a curated dataset of three million manipulation clips with physics-aware annotation, it uses a novel DPO-based post-training framework with decoupled discriminators to suppress unphysical behaviors while preserving visual quality. A parallel context block enables precise spatial action injection for cross-embodiment control. To better evaluate generalization, we introduce EZSbench, the first training-independent embodied zero-shot benchmark combining real and synthetic unseen robot-task-scene combinations. It employs a decoupled protocol to separately assess physical realism and action alignment. ABot-PhysWorld achieves new state-of-the-art performance on PBench and EZSbench, surpassing Veo 3.1 and Sora v2 Pro in physical plausibility and trajectory consistency. We will release EZSbench to promote standardized evaluation in embodied video generation.

Transformer-based video diffusion models rely on 3D attention over spatial and temporal tokens, which incurs quadratic time and memory complexity and makes end-to-end training for ultra-high-resolution videos prohibitively expensive. To overcome this bottleneck, we propose a pure image adaptation framework that upgrades a video Diffusion Transformer pre-trained at its native scale to synthesize higher-resolution videos. Unfortunately, naively fine-tuning with high-resolution images alone often introduces noticeable noise due to the image-video modality gap. To address this, we decouple the learning objective to separately handle modality alignment and spatial extrapolation. At the core of our approach is Relay LoRA, a two-stage adaptation strategy. In the first stage, the video diffusion model is adapted to the image domain using low-resolution images to bridge the modality gap. In the second stage, the model is further adapted with high-resolution images to acquire spatial extrapolation capability. During inference, only the high-resolution adaptation is retained to preserve the video generation modality while enabling high-resolution video synthesis. To enhance fine-grained detail synthesis, we further propose a High-Frequency-Awareness-Training-Objective, which explicitly encourages the model to recover high-frequency components from degraded latent representations via a dedicated reconstruction loss. Extensive experiments demonstrate that our method produces ultra-high-resolution videos with rich visual details without requiring any video training data, even outperforming previous state-of-the-art models trained on high-resolution videos by 0.8 on the VBench benchmark. Code will be available at https://github.com/WillWu111/ViBe.

Diffusion Transformers (DiTs) power high-fidelity video world models but remain computationally expensive due to sequential denoising and costly spatio-temporal attention. Training-free feature caching accelerates inference by reusing intermediate activations across denoising steps; however, existing methods largely rely on a Zero-Order Hold assumption i.e., reusing cached features as static snapshots when global drift is small. This often leads to ghosting artifacts, blur, and motion inconsistencies in dynamic scenes. We propose WorldCache, a Perception-Constrained Dynamical Caching framework that improves both when and how to reuse features. WorldCache introduces motion-adaptive thresholds, saliency-weighted drift estimation, optimal approximation via blending and warping, and phase-aware threshold scheduling across diffusion steps. Our cohesive approach enables adaptive, motion-consistent feature reuse without retraining. On Cosmos-Predict2.5-2B evaluated on PAI-Bench, WorldCache achieves \textbf{2.3$\times$} inference speedup while preserving 99.4% of baseline quality, substantially outperforming prior training-free caching approaches. Our code can be accessed on World-Cache.

Video subtitle removal aims to distinguish text overlays from background content while preserving temporal coherence. Existing diffusion-based methods necessitate explicit mask sequences during both training and inference phases, which restricts their practical deployment. In this paper, we present CLEAR (Context-aware Learning for End-to-end Adaptive Video Subtitle Removal), a mask-free framework that achieves truly end-to-end inference through context-aware adaptive learning. Our two-stage design decouples prior extraction from generative refinement: Stage I learns disentangled subtitle representations via self-supervised orthogonality constraints on dual encoders, while Stage II employs LoRA-based adaptation with generation feedback for dynamic context adjustment. Notably, our method only requires 0.77% of the parameters of the base diffusion model for training. On Chinese subtitle benchmarks, CLEAR outperforms mask-dependent baselines by + 6.77dB PSNR and -74.7% VFID, while demonstrating superior zero-shot generalization across six languages (English, Korean, French, Japanese, Russian, German), a performance enabled by our generation-driven feedback mechanism that ensures robust subtitle removal without ground-truth masks during inference.

Video generation has recently emerged as a central task in the field of generative AI. However, the substantial computational cost inherent in video synthesis makes model distillation a critical technique for efficient deployment. Despite its significance, there is a scarcity of methods specifically designed for video diffusion models. Prevailing approaches often directly adapt image distillation techniques, which frequently lead to artifacts such as oversaturation, temporal inconsistency, and mode collapse. To address these challenges, we propose a novel distillation framework tailored specifically for video diffusion models. Its core innovations include: (1) an adaptive regression loss that dynamically adjusts spatial supervision weights to prevent artifacts arising from excessive distribution shifts; (2) a temporal regularization loss to counteract temporal collapse, promoting smooth and physically plausible sampling trajectories; and (3) an inference-time frame interpolation strategy that reduces sampling overhead while preserving perceptual quality. Extensive experiments and ablation studies on the VBench and VBench2 benchmarks demonstrate that our method achieves stable few-step video synthesis, significantly enhancing perceptual fidelity and motion realism. It consistently outperforms existing distillation baselines across multiple metrics.