Temporal Consistency in Video Generation

AIGC has rapidly expanded from text-to-image generation toward high-quality multimodal synthesis across video and audio. Within this context, joint audio-video generation (JAVG) has emerged as a fundamental task that produces synchronized and semantically aligned sound and vision from textual descriptions. However, compared with advanced commercial models such as Veo3, existing open-source methods still suffer from limitations in generation quality, temporal synchrony, and alignment with human preferences. To bridge the gap, this paper presents JavisDiT++, a concise yet powerful framework for unified modeling and optimization of JAVG. First, we introduce a modality-specific mixture-of-experts (MS-MoE) design that enables cross-modal interaction efficacy while enhancing single-modal generation quality. Then, we propose a temporal-aligned RoPE (TA-RoPE) strategy to achieve explicit, frame-level synchronization between audio and video tokens. Besides, we develop an audio-video direct preference optimization (AV-DPO) method to align model outputs with human preference across quality, consistency, and synchrony dimensions. Built upon Wan2.1-1.3B-T2V, our model achieves state-of-the-art performance merely with around 1M public training entries, significantly outperforming prior approaches in both qualitative and quantitative evaluations. Comprehensive ablation studies have been conducted to validate the effectiveness of our proposed modules. All the code, model, and dataset are released at https://JavisVerse.github.io/JavisDiT2-page.

Text-to-video generation is expensive, so only a few samples are typically produced per prompt. In this low-sample regime, maximizing the value of each batch requires high cross-video diversity. Recent methods improve diversity for image generation, but for videos they often degrade within-video temporal consistency and require costly backpropagation through a video decoder. We propose a joint-sampling framework for flow-matching video generators that improves batch diversity while preserving temporal consistency. Our approach applies diversity-driven updates and then removes only the components that would decrease a temporal-consistency objective. To avoid image-space gradients, we compute both objectives with lightweight latent-space models, avoiding video decoding and decoder backpropagation. Experiments on a state-of-the-art text-to-video flow-matching model show diversity close to strong joint-sampling baselines while substantially improving temporal consistency and color naturalness. Our code is available at https://github.com/XinshuangL/Diverse-Video.

Maintaining spatial world consistency over long horizons remains a central challenge for camera-controllable video generation. Existing memory-based approaches often condition generation on globally reconstructed 3D scenes by rendering anchor videos from the reconstructed geometry in the history. However, reconstructing a global 3D scene from multiple views inevitably introduces cross-view misalignment, as pose and depth estimation errors cause the same surfaces to be reconstructed at slightly different 3D locations across views. When fused, these inconsistencies accumulate into noisy geometry that contaminates the conditioning signals and degrades generation quality. We introduce AnchorWeave, a memory-augmented video generation framework that replaces a single misaligned global memory with multiple clean local geometric memories and learns to reconcile their cross-view inconsistencies. To this end, AnchorWeave performs coverage-driven local memory retrieval aligned with the target trajectory and integrates the selected local memories through a multi-anchor weaving controller during generation. Extensive experiments demonstrate that AnchorWeave significantly improves long-term scene consistency while maintaining strong visual quality, with ablation and analysis studies further validating the effectiveness of local geometric conditioning, multi-anchor control, and coverage-driven retrieval.

Autoregressive (AR) video diffusion models in the Self Forcing family generate a video chunk by chunk, each conditioned on the ones before it, so they can keep rolling out with no preset length. In practice they are trained on short, fixed clips (e.g., 5s). Once a rollout passes this training horizon, the cache fills with self-generated chunks the model never saw in training, and quality degrades: colors oversaturate, identities drift, structures collapse, and motion fades. We frame this failure as a cache-behavior mismatch, a domain shift between the within-horizon caches seen in training and those produced by open-ended testing, and address it with Rolling Sink, a training-free inference-time cache policy. Within the base model's fixed cache budget, Rolling Sink holds the cache close to its stable within-horizon behavior: it restores low-drift content, sliding temporal indices, and evolving semantics, and keeps a small recent context. Attention sink and temporal re-indexing are existing mechanisms. Our contribution is the mismatch formulation, the joint bounded-cache policy, and the Rolling Semantics schedule. On two 5s-trained bases, Self Forcing and Causal Forcing, Rolling Sink attains the best averaged rank on VBench-Long at both 1- and 5-minute rollouts under the same inference setting, at about 0.56% throughput change, and holds GPU memory constant. It also sustains a 30-minute qualitative rollout, 360 times its training horizon. We validate at 1- and 5-minute rollouts and do not claim guaranteed quality at arbitrary duration. Project page: https://rolling-sink.github.io/

Modern generative video models excel at producing convincing, high-quality outputs, but struggle to maintain multi-view and spatiotemporal consistency in highly dynamic real-world environments. In this work, we introduce $\textbf{AnyView}$, a diffusion-based video generation framework for $\textit{dynamic view synthesis}$ with minimal inductive biases or geometric assumptions. We leverage multiple data sources with various levels of supervision, including monocular (2D), multi-view static (3D) and multi-view dynamic (4D) datasets, to train a generalist spatiotemporal implicit representation capable of producing zero-shot novel videos from arbitrary camera locations and trajectories. We evaluate AnyView on standard benchmarks, showing competitive results with the current state of the art, and propose $\textbf{AnyViewBench}$, a challenging new benchmark tailored towards $\textit{extreme}$ dynamic view synthesis in diverse real-world scenarios. In this more dramatic setting, we find that most baselines drastically degrade in performance, as they require significant overlap between viewpoints, while AnyView maintains the ability to produce realistic, plausible, and spatiotemporally consistent videos when prompted from $\textit{any}$ viewpoint. Results, data, code, and models can be viewed at: https://tri-ml.github.io/AnyView/

Despite the rapid progress of video generation models, the role of data in influencing motion is poorly understood. We present Motive (MOTIon attribution for Video gEneration), a motion-centric, gradient-based data attribution framework that scales to modern, large, high-quality video datasets and models. We use this to study which fine-tuning clips improve or degrade temporal dynamics. Motive isolates temporal dynamics from static appearance via motion-weighted loss masks, yielding efficient and scalable motion-specific influence computation. On text-to-video models, Motive identifies clips that strongly affect motion and guides data curation that improves temporal consistency and physical plausibility. With Motive-selected high-influence data, our method improves both motion smoothness and dynamic degree on VBench, achieving a 74.1% human preference win rate compared with the pretrained base model. To our knowledge, this is the first framework to attribute motion rather than visual appearance in video generative models and to use it to curate fine-tuning data.

Autoregressive video diffusion models hold promise for world simulation but are vulnerable to exposure bias arising from the train-test mismatch. While recent works address this via post-training, they typically rely on a bidirectional teacher model or discriminator. To achieve an end-to-end solution, we introduce Resampling Forcing, a teacher-free framework that enables training autoregressive video models from scratch and at scale. Central to our approach is a self-resampling scheme that simulates inference-time model errors on history frames during training. Conditioned on these degraded histories, a sparse causal mask enforces temporal causality while enabling parallel training with frame-level diffusion loss. To facilitate efficient long-horizon generation, we further introduce history routing, a parameter-free mechanism that dynamically retrieves the top-k most relevant history frames for each query. Experiments demonstrate that our approach achieves performance comparable to distillation-based baselines while exhibiting superior temporal consistency on longer videos owing to native-length training.

Generating interactive and dynamic 4D scenes from a single static image remains a core challenge. Most existing generate-then-reconstruct and reconstruct-then-generate methods decouple geometry from motion, causing spatiotemporal inconsistencies and poor generalization. To address these, we present MoGe4D (Motion and Geometry-Aware image-to-4D Synthesis), a geometry-conditioned framework for single-image 4D synthesis that models a scene as dense 4D point trajectories. Instead of treating geometry and dynamics as two disconnected stages, our method starts from an initial geometric prior inferred from the input image and predicts future time-varying trajectories in a diffusion process, improving spatiotemporal coherence while preserving structural stability. To support this task, we first introduce TrajScene-60K, a large-scale dataset of 60,000 video samples with dense 4D point trajectories, addressing the scarcity of high-quality training data for scene-level 4D generation. Built on this, our diffusion-based 4D Scene Trajectory Generator (4D-STraG) predicts geometry-consistent and motion-plausible trajectory fields conditioned on the input image, with a depth-guided motion normalization strategy to reduce scale ambiguity and a Motion Perception Module (MPM) to inject motion-aware priors. We further propose a 4D View Synthesis Module (4D-ViSM) to render the generated 4D representation into videos under arbitrary camera trajectories. Experiments show that MoGe4D produces high-quality 4D scenes with strong temporal coherence, favorable geometry-aware consistency, and compelling novel-view synthesis from a single image. Code: https://github.com/Zhangyr2022/MoGe4D.

Long video generation is a critical step toward building realistic world models, requiring both high visual fidelity and long-range interaction consistency. Recent autoregressive diffusion models enable long-horizon generation through KV cache reuse, yet suffer from two fundamental challenges: failure to preserve long-range interactions due to sliding-window KV cache and error accumulation that progressively degrades generation quality over time. To address these issues, we propose BIFE, a framework that introduces a semantic sparse KV cache for retrieval-based long-range conditioning and a Block Forcing training strategy to enforce cross-block consistency. Together, these designs preserve historical interactions while mitigating drift, enabling stable and coherent minute-long video generation. We also introduce InterVBench, a minute-long video benchmark with fine-grained block-level annotations and Video Drift Error metrics. Extensive experiments on InterVBench and VBench-Long demonstrate that BIFE achieves state-of-the-art performance, including a 22.2% improvement on VDE-Subject and a 19.4% improvement on VDE-Clarity over baselines. Website: https://alibaba-damo-academy.github.io/BIFE. Code: https://github.com/alibaba-damo-academy/BIFE.

Step-by-step painting tutorials are used for learning artistic techniques, but existing video resources lack interactivity and personalization. While recent generative models have advanced artistic image synthesis, they still struggle to generate temporally and structurally consistent videos. Video models excel in consistent frame generation, yet fine-tuning an image-to-video model on forward painting sequences creates a long-range dependency on the first-frame reference, disrupting its pretrained attention pattern and hindering reconstruction. We address this issue by training on reversed sequences that better preserve the reference-attention behavior. Building on this, we present Loomis Painter, a framework for reconstructing coherent painting processes. We further introduce a media-transfer training strategy that decouples input appearance from the target process, enabling generation across different artistic media. Trained on a curated dataset of real-world painting processes, our method outperforms existing baselines across LPIPS, DINO, and CLIP metrics and is consistently preferred in expert and non-expert user studies. Code available at https://github.com/Markus-Pobitzer/wlp.

Diffusion-based text-to-video models are increasingly capable, but mask-based editing over hundreds of frames remains challenging: naïve long-video generation suffers from memory blow-up, window seams, and temporal drift, while existing editors often require specialized modules or heavy fine-tuning. We present Overlapping High-Order Co-Denoising, a lightweight framework that turns a single pre-trained text-to-video model into a unified inpainting-outpainting editor. We train only LoRA adapters using mixed interior and border masks together with a dual-region loss that improves synthesis inside the mask while explicitly preserving known content. At inference, we denoise long latent sequences using overlapping windows, apply second-order Heun sampling within each window, and fuse overlaps with Hamming-weighted blending to reduce boundary artifacts and improve temporal coherence. On InpaintBench (30 real-world videos, 81--300 frames), our method outperforms Wan 2.1 variants and VACE in background faithfulness (SSIM/LPIPS), temporal consistency (tLPIPS), and text alignment (CLIP), and scales to long horizons, demonstrated up to 800 frames, with memory bounded by the chosen window size.

Text-to-image diffusion models have advanced video editing. However, existing video editing methods are limited by their high computational overhead and memory consumption. Furthermore, these approaches often sacrifice visual fidelity, leading to temporal inconsistencies and artifacts such as blurring and mosaic-like patterns. To address these dual challenges and strike a balance between computational efficiency and visual fidelity, we propose Edit-Your-Interest, a lightweight, text-driven, zero-shot video editing method. Edit-Your-Interest introduces a Spatio-Temporal Feature Memory bank (SFM) to cache spatial-attention features from previous frames under a fixed memory budget. The SFM maintains a representative candidate pool, reducing the computational overhead of full-sequence spatio-temporal modeling. To query this memory, we propose Feature Most-Similar Propagation (FMP), a similarity-based retrieval rule that propagates relevant cached tokens to the current frame. Together, SFM and FMP form a bounded-memory retrieval-based propagation design that improves temporal consistency while reducing blurring and mosaic-like artifacts. Finally, we introduce a feature-diversity-based SFM update algorithm to keep the memory representative and non-redundant throughout the video sequence. Furthermore, to enable precise object editing, we leverage cross-attention maps to automatically extract masks for the instances of interest. These masks are seamlessly integrated into the diffusion denoising process, enabling fine-grained control over target objects and allowing Edit-Your-Interest to perform highly accurate edits while robustly preserving the background integrity. Extensive experiments demonstrate that Edit-Your-Interest outperforms state-of-the-art methods in both efficiency and visual fidelity. Project page: https://edit-your-interest.github.io/.

Bridge Matching (BM) learns a stochastic transport between two data distributions and is therefore a natural tool for data-to-data tasks such as video frame interpolation, image-to-video generation, and video super-resolution. Existing bridge models for video, however, use a reference process that treats the frames of a clip independently, so temporal structure enters only through the network or the choice of the bridge endpoints. We propose Time-Correlated Video Bridge Matching (TCVBM), which places temporal structure into the reference process itself: the frames of a clip are coupled by a linear SDE whose drift is a discrete Laplacian along the time axis, and known frames enter as boundary conditions. The transition and bridge distributions remain Gaussian and available in closed form, so training and sampling are as simple as in standard BM, which is recovered in the limit of zero coupling. A spectral analysis shows that the prior damps high temporal frequencies, injects temporally smooth noise, and relaxes towards a boundary-determined profile, such as the linear interpolation between key frames. In controlled comparisons with frame-independent bridges under identical architectures and training budgets, TCVBM improves most automatic metrics for frame interpolation (Moving MNIST, UCF-101, Vimeo-90K) and image-to-video generation (Moving MNIST, UCF-101), e.g., reducing FVD from 688 to 611 for frame interpolation on UCF-101, and human raters prefer its frame interpolations for temporal consistency in 49$\%$ of comparisons versus 15$\%$ for the baseline ($p<10^{-17}$); for super-resolution, it performs on par with BM on automatic metrics.

Existing text-to-video (T2V) evaluation benchmarks, such as VBench and EvalCrafter, predominantly contain either subject-centric prompts or static camera scenes, overlooking the importance of camera motion, essential for producing cinematic shots with realistic scenes. Due to this, the metrics developed in these works often produce incorrect results in the presence of camera motion. To fill this gap, we introduce DynamicEval, a benchmark comprising of systematically curated prompts that emphasize dynamic camera motion, paired with 45k human annotations on video pairs. DynamicEval evaluates two key dimensions of video quality: the consistency of the foreground subjects and the consistency of the background scene. To address the shortcomings of existing metrics under camera motion, we introduce two novel metrics: 1) Tracker-FG, which evaluates subject consistency by assessing the keypoint tracks of the subject. 2) Debiased Motion Smoothness, for background consistency, devoid from issues of occlusions/disocclusions arising from camera and foreground subject movements. Extensive experiments on the large-scale DynamicEval benchmark demonstrate that our proposed metrics achieve stronger correlations with human preferences at both the video level, and the model level with an increase of 0.32 in average Pearson Linear Correlation.

In light of recent progress in video editing, deep learning models focusing on both spatial and temporal dependencies have emerged as the primary method. However, these models suffer from the quadratic computational complexity of traditional attention mechanisms, making them difficult to adapt to long-duration and high-resolution videos. This limitation restricts their applicability in practical contexts such as real-time video processing. To tackle this challenge, we introduce a method to reduce both time and space complexity of these systems by proposing VRWKV-Editor, a novel video editing model that integrates a linear spatio-temporal aggregation module into video-based diffusion models. VRWKV-Editor leverages bidirectional weighted key-value recurrence mechanism of the RWKV transformer to capture global dependencies while preserving temporal coherence, achieving linear complexity without sacrificing quality. Extensive experiments demonstrate that the proposed method achieves up to 3.7x speedup and 60% lower memory usage compared to state-of-the-art diffusion-based video editing methods, while maintaining competitive performance in frame consistency and text alignment. Furthermore, a comparative analysis we conducted on videos with different sequence lengths confirms that the gap in editing speed between our approach and architectures with self-attention becomes more significant with long videos.

We present Vivid-VR, a DiT-based generative video restoration method built upon an advanced T2V foundation model, where ControlNet is leveraged to control the generation process, ensuring content consistency. However, conventional fine-tuning of such controllable pipelines frequently suffers from distribution drift due to limitations in imperfect multimodal alignment, resulting in compromised texture realism and temporal coherence. To tackle this challenge, we propose a concept distillation training strategy that utilizes the pretrained T2V model to synthesize training samples with embedded textual concepts, thereby distilling its conceptual understanding to preserve texture and temporal quality. To enhance generation controllability, we redesign the control architecture with two key components: 1) a control feature projector that filters degradation artifacts from input video latents to minimize their propagation through the generation pipeline, and 2) a new ControlNet connector employing a dual-branch design. This connector synergistically combines MLP-based feature mapping with cross-attention mechanism for dynamic control feature retrieval, enabling both content preservation and adaptive control signal modulation. Extensive experiments show that Vivid-VR performs favorably against existing approaches on both synthetic and real-world benchmarks, as well as AIGC videos, achieving impressive texture realism, visual vividness, and temporal consistency. The codes and checkpoints are publicly available at https://github.com/csbhr/Vivid-VR.

While video generation models excel at producing high-quality monocular videos, generating 3D stereoscopic and spatial videos for immersive applications remains an underexplored challenge. We present a pose-free and training-free method that leverages an off-the-shelf monocular video generation model to produce immersive 3D videos. Our approach first warps the generated monocular video into pre-defined camera viewpoints using estimated depth information, then applies a novel frame matrix inpainting framework. This framework utilizes the original video generation model to synthesize missing content across different viewpoints and timestamps, ensuring spatial and temporal consistency without requiring additional model fine-tuning. Moreover, we develop a \dualupdate~scheme that further improves the quality of video inpainting by alleviating the negative effects propagated from disoccluded areas in the latent space. The resulting multi-view videos are then adapted into stereoscopic pairs or optimized into 4D Gaussians for spatial video synthesis. We validate the efficacy of our proposed method by conducting experiments on videos from various generative models, such as Sora, Lumiere, WALT, and Zeroscope. The experiments demonstrate that our method has a significant improvement over previous methods. Project page at: https://daipengwa.github.io/S-2VG_ProjectPage/

Unlike traditional video editing or inpainting, video semantic mixing fuses a reference concept with a moving target entity to produce a hybrid while preserving the source video's motion and layout. We propose MoCA-Video, a training-free framework that steers a frozen video-diffusion denoising trajectory through concept-localized reference injection. At selected low-noise steps, MoCA-Video uses concept attention to localize the target object and injects the reference latent into the localized region, where object structure has formed but appearance remains editable. A momentum-based correction carries the injected prediction across frames to encourage coherent concept integration through the sequence. We further introduce CASS, a CLIP-based metric that measures the output's directional alignment shift toward the reference and away from the source prompt. Using the denoiser's internal attention avoids an external localization model; in our A100 FP16 setup, MoCA-Video takes 3.2 seconds per output frame, excluding preprocessing. Across the evaluated baselines, MoCA-Video achieves the highest CASS, rel-CASS, and ImageReward, while LPIPS-T and FVD expose separate temporal-coherence and video-quality trade-offs.

Front-view driving video continuation is a critical component for constructing sophisticated world models. However, maintaining long-term coherence faces the fundamental challenge of mitigating ''generative degeneration'' during the auto-regressive process. This phenomenon arises when the limited attention capacity of world models becomes saturated with high-entropy pixel details over extended sequences, leading to a collapse of global scene structure and motion intensity. To address this, we introduce STRIDE (STructural RegIsters for Decoupled Extrapolation), a large vision model architecture with an enforced semantic foresight mechanism. STRIDE decouples video continuation into two sub-tasks via interleaved generation of semantic and RGB tokens: forecasting low-entropy semantic maps, and subsequently generating high-frequency visual appearance conditioned on this structure. The semantic tokens explicitly function as ''structural registers'' -- dedicated memory slots that robustly retain the scene's long-term context and world dynamics. Furthermore, we employ an architecturally decoupled flow matching-based super-resolution module to enhance the final visual fidelity. Extensive experiments in autonomous driving scenarios demonstrate that STRIDE effectively mitigates generative degeneration, achieving minute-level, temporally consistent driving video continuation with strong generalization ability.