arXiv subjects

Wenzhang Sun

Publications and source records attributed to Wenzhang Sun.

16 recordsLinked to original sources

arXiv2026

Stable Curves, Unstable Items: Item-Level Scaling Heterogeneity in Video LLMs

Aggregate scaling curves suggest that Video LLMs improve smoothly or saturate as visual budgets grow. We show that this view can conceal large, opposing changes at the item level. We represent each frozen model--item pair by its response trajectory under controlled visual budgets and derive matched-grid measures of configuration complementarity, harmful transitions, and text overwrite. Across five open Video LLMs from three architecture families, four multiple-choice benchmark splits, open-ended QA and summarization, and fixed-history dialogue generation, no single budget serves all items. On the four-model matched MCQA grid, item-level oracle headroom spans $8.8$--$18.9$ accuracy points and $12.5$--$25.5\%$ of items are correct at a lower budget but wrong at a higher one. Task-appropriate continuous metrics show the same complementarity beyond multiple choice: Token-F1 oracle gaps are $2.7$--$3.7$ score points on MLVU generation and $3.8$--$4.8$ points on AVSD current-turn generation, even when mean quality improves with budget. The effect persists across frame count, spatial resolution, sampling policy, temporal--spatial allocation, and independently executed raw-video and cached pipelines, with per-item rates and membership tracking protocol choices. A controlled sampling intervention recovers $29.0\%$ of terminal regressions, and a structured frame audit identifies several recurring evidence pathways. We release per-item trajectories, protocol provenance, derived annotations, and reproducible analysis code as an auditing artifact. A confidence cascade matches fixed-$128f$ accuracy while reducing average shared frame cost by $31.7\%$, illustrating one operational use of the response matrix.

cs.CV↗

arXiv2026

Holo-World: Unified Camera, Object and Weather Control for Video World Model

Video world models are moving toward preserving an observed world under controllable camera and object motion while allowing its environmental state to change. Yet these controls remain isolated, and weather generation typically relies on a source video or reconstructed scene that already specifies future structure. We study a first-frame-anchored source-to-state setting, where the model starts from a single image and follows explicit camera and object controls and an optional weather instruction, then generates a video that either preserves the source world or transfers it to a target weather state. To address these challenges, we first build HoloStateData, a state video dataset that turns diverse videos into unified control samples for camera, object, and weather supervision. Second, we introduce Holo-World, a unified controllable video world model that jointly controls the scene from a single image. Its Unified Scene Adapter factorizes world preservation and weather transfer into distinct parameter subspaces, using rendered background, geometry buffers, and object controls to maintain controlled scene structure while modeling weather-dependent appearance and particle effects. Additionally, Scene-Weather Decomposed CFG guides scene and weather residuals separately, strengthening target weather effects without over-amplifying the full condition. Quantitative and qualitative experiments demonstrate that Holo-World maintains precise camera and object controls with consistent scene structure while transferring scenes into diverse target weather states, outperforming video-to-video weather editing baselines on weather-state generation. Our project page is available at https://xiangchenyin.github.io/Holo-World/

cs.CV↗

arXiv2026

Preserve, Reveal, Expand: Faithful 4D Video Editing with Region-Aware Conditioning

Existing 4D-driven video diffusion models primarily target plausible generation, but faithful 4D editing requires preserving source-observed regions while synthesizing disoccluded or out-of-view content. We identify Evidence-Role Mismatch: reliable source-backed evidence, unreliable rendered cues, and unsupported regions are entangled in a single conditioning signal, causing preservation drift, ghosting, and unstable extrapolation. We propose PREX (Preserve, Reveal, Expand), a region-aware framework that decomposes the target spatiotemporal volume into Preserve, Reveal, and Expand roles according to observation support and scene extent. PREX builds observation-backed appearance cues with calibrated confidence and injects them into a frozen video diffusion backbone through a region-aware adapter, trained with proxy tasks without requiring paired edited videos. We further introduce PREBench, a diagnostic benchmark with curated edits, region-role masks, and human-aligned metrics that complement global video-quality and 4D-control evaluations. Experiments show that PREX reduces region-structured failures while maintaining strong visual quality and 4D edit control capability. Project Page: https://ricepastem.github.io/PREX-Open

cs.CV↗

arXiv2026

RiO-DETR: DETR for Real-time Oriented Object Detection

We present RiO-DETR: DETR for Real-time Oriented Object Detection, the first real-time oriented detection transformer to the best of our knowledge. Adapting DETR to oriented bounding boxes (OBBs) poses three challenges: semantics-dependent orientation, angle periodicity that breaks standard Euclidean refinement, and an enlarged search space that slows convergence. RiO-DETR resolves these issues with task-native designs while preserving real-time efficiency. First, we propose Content-Driven Angle Estimation by decoupling angle from positional queries, together with Rotation-Rectified Orthogonal Attention to capture complementary cues for reliable orientation. Second, Decoupled Periodic Refinement combines bounded coarse-to-fine updates with a Shortest-Path Periodic Loss for stable learning across angular seams. Third, Oriented Dense O2O injects angular diversity into dense supervision to speed up angle convergence at no extra cost. Extensive experiments on DOTA-1.0, DIOR-R, and FAIR-1M-2.0 demonstrate RiO-DETR establishes a new speed--accuracy trade-off for real-time oriented detection. GitHub Repository: https://github.com/RicePasteM/RiO-DETR.

cs.CV↗

arXiv2026

MUSE: A Multi-agent Framework for Unconstrained Story Envisioning via Closed-Loop Cognitive Orchestration

Generating long-form audio-visual stories from a short user prompt remains challenging due to an intent-execution gap, where high-level narrative intent must be preserved across coherent, shot-level multimodal generation over long horizons. Existing approaches typically rely on feed-forward pipelines or prompt-only refinement, which often leads to semantic drift and identity inconsistency as sequences grow longer. We address this challenge by formulating storytelling as a closed-loop constraint enforcement problem and propose MUSE, a multi-agent framework that coordinates generation through an iterative plan-execute-verify-revise loop. MUSE translates narrative intent into explicit, machine-executable controls over identity, spatial composition, and temporal continuity, and applies targeted multimodal feedback to correct violations during generation. To evaluate open-ended storytelling without ground-truth references, we introduce MUSEBench, a reference-free evaluation protocol validated by human judgments. Experiments demonstrate that MUSE substantially improves long-horizon narrative coherence, cross-modal identity consistency, and cinematic quality compared with representative baselines.

cs.CV↗

arXiv2025

DeCo-VAE: Learning Compact Latents for Video Reconstruction via Decoupled Representation

Existing video Variational Autoencoders (VAEs) generally overlook the similarity between frame contents, leading to redundant latent modeling. In this paper, we propose decoupled VAE (DeCo-VAE) to achieve compact latent representation. Instead of encoding RGB pixels directly, we decompose video content into distinct components via explicit decoupling: keyframe, motion and residual, and learn dedicated latent representation for each. To avoid cross-component interference, we design dedicated encoders for each decoupled component and adopt a shared 3D decoder to maintain spatiotemporal consistency during reconstruction. We further utilize a decoupled adaptation strategy that freezes partial encoders while training the others sequentially, ensuring stable training and accurate learning of both static and dynamic features. Extensive quantitative and qualitative experiments demonstrate that DeCo-VAE achieves superior video reconstruction performance.

cs.CV↗

arXiv2025

PAGS: Priority-Adaptive Gaussian Splatting for Dynamic Driving Scenes

Reconstructing dynamic 3D urban scenes is crucial for autonomous driving, yet current methods face a stark trade-off between fidelity and computational cost. This inefficiency stems from their semantically agnostic design, which allocates resources uniformly, treating static backgrounds and safety-critical objects with equal importance. To address this, we introduce Priority-Adaptive Gaussian Splatting (PAGS), a framework that injects task-aware semantic priorities directly into the 3D reconstruction and rendering pipeline. PAGS introduces two core contributions: (1) Semantically-Guided Pruning and Regularization strategy, which employs a hybrid importance metric to aggressively simplify non-critical scene elements while preserving fine-grained details on objects vital for navigation. (2) Priority-Driven Rendering pipeline, which employs a priority-based depth pre-pass to aggressively cull occluded primitives and accelerate the final shading computations. Extensive experiments on the Waymo and KITTI datasets demonstrate that PAGS achieves exceptional reconstruction quality, particularly on safety-critical objects, while significantly reducing training time and boosting rendering speeds to over 350 FPS.

cs.CV↗

arXiv2025

DrivingScene: A Multi-Task Online Feed-Forward 3D Gaussian Splatting Method for Dynamic Driving Scenes

Real-time, high-fidelity reconstruction of dynamic driving scenes is challenged by complex dynamics and sparse views, with prior methods struggling to balance quality and efficiency. We propose DrivingScene, an online, feed-forward framework that reconstructs 4D dynamic scenes from only two consecutive surround-view images. Our key innovation is a lightweight residual flow network that predicts the non-rigid motion of dynamic objects per camera on top of a learned static scene prior, explicitly modeling dynamics via scene flow. We also introduce a coarse-to-fine training paradigm that circumvents the instabilities common to end-to-end approaches. Experiments on nuScenes dataset show our image-only method simultaneously generates high-quality depth, scene flow, and 3D Gaussian point clouds online, significantly outperforming state-of-the-art methods in both dynamic reconstruction and novel view synthesis.

cs.CV↗

arXiv2025

MotionStrata: Hierarchical Motion Latents for Compact Video Autoencoding

First-frame-conditioned video autoencoders represent a clip with persistent content and a compact motion code. Although this removes much of the appearance redundancy, the remaining motion is typically compressed with a homogeneous latent geometry. Such representations use the same temporal support for broad scene evolution and fine-grained, frame-specific details. We introduce MotionStrata, which organizes a fixed motion budget into Global Motion and Detailed Motion. Temporally compressed Global queries summarize broad evolution, whereas frame-aligned Detailed queries preserve fine-grained structures whose configuration varies across frames. Frequency-guided routing and coarse-to-fine training establish this hierarchy without increasing motion dimensionality. Experiments show that MotionStrata maintains high reconstruction quality under aggressive compression and outperforms uniform and alternative grouped representations. Additional experiments evaluate hierarchical representation, downstream generation, and decoding cost. These results support hierarchical motion organization as a useful design principle for compact video autoencoding.

cs.CV↗

arXiv2025

ChronoTailor: Harnessing Attention Guidance for Fine-Grained Video Virtual Try-On

Video virtual try-on aims to seamlessly replace the clothing of a person in a source video with a target garment. Despite significant progress in this field, existing approaches still struggle to maintain continuity and reproduce garment details. In this paper, we introduce ChronoTailor, a diffusion-based framework that generates temporally consistent videos while preserving fine-grained garment details. By employing a precise spatio-temporal attention mechanism to guide the integration of fine-grained garment features, ChronoTailor achieves robust try-on performance. First, ChronoTailor leverages region-aware spatial guidance to steer the evolution of spatial attention and employs an attention-driven temporal feature fusion mechanism to generate more continuous temporal features. This dual approach not only enables fine-grained local editing but also effectively mitigates artifacts arising from video dynamics. Second, ChronoTailor integrates multi-scale garment features to preserve low-level visual details and incorporates a garment-pose feature alignment to ensure temporal continuity during dynamic motion. Additionally, we collect StyleDress, a new dataset featuring intricate garments, varied environments, and diverse poses, offering advantages over existing public datasets, and will be publicly available for research. Extensive experiments show that ChronoTailor maintains spatio-temporal continuity and preserves garment details during motion, significantly outperforming previous methods.

cs.CV↗

arXiv2025

A Self-supervised Motion Representation for Portrait Video Generation

Recent advancements in portrait video generation have been noteworthy. However, existing methods rely heavily on human priors and pre-trained generative models, Motion representations based on human priors may introduce unrealistic motion, while methods relying on pre-trained generative models often suffer from inefficient inference. To address these challenges, we propose Semantic Latent Motion (SeMo), a compact and expressive motion representation. Leveraging this representation, our approach achieve both high-quality visual results and efficient inference. SeMo follows an effective three-step framework: Abstraction, Reasoning, and Generation. First, in the Abstraction step, we use a carefully designed Masked Motion Encoder, which leverages a self-supervised learning paradigm to compress the subject's motion state into a compact and abstract latent motion (1D token). Second, in the Reasoning step, we efficiently generate motion sequences based on the driving audio signal. Finally, in the Generation step, the motion dynamics serve as conditional information to guide the motion decoder in synthesizing realistic transitions from reference frame to target video. Thanks to the compact and expressive nature of Semantic Latent Motion, our method achieves efficient motion representation and high-quality video generation. User studies demonstrate that our approach surpasses state-of-the-art models with an 81% win rate in realism. Extensive experiments further highlight its strong compression capability, reconstruction quality, and generative potential.

cs.CV↗

arXiv2025

UniCP: A Unified Caching and Pruning Framework for Efficient Video Generation

Diffusion Transformers (DiT) excel in video generation but encounter significant computational challenges due to the quadratic complexity of attention. Notably, attention differences between adjacent diffusion steps follow a U-shaped pattern. Current methods leverage this property by caching attention blocks, however, they still struggle with sudden error spikes and large discrepancies. To address these issues, we propose UniCP a unified caching and pruning framework for efficient video generation. UniCP optimizes both temporal and spatial dimensions through. Error Aware Dynamic Cache Window (EDCW): Dynamically adjusts cache window sizes for different blocks at various timesteps, adapting to abrupt error changes. PCA based Slicing (PCAS) and Dynamic Weight Shift (DWS): PCAS prunes redundant attention components, and DWS integrates caching and pruning by enabling dynamic switching between pruned and cached outputs. By adjusting cache windows and pruning redundant components, UniCP enhances computational efficiency and maintains video detail fidelity. Experimental results show that UniCP outperforms existing methods in both performance and efficiency.

cs.CV↗

arXiv2025

MoEE: Mixture of Emotion Experts for Audio-Driven Portrait Animation

The generation of talking avatars has achieved significant advancements in precise audio synchronization. However, crafting lifelike talking head videos requires capturing a broad spectrum of emotions and subtle facial expressions. Current methods face fundamental challenges: a) the absence of frameworks for modeling single basic emotional expressions, which restricts the generation of complex emotions such as compound emotions; b) the lack of comprehensive datasets rich in human emotional expressions, which limits the potential of models. To address these challenges, we propose the following innovations: 1) the Mixture of Emotion Experts (MoEE) model, which decouples six fundamental emotions to enable the precise synthesis of both singular and compound emotional states; 2) the DH-FaceEmoVid-150 dataset, specifically curated to include six prevalent human emotional expressions as well as four types of compound emotions, thereby expanding the training potential of emotion-driven models. Furthermore, to enhance the flexibility of emotion control, we propose an emotion-to-latents module that leverages multimodal inputs, aligning diverse control signals-such as audio, text, and labels-to ensure more varied control inputs as well as the ability to control emotions using audio alone. Through extensive quantitative and qualitative evaluations, we demonstrate that the MoEE framework, in conjunction with the DH-FaceEmoVid-150 dataset, excels in generating complex emotional expressions and nuanced facial details, setting a new benchmark in the field. These datasets will be publicly released.

cs.CV↗

arXiv2024

UniAvatar: Taming Lifelike Audio-Driven Talking Head Generation with Comprehensive Motion and Lighting Control

Recently, animating portrait images using audio input is a popular task. Creating lifelike talking head videos requires flexible and natural movements, including facial and head dynamics, camera motion, realistic light and shadow effects. Existing methods struggle to offer comprehensive, multifaceted control over these aspects. In this work, we introduce UniAvatar, a designed method that provides extensive control over a wide range of motion and illumination conditions. Specifically, we use the FLAME model to render all motion information onto a single image, maintaining the integrity of 3D motion details while enabling fine-grained, pixel-level control. Beyond motion, this approach also allows for comprehensive global illumination control. We design independent modules to manage both 3D motion and illumination, permitting separate and combined control. Extensive experiments demonstrate that our method outperforms others in both broad-range motion control and lighting control. Additionally, to enhance the diversity of motion and environmental contexts in current datasets, we collect and plan to publicly release two datasets, DH-FaceDrasMvVid-100 and DH-FaceReliVid-200, which capture significant head movements during speech and various lighting scenarios.

cs.CV↗

arXiv2024

TV-3DG: Mastering Text-to-3D Customized Generation with Visual Prompt

In recent years, advancements in generative models have significantly expanded the capabilities of text-to-3D generation. Many approaches rely on Score Distillation Sampling (SDS) technology. However, SDS struggles to accommodate multi-condition inputs, such as text and visual prompts, in customized generation tasks. To explore the core reasons, we decompose SDS into a difference term and a classifier-free guidance term. Our analysis identifies the core issue as arising from the difference term and the random noise addition during the optimization process, both contributing to deviations from the target mode during distillation. To address this, we propose a novel algorithm, Classifier Score Matching (CSM), which removes the difference term in SDS and uses a deterministic noise addition process to reduce noise during optimization, effectively overcoming the low-quality limitations of SDS in our customized generation framework. Based on CSM, we integrate visual prompt information with an attention fusion mechanism and sampling guidance techniques, forming the Visual Prompt CSM (VPCSM) algorithm. Furthermore, we introduce a Semantic-Geometry Calibration (SGC) module to enhance quality through improved textual information integration. We present our approach as TV-3DG, with extensive experiments demonstrating its capability to achieve stable, high-quality, customized 3D generation. Project page: \url{https://yjhboy.github.io/TV-3DG}

cs.CV↗

arXiv2024

DH-FaceVid-1K: A Large-Scale High-Quality Dataset for Face Video Generation

Human-centric generative models are becoming increasingly popular, giving rise to various innovative tools and applications, such as talking face videos conditioned on text or audio prompts. The core of these capabilities lies in powerful pre-trained foundation models, trained on large-scale, high-quality datasets. However, many advanced methods rely on in-house data subject to various constraints, and other current studies fail to generate high-resolution face videos, which is mainly attributed to the significant lack of large-scale, high-quality face video datasets. In this paper, we introduce a human face video dataset, \textbf{DH-FaceVid-1K}. Our collection spans 1,200 hours in total, encompassing 270,043 video clips from over 20,000 individuals. Each sample includes corresponding speech audio, facial keypoints, and text annotations. Compared to other publicly available datasets, ours distinguishes itself through its multi-ethnic coverage and high-quality, comprehensive individual attributes. We establish multiple face video generation models supporting tasks such as text-to-video and image-to-video generation. In addition, we develop comprehensive benchmarks to validate the scaling law when using different proportions of proposed dataset. Our primary aim is to contribute a face video dataset, particularly addressing the underrepresentation of Asian faces in existing curated datasets and thereby enriching the global spectrum of face-centric data and mitigating demographic biases. \textbf{Project Page:} https://luna-ai-lab.github.io/DH-FaceVid-1K/

cs.CV↗