Skip to content

🎬 Video Generation

🧠 NeurIPS2026 · 90 paper notes

📌 Same area in other venues: 🎞️ ECCV2026 (176) · 📷 CVPR2026 (188) · 🔬 ICLR2026 (99) · 💬 ACL2026 (4) · 🧪 ICML2026 (32) · 🤖 AAAI2026 (11)

🔥 Top topics: Video Generation ×36 · Diffusion Models ×14 · Agents ×3 · Self-Supervised Learning ×3 · Multimodal/VLM ×3

AdaState: Self-Evolving Anchors for Streaming Video Generation

AdaState eliminates the frozen scene problem in streaming video diffusion by replacing static first-frame attention sinks with an adaptive latent state that is jointly denoised but never rendered, casting iterative denoising into an implicit recurrence carried by the KV cache to break the consistency-dynamics tradeoff.

Addressable Memory for Video World Models

Addressing the dual failure modes of out-of-distribution temporal RoPE offsets and phase cancellation during cache compression in autoregressive video world models, this paper introduces WorldTrace, a training-free memory framework that maintains long-horizon visual persistence and episodic recall under a bounded cache budget by assigning in-distribution virtual positions to summary slots and manipulating keys in canonical space.

Advancing Narrative Long Video Generation via Training-Free Identity-Aware Memory

IAMFlow introduces a training-free autoregressive long video generation framework that constructs an explicit global entity registry to track persistent character identities across evolving multi-prompt streams, combining attention-key greedy retrieval, adaptive prompt transitions, and asynchronous visual verification to achieve superior consistency while accelerating inference by 1.69×.

AnyMo: Scaling Any-Modality Conditional Motion Generation with Masked Modeling

This paper introduces OmniHuMo, a large-scale human motion dataset comprising over 5,000 hours of motion sequences with aligned text, speech, music, and trajectory annotations, alongside AnyMo, a unified framework combining a Residual FSQ tokenizer and a scalable masked Transformer that enables high-fidelity 3D motion synthesis under arbitrary modality combinations.

AtlasVid: Efficient Ultra-High-Resolution Long Video Generation via Decoupled Global-Local Modeling

AtlasVid decouples ultra-high-resolution long video synthesis into coarse global semantic proxy generation and locality-preserving high-resolution detail denoising, enabling a model trained only at 720P on two GPUs to extrapolate seamlessly to 4K/8K, 321-frame generation with a 60.9× speedup over full attention.

AUTEUR: Language-Driven Cinematographic Framing for Human-Centric Video Generation

Auteur introduces an actor-relative camera framing parameterization and cinematographic DSL, where a fine-tuned multimodal LLM generates sparse directorial keyframe programs from natural language that are deterministically decoded into continuous 6-DoF trajectories, achieving superior framing adherence and drastically reducing out-of-frame errors.

AVENUE: Audio-Video Editing, Understanding, and Evaluation

To overcome the limited edit diversity and modality-blind evaluation of existing audio-video editing systems, AVENUE introduces a benchmark of 1,291 video clips and 7,957 instructions across audio-targeted, video-targeted, and AV-coupled types, paired with a sample-specific MLLM-as-a-judge protocol grounded in explicit change and preservation criteria that systematically diagnoses cross-modal leakage across joint, sequential, and separate editing paradigms.

Bernini: Latent Semantic Planning for Video Diffusion

Bernini unifies video generation and editing by decoupling high-level semantic planning in the MLLM's native ViT embedding space from pixel rendering in a DiT, augmented by Segment-Aware 3D RoPE and Chain-of-Thought reasoning.

Beyond Distribution Matching: Self-Supervised Representation Forcing for Few-Step Video Generation

Addressing the severe spatial blurring and temporal flickering in few-step video distillation caused by output-only supervision, Self Flow Forcing (SF2) introduces self-supervised representation alignment via asymmetric trajectory noise and a self-bootstrapped EMA teacher without changing the output objective.

Beyond Distribution Matching: Self-Supervised Representation Forcing for Few-Step Video Diffusion

Addressing the spatial degradation and temporal flickering caused by output-only supervision in Distribution Matching Distillation (DMD) under aggressive step compression, this paper introduces Self Flow Forcing (SF2), establishing an information asymmetry via heterogeneous noise levels from a shared trajectory to force the student to predict deep EMA teacher representations, substantially enhancing spatio-temporal coherence with negligible overhead.

Bidirectional Sparse Attention for Faster Video Diffusion Training

Addressing the computational bottleneck where full attention accounts for over 90% of training costs in long-sequence Video Diffusion Transformers (DiTs), this paper introduces Bidirectional Sparse Attention (BSA)—the first orthogonally decoupled query-and-KV sparsification framework that prunes redundant spatiotemporal queries and adaptively filters salient KV blocks via score distribution statistics, achieving up to 20× FLOPs reduction and 17.79× faster attention training with zero quality loss.

Breaking the Uniformity Trap: Scaling Video Diffusion Model via SplitMoE

SplitMoE divides the sparse experts of a video diffusion Transformer into semantic and generic groups, guides semantic routing with clean VAE features, and improves video quality over same-source baselines with 27B total and 14B activated parameters, rather than forcing every expert to process an equal number of tokens.

CanvasMAR: Improving Masked Autoregressive Video Prediction With Canvas

Addressing severe distortions in masked autoregressive (MAR) video prediction under aggressive few-step sampling due to the absence of a structured global prior, CanvasMAR predicts a one-step blurred global canvas as a non-uniform mask, paired with motion-aware adaptive sampling and compositional CFG to rival diffusion-level fidelity with far fewer sampling steps.

CausalCine: Real-Time Autoregressive Generation for Multi-Shot Video Narratives

Addressing motion stagnation and semantic drift in autoregressive long-video generation across multi-shot narratives, CausalCine trains a causal base model on native multi-shot sequences, introduces parameter-free Content-Aware Memory Routing (CAMR) with Block-Relative RoPE, and distills the model into a 4-step generator via DMD with adversarial regularization, unlocking 16 FPS interactive streaming generation.

Character Mixing for Video Generation

Addressing the core bottlenecks where characters from disparate worlds and visual styles fail to co-occur and interact naturally, this paper proposes a fine-tuning framework integrating Cross-Character Embedding (CCE) and Cross-Character Augmentation (CCA) that eliminates cross-domain style delusion while achieving high-fidelity multi-character interaction video synthesis.

CineOrchestra: Unified Entity-Centric Conditioning for Cinematic Video Generation

CineOrchestra unifies visual subjects, camera movements, and shot transitions into timestamped entity-centric conditioning primitives, leveraging parameter-free interval-sampled temporal RoPE and 2D entity-temporal cross-attention RoPE within a single video DiT to jointly achieve multi-subject personalization, temporal event timing, multi-shot composition, and directorial camera control.

CityRAG: Stepping Into a City via Spatially-Grounded Video Generation

CityRAG retrieves geo-registered street view panorama sequences as an external spatial memory, using temporally unaligned training within a Wan 2.1 DiT to semantically disentangle static geometry from transient environmental factors, enabling minutes-long 3D-consistent walkthrough video generation that adheres to user trajectories and physical real-world layouts.

Coarse-to-Real: Generative Rendering for Populated Dynamic Scenes

C2R introduces a two-stage generative rendering framework that takes lightweight coarse 3D simulations to control scene layout, camera trajectories, and crowd dynamics, transforming highly abstract 3D proxy renders into high-fidelity, temporally consistent real-world urban videos via self-supervised implicit spatio-temporal features and a synthetic-real domain-hedging strategy.

CoDMD: Copula-aware Distribution Matching Distillation for Fast Video Generation

Addressing the severe motion degradation and mode collapse of standard Distribution Matching Distillation (DMD) under few-step video generation budgets, this paper introduces a copula-aware perspective via Sklar's theorem and proposes CoDMD—a lightweight relational regularizer that reuses score predictions across batch elements and temporal frames to achieve 25× acceleration and state-of-the-art 4-step synthesis.

CompAdapt: Adaptable Composite Motion Modeling for Physics-Consistent Text-to-Video Generation

CompAdapt introduces a decoupled language-to-physics-to-video pipeline for physics-consistent video synthesis, leveraging dual-branch LLM parsing, 10D composite dynamics (MTC superposition and MOT collision impulse mapping), spatially warped latent diffusion guidance, and one-shot dynamics prior matching for novel physical environments.

Decoupling Action from Egocentric Observation for World Simulation

EgoACT proposes a training-free test-time optimization framework for egocentric action transfer that decouples action semantics from scene appearance via early velocity-guided structure anchoring and mid-to-late sparse correspondence calibration, accompanied by the comprehensive benchmark EgoActionBench.

Diffusion-DRF: Free, Rich, and Differentiable Reward for Video Diffusion Fine-Tuning

Diffusion-DRF introduces a preference-data-free and reward-model-training-free alignment framework that uses a frozen off-the-shelf VLM to supply multi-dimensional structured queries and dense free-form explanations, optimizing video diffusion models via end-to-end differentiable backpropagation to overcome reward hacking and optimization collapse.

Do Joint Audio-Video Generation Models Understand Physics?

This paper introduces AV-Phys Bench, the first benchmark for evaluating physical commonsense in joint audio-video generation across steady-state, event transition, and environment transition dynamics, along with AV-Phys Agent that integrates deterministic audio DSP measurement tools to expose severe physical consistency failures in state-of-the-art models.

Dynamic Video Generation: Shaping Video Generation Across Time and Space

DVG (Dynamic Video Generation) addresses the heavy spatio-temporal computational bottleneck in video diffusion models by inferring spatial and temporal content demands from early latent sketches, automatically selecting content-aware joint compression strategies, and reshaping latents with anchor-based interpolation and coordinate-hashed re-noising, achieving up to 18× lossless acceleration across diverse architectures without retraining.

DynamicRad: Content-Adaptive Sparse Attention for Long Video Diffusion

DynamicRad grounds content-adaptive attention within a hardware-friendly radial locality prior, introducing dual-mode candidate selection, offline Bayesian Optimization on a physics proxy, and prompt-driven semantic motion routing to achieve over 80% sparsity and 1.7×–2.5× inference speedups on long video diffusion backbones while maintaining temporal coherence.

Dynamics-Aware Sparse Attention for Efficient Autoregressive Video Diffusion

Addressing the severe memory explosion and compute latency in long autoregressive video diffusion, Dynamics-Aware Sparse Attention (DASA) decouples video dynamics from static backgrounds via frame-wise memory management and region-level computation early-exit, paired with dynamics-guided distillation to achieve a 2.16× speedup at 1080P with negligible quality loss.

DynaTokens: Teaching Dynamics to Camera-Controlled Video Models at Test Time

DynaTokens addresses the mutual entanglement between camera trajectory and scene dynamics in camera-controlled video models by introducing lightweight, scene-specific learnable tokens via cross-attention at test time, decoupling global camera movement from spatially localized dynamics using only ~15 training trajectories.

EditDistill: Is It Possible to Guide Video Editing with Image Editing?

EditDistill distills an image edit pair into a compact 1024-dimensional transformation descriptor and injects it into a frozen Wan2.1-I2V-14B video diffusion model via lightweight gated multi-branch modules and DINO structural losses, achieving mask-free, temporally coherent video editing across local modifications, dynamic effects, and global relighting.

EntityBench: Towards Entity-Consistent Long-Range Multi-Shot Video Generation

To tackle identity drift across characters, objects, and locations in long-range multi-shot video generation, this paper constructs EntityBench—a benchmark of 140 narrative episodes (2,491 shots) with explicit per-shot entity schedules—introduces a three-pillar evaluation framework with 51 metrics and a fidelity gate, and presents EntityMem, a training-free baseline powered by persistent pre-generated entity memory.

Equilibrium Forcing: Adaptive Video Generation Without Noise Conditioning

Addressing error accumulation and schedule divergence caused by rigid external noise conditioning in autoregressive video generation, this paper proposes Equilibrium Forcing (EqF), a noise-unconditional framework that unifies denoising fields across noise levels and leverages internal feature readout for closed-loop, Nesterov-accelerated, budget-adaptive sampling.

FaithfulFaces: Pose-Faithful Facial Identity Preservation for Text-to-Video Generation

To tackle identity distortion and facial structure collapse caused by substantial head rotations and occlusions in complex dynamic scenes, FaithfulFaces introduces a pose-shared identity aligner paired with explicit Euler angle embeddings to extract global facial pose priors from a single reference image, achieving state-of-the-art identity consistency in text-to-video generation.

FashionChameleon: Towards Real-Time and Interactive Human-Garment Video Customization

FashionChameleon introduces the first real-time and interactive human-garment video customization framework based on streaming autoregressive diffusion, integrating an in-context learning teacher model, gradient-reweighted distribution matching distillation, and training-free KV cache rescheduling to achieve 23.8 FPS 720p generation with seamless on-the-fly garment switching on a single H200 GPU.

FlockWorld: Toward Population-Scale Multi-Agent Video World Models through Artificial Life Simulation

Introduces FlockWorld, an artificial-life environment based on Reynolds's Boids, and FlockDiT, a multi-agent flow-matching video transformer that jointly generates synchronized egocentric rollouts across 2 to 30 camera agents, demonstrating key fidelity gains from joint modeling and diffusion forcing.

Focused Forcing: Content-Aware Per-Frame KV Selection for Efficient Autoregressive Video Diffusion

Focused Forcing introduces a training-free KV cache compression framework that pairs per-frame attention scoring with key diversity to decouple temporal bias, allocating head-adaptive budgets via distribution matching loss to deliver up to 1.48× end-to-end acceleration across autoregressive video diffusion models while boosting visual fidelity.

FreeSpec: Training-Free Long Video Generation via Singular-Spectrum Reconstruction

Addressing the spectral concentration issue where extended self-attention windows force latent energy into dominant low-rank singular components—smoothing spatial details and suppressing temporal dynamics—FreeSpec proposes a training-free singular-spectrum reconstruction framework that uses the global branch as low-rank spectral guidance and the local branch as a high-rank reconstruction basis, achieving superior motion dynamics and temporal smoothness across 4× length extensions.

From Articulated Kinematics to Routed Visual Control for Action-Conditioned Surgical Video Generation

To resolve the representational mismatch between low-dimensional robot kinematics and high-dimensional surgical scene evolution, KVLR lifts articulated controls into a 9-channel pixel-aligned Kinematic-to-Visual Action Field (KVA-Field) and adaptively routes computation across physical modalities and motion scales, delivering state-of-the-art action faithfulness and visual fidelity alongside a \(4\times\) latency reduction via significance-driven sparse distillation.

From Zero to Hero: Training-Free Custom Concept Spawning in World Models

The paper presents SPAWN, a training-free method for custom concept spawning in autoregressive world models that temporarily swaps the pinned anchor slot in context memory with an external concept latent during a brief injection window, allowing user-specified visual concepts to integrate seamlessly and persist across camera rollouts.

GameVerse: A Minute-Scale Gameplay Dataset for Long-Horizon Interactive World Modeling

GameVerse establishes a minute-scale interactive gameplay video dataset comprising 69.7K clips and 1,655 hours (including 1,358 hours of continuous two-minute sequences), combining scene-consistent segmentation, multi-stage visual/temporal quality filtering, jump- and drift-filtered ViPE camera trajectories, and four-level decoupled semantic captions to substantially improve temporal coherence and geometric controllability in 60-second world model rollouts.

GLARE: Generating Listening Heads with Appropriate Reactions

GLARE models listener feedback such as nodding and smiling as typed temporal events, improving reaction consistency on RealTalk and Seamless through prosody-conditioned flow matching and training-time reaction supervision, while its metrics measure single-reference consistency rather than uniquely correct social responses.

\(h\)-control: Training-Free Camera Control via Block-Conditional Gibbs Refinement

H-Control: Training-Free Camera Control via Block-Conditional Gibbs Refinement

Addressing the trade-off where hard replacement creates boundary artifacts while soft guidance drifts off trajectory, H-Control introduces an inner-loop block-conditional pseudo-Gibbs refinement on the unobserved latents within a training-free sampler, achieving SOTA FVD and trajectory alignment on RealEstate10K and DAVIS without any model fine-tuning.

Holo4D: Holistic 4D Reconstruction as Geometric Control for Video Diffusion

Holo4D leverages persistent 3D trajectories and source appearance from feed-forward 4D reconstruction as a unified 4D scaffold, injecting projected trajectory features into a pretrained video diffusion model along the channel dimension to achieve precise camera control and novel-view synthesis without retraining for different reconstructors.

HOMIE: Human-Object Centric Video Personalization via Multimodal Intelligent Enhancement

HOMIE addresses human-object centric video personalization by retaining the native text encoder for base controllability while leveraging an MLLM via Global Multimodal Guidance and Modality-Reference Embeddings to unify inter-subject interactions and intra-subject multi-view/OCR fidelity enhancements.

I2V-DETACH: Source Grounding Detachment for Unauthorized Image-to-Video Generation

Addressing the rising risk of unauthorized Image-to-Video (I2V) misuse, I2V-Detach proposes an image immunization framework that detaches source grounding by suppressing source-to-video cross-attention while amplifying intra-video dynamic interactions.

I2V-Detach: Source Grounding Detachment for Unauthorized Image-to-Video Generation Defense

To defend against unauthorized Image-to-Video (I2V) misuse, I2V-Detach introduces source grounding detachment by suppressing source-to-video query-key attention retrieval and reinforcing intra-video interactions, fundamentally severing the generated video from the protected source content.

LaMo: Self-Supervised Latent Motion Priors for Physical Realism in Video Generation

Addressing the lack of physical and motion consistency in video diffusion models, LaMo formulates a self-supervised latent motion prior directly over frame-to-frame latent increments, realized via a macro Motion Drift Loss during training and a micro Motion Prior Guidance during sampling, outperforming externally supervised physics baselines.

Learning to Correct Geometry in Generated Videos

To tackle warped structures, inconsistent parallax, and scene layout drift in generated videos under camera motion, this paper presents Video Geometry Corrector (VGC), a post-hoc video-to-video diffusion framework trained on synthetically constructed paired data (VG-Pairs) with auxiliary optical flow cues, achieving strong cross-generator geometric consistency without modifying source generation models.

LongScape: Advancing Long-Horizon Embodied World Models with Context-Aware MoE

LongScape proposes an embodied world model combining intra-chunk diffusion denoising with inter-chunk causal autoregression, introducing an action-prior-guided variable-length chunking mechanism and a Context-Aware Mixture-of-Experts (CMoE) framework to achieve visually coherent long-horizon generation across 15 iterative rollouts.

Meta Inverse Prompting for Video Generative Models

This paper introduces Meta Video Inverse Prompting (MVIP), the first practical framework for video inverse prompting that meta-learns a shared soft-prompt initialization across video tasks, enabling fast prompt inference that balances temporal dynamics and reconstruction fidelity without costly test-time iterative optimization.

MEV: A Multi-Event Video Dataset for Long-Take Generation

Addressing the lack of fine-grained temporal boundaries and hierarchical supervision in long-take video generation, MEV provides over 8K high-definition videos with 25K+ atomic event segments, a comprehensive global-event annotation suite, and the event-centric MEV-Eval protocol, enabling Wan2.1 fine-tuning for coherent long-take multi-event synthesis.

Motion Forcing: Decoupling Ego and Object Motion via Sparse Inputs for Structured Video Generation

Motion Forcing first converts sparse object trajectories and camera motion into dynamic depth, then renders RGB video with a shared diffusion backbone, achieving FVD 157.8, FVMD 205.2, and Physics-IQ 33.2 on Waymo while trading some distributional visual similarity for better motion consistency and physical plausibility.

MotionGrounder: Grounded Multi-Object Motion Transfer via Diffusion Transformer

Addressing the single-object limitation and lack of semantic spatial binding in Diffusion Transformer (DiT)–based video motion transfer, MotionGrounder introduces Flow-based Motion Signals (FMS) and Object-Caption Alignment Loss (OCAL) to achieve training-free, controllable multi-object motion transfer.

MotionGrounder: Grounded Multi-Object Motion Transfer via Diffusion Transformers

Addressing the fundamental bottlenecks where existing DiT-based video motion transfer methods remain constrained to single-object scenes and lack fine-grained semantic grounding, this paper presents MotionGrounder, the first training-free framework enabling multi-object grounded motion transfer by pairing a Flow-based Motion Signal (FMS) for stable motion priors with an Object-Caption Alignment Loss (OCAL) for explicit spatial-textual binding.

NEvo: Neural-Guided Evolutionary Video Synthesis for Dynamic Visual Selectivity

Addressing the bottleneck where traditional model-driven neural stimulus synthesis is confined to static images, NEvo presents the first framework combining multi-layer dynamic brain encoding models with two-stage structured prompt evolutionary search, synthesizing dynamic video probes that surpass natural videos and classical functional localizers across the human visual cortex while mapping the social dynamic functional gradient of the lateral pathway.

OmniMem: Scalable and Adaptive Memory Retrieval for Long Video Generation

Addressing the twin bottlenecks of local bias and union explosion in chunk-based autoregressive video generation, OmniMem proposes an explicit full-range memory retrieval framework combining adaptive window exclusion, query-shared retrieval, and per-head scattered access, boosting Dynamic Degree by 52.3% with only 1.7% extra VRAM overhead.

OmniVerifier-C1: Decomposed Graded Verifier for Rule-Governed Generative World Modeling

The paper introduces OmniVerifier-C1, a decomposed graded visual verifier that replaces coarse binary judgments with fine-grained sub-question decomposition and continuous scoring optimized via decoupled reinforcement learning, and further proposes a pairwise ground-truth reference paradigm that drastically enhances reward reliability for rule-governed generative world modeling.

ONE-SHOT: Compositional Human-Environment Video Synthesis via Spatial-Decoupled Motion Injection and Hybrid Context Integration

Addressing the trade-off between structural control and generative flexibility, cumbersome 3D pre-processing, and long-horizon drift in human-centric video synthesis, ONE-SHOT introduces a canonical-space motion injection mechanism, Dynamic-Grounded-RoPE, and hybrid context integration, achieving minute-level compositional video generation with parameter-efficient LoRA tuning.

OpenCoF: Learning to Reason Through Video Generation

Introduces OpenCoF-17K, a diverse 17K-video temporal reasoning dataset spanning 11 task families, alongside complementary Visual (vt) and Textual (tt) Reasoning Tokens that significantly enhance Chain-of-Frame reasoning in the Wan2.2-I2V-A14B video diffusion model.

PanoWorld: Geometry-Consistent Panoramic Video World Modeling

PanoWorld reformulates panoramic video generation as geometry- and dynamics-consistent latent world-state modeling, incorporating spherical adaptation (latitude-aware RoPE, area weighting, circular shift) and offline pseudo-labeled depth and 3D trajectory consistency losses onto a pretrained perspective DiT to synthesize coherent 360° video from a single image and prompt.

PermaVid: Consistent Video Generation Across Edits via Disentangled Context Memory

PermaVid disentangles spatial context into an RGB appearance memory and a depth geometry memory, combined with an edit-aware invalidation and retrieval mechanism, to maintain long-term spatiotemporal consistency across viewpoint revisits and video edits.

Persistent Planar Memory for Video World Models

PPMem adopts compact parametric plane primitives as persistent 3D spatial memory for video world models, jointly predicting video and metric planar geometry in an autoregressive chunked diffusion framework and conditioning future chunks via forward-rendered depth, maintaining cross-view visual and geometric coherence across hundreds of frames with under 5 MB of 3D memory.

PhyMetric: Diagnosing Physical Plausibility in Text-to-Video Generation via Scene-Level QA

PhyMetric reframes physical plausibility assessment for text-to-video (T2V) models into 8,920 balanced binary yes/no questions over 1,115 scenes across 50 categories and six physical domains, combining structured physical priors with question-aware frame sampling to eliminate evaluator drift and provide interpretable failure diagnosis.

PhyProbe: Rethinking Physical Consistency Evaluation in Generated Videos

PhyProbe trains a lightweight scoring head on a frozen video encoder, combining ranking, severity regression, and endpoint anchoring from heterogeneous sources to score individual videos; it achieves the highest pairwise accuracy on three of four physical benchmarks, but remains a proxy for perceived plausibility rather than a test of physical laws.

Physics-Informed Video Generation via Mixture-of-Experts Latent Alignment

To tackle physically implausible dynamics in large-scale video generators without explicit simulation, PILA establishes an operational physical attribute bank anchored by observable motion in the frozen latent space and activates category-specific MoE residual operators to decode a lightweight vector-field correction.

PickMoment: Single-Image-to-Video via Learning Continuous-Time Reblurring

PickMoment establishes an algebraic duality between the physical integral formation of motion blur and the average velocity field of MeanFlow, training a single deterministic model to predict interval-mean blurs over arbitrary exposure sub-intervals and unifying deblurring, continuous-time video generation, and pick-a-moment recovery in a single forward pass.

PISCO: Precise Video Instance Insertion with Sparse Control

PISCO propagates a few instance keyframes into existing footage through variable-density conditioning, pre-encoding frame completion and post-encoding masking, and depth and appearance augmentation; on PISCO-Bench, first-and-last-frame control with its 14B model reduces whole-video FVD from VACE's 371 to 204, although the compared methods receive different inputs.

Proprio: Latent Self-Scoring and Inference-Time Refinement for Physically Plausible Whole-Body Pose Estimation

This paper proposes Proprio, a training-free framework that leverages a frozen video generator's flow prediction residual under controlled latent perturbations as an intrinsic self-score of physical plausibility, guiding test-time candidate search and gradient-based initial noise refinement.

Q-ARVD: Quantizing Autoregressive Video Diffusion Models

Q-ARVD is the first dedicated post-training quantization framework for autoregressive video diffusion models, eliminating compounding temporal errors and heterogeneous weight outliers via final-quality guided frame weighting and adaptive dual-scale quantization to achieve near-lossless INT8/INT4 streaming inference.

RAVEN: Real-time Autoregressive Video Extrapolation with Consistency-model GRPO

Addressing the train-test history discrepancy and the lack of end-to-end history supervision in causal autoregressive video diffusion distillation, RAVEN repacks self-rollouts into an interleaved sequence to supervise cached history representations and introduces CM-GRPO to apply online RL directly on the consistency Gaussian transition kernel, achieving leading real-time streaming video generation.

Rebalancing Reference Frame Dominance to Improve Motion in Image-to-Video Models

DyMoS subtracts an adjustable bias from non-reference-query-to-reference-key self-attention logits in the text-conditioned branch during early image-to-video denoising, raising Wan 2.2's VBench Dynamic Degree from 51.7 to 64.8, although stronger dynamics do not automatically imply more realistic motion or higher reference fidelity.

RISE-Video: Can Video Generators Decode Implicit World Rules?

Presents RISE-Video, a cognitive reasoning-centric benchmark for Text-Image-to-Video (TI2V) evaluation across 8 knowledge domains with 467 expert-annotated samples, decoupling reasoning alignment, temporal consistency, physical rationality, and visual quality to expose severe reasoning deficits in state-of-the-art video generators.

RoboAlign-R1: Distilled Multimodal Reward Alignment for Robot Video World Models

Addressing reward misalignment with control utility and long-horizon autoregressive error compounding in robot video world models, RoboAlign-R1 introduces a distilled multimodal reward model for efficient GRPO post-training alongside a training-free sliding window re-encoding decoding strategy, significantly improving physical fidelity and rollout stability.

SierpinskiCam: Camera-Controlled Video Retaking with Sierpinski Triangle Pattern Cues

Addressing the failure of geometry-guided video retaking under large viewpoint departures, SierpinskiCam augments forward-warped 3D proxies with multi-scale trackable Sierpinski fractal dome textures and introduces position-disentangled Negative RoPE (NegRoPE) for architecture-preserving reference conditioning.

SocialDirector: Training-Free Social Interaction Control for Multi-Person Video Generation

Addressing actor-action mismatch and wrong action target issues in multi-person video generation, SocialDirector introduces a training-free controller that modulates cross-attention maps via spatiotemporal masking and directional reweighting, achieving high-fidelity social interaction control with less than 3% computational overhead.

SOLARIS: Building a Multiplayer Video World Model in Minecraft

SOLARIS presents the first multiplayer video world model capable of simulating consistent multi-view perspectives in an unbounded 3D Minecraft world, powered by a containerized multi-agent data engine SolarisEngine, an interleaved multi-player self-attention DiT, and a novel Checkpointed Self Forcing training scheme.

Sparkle: Realizing Lively Instruction-Guided Video Background Replacement via Decoupled Guidance

Addressing the static backgrounds and structural collapse inherent to conventional mixed generation pipelines, this paper proposes a five-stage data synthesis pipeline driven by decoupled foreground and background dual guidance, curating ~140K high-quality video pairs in the Sparkle dataset alongside the Sparkle-Bench benchmark to substantially boost downstream editing model performance.

Sparse Video Generation Propels Real-World Beyond-the-View Vision-Language Navigation

To tackle dead-end trapping and unexpected spinning caused by short-horizon supervision in LLM-based navigation, this paper introduces SparseVideoNav—the first world model navigation system driven by sparse video generation—achieving a 20-second prediction horizon and sub-second continuous action inference via a four-stage progressive training and distillation pipeline.

SpecLoR: Spectral Lookahead Rectification for Motion-Coherent Text-to-Video Generation

To eliminate spatiotemporal inconsistencies and physical artifacts caused by accumulated ODE discretization and velocity approximation errors in Flow Matching video generation, SpecLoR decouples the 3D spatiotemporal spectrum via single-step lookahead projection during early sampling, locking the sensitive geometric phase while rectifying the macroscopic amplitude toward natural video priors, steering the drifted trajectory back onto the natural video manifold with only 4 extra NFEs.

SPIRAL: Self-Evolving Action-Conditioned Video Generation via Reflective Planning

SPIRAL introduces an agentic closed-loop framework coordinating planning, step-wise generation, multi-dimensional critique, and long-horizon memory under a think-act-reflect paradigm, enabling underlying video diffusion models to self-evolve continuously via curriculum GRPO.

Teaching Video Generators to Remember: Eliciting Dynamic Memory for Out-of-Sight State Evolution

ReMind trains video generators to recover states across interruptions using event frame graphs, reliable-anchor training, and KV caches that retain original spatiotemporal addresses, reaching the highest STEVO-Bench Total of 300.9, but only 14.5% State Progress; the evidence primarily supports short-horizon recovery rather than long-term unobserved physical evolution.

Temporal Concentration from Rollout Errors: Implicit Preference Optimization for Text-to-Video Generation

Addressing the preference attribution bottleneck and temporal credit misallocation in video alignment, this paper proposes concentrated Implicit Preference Optimization (cIPO), which constructs implicit preference pairs via diffusion denoising rollouts and concentrates DPO updates onto contiguous high-error temporal windows.

The Invisible Hand of Physics: When Video Diffusion Models Know More Than They Show

By approximately inverting the flow-matching velocity field backward from clean videos to noise, this paper discovers that video diffusion models linearly encode physical plausibility and continuous kinematic variables inside their intermediate transformer blocks (reaching 81.27% accuracy on IntPhys), proving that physical representations emerge naturally as a byproduct of generative denoising without explicit predictive training.

The Surprising Effectiveness of Video Diffusion Models for Hand Motion Reconstruction

ViDiHand is the first method to repurpose an internet-scale pretrained video diffusion model as a representation backbone for 4D two-hand motion reconstruction, adapting it via a hand-overlay rendering pretext task and pairing it with a dual-branch decoder and closed-form mixed-projection solver to recover metric-scale 4D hand trajectories directly on full frames without detectors, infillers, or test-time optimization.

TrajLoc: Trajectory-Attention Localization for Multi-Object Motion Control

TrajLoc introduces a sparse multi-object trajectory control paradigm without dense conditioning tensors or auxiliary modules, directly replacing cross-attention weights of object text tokens with spatiotemporal Gaussian heatmaps in DiT attention layers alongside learned trajectory and appearance embeddings, achieving robust, occlusion-resilient motion control for up to 20 objects.

TrajLoc: Trajectory-Attention Localization for Multi-Object Motion Control

TrajLoc departs from traditional dense conditioning tensors by directly substituting cross-attention weight columns with spatiotemporal Gaussian heatmaps centered along target trajectories, paired with dedicated trajectory and appearance tokens in text prompt space, achieving state-of-the-art multi-object motion control without extra dense parameters.

Transforming Image Editors into Video Editors

The paper proposes Anchor-based Video Editing (AVE), a two-stage framework that decomposes video editing into composed keyframe editing and motion-guided interpolation, inheriting the capabilities of mature image editors while training only a 710M-parameter motion encoder.

TriTD: Tri-Partite Trajectory–Distribution Distillation for Real-Time Autoregressive Video Generation

Addressing the quality-efficiency bottleneck in few-step autoregressive video diffusion, TriTD establishes closed-form cumulative error bounds and introduces a tri-partite trajectory-distribution distillation framework, achieving 26.8 FPS real-time video generation on a single H100 GPU with 4 NFEs and state-of-the-art visual quality.

Ultra Flash: Scaling Real-Time Streaming Video Generation to High Resolutions

Ultra Flash introduces a cascaded streaming framework combining an architecture-preserving generative super-resolution paradigm, a lightweight causal latent upsampler, and cascaded self-forcing preference optimization, achieving real-time streaming video generation at ~30 FPS at 1K and ~18 FPS at 2K resolution on a single GPU.

VINCIE-NExT: Unlocking Video Editing from Images via In-Context Modeling

VINCIE-NExT decomposes video editing into a structured \(\text{Video} \to \text{Image} \to \text{Image} \to \text{Video}\) sub-task chain, transferring mature image editing priors to videos via in-context visual demonstrations, Time Dual-Frequency 3D RoPE, and a two-stage Chain-of-Editing inference framework.

ViTeX-Bench: Benchmarking High-Fidelity Video Scene Text Editing

ViTeX-Bench uses paired training data from real videos, a frozen evaluation split, and 13 metrics across three axes to distinguish text correctness, motion stability, and background preservation, while providing a motion-aligned glyph-conditioned reference editor, ViTeX-Edit-14B, with the highest mean character accuracy among the evaluated video-native editors rather than the best performance on every metric.