Overview

Video diffusion models synthesize or transform frame sequences by denoising video latents, often in diffusion transformers, conditioned on text, images, reference clips, or structured controls. Researchers use temporal attention, autoregressive chunking, and pose, camera, depth, or physics cues to improve motion, identity, geometry, and interaction; related systems support editing, restoration, avatars, and world-model prediction. Evaluation spans prompt alignment and visual quality, using benchmarks such as VBench and FVD, as well as controllability, physical plausibility, and downstream action performance. A central challenge is balancing fidelity and precise control over long videos with the speed and memory needed for streaming, while limiting drift and physically implausible events.

Latest papers

812 papers72 papers in the last four weeks, down 36% on the four weeks before. 0.5% of all new papers.

Nov 19, 2026

AURA: AUdio-dRiven streaming Avatar

Current audio-driven video diffusion models can synthesize realistic talking-head videos but generally require dozens of diffusion denoising steps over multi-second video chunks, making them unsuitable for streaming applications such as voice assistants and virtual avatars, where low-latency and online causal generation are required properties. AURA generates video chunk by chunk with a rolling KV cache, requiring only the audio of the current chunk to start generating, rather than a multi-second audio look-ahead like bidirectional models. It comes in two sizes: AURA 1.3B streams 576p video in real time on a single GPU (20.7 fps for a 16 Hz playback rate), making it suited to interactive avatar applications, and AURA 5B streams 720p video at 9.7 fps. Our two-stage distillation pipeline first aligns the causal student with the teacher, by training it to match the teacher's trajectories with fewer diffusion steps and a chunk-wise sliding window causal attention mask, then applies Distribution Matching Distillation (DMD) with backward simulation to close the remaining quality gap. On standard talking-head benchmarks, AURA 1.3B runs 23x faster than its own bidirectional teacher and reaches a quality comparable to TalkVerse at a lower resolution, and AURA 5B matches its TalkVerse teacher at 720p while running 17x faster. To support reproducibility and further research on streaming avatar generation, we release the weights of both models along with a dataset of 30K synthetic videos and their precomputed teacher trajectories.
Nov 19, 2026

ViDecide: Teaching Video Diffusion Models to Reason and Plan Step by Step

Current leading video generation models can synthesize highly realistic videos, while still frequently violating physical and logical constraints in multi-step reasoning and planning, even after training with large-scale reasoning data or existing diffusion RL methods. We attribute these failures to the conventional video training paradigm, which treats a solution as a fixed-length clip rather than a chain of decisions. We present ViDecide, a decision-centric framework that trains video diffusion models to predict state transitions causally and decide when to stop, with a decision-focused training schedule. We further propose Corrective Fine-tuning (CFT), which corrects failed decisions by mapping their noise to the model's own successful continuations, achieving stable improvement without policy-gradient optimization. Across abstract games, planning tasks, and a new benchmark of multi-step robot planning, ViDecide achieves an average strict success rate of 83.8%, far above 23.3% for standard fine-tuning and 67.8% for the strongest prior method, and generalizes to larger puzzles and unseen layouts. Our framework therefore unlocks greater potential for video generation models as reliable reasoners and planners in the visual world. Project page: https://kevin-thu.github.io/ViDecide.
Nov 19, 2026

Online Temporal Rotations for 2 Bit KV-Cache Quantization in Video Diffusion Transformers

Video diffusion transformers reuse key--value (KV) caches across denoising steps, which quickly becomes a bottleneck for long video generation. We find that quantizing temporally distant cache vectors leads to a large range and high quantization error, but that nearby video tokens in the KV-cache share a persistent pattern per channel. These patterns also vary across videos, motivating rotations that are adapted to each video's unique cache. We propose Temporal Rotation Quantization (TRQ), a method that rotates frames in the cache that are adjacent in time before quantization. For each block of temporally adjacent cache vectors, TRQ fits an orthogonal transform that maps the block's per-channel means to a vector that is close to a constant across channels. To keep the quantization format lightweight, we parameterize this rotation as a tree of 2×22 \times 2 planar rotations, each stored with only two bits, followed by a fixed Hadamard transform shared across all blocks. The rotations are fit without gradients or changes to model weights, and we prove that a simple greedy fit is globally optimal for our 'mean-flattening' objective. The transforms are fit online for each new video, and we release fused Triton kernels to make this process efficient at inference. At 2-bit precision and matched compression, TRQ improves VBench subject consistency over the state-of-the-art method QVG from 90.38 to 92.02 on LongCat-Video-13B and from 86.33 to 87.61 on HY-WorldPlay-8B, as well as improving other metrics.
Nov 19, 2026

FlexVSR: Continuous-Scale Super-Resolution for Streaming Video Generation

Real-time streaming video generation has gained rapid momentum in live content creation and interactive media. Deploying such systems in practice requires flexible scaling factors to accommodate diverse source resolutions, target displays, and computational budgets. However, extending fixed-scale generative video super-resolution (VSR) to varying target scales poses two challenges: downsampling from a larger fixed scale incurs redundant cost, while re-optimizing the diffusion transformer (DiT) is costly and alters the shared generative prior. To address both, we propose FlexVSR, a decoupled framework for continuous-scale streaming generative VSR with supervision at only a few scales. Our key insight is that naively resampling fixed-scale latents onto a new target grid induces scale-dependent input mismatch, which can be effectively addressed through parameter-efficient, scale-aware adaptation of the latent interface without updating the DiT. To this end, FlexVSR introduces ReLSA (Refined Latent Scale Adaptation), a lightweight latent adapter that constructs and refines target-grid latents before they enter the DiT, enabling a shared frozen VSR prior to generalize across continuous scales with only 0.274M trainable parameters and an adaptation cost of 7.3 H100 GPU-hours, while introducing no additional temporal dependencies and preserving the native causal streaming protocol. Extensive experiments demonstrate competitive or superior reconstruction and temporal quality, lower latency than most native-scale streaming baselines and offline methods, and effective generalization to unseen continuous scales, filling the continuous-scale gap in real-time streaming generative VSR.
Nov 19, 2026

StreamEgo: Streaming Exocentric-to-Egocentric Video Generation

Exocentric-to-egocentric (Exo2Ego) video generation transforms third-person recordings into first-person viewpoints, unlocking immersive AR/VR experiences and scalable demonstration synthesis for embodied robotics. Extending it to a streaming setting is expected to further enable live applications. However, existing methods suffer from degraded manipulation regions, overfitting and confinement to fixed-length clips with prohibitive latency. To address these limitations, we present StreamEgo, a real-time framework for unbounded, generalizable Exo2Ego video generation. To resolve intricate manipulation dynamics while supporting the streaming pipeline, a streaming dual-branch prior provides pixel-aligned guidance for translating appearance and skeletal kinematics. We also introduce a region-aware loss curriculum to improve generalization and stabilize multi-objective diffusion training. Furthermore, we distill the model into a chunk-wise autoregressive model, augmented with a region-aware corruption mechanism and a first-chunk curriculum, enabling high-quality, generalizable real-time streaming Exo2Ego generation at 24.3 FPS on a single H100 GPU.
Nov 19, 2026

MultiBinder: Preventing Multi-Subject Identity Leakage in Personalized Video Generation

Personalized video generation faces a significant scalability challenge: as the number of custom subjects increases, existing methods suffer from severe identity leakage and semantic entanglement, resulting in videos with blended attributes and incoherent narratives. We identify that this stems from a lack of explicit binding between textual entity phrases and their corresponding visual patterns prior to fusion in diffusion models. To address this, we propose MultiBinder, a novel framework built on a bind-then-narrate'' paradigm. Our approach introduces a two-stage conditioning process before video diffusion. First, Cross-modal Entity Binding devises an entity-aware Q-Former to extract identity-centric features from each subject's reference image and tightly align them with its unique entity phrase, yielding disentangled identity tokens.'' Second, a Multimodal Large Language Model (MLLM) acts as a narrative director'' in the Narrative-Aware Subject Arrangement stage, capturing the action of each subject from the text prompt and synthesizes contextualized subject features that encapsulate who is doing what'' and preserve role-specific distinctions. These features are then injected into a video diffusion model via dedicated conditioning pathways, ensuring precise attribute allocation. Extensive experiments on both general and newly curated multi-subject benchmarks show that MultiBinder significantly outperforms existing methods in multi-subject identity fidelity.
Nov 19, 2026

Frequency-Synchronized Boundary Coupling for Training-Free Multi-Prompt Long Video Generation

Text-to-video generation has made remarkable progress with diffusion models and transformer-based video architectures, yet generating coherent long videos from evolving textual descriptions remains challenging. Training long-video models from scratch requires substantial computation and large-scale multi-event data, while training-free multi-prompt generation must activate new prompt semantics without breaking temporal continuity. We identify a frequency-agnostic boundary coupling problem in existing transition strategies: weak coupling causes abrupt changes, flicker, and background jumps, whereas overly strong or coarse coupling may over-preserve source semantics and produce ghosting or duplicated subjects around prompt boundaries. To address this problem, we propose FreqSync, a training-free framework that formulates prompt transitions as frequency-selective boundary coupling. FreqSync combines Frequency-Synchronized Source-Conditioned Attention (FS-SCA) with Residual-Clipped Spectral Stitching (RCSS), enabling adjacent segments to share stable low- and mid-frequency structure and motion cues while keeping high-frequency prompt-specific details target-dominant under explicit residual constraints. Extensive experiments show that FreqSync achieves state-of-the-art transition smoothness and prompt alignment while maintaining strong perceptual video quality, demonstrating its effectiveness for coherent multi-prompt long video generation.
Nov 19, 2026

ContextRenderer: Dynamic-Context Video Generation for Interactive Game Worlds

Interactive game world models separate executable world evolution from visual generation, using video models to render engine-provided scene representations into detailed observations. During sustained gameplay, however, existing generative renderers struggle to follow changing conditions while preserving visual continuity across diverse proxies and target styles. We therefore introduce ContextRenderer, an autoregressive video diffusion framework designed for both dynamic conditioning and cross-proxy generalization. It couples Dynamic-Context Learning (DCL) for chunk-level interaction supervision with Recurrent History Compression (RHC) to follow condition updates while preserving visual continuity within a bounded token budget. Across proxy representations, Cross-Representation Flow Alignment (CRFA) connects proxy reconstruction with appearance and interaction variation through a shared high-fidelity reconstruction. WorldPrism, our interaction-centric dataset constructed from agent-built worlds, AAA gameplay, and open-source videos, supplies the corresponding renderings and chunk-level annotations for these learning tasks. Experiments on GWM-Bench and Solar-Bench show improved reconstruction and visual quality, while qualitative results demonstrate minute-long interactive rendering under changing prompts and subject references. Project page: https://rainzor.github.io/contextrenderer/
Nov 19, 2026

What One Sample Cannot See: Why Fewer Denoising Steps Can Score Higher on Physics Benchmarks

When a storm can head north or south, a forecast down the middle may score better under squared error than either plausible path. The same paradox appears when video generators are tested as world models: fewer denoising steps can raise physics scores even as videos blur or stop moving early. Why? Physics-IQ compares each generated video with one recording, although a scene can have multiple valid futures. An exact decomposition shows its scores can reward reduced variation among generated futures, even when their distribution becomes less accurate. We introduce Kernel Physics-IQ to remove this incentive in expectation, needing only two independent samples per condition. Two-step sampling lowers Kernel Physics-IQ in all five image-to-video models. For CogVideoX, the official score rises by 0.055 while ours falls by 0.131. Changing the sampling schedule reveals further differences official scores miss. In controlled physical systems with known outcome probabilities, our metric favors the true distribution over collapsed predictions. In short, a world model should be judged not only by how closely one video matches what happened, but by how faithfully its predictions represent what could have happened.
Nov 18, 2026

H3Remover: Reinforcing Video Object Removal without Paired Data

Pretrained video editors possess rich removal and completion priors, yet may retain targets or alter unrelated content. We present H3Remover, which reinforces MiniMax-H3 Base Ref2VA for video object removal without paired object-free training targets. To discourage superficial edits such as recoloring, a target-removal reward favors calibrated depth changes consistent with revealing farther content. To limit unintended scene changes, a background-geometry reward encourages preservation of relative depth ordering. To preserve background appearance without retaining unwanted shadows or reflections, an appearance reward constrains RGB changes in source-derived protected regions while exempting targets and estimated effects. Frozen HPSv3 complements these constraints with perceptual-quality feedback for synthesized content. Evaluation regions remain fixed across candidates, and DiffusionART optimizes LoRA adapters with the backbone frozen. Experiments show substantial reconstruction gains over original H3, with qualitative comparisons demonstrating more effective object removal and better preservation of surrounding scene content.
Nov 18, 2026

Every Query Routes Itself: Accelerating Video Diffusion Transformers with Per-Query Routed Sparse Attention

Self-attention over long token sequences dominates the inference cost of video diffusion transformers. Clustering-based sparse attention clusters both queries and keys and lets all queries in a cluster share one selection of key clusters. We show that this sharing limits accuracy: with key clusters and density fixed, attention grows more accurate as the query clusters are refined, and is most accurate when each query selects on its own, which needs no query clustering at all. We therefore propose Per-Query Routed Sparse Attention (PQRS), a training-free method that clusters only keys, lets each query select its own key clusters, and compensates for the unselected ones with their centroids. Per-query selections leave no query tile with a shared key list, so PQRS traverses key clusters instead of queries, buckets queries with identical selections into dense tasks, and merges their partial results in one pass under a fixed softmax offset carried from the previous denoising step. On four text-to-video and image-to-video models from the Wan and HunyuanVideo families, PQRS is the fastest of all the sparse attention methods we evaluate and the closest to the dense output, reaching 28.928.9 dB PSNR on Wan2.1-T2V and 35.435.4 dB on HunyuanVideo at 1.661.66--2.10×2.10\times end-to-end speedup.
Nov 18, 2026

Probing Physics in Motion Diffusion Models Against Their Own Estimate

Recent studies report that video diffusion models "know more than they show": physical properties decode better from their activations than from their inputs or outputs. But a diffusion model computes its clean-sample estimate from those same activations, so such a comparison measures how easily a readout extracts the information. Whether the model hides anything is a separate question. We set out to find hidden physical knowledge in kinematic text-to-motion diffusion models, where the physics of every intermediate estimate can be computed exactly, and recorded every layer along the sampling trajectories of MDM (two checkpoints) and MotionDiffuse. Against the usual baselines the layers do look ahead of the model's own estimate on balance and flight. With every readout tuned the same way, the estimate closes the gap: for balance and flight, adding a layer gains at most 0.01, a size that a positive control shows we would detect. The fall rate of jumps is the one clear exception. Early in denoising, MDM's layers add about 0.08 in R2R^2 (up to 0.11 in a refit) that the estimate does not yet show, and the gain fades as sampling proceeds; on MotionDiffuse it is small and not robust. The estimate also flags floating samples without training, but aborting them saves under a tenth of the compute. In a video model, plausibility that linear probes miss in Wan 2.1's clean latent becomes readable after one fixed nonlinearity.
Nov 18, 2026

Spark-H3: Adaptive Block Sparse Attention for Video Diffusion Models

Long video sequences make attention a substantial cost in video diffusion models. Block sparse attention (BSA) reduces this cost by computing selected blocks exactly while compressing the remaining blocks. However, fixed token grouping can place tokens with dissimilar attention preferences in the same group, while uniform key averaging can underestimate the attention mass of compressed blocks. We introduce Spark-Attn, a training-free BSA method comprising Spark-Reblock and Spark-Reweight. Spark-Reblock recursively groups tokens by attention preference while preserving regular block shapes for efficient GPU execution. Spark-Reweight uses the global mean query as an anchor to construct attention-weighted key-value representations and logit biases for the compressed blocks, calibrating their attention mass. We instantiate Spark-Attn in MiniMax-H3 as Spark-H3. At 10% exact-block density, Spark-H3 delivers a 2.49×\times DiT speedup and reaches 24.81 dB PSNR against the corresponding dense outputs on MiniMax-H3 (20 denoising steps, 345 frames, 1344×\times768). Increasing the density to 20% further improves PSNR to 26.76 dB while still achieving a 2.26×\times DiT speedup. These results demonstrate that Spark-H3 achieves a state-of-the-art quality-efficiency trade-off for training-free BSA on video diffusion models.
Nov 17, 2026

Autoregressive View Model

We present the Autoregressive View Model (AVM), which generates a scene as a stream of views without constructing an explicit 3D representation. Through a unified in-context interface, AVM conditions on images or videos to generate short clips along a requested camera path. Visual context grounds reconstruction and interpolation, while a learned prior enables extrapolation into unseen regions. AVM retrieves relevant views from inputs and generated history to form a bounded active context. We fine-tune a pretrained bidirectional video diffusion transformer with flow matching in an N-in-1-out formulation. We then tune a causal model on its own rollouts to improve photorealism, reduce denoising steps, and limit long-horizon drift. We scale training to approximately 8M scenes and video clips. We evaluate generation capacity and scene persistence across benchmarks for camera control, text control, visual quality, dynamics, sparse and 4D reconstruction, and long-horizon consistency. The results demonstrate that AVM adapts to diverse input and output settings through one in-context interface.
Nov 17, 2026

FastFrames: Video Generation at Interactive Latency

Autoregressive video diffusion models generate frames based on new observations and inputs, making them a promising medium for interactive visual experiences. However, existing models are orders of magnitude too slow for fluid interaction. This paper introduces FastFrames, a family of fast autoregressive video diffusion models that achieves interactive latency without compromising quality. We profile the inference pipeline of state-of-the-art models and find that operations often treated as fixed implementation details, such as VAE decoding and KV cache updates, in fact account for over half of end-to-end inference latency. We challenge the assumption that these components are indispensable and replace them with lightweight alternatives. Our 1.3B variant runs at 36 ms per frame, or 28 responses per second (RPS), on a single B200 GPU while matching the generation quality of a state-of-the-art model that is nearly 8x slower. Our 14B variant runs at 45 ms per frame, or 22 RPS, on an 8x B200 GPU node and yields significantly better visual quality and realism. We show that FastFrames enables real-time interaction in applications such as camera-conditioned and physics-conditioned video generation.
Nov 17, 2026

When a Larger Guidance Scale Repairs Text-to-Video Failures and When It Does Not

When a text-to-video diffusion model ignores part of a prompt, the usual first fix is to raise the classifier-free guidance scale. This paper is a measurement study of when that fix works. On Mochi, the same increase repairs failures on rare compositions ("a snowy beach") beyond what generating again does, does not repair negations ("a highway with no vehicles"), where it fixes fewer failures than generating again does, and changes nothing for events that must happen late in a clip ("a sandcastle that a wave washes flat"). Which of these happens can be known in advance, from the prompt and one inexpensive score. The score, which we call demand, measures how strongly the denoiser reacts to the part of the prompt that the failed video leaves out, and it is read from the first quarter of denoising, before any video exists. Our main finding is a condition: raising the scale is effective only when the prompt contains no negation and no event tied to a specific time, and when demand is sufficiently high, with recovery clearly higher above a fitted threshold than below it. Published corrections for the three failure types, used as fixed probes, behave as this account predicts. A short decision rule follows: before generating again, a user can tell whether the most common fix is worth trying, and what to use when it is not.
Nov 17, 2026

Harnessing Image Diffusion Prior for Photo-Realistic Video Restoration

Recent image generation models produce photo-realistic results with rich details, but transferring their powerful priors to video restoration remains difficult due to temporal inconsistency from stochastic detail synthesis. Existing video restoration methods improve temporal coherence but often sacrifice visual fidelity, leaving a gap between image-level quality and video-level stability. We propose a generative video restoration method that bridges this gap by enabling temporally consistent detail synthesis from image generation priors. Our method combines noise energy rebalancing to suppress structure-disturbing low-frequency stochasticity, motion-aligned noise warping to propagate details along object motion, a self-supervised denoising encoder to mitigate encoder-induced temporal uncertainty, and decoupled temporal attention to model cross-frame dependencies. Together, these components preserve fidelity, enhance fine-grained realism, and maintain temporal consistency. Extensive experiments show that our method substantially outperforms existing video restoration approaches in visual quality, detail richness, and temporal stability.
Nov 17, 2026

Streaming Generative Video Compression with Embedding Alignment and In-Context Propagation

Video compression at ultra-low bitrates presents substantial challenges in preserving perceptual quality. Although pretrained video diffusion transformers (DiTs) provide powerful generative priors, existing generative video compression (GVC) methods built on them typically rely on computationally expensive VAE decoders to reconstruct video frames. Moreover, jointly refining multiple frames to exploit the inherent spatiotemporal modeling capabilities of video DiTs introduces additional buffering delay. To address these challenges, we propose SGVC, a streaming generative video compression framework that adapts a pretrained latent video DiT into a causal pixel-space refiner. Specifically, SGVC projects decoded features from a contextual pixel codec directly into the DiT input embedding space. To bridge the representation gap between codec-derived embeddings and pretrained DiT embeddings, we introduce Embedding Alignment, which aligns the projected embeddings with input embeddings extracted from the original frames during training. The refined DiT hidden states are then mapped directly to the pixel domain through a lightweight reconstruction path, bypassing the pretrained VAE decoder. To further exploit the spatiotemporal priors of the video DiT while maintaining causal decoding, we introduce In-Context Propagation, which reuses layer-wise K/V states across frames to exploit historical generative context, improving temporal consistency without requiring clip-wise buffering. Extensive experiments demonstrate that SGVC achieves competitive rate-perception performance, with average BD-rate savings of 71.97% over S2^2VC in terms of DISTS. Project Page: https://chigland.github.io/SGVC/
Nov 17, 2026

Evolve in Panorama, Render in View: Panoramic World State Evolution for Video World Models

Video world models should maintain a world that keeps evolving as the viewpoint changes. Yet when the camera turns away and returns, current models often freeze the unobserved scene, lose objects, or change their appearance, because content outside the view has no representation that keeps changing. We present PanoEvo, which uses a Panoramic Evolver to evolve the world state in the latent space of a panoramic video diffusion model, where every viewing direction is present at every time step, so the scene evolves whether or not it is observed. Existing panoramic video models are not trained to evolve a scene from a single image or the objects within it, so we build PanoEvo-Data, an event-centric dataset from real 360° videos, and train the Panoramic Evolver coarse-to-fine as an image-to-panoramic-video model. Since each view covers only a small part of the low-resolution panorama, directly projected views are blurry and often show low-quality details or structural errors. A Perspective Renderer, built on a perspective video diffusion model, reads this state via latent-space projection and direction-restricted cross-attention to render sharp, high-quality views. PanoEvo achieves the best overall score among 23 publicly evaluated models on WRBench and the highest total score on STEVO-Bench. Code, models, and data will be released.
Nov 17, 2026

Decoding Without a Decoder: Direct Pixel Prediction from Diffusion Transformers

Latent diffusion models rely on a decoder to map denoised latents to pixels, adding complexity and inference cost. We introduce Decoding Without a Decoder (DWD), which lets a finetuned diffusion transformer produce final pixels directly through a linear head. Unlike previous methods, where the denoising output leads to final pixels, DWD separates the two roles: the original denoising head continues to drive sampling, then the new DWD head predicts the pixels at the final step. We extend pretrained latent models with DWD efficiently by finetuning only the last few transformer blocks, training the DWD head with pixel quality losses while anchoring the denoising head to the pretrained model to preserve the sample trajectory. Across latent image and video generators, DWD removes the decoder while maintaining or even improving pixel quality. On Wan video generation, it visibly reduces detail distortion and flickering, lowering warping error by 23%. Remarkably, the same separation also improves pixel diffusion models that do not even use a decoder: a final-step DWD head produces substantially better texture detail in text-to-image generation, and improves ImageNet FID from 1.55 to 1.43, a new state of the art among pixel flow models.
Nov 17, 2026

BATMoS: Budget-Aware Guidance Allocation for Efficient Video Diffusion

The effectiveness of a video diffusion model's inference budget depends on where computation is spent along the denoising trajectory. Across eight equal-budget inference configurations, VBench total scores span 4.80 points, highlighting the importance of explicit allocation design. Our key insight is that guidance evaluations are not equally valuable across diffusion time: retaining the sampling updates while selectively reusing guidance can preserve generation quality with fewer denoiser evaluations. We introduce BATMoS (Budget Aware Temporal Mixture of Strategies), which routes between fresh classifier-free guidance and cached guidance using a single pretrained denoiser. A magnitude-based temporal rule assigns guidance reuse to low-magnitude blocks and reserves fresh guidance for the remaining blocks, with a single integer controlling the computation budget. The resulting schedule requires no additional router training or per-budget threshold search. Evaluated on CogVideoX-2B with 4,720 videos per configuration across all sixteen official VBench dimensions, BATMoS achieves the highest total score among the eight allocation policies and comes within 0.10 points of full classifier-free guidance using 20% fewer branch evaluations. BATMoS also composes with feature reuse, reducing computation from 80 to 62 effective branch evaluations with a 0.11-point decrease in total score. These findings make temporal guidance allocation a practical mechanism for extracting more generation quality from a fixed inference budget.
Nov 16, 2026

MovieRoll: A Systematic Framework for Rolling Out Long-Form Videos

Streaming video generation must sustain quality beyond the training horizon while handling self-generated history and growing cache costs. To address these challenges, we present MovieRoll, a base-to-distilled framework that connects streaming fine-tuning, long-rollout distillation, and long-horizon inference. We first establish a robust multi-step base through Window-wise Teacher Forcing (WinTF), which jointly denoises chunks at progressive noise levels with bidirectional attention within each window and causal access to preceding history. To retain historical context efficiently, hierarchical KV-cache compression (HieraKV) exploits layer-wise attention patterns, retaining full history in shallow layers and sink--recent history in deeper layers. This reduces cache memory by up to 78% while improving generation quality. Building on this base, cumulative long tuning with distribution matching distillation (CuLT-DMD) holds parameters fixed throughout each long self-rollout and averages local gradients before updating. Each update therefore combines supervision across temporal positions while retaining only one local computation graph, enabling efficient long-rollout training of a few-step generator. To extend generation beyond the training horizon, KV recomputation bounds context for both models, while guidance annealing (AnnealCFG) reduces guidance toward the clean endpoint to stabilize MovieRoll-Base. Complementing these modeling advances, FBCA attention and PS-VAE decoding support efficient training and streaming output. On HeliosBench, MovieRoll-Base remains competitive at 30s after training on videos of up to 15s, and MovieRoll-Distilled achieves the highest total scores among the compared methods across single- and multi-event evaluations up to 60s, combining sustained naturalness with active motion.
Nov 16, 2026

ScoutCache: Efficient Few-Step Autoregressive Video Diffusion via Conditional Readout and Selective Caching

Few-step autoregressive video diffusion enables efficient long video generation, but inference still incurs substantial computational costs from repeated denoising and KV cache operations. Existing acceleration methods reduce parts of this overhead, while historical KV reading and cache construction remain costly across video chunks. To address these challenges, we propose ScoutCache, a training-free framework that jointly optimizes denoising, historical KV reading, and cache construction. We observe that shallow feature changes provide useful information for estimating clean latent updates, while historical KV dependence varies across Transformer layers and attention heads but remains relatively stable across video chunks. Based on these observations, we develop Conditional Denoising Readout (CDR) to estimate intermediate denoising updates from shallow feature changes and known noise variations. We further introduce Output Risk Guided Historical KV Reading (ORHR) to allocate historical reading budgets based on attention output perturbations, and Content and Coverage Guided Sparse Cache Construction (CCSC) to reduce cache construction costs by selecting informative spatial tokens while maintaining spatial coverage. Extensive experiments on Self-Forcing and LongLive using VBench and MovieGen demonstrate that ScoutCache achieves up to 2.28×\times and 1.86×\times speedups, respectively, while consistently improving generation quality without modifying pretrained model parameters.
Nov 16, 2026

BlockAlign: Preference Optimization for Trajectory-Distilled Video Generation

Directly aligning trajectory-distilled video generators slashes rollout costs and optimizes the deployed few-step model without teacher fine-tuning or re-distillation. However, conventional flow-matching preference objectives assume single-head local velocity supervision, mismatching the finite-interval, block-average transition interface of trajectory-distilled generators such as Parallel Decoding Distillation (PDD) and Parametric Trajectory Distillation (PTD). Naively supervising individual unfused sub-interval velocity heads across all sub-steps rather than the fused block-average velocity at block starts (DirectAlign) induces a severe training--inference mismatch that degrades preference below the unaligned baseline. We introduce BlockAlign, an interface-consistent preference optimization framework that constructs states at valid block boundaries, queries the fused block-level velocity predictor, and supervises span-normalized transition residuals---seamlessly supporting DPO-, NFT-, and rollout-style objectives while preserving the native few-step sampler. We further build a unified post-training pipeline integrating offline pairs, reward-free semi-online continuations, and online rollouts with hierarchical preference selection and objective-aware block scheduling. Experiments on multi-reference video generation across MiniMax-H3 Ref2VA and an in-house model RefVid-V4 show that BlockAlign consistently outperforms unaligned distilled models and DirectAlign in structural plausibility, reference consistency, and visual quality. Notably, our approach unlocks PDD's 5x few-step rollout speedup for 3x faster end-to-end online training, achieving statistically significant human preference gains over PDD baselines (54.0% vs. 33.3% wins/losses, or 61.8% non-tie win rate on RefVid-V4; 51.3% vs. 34.7%, or 59.7% non-tie win rate on MiniMax-H3) and up to an 82.7% win rate against DirectAlign.
Nov 15, 2026

PoCoAttn: Position-Content Collaborative Sparse Attention for Fast and High-Fidelity Video Generation

Video diffusion transformers rely on 3D full attention over tens of thousands of spatiotemporal tokens, and its quadratic cost dominates inference latency. Sparse attention reduces this cost, but existing methods face a dilemma: static methods apply predefined masks and cannot capture input-dependent attention, whereas dynamic methods reconstruct the sparse mask repeatedly, paying online to rediscover positional structure that is largely stable across inputs. To this end, we propose PoCoAttn, a training-free sparse attention framework that separates input-stable from input-dependent attention and resolves the former offline and the latter online. PoCoAttn features: (1) A relative-offset analysis of 3D query-key pairs, which shows that video attention decomposes into an input-stable positional prior that carries most of the attention mass and a small content-dependent residual. (2) A position-content collaborative sparse attention mechanism, which retrieves positional keys from offline-calibrated tables, selects content keys online with a head-adaptive ratio between the two sources, and compensates for the discarded keys to approximate full attention. (3) A step-adaptive budget allocation algorithm, which models how early-step sparse attention errors are amplified along the denoising trajectory and derives the optimal sparsity budget per step. Extensive experiments across four video diffusion models and three tasks show that PoCoAttn consistently achieves a better speed-fidelity Pareto frontier than state-of-the-art (SOTA) methods, improving PSNR by 1.9 to 5.0 dB over baselines and delivering up to 4.1×\times attention and 2.0×\times denoising speedups over dense FlashAttention-3 attention.
Nov 13, 2026

Restoring Diversity in Distilled Image and Video Generators by Probing Activations

Distilling diffusion models into few-step student models substantially reduces sampling latency, but induces severe diversity collapse: samples generated from different random seeds show minimal variation. In this work, we investigate the root cause of this failure by probing the internal representations of student models relative to their teachers. We find that distilled students maintain total activation variance comparable to their teachers, but concentrate it into substantially fewer directions. This directional collapse is localized to an identifiable intermediate block at the initial denoising step. Consequently, diversity collapse stems from a misallocation of activation variance rather than an inherent loss of generative capacity. Guided by this finding, we propose ReDiverse to reinstate the missing variance directions. At training time, ReDiverse-FT fine-tunes a lightweight adapter that regresses student activations onto the teacher's, reviving the teacher's feature variance characteristic. At inference time, ReDiverse-TF introduces a training-free mechanism that steers activations along precomputed teacher principal directions. Across image and video distillation benchmarks, ReDiverse consistently restores sample diversity while preserving generation quality, further supported by the user study.
Nov 13, 2026

JVT: Just Video Transformer makes pixel-space video diffusion simple

Direct diffusion modeling of high-dimensional raw video data is challenging, and existing pixel-space methods often rely on specialized architectures. We introduce JVT (Just Video Transformer), a simple large-patch DiT-style Transformer for pixel-space video diffusion, without a pretrained autoencoder, auxiliary detailers, cascaded stages, or U-Net-style skip connections. We systematically examine video patch embeddings, register tokens, and adaptive layer normalization to identify design choices that balance generation quality and computational efficiency. The resulting JVT design achieves generation quality competitive with latent diffusion baselines in both class-conditioned and unconditional settings. The pixel-space formulation of JVT also naturally supports joint generation of pixel-aligned modalities. We demonstrate this capability through camera-conditioned RGB-disparity generation, which improves camera following relative to RGB-only baselines and yields explicit scene geometry without an additional reconstruction model at inference. These results highlight the potential of simple DiT-style Transformers for pixel-space video diffusion, and we hope our findings offer useful insights and practical guidance for future research.
Nov 13, 2026

LongCat-Video-Avatar 1.6 Technical Report

Recent advances in audio-driven avatar video generation have enabled increasingly realistic synthesis, yet achieving precise lip-audio synchronization, natural, human-like motion, and visual consistency over long sequences remains challenging. We present LongCat-Video-Avatar 1.6, which builds on LongCat-Video-Avatar 1.5 to improve expressive naturalness and reduce error accumulation during extended generation, e.g., exaggerated lip movements and overly red, temporally unstable lip coloration, enhance the naturalness of facial expressions and coordinated head and body movements. These improvements are driven by a refined post-training pipeline that integrates reinforcement learning, improved classifier-free guidance strategies for DMD-based few-step distillation, long-video rollouts, and on-policy distillation. Together, these refinements improve expressive realism and enable the generation of videos exceeding 10 minutes while maintaining visual quality and temporal consistency. Arena evaluations show that LongCat-Video-Avatar 1.6 outperforms its predecessor and competing commercial audio-driven avatar models, demonstrating the effectiveness of the improved post-training pipeline.
Nov 13, 2026

Token2Tile: Token-Wise Sparse Attention with Tile Execution through a Low-Rank Space for Video DiTs

Attention is essential for modeling spatio-temporal dependencies in video diffusion transformers, but its quadratic cost makes it a major inference bottleneck. Its high sparsity offers an opportunity to accelerate generation by focusing computation on important interactions. However, accuracy and execution favor different granularities: each query attends to its own key tokens, whereas GPU kernels work on tiles. We present Token2Tile, a training-free sparse attention operator that combines per-query scoring and a per-query residual with tile execution. An online low-rank space, computed directly from queries and keys to preserve attention logits, makes clustering, scoring, and residual approximation efficient. Token2Tile unites query-specific selections to allocate exact work across query clusters, then packs the selected keys into GPU tiles. On MiniMax-H3 and Wan2.1-14B, Token2Tile delivers 1.48-1.54x end-to-end speedup over dense FlashAttention-3 with VBench scores within 0.5 points of dense. At comparable speeds on H3, it improves PSNR over SOL and SubBlock by 5.5-6.8 dB; on Wan, it exceeds SOL by 4.6-4.8 dB while running faster.
Nov 12, 2026

Back to Pixels: Lifting Latent Video Diffusion into Pixel Space

Frontier video generation models work in a VAE latent space that compresses video in both space and time. This compression made training tractable, but now limits high-resolution, controllable, and interactive generation -- doubling resolution quadruples token count, conditions apply to groups of frames rather than individual ones, and each response must wait for the rest of its multi-frame block. We present PixVid, a simple yet effective method that lifts a pretrained latent video model to directly generate every frame in pixel space. PixVid keeps the diffusion transformer and replaces the VAE decoder with a convolutional head that renders its outputs into pixels. We find that predicting clean video rather than velocity, together with a larger noise shift, is key to quality. Finetuned from Wan2.1-14B, PixVid is competitive with the latent baseline within ∼1.3\sim 1.3M training clips. Working framewise in pixels, it generates at 2×2\times higher resolution with the same per-frame token count, keeps depth-conditioned video editing aligned under fast motion, and responds to each action at the next frame. Project page: https://chenshuo20.github.io/PixVid/