多模态生成与编辑

Existing zero-shot video editing methods rely on pre-trained diffusion models, successfully achieving spatial control and basic temporal consistency but fundamentally fail to preserve the video's original temporal structure.This distinction is critical: temporal consistency ensures visual smoothness, but temporal structure dictates the video's high-level narrative, rhythm, and semantic flow. Without this preservation, the edited output, especially for long videos with complex semantic variations, becomes narratively incoherent and semantically ambiguous. To address this limitation, we introduce a novel zero-shot editing approach that, for the first time, explicitly focuses on preserving the source video's temporal structure. We achieve this by adaptively partitioning the video into semantically distinct clips based on feature similarity and selecting a representative anchor frame for each clip. To enhance both intra-clip fidelity and computational efficiency, we design a clip-adaptive token merging strategy which leverages the anchor's semantic dominance to stabilize the editing. Furthermore, we employ an alternating combination strategy that ensures seamless inter-clip transitions while maintaining semantic distinction. Extensive experiments demonstrate that our method achieves state-of-the-art results, successfully balancing the preservation of original temporal structure with computational efficiency, and setting a new benchmark for zero-shot video editing fidelity.

Visual world models have shown great potential in learning complex system dynamics. Recent advancements leverage these models as transition functions within Model Predictive Control (MPC) frameworks to solve various control tasks. When applied to robotics, however, they are limited to single-stage tasks such as reaching or grasping, and struggle with multi-stage ones that demand complex sequential planning. In this work, we introduce WorldDP, a world model framework designed for multi-stage robotic manipulation. Our hierarchical approach utilizes a high-level world model as a transition function to optimize for feasible subgoals during runtime, which are subsequently reached by a low-level Diffusion Policy. To further aid in learning dynamics and planning, we incorporate object-centric representations that decouple environmental entities and enable us to plan sequentially with respect to each. Evaluated across several robotics benchmarks, WorldDP consistently outperforms existing baselines, validating that coupling the world model's physically grounded planning with diffusion policy's efficient execution yields superior multi-stage performance.

Representation alignment with pretrained vision models has recently shown strong potential for accelerating diffusion transformer training. By aligning intermediate diffusion features with clean-image representations from self-supervised vision encoders, existing methods improve convergence and generation quality. However, such alignment also introduces a non-trivial constraint: diffusion models operate on noisy inputs whose usable information varies across timesteps, while the reference features are extracted from clean images. In this paper, we revisit this mismatch from a token-level perspective. We find that, under full-token representation alignment, tokens with large alignment-gradient norms exhibit a stable spatial preference, suggesting that the alignment objective does not affect all tokens uniformly and may encourage the model to rely on the complete set of clean-image tokens. To address this issue, we propose MaskAlign, a token-subset representation alignment method that applies alignment to randomly sampled token subsets during training. By exposing the model to different token subsets across iterations, MaskAlign reduces the dependence of representation alignment on the complete token set and encourages alignment behavior that is more stable under token-subset perturbations. To mitigate the information loss caused by directly dropping tokens, we further introduce a lightweight pre-mask token mixing block that shares information across tokens before masking.

Current image editing software often hinges on fixed filters or expert tuning, leaving a gap between amateur users' intent and outcomes. Creations by generative models may contain artifacts, implausible details, or stylistic drift away from photorealism and offer little insight into why an edit was made. We propose IEA, a conversational Image Editing Agent that learns to operate parameterized tools in an explicit, interpretable action space. IEA is trained via a three-stage multitask pipeline: (1) SFT on distilled expert edits, (2) GRPO with rewards for likeness improvement, tool usefulness, and intent summarization, and (3) large-scale synthetic fine-tuning to jointly master image editing, refinement, and user intent summarization. By manipulating 16 editing tools step by step, IEA produces transparent edit traces that can be inspected and debugged. In quantitative experiments, it attains a lower pixel distance on the edit task and a higher ROUGE-L on the summary task than strong baselines. In user studies, it ranks best among tool-calling methods for instruction following while surpassing generative methods in overall perceptual quality. Our results validate interpretable, tool-centric VLMs as a reliable path to human instruction-guided image retouching.

A learned world model provides a powerful physical intuition for evaluating future states. But its effectiveness in continuous control also depends critically on how candidate actions are generated for model-based planning. Rather than solely asking how accurately a model can simulate the future, we ask: which candidate actions are worth evaluating in the first place? Existing planners typically search arbitrarily or use expert demonstrations only to initialize a sampling mean, discarding the expert's state-conditioned confidence. Properly guiding this search requires a robust action prior, yet current approaches often rely on independent visual encoders or large-scale VLMs to obtain one. We argue that this architectural bloat is unnecessary: the exact same data - and the learned representations of the world model itself - inherently encode the agent's action intuition. We introduce PRISM, a task-agnostic framework that extracts both from a single dataset while maintaining strict architectural simplicity. Building on a standard JEPA-style latent world model, PRISM attaches a lightweight MLP directly to its frozen encoder to predict a state-conditioned Gaussian prior. At plan time, PRISM fuses this prior into the planner's sampling distribution via a precision-weighted Product-of-Gaussians update. This parameter-free, closed-form integration steers the sampling process, making the prior confident where it is and ceding control where it is not. PRISM improves success rates by 35 percentage points over vanilla world-model-based MPC on Cube and 32 percentage points on PushT, without introducing significant inference overhead.

Controllable video world models target interactive world exploration, where models must faithfully execute explicit action commands while preserving visual quality and temporal coherence. However, most existing approaches rely on continuous camera trajectories as action conditions, which often lead to unreliable action following, especially under complex motion sequences. In this work, we identify action representation entanglement as a key bottleneck in controllable video generation, and show that continuous camera representations lead to high feature similarity across distinct motion patterns, degrading action controllability. Based on this insight, we propose DisCo, a controllable video world model that conditions generation on a compact set of discrete action primitives to improve action separability. We further introduce DisCoBench, a comprehensive benchmark for evaluating the ability of models in short-term, long-horizon, and highly dynamic exploration scenarios. Extensive experiments demonstrate that DisCo achieves significantly more reliable action following while preserving visual quality.

Recent advances in Diffusion Transformers have driven rapid progress in video generation and editing, yet these capabilities are still handled by separate, task-specific models. Building a unified framework that supports diverse video tasks remains an open challenge: existing unified attempts either require dedicated auxiliary encoders or lack explicit mechanisms to distinguish heterogeneous conditioning tokens, struggling when the number and type of visual conditions vary across tasks. We propose TIDE, a unified framework that integrates instruction-based editing, reference-guided editing, and multi-reference generation. At its core, we introduce per-token task embeddings that assign each input token a task-specific identifier, enabling the model to explicitly disambiguate target, source, and reference tokens. To simultaneously capture high-level semantic understanding and fine-grained structural fidelity, we design a dual-path conditioning scheme that couples a vision-language model with a VAE latent path for complementary signals. We further devise a multi-task progressive training strategy that incrementally introduces tasks of increasing complexity, effectively harmonizing diverse objectives and enabling smooth generalization across heterogeneous task distributions. Extensive experiments on multiple video editing and generation benchmarks demonstrate that TIDE achieves state-of-the-art performance across all evaluated tasks. Our project page is available at https://LittleWork123.github.io/tide.

Bimanual dexterous tool use remains challenging for robots due to high-dimensional hand configurations and complex hand-tool-object dynamics and contact. Most existing control policies depend on future configuration references provided from demonstrations, while future action-conditioned world models require slow online planning over high-dimensional action sequences. A significant challenge is generating a dynamically consistent future reference trajectory without relying on privileged states from demonstrations or slow counterfactual planning. We propose DexFuture, a hierarchical system that couples a high-level Future-State Visuomotor Target Predictor with a low-level Target-Conditioned Structured Dexterous Policy. Conditioned on egocentric RGB, proprioceptive and geometric history, the high-level predictor constructs structured hand-tool-object visuomotor embeddings and uses a horizon-conditioned transformer to generate a multi-step future target trajectory. Then, the low-level policy tracks them with a target-conditioned per-link transformer. This hierarchy decouples coarse future reference generation from fine-grained action control, and slow long-horizon semantic prediction from high-frequency execution. On OakInk2 bimanual tool-use tasks, DexFuture achieves 90% of the privileged-oracle performance, compared to 7% for a no-reference policy. DexFuture operates at 60 Hz, approximately 250 times faster than DexWM-style Cross-Entropy Method (CEM) planning with a future action-conditioned world model.

Text-guided image editing has advanced rapidly with diffusion models and unified multimodal foundation models. However, most existing methods remain confined to single-turn settings, overlooking the more realistic scenario of multi-turn in-context editing, where users iteratively refine an image through a sequence of instructions. In this setting, a model must follow each new instruction while preserving accumulated session-level constraints, challenged by two coupled failure modes: long-context dilution, where sparse textual constraints become difficult to recover from growing interleaved image-text histories, and state contamination, where earlier editing mistakes degrade subsequent generations. We introduce Edit-R2, a novel reinforcement learning post-training framework for unified multimodal models. Edit-R2 reconstructs the operative session intent, which effectively consolidates scattered historical constraints into an explicit reasoning trace before each editing turn. It further enables multi-turn RL over both reasoning and generation through a unified objective that jointly optimizes intent reconstruction generation in discrete text space and flow-matching image generation in continuous latent space, while a trajectory filtering mechanism suppresses corrupted rollouts to stabilize training under state contamination. To support systematic evaluation, we introduce MICE-Bench, a large-scale benchmark for multi-turn in-context editing with automated metrics for instruction following (IF), content consistency (CC), and global awareness (GA) over accumulated session constraints. Experiments show that Edit-R2 substantially improves multi-turn in-context editing and achieves competitive performance compared against strong baselines.

Diffusion-based vision-language-action (VLA) models often inherit the image-generation view: actions are generated by iterative denoising. We argue that VLA action generation has a different condition-target structure: the policy is conditioned on rich observations, language, and state, but predicts only a compact, low-dimensional action chunk. Under this asymmetry, strong one-step action generation should not necessarily require the advanced one-step methods developed for image synthesis. We keep standard velocity prediction and add no teacher model, distillation stage, or auxiliary objective; in our main recipe, we simply bias the training time distribution toward high-noise states. We first isolate the effect in a controlled MNIST grid-to-sequence task, then test it with extensive robot-policy experiments. Across standard LIBERO, LIBERO-Plus, and LIBERO-Pro, one-step policies trained with high-noise biased schedules generally match ten-step decoding under the same recipe, and on standard LIBERO can exceed ten-step policies trained with a uniform time distribution. A real-robot bimanual YAM RSS evaluation gives a small-sample cross-architecture check of the same sampler trend. On a 1.4B VLM model with a 30M action head, one-step decoding reaches 95.6% on LIBERO-Long. These results show that strong one-step VLA action generation can emerge from standard diffusion training, without importing the full few-step diffusion machinery developed for image generation.

We propose world-language-action (WLA) models as a new class of embodied foundation models. WLA takes textual instructions, images, and robot states as inputs to jointly predict textual subtasks, subgoal images, and robot actions, conjoining the world modeling interface to learn from extensive egocentric videos as in the world-action model (WAM) and the language reasoning capacities to solve complex long-horizon tasks as in vision-language-action (VLA) models. At the core of WLA lies an autoregressive (AR) Transformer backbone, instead of a bidirectional diffusion Transformer as in WAMs, to predict the next state, comprising the semantic-level textual intention and complementary fine-grained physical dynamics. The physical dynamics are supervised by the world modeling objective based on a dedicated World Expert, and are leveraged to ease the characterization of the state-action correlation for the Action Expert. WLA leverages meta-queries to make the world prediction implicitly impact the action generation so that the former can be disabled during inference. The world prediction can also be activated to enable test-time scaling for improved robot control. Our WLA-0 prototype, with 2B active parameters, achieves 40 ms per inference on an NVIDIA RTX 5090. Evaluations across simulated and real-world environments demonstrate that WLA-0 achieves state-of-the-art multi-task and long-horizon learning abilities, e.g., 92.94% success rate on RoboTwin2.0 Clean and 56.5% success rate on RMBench. WLA-0 also holds the promise to learn novel tasks directly from cross-embodiment robot videos without action annotations.

End-to-end Vision-Language-Action (VLA) models have shown promise in UAV navigation. However, existing approaches typically rely on historical observations to directly predict actions, often struggling in dense urban environments where severe occlusions and sharp turns result in drastic viewpoint transitions. We argue that the ability to "imagine" future states — inherent in World Models — is critical for robust decision-making under such partial observability. To address this, we construct a challenging Urban Canyon Traversal Benchmark, specifically designed to evaluate spatial understanding in scenarios characterized by severe occlusions and drastic viewpoint transitions. To this end, we propose WorldFly, a novel world-model-based VLA framework that employs a dual-branch coupled flow matching mechanism to jointly generate future video predictions and navigation actions, thereby explicitly guiding the agent's policy via spatial imagination. Extensive evaluations on our benchmark demonstrate that WorldFly outperforms other baselines, particularly in unseen environments, validating the effectiveness of integrating world models into embodied aerial agents.

Aggressive distillation of the diffusion U-Net inverts the per-frame bottleneck of real-time text-to-image pipelines: once the denoiser is a 4-step or 1-step distilled student, the text encoder becomes the critical path. This inversion is most acute in vision-aware edit diffusion, where the encoder is a multimodal large language model (MLLM). We study the case of a 0.39B distilled edit U-Net paired with a 2.13B MLLM text encoder (Qwen3-VL) and present a streaming pipeline targeted at this regime built around three engineering mechanisms: asymmetric side-stream / main-stream CUDA pipelining with batched text-encoder amortisation (and optional static-prompt caching), a compile-friendly ControlNet-LLLite reformulation that folds the entire U-Net + adapter stack into a single fused graph, and a periodic conditioning-refresh schedule with a hook subset that amortises the per-frame conditioning cost. On a single consumer RTX 3090 Ti at 512x512 the pipeline sustains 27.4 fps over a 480-frame run at batch size B=8 and 29.6 fps at B=16, with end-to-end p50 latency of approximately 0.5 and 1.0 seconds respectively; the same operating point measures 54.9 fps on RTX 4090 and 74.1 fps on RTX 5090. We report video-rate streaming throughput rather than interactive low latency, and locate our numbers against same-stack StreamDiffusion re-runs as systems context, not as a benchmark superiority claim. For the trained oil-painting style, the released temporal adapter generalises within in-clip noise to 19 unused DAVIS-2017 sequences and 15 non-DAVIS clips from seven sources; prompt-level generalisation to unseen style families is bounded and reported separately.

Video generation models based on Diffusion Transformers (DiTs) have achieved remarkable performance in video synthesis, yet they suffer from high inference latency and computational costs due to the quadratic complexity of 3D attention. Existing acceleration methods primarily reduce computational complexity within each individual denoising steps through techniques such as sparse attention and KV-caching. However, they rigidly adhere to the inherent constraint of the standard diffusion pipeline: every frame in the target video sequence must be subjected to a complete, dense denoising process across all diffusion timesteps. We observe that due to the corresponding contents and motions among adjacent frames, when keyframes with critical semantic transitions are anchored, the intermediate states of others often follow more predictable trajectories, which indicates that such uniform, dense denoising process is inherently redundant for natural video data. To this end, we introduce RhymeFlow, a training-free framework that decouples the denoising trajectories of different frames. Specifically, we first identify a sparse set of pivotal key frames that dominate the latent semantic evolution. Then, only these keyframes undergo dense, step-by-step denoising to ensure structural integrity, while non-keyframes progressively skip denoising steps to minimize computational cost. Since skipped intermediate states of non-keyframes break the temporal coherence in keyframe denoising steps, leading to visual degradation, we further introduce a latent trajectory projection module, which enables keyframes to interact with a complete and temporally consistent sequence representation. Extensive experiments on current DiT-based video generation models demonstrate our method outperforms existing baselines with higher inference speed and better visual quality.

Vision-language-action (VLA) policies operate in a closed loop in real-world robot tasks: a robot observes the scene, executes an action chunk, and conditions its next decision on the resulting observation. However, most existing world models for robot action evaluation are limited to open-loop prediction along pre-collected action trajectories. This prevents them from supporting closed-loop VLA evaluation, where each action chunk must be conditioned on the observation generated by the previous execution. To address this gap, we propose PiL-World, a chunk-wise world model designed for policy-in-the-loop VLA evaluation. Given the current observation and the action trajectory rolled out by a VLA policy, PiL-World generates multi-view future observations that are consistent with the VLA rollout and match the image inputs required by the policy. By alternating between VLA inference and world-model prediction, PiL-World enables closed-loop evaluation without real robot execution at every step. To improve rollout fidelity, PiL-World conditions video generation on action-derived visual control from head-view robot motion and latent histories that encode task execution context, while jointly predicting complementary multi-view observations. Beyond successful teleoperated demonstrations, it also learns from failed execution trajectories, helping the imagined rollouts better match the distribution of real policy executions. We evaluate PiL-World on three real dual-arm manipulation tasks. PiL-World generates imagined rollouts that are highly consistent with real robot executions. More importantly, compared with the baseline, it reduces the error between VLA success rates measured in real-world rollouts and those estimated through closed-loop world-model evaluation from 63.2% to 12.0%.

Evaluating large language model (LLM) agents in multi-turn interactive environments is expensive and risky, as it requires online environment interaction. We propose ADWM (Autoregressive Diffusion World Model), an evaluation framework that estimates the performance of a new LLM agent policy purely from pre-collected trajectories. The core idea is to learn a latent diffusion world model that simulates how the environment responds to the evaluation policy, without ever executing it in the real environment. Existing diffusion-based OPE methods guide full trajectories in a single pass by jointly diffusing states and actions, an assumption that breaks down for LLM agents whose actions are discrete text that must be sampled from the policy after observing the environment. Unlike autoregressive world models that suffer from compounding errors, ADWM models each transition as an independent denoising process, enabling reliable step-by-step rollouts where the world model and agent alternate in causal order. Crucially, the LLM agent under evaluation directly guides the diffusion generation at each step via a policy-conditioned score function, ensuring that simulated trajectories accurately reflect its decision-making patterns. Empirically, ADWM achieves accurate value estimates and evaluation reliability across diverse multi-turn agent tasks, demonstrating its promise as a practical framework for offline LLM agent evaluation.

Developing unified video generation and editing models capable of interpreting interleaved multimodal inputs is a promising yet challenging frontier field. Existing unified frameworks predominantly rely on massive models (typically 13B parameters or more) and incorporate source video conditions for editing by concatenating sequence tokens. This concatenation inevitably doubles the sequence length, quadrupling the computational complexity of the self-attention mechanism and introducing prohibitive overhead. To address these bottlenecks, we present LoomVideo, a highly efficient 5B-parameter unified architecture for both video generation and editing. LoomVideo replaces the standard text encoder with a Multimodal Large Language Model (MLLM) and employs Deepstack injection mechanism to align multi-layer MLLM features with the Diffusion Transformer (DiT). Crucially, we introduce a zero-overhead Scale-and-Add conditioning approach for video editing. By scaling and directly adding the clean source video latent to the noised target latent, this elegant design eliminates the need for token concatenation, drastically reducing computational cost while maintaining robust capabilities for complex, non-rigid edits. Furthermore, a Negative Temporal RoPE strategy is seamlessly integrated to handle multiple reference images. Extensive experiments demonstrate that our compact 5B model achieves state-of-the-art or highly competitive performance across comprehensive benchmarks, exhibiting exceptional superiority in e-commerce and fashion generation scenarios. Benefiting from the zero-overhead conditioning mechanism, LoomVideo achieves at least a 5.41x acceleration in inference speed compared to models of similar capabilities, paving the way for highly practical and efficient video foundation models.

Autoregressive (AR) models have demonstrated remarkable performance in generating high-fidelity images. However, their inherently sequential next-token prediction leads to significantly slower inference. Recent studies have introduced Jacobi-style decoding to accelerate autoregressive image generation. Extending the draft sequence initially improves efficiency, yet the acceleration quickly saturates as error propagation in the one-dimensional sequence hinders convergence. Observing that images exhibit strong local spatial correlations, we propose Parallel Jacobi Decoding (PJD), a training-free decoding approach that expands draft tokens in the two-dimensional spatial domain to enable efficient spatially parallel refinement. PJD adjusts the attention mask to mitigate error accumulation and improve convergence stability. Extensive experiments on diverse datasets show that PJD achieves 4.8x-6.4x acceleration across multiple autoregressive image generation models while maintaining competitive generation quality.

T2I models cannot effectively capture sentiment from various types of text, including diaries, as they primarily focus on visual object-related patterns rather than contextual emotional understanding. This paper proposes an emotion-aware text-to-image pipeline that generates children's hand drawing style images from short Korean diary entries. The proposed pipeline employs Qwen3-8B for recognising implicit sentiment from short diaries, and Stable Diffusion 3.5 Medium fine-tuned with LoRA on children's drawing images with emotion-based trigger words for image generation. Additionally, this paper presents experiments examining the effect of emotion trigger words on generated images and discusses the limitations of CLIP Score as an evaluation metric for emotion-aware image generation.

Typography generation in diffusion models faces a persistent trade-off: enabling precise font control typically degrades text legibility, while maintaining readability often sacrifices typographic fidelity. We present FontFusion, a plug-and-play conditioning framework for Diffusion Transformer (DiT) architectures that resolves this dilemma through three core innovations: (1) a hierarchical token representation establishing explicit text-font relationships at multiple granularities, (2) position-aware embeddings creating spatial bindings between typography and image content, and (3) a multi-level token dropping strategy improving both computational efficiency and generalization to unseen fonts. Our systematic evaluation of font embedding spaces reveals that a dual encoder combining DeepFont and DINOv2 outperforms any single encoder for typography tasks. FontFusion demonstrates 76% relative improvement on challenging decorative fonts over single-encoder baselines and font consistency gains exceeding approximately 68-76% over unconditioned models, while integrating into existing DiT architectures without retraining.

Recent advances in generative image editing have improved the realism and controllability of localized image manipulation, raising new challenges for image manipulation detection and localization (IMDL). However, existing IMDL benchmarks still have limitations in visual realism, manipulation diversity, and generator coverage, making it difficult to reflect recent trends in image manipulation. To address these limitations, we introduce Impostor, a high-quality AI-edited image manipulation localization dataset containing 100K manipulated images. Impostor is constructed by CraftAgent, a closed-loop agent framework that integrates scene perception, editing planning, manipulation execution, quality validation, and iterative reflection to automatically generate diverse and visually realistic manipulated images. Moreover, Impostor contains images generated by seven recent AIGC models across three manipulation types and includes multiple manipulated regions, providing a more comprehensive benchmark for AIGC-based IMDL. Furthermore, we propose PhaseAware-Net (PANet), a semantic-forensic framework that introduces local phase modeling and semantic-forensic consistency learning to better localize semantically plausible yet forensically disrupted manipulated regions. Extensive experiments show that Impostor poses significant challenges to existing large vision-language models (LVLMs) and specialized IMDL methods, while PANet achieves superior performance on Impostor and multiple public benchmarks.

Multimodal Large Language Models (MLLMs) are increasingly used for video understanding, yet their reliability under multi-video inputs remains poorly understood. We study positional bias in multi-video summarization, where the quality of a per-video summary can change with the video's input slot even when the underlying content is unchanged. We construct a benchmark from ActivityNet and News videos, covering Cooking, Domestic, Leisure, and News settings with two- and four-video inputs. We evaluate nine open-source and proprietary MLLMs and measure position effects with three complementary metrics: Coverage, Directional Positional Bias (DPB), and Middle-Edge Gap (MEG). Our results show that positional effects are domain- and model-dependent: signed directional bias can be small even when middle positions underperform, and increasing visual or generation budget does not uniformly remove the imbalance. We further analyze prompt-level mitigation methods. Together, the results show that multi-video summarization remains sensitive to input protocol and position, motivating more robust order-invariant multimodal systems.

Video generation models have made impressive strides in synthesizing visually compelling content, yet their outputs remain confined to the virtual domain. A natural question follows: how well do these models reflect the physical world when their generated videos leave the screen and enter reality? We propose robotic manipulation as a concrete, measurable window onto this question: if a model has truly internalized physical laws, the motion it depicts should translate into executable robot behavior. We introduce Dream$.$exe, an evaluation framework that operationalizes this criterion through a video-to-execution pipeline. Given a scene image and a task description, Dream$.$exe synthesizes a manipulation video, converts the generated motion into robot trajectories, and executes them in a physics simulator, yielding a grounding signal that purely visual metrics cannot offer. Using this pipeline, we evaluate 8 models spanning frontier closed-source generators, open-source generators, and robot-specific models. Our benchmark covers 101 manually curated manipulation tasks at three levels of physical complexity, measured across visual quality, trajectory fidelity, and execution success. Encouragingly, several models achieve measurable execution success, suggesting that generative priors learned from internet-scale data already encode meaningful physical knowledge. Yet visual quality proves a poor predictor of executability, exposing a dimension of model capability that standard visual evaluations do not capture. Dream$.$exe will be open-sourced at https://github.com/showlab/Dream.exe.

Agile quadrotor flight in cluttered scenes requires more than a reactive mapping from a depth image to a control command: the vehicle must remember which regions have been observed, infer nearby occupied space, and act under partial visibility and tight latency. In this paper, we present Mapping-Aware Dreamer (MAD), a geometry-aware world model for vision-based quadrotor flight. Instead of using raw-image reconstruction as the main self-supervised objective, MAD learns recurrent latent dynamics that reconstruct robocentric occupancy and visibility grid maps together with proprioceptive states. This design forces the latent state to encode local geometry, visibility history, and ego-motion in a form that is directly relevant to collision avoidance. MAD is trained in DiffAero using a GPU-parallel map-construction module that provides high-throughput supervision for occupancy and visibility. The learned representation is used in three policy-learning modes: imagination-based MAD-Dreamer and feature-extractor variants based on PPO and SHAC. Across visual navigation and racing tasks, MAD-based agents achieve higher success rates, faster flight, and better cross-task transfer than corresponding vision-only baselines. The model also produces interpretable map predictions and accurate ego-motion estimates from depth observations. We further deploy the learned policy on a physical quadrotor with an Intel RealSense D435i and demonstrate safe indoor and outdoor flight under limited sensing, reaching 9.66 m/s in simulation and 5.05 m/s in real-world forest experiments. These results show that mapping-aware world models provide a practical middle ground between modular aerial navigation and end-to-end learning.

Modern video diffusion models generate increasingly realistic and temporally coherent videos, motivating their use as candidate world simulators. Yet it remains unclear whether these models internally encode physical structure, or merely reproduce motion patterns seen during training. We study this question by probing video diffusion models along latent trajectories corresponding to real videos with known physical plausibility. To obtain such trajectories, we approximately invert the deterministic sampling process by integrating the learned velocity field backward from a clean video latent to noise, giving access to the model's intermediate states and attention maps. Using these recovered trajectories, we show that physical plausibility is linearly decodable from diffusion transformer states across IntPhys and InfLevel, reaching around 81.27% average accuracy and outperforming dedicated representation-learning baselines such as V-JEPA and VideoMAE. Surprisingly, this signal is absent from the VAE latent input and emerges inside the denoising transformer itself, despite the model not being trained with a self-supervised predictive objective. These findings suggest that physically meaningful representations can arise as a byproduct of generative denoising.

World models, learned generative models that predict how an environment evolves, have become a promising tool for sample-efficient robot learning. Yet how robust they are to environmental variability remains poorly understood. To address this, we conduct a systematic study using vision-based quadrotor navigation as a testbed problem, training DreamerV3-based world models under varying levels of environmental randomness and evaluating them across all levels through cross-environment validation, spanning both Self-Supervised Learning (SSL) pretraining and Reinforcement Learning (RL) fine-tuning. We then deploy all world models and associated navigation policies on a real quadrotor in unseen environments, including an open-loop run where the model receives just 2.5s of real sensory input before all sensors are cut off, leaving the system to navigate entirely in imagination over a 12m traverse. Our results show that world model robustness during SSL pretraining is a strong predictor of sim-to-real transfer: every model that generalized well in cross-environment SSL validation deployed successfully in the real world, passing through gaps as narrow as 0.67m, whereas the model that dominated simulation policy evaluation failed on the real platform. We further identify (a) the discrete latent size and (b) the training-sequence length as the dominant factors governing world model quality.

Video diffusion transformers have achieved state-of-the-art visual quality, but their high inference cost remains a major bottleneck for real-time applications. Recent distillation frameworks produce autoregressive video diffusion models with reduced latency, yet these models still use a fixed number of denoising steps per frame, wasting computation on predictable frames and under-refining challenging ones. We present DSA, a confidence-guided adaptive computation framework for AR video diffusion. DSA introduces a lightweight confidence head, trained jointly with the generator under a distribution-matching distillation objective, to estimate per-frame denoising reliability. At inference, this confidence signal dynamically adjusts the number of diffusion steps: simple frames terminate early for speed, while complex frames receive additional refinement. Our method requires no extra video data, no heuristics, and little architectural modification. Experiments show that DSA achieves real-time autoregressive video generation, reaching 22.63 FPS with sub-second latency on H100 GPUs, while maintaining competitive or superior VBench quality compared to recent autoregressive and bidirectional video diffusion models. Our results demonstrate that confidence-guided adaptive sampling provides an effective and practical path toward interactive video generation.

Large-scale video generation models have made remarkable progress in semantic consistency and visual quality, producing videos that are increasingly coherent and visually convincing. Nevertheless, the dynamics induced by pixel-level fitting do not naturally accommodate the regularities that govern real-world motion and interaction, resulting in persistent shortcomings in physical plausibility. To address this limitation, we propose PILA (Physics-Informed Latent Alignment), a framework that injects physics-structured latent guidance into the frozen flow-matching dynamics of pretrained video models. Specifically, PILA first employs anchored field estimation to map frozen-generator latents into an operational physical attribute bank organized by field-proxy slots, using observable motion as a kinematic anchor for constructing less directly observed proxies. To handle the heterogeneity of real-world dynamics, PILA adopts a mixture-of-experts design over physical categories. Label-prior masked expert routing selects category-specific operator experts, whose refinements are regularized by operational residuals abstracted from physical relations. Finally, the refined proxies are fused into the physical attribute bank and decoded into a correction to the flow-matching vector field, injecting physics-aware guidance while preserving the visual prior of the pretrained backbone. With staged adapter training on Wan 2.1-1.3B and direct transfer of the learned adapter to Wan 2.2-14B, PILA achieves state-of-the-art results on VBench-2.0, VideoPhy-2, and PhyGenBench in both visual quality and benchmark-measured physical plausibility.

We present OSCAR, a precise action-conditioned video world model that generalizes across different robot embodiments and enables robot policy evaluation. Existing video world models face three main challenges for real-world robot evaluation: limited scenario diversity in current robot training datasets, imprecise action following, and poor generalization across embodiments for broad adoption. We tackle these challenges from two perspectives. At its core is a large-scale standardized data pipeline that curates, filters, and deduplicates broad robotics and egocentric human datasets, yielding a clean joint-training dataset that spans diverse tasks, scenarios, actions, and robot embodiments. To condition the video model, we adopt 2D kinematic skeleton rendering as a unified conditioning representation that generalizes across different robot arms or even human hands. We finetune the Cosmos-Predict2.5-2B model on a single GH200 GPU. Our model achieves significant improvement on action following, appearance quality, and motion consistency, compared to existing baselines, which either have a much larger model size or require more GPUs. We further deploy OSCAR to evaluate robot policies from RoboArena. Extensive experiments demonstrate the significant correlation between our virtual policy evaluation in OSCAR and real-world evaluation, paving the way for the future where robot policies can be purely evaluated in virtual generated worlds.

We present Echo Infinity, an autoregressive (AR) framework towards real-time infinite video generation that employs a learnable evolving memory to dynamically filter, abstract, and compress any-length history at constant cost. Existing methods mainly curate memory with predefined KV-cache schedules, fixed-ratio heuristic compression, or inference-time RoPE adaptation. These designs inevitably lose historical information and amplify compounding errors due to their limited cache window and ignorance of autoregressive generation noise. Inspired by human memory consolidation, Echo-Infinity replaces handcrafted memory curation with learnable Memory Query, which are updated by attention and a gating mechanism when past frames are evicted from the local window. The queries are optimized end-to-end with the video diffusion transformers (DiTs), forming an evolving memory that supports arbitrary compression ratios with constant computation independent of video length. They also act as a generalizable generation prior, improving quality even when only the optimized initial state is used. We further introduce Unified Relative RoPE Recipe, which anchors the sink frames to start from id 0 and lets the newest frame id grow at most to the DiTs' pretrained maximum temporal RoPE id throughout training and inference, freeing the model from the finite RoPE constraint and closing the train-test RoPE extrapolation gap. In long and short video generation, Echo-Infinity achieves state-of-the-art performance, and, to our knowledge, demonstrates promising 24-hour (>1.3 M frames) real-time rollouts for the first time, suggesting a practical path toward infinite video generation.