Text- and image-conditioned world generators can create visually rich 3D environments, yet these worlds often remain silent or rely on soundtracks synthesized solely from text or rendered video. Although such audio can convey what should be heard, it lacks an explicit representation of where sound sources are located and how their perceived sound should vary with listener movement. We introduce Audible World Models, a training-free framework that incorporates sound into the generated world state. Starting from a text prompt, our system constructs a panoramic 3D proxy, separates it into semantic layers, identifies sound-producing foreground objects and ambient background regions, and synthesizes dry audio for each sound label. It then anchors these sources to reconstructed geometry and renders listener-dependent spatial audio using geometric acoustic propagation. By explicitly linking semantics, geometry, and sound propagation, the framework maintains persistent source locations while adapting the rendered audio to changes in listener viewpoint and motion. Experiments across 80 generated scenes demonstrate substantial gains in spatial consistency over text-, video-, and panorama-conditioned baselines, while preserving competitive semantic alignment. VLM-based assessments and human evaluations further indicate that our soundtracks are preferred for their audio-visual consistency, spatial plausibility, and motion-dependent behavior.
展开完整摘要收起摘要↓
Text- and image-conditioned world generators can create visually rich 3D environments, yet these worlds often remain silent or rely on soundtracks synthesized solely from text or rendered video. Although such audio can convey what should be heard, it lacks an explicit representation of where sound sources are located and how their perceived sound should vary with listener movement. We introduce Audible World Models, a training-free framework that incorporates sound into the generated world state. Starting from a text prompt, our system constructs a panoramic 3D proxy, separates it into semantic layers, identifies sound-producing foreground objects and ambient background regions, and synthesizes dry audio for each sound label. It then anchors these sources to reconstructed geometry and renders listener-dependent spatial audio using geometric acoustic propagation. By explicitly linking semantics, geometry, and sound propagation, the framework maintains persistent source locations while adapting the rendered audio to changes in listener viewpoint and motion. Experiments across 80 generated scenes demonstrate substantial gains in spatial consistency over text-, video-, and panorama-conditioned baselines, while preserving competitive semantic alignment. VLM-based assessments and human evaluations further indicate that our soundtracks are preferred for their audio-visual consistency, spatial plausibility, and motion-dependent behavior.
Part-level 3D assets are essential for editing, reassembly, and interaction, yet recovering such structure from a single image remains challenging due to occlusion, ambiguous boundaries, and the need for coherent multi-part reasoning. Existing approaches struggle to achieve both controllable part-level generation and coherent multi-part structure, as part identity and spatial allocation are typically inferred implicitly. We present Seg3DParts, a segmentation-grounded framework for controllable part-level 3D generation from a single image. By treating segmentation as an explicit grounding signal, our method defines part identity during generation, enabling each component to be anchored to a corresponding image region. To ensure coherent assemblies, we introduce structured cross-part interaction that allows components to exchange global context throughout the generative process. As a result, Seg3DParts directly generates well-aligned part meshes in a shared canonical space without post-hoc alignment, supporting flexible and controllable decomposition. We further introduce PartObjectNet, a large-scale dataset with over 200K objects and 1M annotated parts. Experiments demonstrate that Seg3DParts achieves superior geometry quality, cross-part coherence, and part-level controllability over existing methods.
展开完整摘要收起摘要↓
Part-level 3D assets are essential for editing, reassembly, and interaction, yet recovering such structure from a single image remains challenging due to occlusion, ambiguous boundaries, and the need for coherent multi-part reasoning. Existing approaches struggle to achieve both controllable part-level generation and coherent multi-part structure, as part identity and spatial allocation are typically inferred implicitly. We present Seg3DParts, a segmentation-grounded framework for controllable part-level 3D generation from a single image. By treating segmentation as an explicit grounding signal, our method defines part identity during generation, enabling each component to be anchored to a corresponding image region. To ensure coherent assemblies, we introduce structured cross-part interaction that allows components to exchange global context throughout the generative process. As a result, Seg3DParts directly generates well-aligned part meshes in a shared canonical space without post-hoc alignment, supporting flexible and controllable decomposition. We further introduce PartObjectNet, a large-scale dataset with over 200K objects and 1M annotated parts. Experiments demonstrate that Seg3DParts achieves superior geometry quality, cross-part coherence, and part-level controllability over existing methods.
Despite rapid progress in video generation models, they still exhibit obvious motion deficiencies, often manifested as incorrect object motion. However, most existing video quality evaluations focus on aesthetic quality or text-video alignment. To address this gap, we study object-centric motion fidelity assessment, evaluating target objects along object consistency, motion continuity, and physical plausibility. To achieve this, we first introduce VidMotion, a diagnostic dataset of 6,879 videos with designated moving objects and fine-grained annotations including dimension-wise scores and failure causes. We further propose MotionInsight, a diagnostic evaluator that shifts assessment from implicit RGB-frame observation to explicit motion-space diagnosis. By constructing motion-aware representations, MotionInsight makes subtle motion deficiencies more observable. We also introduce motion-specific rewards during GRPO to transform observed motion into a diagnostic assessment. Experiments demonstrate that MotionInsight provides an effective basis for diagnosing object motion deficiencies, producing human-aligned scores along three dimensions and grounded explanations.
展开完整摘要收起摘要↓
Despite rapid progress in video generation models, they still exhibit obvious motion deficiencies, often manifested as incorrect object motion. However, most existing video quality evaluations focus on aesthetic quality or text-video alignment. To address this gap, we study object-centric motion fidelity assessment, evaluating target objects along object consistency, motion continuity, and physical plausibility. To achieve this, we first introduce VidMotion, a diagnostic dataset of 6,879 videos with designated moving objects and fine-grained annotations including dimension-wise scores and failure causes. We further propose MotionInsight, a diagnostic evaluator that shifts assessment from implicit RGB-frame observation to explicit motion-space diagnosis. By constructing motion-aware representations, MotionInsight makes subtle motion deficiencies more observable. We also introduce motion-specific rewards during GRPO to transform observed motion into a diagnostic assessment. Experiments demonstrate that MotionInsight provides an effective basis for diagnosing object motion deficiencies, producing human-aligned scores along three dimensions and grounded explanations.
作者Daniel Pröll, Thomas Kraxner, Tobias Schaefer, Sebastian Hegenbart
Defect detection systems for industrial condition monitoring can only be relied upon if they are validated, yet defective samples are rare and, for a specific asset, often nonexistent. We present a framework that synthesizes severity-graded defects on real non-defective images without any defect references for the target asset, that can be used for model selection and validation. A defect taxonomy for common failure modes is distilled from literature into prescriptive prompts at varying defect severities. Regions of interest are cropped from in defect-free images and edited with a pre-trained image generation model ("FLUX.2 [klein]"). Color-matching and blending are employed to improve structural coherence with the original image. Generations are filtered out by a scorer and by estimated detection difficulty. Model selection experiments on MVTecAD show image AUROC choice regret over model selection can be nearly halved compared to the best fixed model chosen with access to test data. Experiments show the need for severity-graded anomaly synthesis. A case study investigates the proposed method for in-situ monitoring of Pelton turbine runners in hydropower, where real defect images are rare and expensive to collect. A PatchCorebased anomaly detection model is fit on Pelton turbine images and selected and validated using synthetic images, showing strong detection performance (94 % correct detection at optimal threshold and AUROC 0.97). The model reliably detects moderate and advanced defects, while early-stage defects remain challenging, indicating the synthetic data meaningfully stresses detector sensitivity.
展开完整摘要收起摘要↓
Defect detection systems for industrial condition monitoring can only be relied upon if they are validated, yet defective samples are rare and, for a specific asset, often nonexistent. We present a framework that synthesizes severity-graded defects on real non-defective images without any defect references for the target asset, that can be used for model selection and validation. A defect taxonomy for common failure modes is distilled from literature into prescriptive prompts at varying defect severities. Regions of interest are cropped from in defect-free images and edited with a pre-trained image generation model ("FLUX.2 [klein]"). Color-matching and blending are employed to improve structural coherence with the original image. Generations are filtered out by a scorer and by estimated detection difficulty. Model selection experiments on MVTecAD show image AUROC choice regret over model selection can be nearly halved compared to the best fixed model chosen with access to test data. Experiments show the need for severity-graded anomaly synthesis. A case study investigates the proposed method for in-situ monitoring of Pelton turbine runners in hydropower, where real defect images are rare and expensive to collect. A PatchCorebased anomaly detection model is fit on Pelton turbine images and selected and validated using synthetic images, showing strong detection performance (94 % correct detection at optimal threshold and AUROC 0.97). The model reliably detects moderate and advanced defects, while early-stage defects remain challenging, indicating the synthetic data meaningfully stresses detector sensitivity.
Teaching humanoids loco-manipulation skills, such as carrying diverse objects, via visual imitation is a promising path toward generalist robots. However, collecting diverse, high-quality interaction videos, such as clips that clearly show a person's full body and unoccluded interactions with objects, poses a practical barrier to scaling this approach. We propose PRISM, a real-to-sim-to-real framework that overcomes this limitation by amplifying a handful of real videos into a large, diverse training set. PRISM first generates hundreds of diverse "counterfactual" human-object interaction videos via video-to-video (V2V) generation from a few exemplar real videos. Our contact-anchored real-to-sim pipeline then reconstructs both human and object motions, retargeting this imperfect video data into physically plausible trajectories. The intra-class variability across these counterfactual videos lets us train a single policy that generalizes to unseen objects within each category. We demonstrate the full pipeline by deploying this policy on a real robot without any real-world fine-tuning. Using only onboard depth observations, our humanoid picks up, carries, and drops objects, including boxes, barrels, bins, and balls, across novel instances, sizes, and initial configurations.
展开完整摘要收起摘要↓
Teaching humanoids loco-manipulation skills, such as carrying diverse objects, via visual imitation is a promising path toward generalist robots. However, collecting diverse, high-quality interaction videos, such as clips that clearly show a person's full body and unoccluded interactions with objects, poses a practical barrier to scaling this approach. We propose PRISM, a real-to-sim-to-real framework that overcomes this limitation by amplifying a handful of real videos into a large, diverse training set. PRISM first generates hundreds of diverse "counterfactual" human-object interaction videos via video-to-video (V2V) generation from a few exemplar real videos. Our contact-anchored real-to-sim pipeline then reconstructs both human and object motions, retargeting this imperfect video data into physically plausible trajectories. The intra-class variability across these counterfactual videos lets us train a single policy that generalizes to unseen objects within each category. We demonstrate the full pipeline by deploying this policy on a real robot without any real-world fine-tuning. Using only onboard depth observations, our humanoid picks up, carries, and drops objects, including boxes, barrels, bins, and balls, across novel instances, sizes, and initial configurations.
We study whether scaling video generation enables models to reason about hidden information from the past frames, and at what computational cost. Our controlled benchmark requires predicting nine prescribed moves of an initially solved 2x2x2 Rubik's Cube from a fixed view of three faces. Correct predictions require inferring how actions change hidden states, and the simulator provides exact ground truth for evaluation. Models learn plausible cube geometry early, while correct sticker configurations require substantially more training. Although validation MSE follows approximate power-law scaling, lower MSE loss does not reliably indicate downstream reasoning capabilities. Smaller autoregressive models achieve higher state accuracy with limited compute, while larger models reach higher accuracy after more training. At roughly 0.1 PF-days, the 70M-parameter model correctly predicts the visible sticker configuration in 44.6% of post-action frames, compared with 0.3% for the 1B model, which reaches 83.7% at 3.14 PF-days. Symbolic state supervision raises the 20M model's frame accuracy from 31.1% to 67.3% at the same training-data budget, suggesting that learning representations of state changes can complement scaling.
展开完整摘要收起摘要↓
We study whether scaling video generation enables models to reason about hidden information from the past frames, and at what computational cost. Our controlled benchmark requires predicting nine prescribed moves of an initially solved 2x2x2 Rubik's Cube from a fixed view of three faces. Correct predictions require inferring how actions change hidden states, and the simulator provides exact ground truth for evaluation. Models learn plausible cube geometry early, while correct sticker configurations require substantially more training. Although validation MSE follows approximate power-law scaling, lower MSE loss does not reliably indicate downstream reasoning capabilities. Smaller autoregressive models achieve higher state accuracy with limited compute, while larger models reach higher accuracy after more training. At roughly 0.1 PF-days, the 70M-parameter model correctly predicts the visible sticker configuration in 44.6% of post-action frames, compared with 0.3% for the 1B model, which reaches 83.7% at 3.14 PF-days. Symbolic state supervision raises the 20M model's frame accuracy from 31.1% to 67.3% at the same training-data budget, suggesting that learning representations of state changes can complement scaling.
作者Jiawei Zhang, Shuhao Liu, Rong Huang, Yuancheng Li, Zhihui Li, Xiaojun Chang, Changlin Li
One-dimensional (1D) variable-length visual tokenizers enable adaptive compression by varying the number of tokens, allowing downstream autoregressive (AR) models to flexibly trade off generation quality against computational cost using a single tokenizer. However, existing approaches based on nested dropout often fail to fully exploit the representational capacity of the tokenizer, resulting in suboptimal performance in both image reconstruction and generation. In this work, we introduce NesTok, a nested self-alignment framework tailored to dynamic visual tokenizers. NesTok introduces cross-length training, which jointly optimizes reconstruction across token lengths while using the full-length sequence to guide shorter counterparts, enabling shorter token sequences to approach the reconstruction quality of full-length sequences. On ImageNet, NesTok improves substantially over standard training and achieves an rFID score of 0.98. On downstream image generation, it achieves the state-of-the-art gFID score of 1.46 on ImageNet 256$\times$256 among existing variable-length autoregressive image generation methods. Code will be available at https://github.com/Jiawei804/NesTok.
展开完整摘要收起摘要↓
One-dimensional (1D) variable-length visual tokenizers enable adaptive compression by varying the number of tokens, allowing downstream autoregressive (AR) models to flexibly trade off generation quality against computational cost using a single tokenizer. However, existing approaches based on nested dropout often fail to fully exploit the representational capacity of the tokenizer, resulting in suboptimal performance in both image reconstruction and generation. In this work, we introduce NesTok, a nested self-alignment framework tailored to dynamic visual tokenizers. NesTok introduces cross-length training, which jointly optimizes reconstruction across token lengths while using the full-length sequence to guide shorter counterparts, enabling shorter token sequences to approach the reconstruction quality of full-length sequences. On ImageNet, NesTok improves substantially over standard training and achieves an rFID score of 0.98. On downstream image generation, it achieves the state-of-the-art gFID score of 1.46 on ImageNet 256$\times$256 among existing variable-length autoregressive image generation methods. Code will be available at https://github.com/Jiawei804/NesTok.
End-to-end autonomous driving policies are commonly trained via imitation learning on logged demonstrations without observing the consequences of their own actions, leading to causal confusion in closed-loop real-world deployment. To address this issue, reinforcement learning (RL) post-training offers a promising alternative by leveraging world models as interactive training environments to enable future scene generation for policy improvement. Nevertheless, existing approaches either rely on reconstruction-based simulators, offering limited counterfactual interaction, or adopt synthetic simulators to enable long-horizon closed-loop interaction at the cost of a substantial sim-to-real gap. Recently, video world models have exhibited the ability to generate realistic multi-step future rollouts but may not faithfully reflect action conditions, resulting in action-vision mismatch. In this paper, we introduce RoXDrive, a plug-and-play closed-loop RL framework that enables reliable policy optimization by identifying action-faithful world-model rollouts, consisting of two stages: 1) Model pre-training: In addition to imitation-based policy pre-training, we devise an Action-Vision Faithfulness Evaluator for inverse dynamics estimation with our geometry-aware auxiliary trajectory supervision, enabling long-horizon assessment of whether visual dynamics faithfully reflect the conditioning ego actions. 2) Action-faithful RL post-training: Agents iteratively interact with world models to form long-horizon scene rollouts, retaining only action-faithful ones for dense safety-aware scoring and scene-level closed-loop RL post-training. Extensive experiments on nuScenes and an in-house dataset with over 130K training scenarios demonstrate consistent gains across planners, reducing safety violations by 27.6% with DiffusionDrive on nuScenes and 33.7% with Qwen3-VL on the internal data.
展开完整摘要收起摘要↓
End-to-end autonomous driving policies are commonly trained via imitation learning on logged demonstrations without observing the consequences of their own actions, leading to causal confusion in closed-loop real-world deployment. To address this issue, reinforcement learning (RL) post-training offers a promising alternative by leveraging world models as interactive training environments to enable future scene generation for policy improvement. Nevertheless, existing approaches either rely on reconstruction-based simulators, offering limited counterfactual interaction, or adopt synthetic simulators to enable long-horizon closed-loop interaction at the cost of a substantial sim-to-real gap. Recently, video world models have exhibited the ability to generate realistic multi-step future rollouts but may not faithfully reflect action conditions, resulting in action-vision mismatch. In this paper, we introduce RoXDrive, a plug-and-play closed-loop RL framework that enables reliable policy optimization by identifying action-faithful world-model rollouts, consisting of two stages: 1) Model pre-training: In addition to imitation-based policy pre-training, we devise an Action-Vision Faithfulness Evaluator for inverse dynamics estimation with our geometry-aware auxiliary trajectory supervision, enabling long-horizon assessment of whether visual dynamics faithfully reflect the conditioning ego actions. 2) Action-faithful RL post-training: Agents iteratively interact with world models to form long-horizon scene rollouts, retaining only action-faithful ones for dense safety-aware scoring and scene-level closed-loop RL post-training. Extensive experiments on nuScenes and an in-house dataset with over 130K training scenarios demonstrate consistent gains across planners, reducing safety violations by 27.6% with DiffusionDrive on nuScenes and 33.7% with Qwen3-VL on the internal data.
作者Mingu Kang, Yoori Oh, Sookyung Kim, Joonseok Lee
Latent world models learn action-conditioned dynamics in representation space and often score candidate actions by Euclidean distance to a goal representation. Joint training typically regularizes the representation to prevent collapse, but the resulting representation geometry also determines how terminal errors are weighted during planning. We show that accurate prediction and noncollapsed representations do not guarantee a task-aligned latent planning cost: isotropic Gaussian regularization can induce a geometry that ranks feasible outcomes differently from the task cost. To address this mismatch, we introduce AnisoWM with $Λ$Reg, which replaces the fixed isotropic Gaussian target with a learnable diagonal covariance under fixed-trace and anisotropy constraints. The prediction objective, predictor architecture, and Euclidean planner remain unchanged; the target is used only during training. Our analysis characterizes the prediction-driven allocation of target variance, its dependence on the training distribution, and the conditions under which the induced metric reduces planning regret. Across four visual control environments, AnisoWM improves planning success over LeWorldModel in all four. Its latent planning cost also shows better agreement with task outcomes. Project website: https://rkdrn79.github.io/AnisoWM-page/
展开完整摘要收起摘要↓
Latent world models learn action-conditioned dynamics in representation space and often score candidate actions by Euclidean distance to a goal representation. Joint training typically regularizes the representation to prevent collapse, but the resulting representation geometry also determines how terminal errors are weighted during planning. We show that accurate prediction and noncollapsed representations do not guarantee a task-aligned latent planning cost: isotropic Gaussian regularization can induce a geometry that ranks feasible outcomes differently from the task cost. To address this mismatch, we introduce AnisoWM with $Λ$Reg, which replaces the fixed isotropic Gaussian target with a learnable diagonal covariance under fixed-trace and anisotropy constraints. The prediction objective, predictor architecture, and Euclidean planner remain unchanged; the target is used only during training. Our analysis characterizes the prediction-driven allocation of target variance, its dependence on the training distribution, and the conditions under which the induced metric reduces planning regret. Across four visual control environments, AnisoWM improves planning success over LeWorldModel in all four. Its latent planning cost also shows better agreement with task outcomes. Project website: https://rkdrn79.github.io/AnisoWM-page/
作者Jiachang Zhang, Teng Hu, Bohao Feng, Songhang Shen, Wenqiang Wang, Hongqian Deng, Ran Yi
Video diffusion transformers (DiTs) increasingly adopt mixture-of-experts (MoE) architectures to reduce active computation, but their full expert storage remains costly. Existing one-shot pruning criteria mainly rely on static activation or routing statistics and cannot capture layer-level re-routing after expert deletion. We introduce DIET, a training-free expert pruning framework based on deletion responses. A single all-expert calibration pass records expert outputs and router states for matched conditional and unconditional tokens. Candidate deletions are then replayed from cached tensors, requiring no additional model forward passes. The resulting deletion-response signatures characterize each expert by the changes induced when it is removed. DIET selects retained experts by minimizing Overall Diversity Loss (ODL), which preserves directional coverage in signature space, and combines intra-layer local search with an inter-layer regression-guided budget search to allocate experts across layers. On LingBot-Video 30B-A3B, pruning 50% of experts (6,144 to 3,072) reduces the checkpoint from 57 GB to 30 GB and enables single-card deployment on a 48 GB GPU without fine-tuning. Under a fixed 284-case VBench protocol, the VBench Total increases from 0.7941 to 0.8115. Across tested retention budgets, DIET consistently outperforms competitive pruning baselines adapted from large language models.
展开完整摘要收起摘要↓
Video diffusion transformers (DiTs) increasingly adopt mixture-of-experts (MoE) architectures to reduce active computation, but their full expert storage remains costly. Existing one-shot pruning criteria mainly rely on static activation or routing statistics and cannot capture layer-level re-routing after expert deletion. We introduce DIET, a training-free expert pruning framework based on deletion responses. A single all-expert calibration pass records expert outputs and router states for matched conditional and unconditional tokens. Candidate deletions are then replayed from cached tensors, requiring no additional model forward passes. The resulting deletion-response signatures characterize each expert by the changes induced when it is removed. DIET selects retained experts by minimizing Overall Diversity Loss (ODL), which preserves directional coverage in signature space, and combines intra-layer local search with an inter-layer regression-guided budget search to allocate experts across layers. On LingBot-Video 30B-A3B, pruning 50% of experts (6,144 to 3,072) reduces the checkpoint from 57 GB to 30 GB and enables single-card deployment on a 48 GB GPU without fine-tuning. Under a fixed 284-case VBench protocol, the VBench Total increases from 0.7941 to 0.8115. Across tested retention budgets, DIET consistently outperforms competitive pruning baselines adapted from large language models.
作者Yu Yuan, Yawen Lu, Guoxian Song, Kevin Duarte, Ratheesh Kalarot, Di Chang, Xijun Wang, Stanley H. Chan
Most video editing methods focus on changing the appearance of the source video, while offering limited control over its dynamics. We introduce Reimagine Video Dynamics (RVD), a framework that disentangles a compact, editable dynamics token from visual context. We learn this token through self-supervised reconstruction: given the first frame as visual context, a renderer must recover the original video from the dynamics token, encouraging it to capture how the scene evolves rather than how it looks. This disentanglement allows video dynamics to be edited directly while preserving visual context. We develop a language-guided dynamics-token editor that transforms source dynamics into target dynamics, and train it with a scalable counterfactual video-pair pipeline and a two-stage training strategy. Extensive experiments show that RVD enables effective video dynamics editing, training-free retiming, and appearance-controlled re-rendering.
展开完整摘要收起摘要↓
Most video editing methods focus on changing the appearance of the source video, while offering limited control over its dynamics. We introduce Reimagine Video Dynamics (RVD), a framework that disentangles a compact, editable dynamics token from visual context. We learn this token through self-supervised reconstruction: given the first frame as visual context, a renderer must recover the original video from the dynamics token, encouraging it to capture how the scene evolves rather than how it looks. This disentanglement allows video dynamics to be edited directly while preserving visual context. We develop a language-guided dynamics-token editor that transforms source dynamics into target dynamics, and train it with a scalable counterfactual video-pair pipeline and a two-stage training strategy. Extensive experiments show that RVD enables effective video dynamics editing, training-free retiming, and appearance-controlled re-rendering.
Visual reward models are essential for evaluating and improving visual generation models, yet existing approaches typically map task conditions and candidate outputs directly to scalar rewards, leaving implicit what should be evaluated for each individual case. We introduce Think Before You Score, a paradigm that explicitly determines what matters for each case before judging how well the candidate performs. Following this principle, we propose the Thinking Reward Model (TRM), which formulates case-adaptive rubrics, performs rubric-guided assessment, and produces fine-grained pointwise rewards. We further observe that conventional pairwise preference optimization can induce score polarization, and introduce Pairwise Dual-Group Relative Policy Optimization (PD-GRPO), which leverages pairwise supervision to improve reward discrimination while preserving fine-grained pointwise scoring. Extensive experiments on image generation and editing reward-modeling benchmarks demonstrate that TRM achieves state-of-the-art performance among open-source reward models while remaining highly competitive with proprietary alternatives. Moreover, using TRM as a reward for reinforcement learning consistently improves diverse visual generation models, demonstrating that its fine-grained, case-adaptive rewards translate into effective optimization signals for visual generation.
展开完整摘要收起摘要↓
Visual reward models are essential for evaluating and improving visual generation models, yet existing approaches typically map task conditions and candidate outputs directly to scalar rewards, leaving implicit what should be evaluated for each individual case. We introduce Think Before You Score, a paradigm that explicitly determines what matters for each case before judging how well the candidate performs. Following this principle, we propose the Thinking Reward Model (TRM), which formulates case-adaptive rubrics, performs rubric-guided assessment, and produces fine-grained pointwise rewards. We further observe that conventional pairwise preference optimization can induce score polarization, and introduce Pairwise Dual-Group Relative Policy Optimization (PD-GRPO), which leverages pairwise supervision to improve reward discrimination while preserving fine-grained pointwise scoring. Extensive experiments on image generation and editing reward-modeling benchmarks demonstrate that TRM achieves state-of-the-art performance among open-source reward models while remaining highly competitive with proprietary alternatives. Moreover, using TRM as a reward for reinforcement learning consistently improves diverse visual generation models, demonstrating that its fine-grained, case-adaptive rewards translate into effective optimization signals for visual generation.
Latent world models often struggle with long-horizon planning despite accurate short-term predictions. Recursive rollouts accumulate errors, while distance concentration in high-dimensional latent spaces can weaken goal discrimination. We introduce the Dual-Latent World Model (Dual-WM), which separates local execution and long-range planning through distinct state representations and dynamics models. The low-level model predicts action-conditioned transitions, while the high-level model uses learned macro-actions to plan over longer temporal spans. We also propose Long-Horizon Representation Learning with Weighted Rollout (LoRe), which supervises self-generated predictions at both levels. An analysis of recursive error propagation motivates exponential horizon weights with separate decay rates for the two temporal scales. During planning, the high-level model generates latent subgoals that the low-level model refines into actions for precise execution. We evaluate from-scratch Dual-WM on five goal-conditioned visual control tasks against the task-wise strongest baselines without actor-guided proposals. At goal offsets of 50 and 100 environment steps, mean success increases from 75.9% to 84.4% and from 61.4% to 69.5%, respectively. At offset 100, Dual-WM outperforms these baselines on all five tasks and improves mean success over LeWM by 30.8 percentage points. Ablations and supporting analyses provide evidence of more informative representations for goal evaluation and greater consistency under recursive prediction. These results highlight the value of separating temporal roles and training across multiple horizons for reliable latent planning. Our core implementation is available at https://github.com/DeLin1001/Dual-WM-Official.
展开完整摘要收起摘要↓
Latent world models often struggle with long-horizon planning despite accurate short-term predictions. Recursive rollouts accumulate errors, while distance concentration in high-dimensional latent spaces can weaken goal discrimination. We introduce the Dual-Latent World Model (Dual-WM), which separates local execution and long-range planning through distinct state representations and dynamics models. The low-level model predicts action-conditioned transitions, while the high-level model uses learned macro-actions to plan over longer temporal spans. We also propose Long-Horizon Representation Learning with Weighted Rollout (LoRe), which supervises self-generated predictions at both levels. An analysis of recursive error propagation motivates exponential horizon weights with separate decay rates for the two temporal scales. During planning, the high-level model generates latent subgoals that the low-level model refines into actions for precise execution. We evaluate from-scratch Dual-WM on five goal-conditioned visual control tasks against the task-wise strongest baselines without actor-guided proposals. At goal offsets of 50 and 100 environment steps, mean success increases from 75.9% to 84.4% and from 61.4% to 69.5%, respectively. At offset 100, Dual-WM outperforms these baselines on all five tasks and improves mean success over LeWM by 30.8 percentage points. Ablations and supporting analyses provide evidence of more informative representations for goal evaluation and greater consistency under recursive prediction. These results highlight the value of separating temporal roles and training across multiple horizons for reliable latent planning. Our core implementation is available at https://github.com/DeLin1001/Dual-WM-Official.
作者I. Samuel Akinwande, Mykel J. Kochenderfer, Clark Barrett
Verifying a vision-based neural feedback system requires a model of the observations its controller acts upon. Such a model must capture the variation the sensor produces, while remaining tractable for closed-loop analysis. Generative adversarial networks (GANs) have served as perception surrogates, but they are large, reproduce complex scenes poorly, and are hard to verify. We explore stochastic world models as a richer class of perception surrogates. We train a world model with physically grounded latents, built from operations that standard verifiers bound. It reproduces held-out frames more faithfully than GAN surrogates with up to 130 times as many parameters. To verify these surrogates, we develop a procedure that combines falsification, adaptive refinement, symbolic, and backward analyses. On an emergency braking benchmark with a GAN surrogate, our procedure resolves the entire state space, 38% of which the state-of-the-art verifier left unresolved. On the RGB version of the benchmark, where no verification results have previously been reported, our procedure resolves over 80% of the state space with a world model surrogate.
展开完整摘要收起摘要↓
Verifying a vision-based neural feedback system requires a model of the observations its controller acts upon. Such a model must capture the variation the sensor produces, while remaining tractable for closed-loop analysis. Generative adversarial networks (GANs) have served as perception surrogates, but they are large, reproduce complex scenes poorly, and are hard to verify. We explore stochastic world models as a richer class of perception surrogates. We train a world model with physically grounded latents, built from operations that standard verifiers bound. It reproduces held-out frames more faithfully than GAN surrogates with up to 130 times as many parameters. To verify these surrogates, we develop a procedure that combines falsification, adaptive refinement, symbolic, and backward analyses. On an emergency braking benchmark with a GAN surrogate, our procedure resolves the entire state space, 38% of which the state-of-the-art verifier left unresolved. On the RGB version of the benchmark, where no verification results have previously been reported, our procedure resolves over 80% of the state space with a world model surrogate.
作者Haoyi Jiang, Liu Liu, Xinjiang Wang, Zhihao Sun, Zequn Chen, Sen Wang, Xinjie Wang, Xia Chen, Jingfeng Yao, Weiheng Zhao, Shanglin Yuan, Zhizhong Su, Wei Sui, Wenyu Liu, Xinggang Wang
World-action models jointly learn robot policies and predict future observations, making the representation space an interface between control and prediction. We study the design of this space through controlled comparisons, finding that neither reconstruction fidelity nor pre-trained perceptual features alone ensure effective policy learning. These findings motivate ReWAM, a representation-centric world-action model built on pre-trained DINO features. Feature Calibration and a Temporal Representation Bottleneck organize these features into compact world states suited to dynamics modeling. Action-Grounded Representation Shaping routes only action-loss gradients to the bottleneck, thereby letting the policy shape what the representation encodes while the world model learns how it evolves. Without generative video pre-training, ReWAM achieves 93.6% success on RoboTwin 2.0. On RoboDojo, it achieves an average score of 12.29 and a success rate of 8.28% using approximately 600 hours of embodied pre-training data.
展开完整摘要收起摘要↓
World-action models jointly learn robot policies and predict future observations, making the representation space an interface between control and prediction. We study the design of this space through controlled comparisons, finding that neither reconstruction fidelity nor pre-trained perceptual features alone ensure effective policy learning. These findings motivate ReWAM, a representation-centric world-action model built on pre-trained DINO features. Feature Calibration and a Temporal Representation Bottleneck organize these features into compact world states suited to dynamics modeling. Action-Grounded Representation Shaping routes only action-loss gradients to the bottleneck, thereby letting the policy shape what the representation encodes while the world model learns how it evolves. Without generative video pre-training, ReWAM achieves 93.6% success on RoboTwin 2.0. On RoboDojo, it achieves an average score of 12.29 and a success rate of 8.28% using approximately 600 hours of embodied pre-training data.
Inpainting 3D Gaussian Splatting scenes, a key challenge in 3D editing, requires generating plausible content within a masked region of 3D space. Prior approaches rely on 2D diffusion models to produce one or several inpainted reference views, making them susceptible to challenges associated with multi-view inconsistency and lengthy optimization times. Departing from these approaches, we introduce a reference-free Gaussian splatting inpainting method operating natively in 3D. Our method leverages the embedding space of a large, pre-trained 3D prior, combined with a structure completion network to feed a generative prior which reconstructs the missing region's geometry and appearance. Performing content generation entirely in 3D, it avoids the need to reconcile inconsistencies of multiple inpainted reference images, and is faster than related 2D-based methods. We demonstrate the effectiveness of our method qualitatively and quantitatively, through extensive experiments and a user study. To the best of our knowledge, this work is the first Gaussian splatting inpainting method to operate in the learned representation space of a 3D-native generative prior without relying on inpainted reference views.
展开完整摘要收起摘要↓
Inpainting 3D Gaussian Splatting scenes, a key challenge in 3D editing, requires generating plausible content within a masked region of 3D space. Prior approaches rely on 2D diffusion models to produce one or several inpainted reference views, making them susceptible to challenges associated with multi-view inconsistency and lengthy optimization times. Departing from these approaches, we introduce a reference-free Gaussian splatting inpainting method operating natively in 3D. Our method leverages the embedding space of a large, pre-trained 3D prior, combined with a structure completion network to feed a generative prior which reconstructs the missing region's geometry and appearance. Performing content generation entirely in 3D, it avoids the need to reconcile inconsistencies of multiple inpainted reference images, and is faster than related 2D-based methods. We demonstrate the effectiveness of our method qualitatively and quantitatively, through extensive experiments and a user study. To the best of our knowledge, this work is the first Gaussian splatting inpainting method to operate in the learned representation space of a 3D-native generative prior without relying on inpainted reference views.
Distribution matching distillation (DMD) provides a general framework for few-step diffusion generation, but its modern text-to-image instantiations have been developed primarily around latent diffusion. It therefore overlooks key properties and design opportunities of native RGB. We revisit two DMD interfaces for pixel-space teachers. On the teacher-matching side, diagnostics show low-noise RGB matching is dominated by a local-texture cue, motivating a fixed high-noise matching band. On the real-data side, native clean-RGB outputs allow guidance from an external visual representation without traversing a decoder or sharing the heavy fake-score critic. DINO-Adv removes this critic from the adversarial gradient path and supplies local parametric patch guidance. For distribution-level guidance, we introduce AF-Loss, a parameter-free auxiliary semantic distribution-field objective designed for text-to-image DMD. It operates on detached rolling real and generated supports in the shared DINOv2 space while preserving prompt-conditioned teacher supervision. AF-Loss adds no learnable parameters or inference-time computation. Together these designs form DMA$^2$. Across DPG-Bench, GenEval, VQAScore, and COCO30K, the four-step DMA$^2$ student performs better than the 25-step teacher and evaluated few-step distillers.
展开完整摘要收起摘要↓
Distribution matching distillation (DMD) provides a general framework for few-step diffusion generation, but its modern text-to-image instantiations have been developed primarily around latent diffusion. It therefore overlooks key properties and design opportunities of native RGB. We revisit two DMD interfaces for pixel-space teachers. On the teacher-matching side, diagnostics show low-noise RGB matching is dominated by a local-texture cue, motivating a fixed high-noise matching band. On the real-data side, native clean-RGB outputs allow guidance from an external visual representation without traversing a decoder or sharing the heavy fake-score critic. DINO-Adv removes this critic from the adversarial gradient path and supplies local parametric patch guidance. For distribution-level guidance, we introduce AF-Loss, a parameter-free auxiliary semantic distribution-field objective designed for text-to-image DMD. It operates on detached rolling real and generated supports in the shared DINOv2 space while preserving prompt-conditioned teacher supervision. AF-Loss adds no learnable parameters or inference-time computation. Together these designs form DMA$^2$. Across DPG-Bench, GenEval, VQAScore, and COCO30K, the four-step DMA$^2$ student performs better than the 25-step teacher and evaluated few-step distillers.
作者Sieun Hyeon, Yejoon Lee, Mintaek Lim, Woojin Kim, Jaeik Kim, Jaeyoung Do
Omnimodal evaluation should go beyond independent text, image, and speech production: individually plausible outputs may not express a coherent shared event. We introduce Omni-StoryBench, a story-grounded omnimodal benchmark evaluating whether models can coherently continue stories across image, narration, and speech. Each instance provides a current storybook page and structured next-page conditions, requiring models to generate the next illustration, narration, and spoken character utterance. Omni-StoryBench contains 900 rigorously validated story transitions from openly licensed children's books, with ground-truth next-page references and speech metadata. We evaluate systems with modality-specific metrics and consistency-centered LLM-as-a-judge rubrics for context preservation, condition following, reference consistency, and cross-modal coherence. Across 32 baseline configurations spanning orchestration, semi-orchestration, and native any-to-any paradigms, we find orchestration with strong VLM planning most reliable, while current native omnimodal models often struggle with output completeness and controllability. Our analysis shows text-side performance is associated with image and speech quality, but image generation and visual continuity form the clearest observed bottleneck among the evaluated configurations. These results position Omni-StoryBench as a system-level benchmark measuring coherent omnimodal generation beyond isolated modality quality.
展开完整摘要收起摘要↓
Omnimodal evaluation should go beyond independent text, image, and speech production: individually plausible outputs may not express a coherent shared event. We introduce Omni-StoryBench, a story-grounded omnimodal benchmark evaluating whether models can coherently continue stories across image, narration, and speech. Each instance provides a current storybook page and structured next-page conditions, requiring models to generate the next illustration, narration, and spoken character utterance. Omni-StoryBench contains 900 rigorously validated story transitions from openly licensed children's books, with ground-truth next-page references and speech metadata. We evaluate systems with modality-specific metrics and consistency-centered LLM-as-a-judge rubrics for context preservation, condition following, reference consistency, and cross-modal coherence. Across 32 baseline configurations spanning orchestration, semi-orchestration, and native any-to-any paradigms, we find orchestration with strong VLM planning most reliable, while current native omnimodal models often struggle with output completeness and controllability. Our analysis shows text-side performance is associated with image and speech quality, but image generation and visual continuity form the clearest observed bottleneck among the evaluated configurations. These results position Omni-StoryBench as a system-level benchmark measuring coherent omnimodal generation beyond isolated modality quality.
作者Xiaokang Ye, Siddhant Hitesh Mantri, Zimeng Chen, Edward Zhang, Zhaoxu Zheng, Yuanheng Li, Yizhao Chen, Tianyang Huang, Lianhui Qin
Frontier coding agents can now write and execute code that authors 3D environments, but whether they reliably understand 3D structure and precisely control scene state remains unclear. The generated 3D scene is a persistent, executable artifact: a convincing render can hide incorrect spatial relations, intersecting objects, or unintended modifications. We introduce Code4Scene, a benchmark of 190 Unreal Engine cases built from human-assembled scenes that evaluates coding agents on two complementary settings under a shared execution interface. Construction tests scene-level spatial reasoning from open-ended language specifications, where many realizations are valid; editing tests precise control of scene state, where the agent must recover the target scene from reference images while preserving everything else. Rather than scoring code or rendered views, Code4Scene evaluates the generated engine-native scene for task fulfillment, artifact integrity, and static physical validity, with edits additionally compared against withheld ground truth. Across 14 coding-agent configurations on the 95-case public set, construction and editing performance are strongly correlated but not interchangeable (Spearman $ρ= 0.78$): Claude Fable 5.1 leads construction, Gemini 3.8 Flash leads editing, and GPT-6 Astra narrowly leads overall. Spatial Composition is the weakest construction category for every agent, while editing remains imprecise: the best Repair F1 is only 0.527, and 35.8% of edits that fully recover the target still introduce unintended changes elsewhere in the scene. These results expose a gap between plausible 3D generation and reliable spatial reasoning and state control.
展开完整摘要收起摘要↓
Frontier coding agents can now write and execute code that authors 3D environments, but whether they reliably understand 3D structure and precisely control scene state remains unclear. The generated 3D scene is a persistent, executable artifact: a convincing render can hide incorrect spatial relations, intersecting objects, or unintended modifications. We introduce Code4Scene, a benchmark of 190 Unreal Engine cases built from human-assembled scenes that evaluates coding agents on two complementary settings under a shared execution interface. Construction tests scene-level spatial reasoning from open-ended language specifications, where many realizations are valid; editing tests precise control of scene state, where the agent must recover the target scene from reference images while preserving everything else. Rather than scoring code or rendered views, Code4Scene evaluates the generated engine-native scene for task fulfillment, artifact integrity, and static physical validity, with edits additionally compared against withheld ground truth. Across 14 coding-agent configurations on the 95-case public set, construction and editing performance are strongly correlated but not interchangeable (Spearman $ρ= 0.78$): Claude Fable 5.1 leads construction, Gemini 3.8 Flash leads editing, and GPT-6 Astra narrowly leads overall. Spatial Composition is the weakest construction category for every agent, while editing remains imprecise: the best Repair F1 is only 0.527, and 35.8% of edits that fully recover the target still introduce unintended changes elsewhere in the scene. These results expose a gap between plausible 3D generation and reliable spatial reasoning and state control.
Latent world models are trained to predict what happens next, so nothing in their objective separates what an action caused from what merely co-occurred with it. Object-masking models such as C-JEPA intervene on what the predictor can see; we intervene on what physically happens. From one saved simulator state we run the dynamics under an action $a$ and under a reference action $a_{\varnothing}$, and train the model to predict the difference $Δz=z^{a}-z^{a_{\varnothing}}$ between the two latent futures. The resulting objective, Do-JEPA, has an effect loss, a support loss (where the action enters), a propagation loss (where its effect travels) and invariance losses (what must not change). In a synthetic system with object-aligned variables, support supervision finds the directly intervened object in 99.95% of test cases, where a sparse action mask sends the action to a nuisance slot in every case, and response-onset supervision recovers the ring-shaped propagation graph (edge AUROC 0.975 vs. 0.624). From pixels, the effect loss beats a control trained on exactly the same data: it lowers latent effect error by 28.4% on an end-to-end LeWM model and physical effect error by 13.5% when trained and tested on natural action sequences, and on three independently generated CausalWorld benchmarks it lowers responsive effect error by about 20% under physics shifts and the latent context sensitivity of predicted effects by 66%. Trained from scratch it costs factual accuracy; fine-tuning an existing model with it removes this cost. Together, these results show that intervening on the world, rather than on what the model sees, helps latent world models predict what their actions cause.
展开完整摘要收起摘要↓
Latent world models are trained to predict what happens next, so nothing in their objective separates what an action caused from what merely co-occurred with it. Object-masking models such as C-JEPA intervene on what the predictor can see; we intervene on what physically happens. From one saved simulator state we run the dynamics under an action $a$ and under a reference action $a_{\varnothing}$, and train the model to predict the difference $Δz=z^{a}-z^{a_{\varnothing}}$ between the two latent futures. The resulting objective, Do-JEPA, has an effect loss, a support loss (where the action enters), a propagation loss (where its effect travels) and invariance losses (what must not change). In a synthetic system with object-aligned variables, support supervision finds the directly intervened object in 99.95% of test cases, where a sparse action mask sends the action to a nuisance slot in every case, and response-onset supervision recovers the ring-shaped propagation graph (edge AUROC 0.975 vs. 0.624). From pixels, the effect loss beats a control trained on exactly the same data: it lowers latent effect error by 28.4% on an end-to-end LeWM model and physical effect error by 13.5% when trained and tested on natural action sequences, and on three independently generated CausalWorld benchmarks it lowers responsive effect error by about 20% under physics shifts and the latent context sensitivity of predicted effects by 66%. Trained from scratch it costs factual accuracy; fine-tuning an existing model with it removes this cost. Together, these results show that intervening on the world, rather than on what the model sees, helps latent world models predict what their actions cause.
作者Xuanyu Zhu, Yan Bai, Yang Shi, Yihang Lou, Yuanxing Zhang, Tengfei Liu, Jing Jin, Yuan Zhou
Pretrained visual representations support image generation, but may not fully preserve the fine-grained details needed for faithful reconstruction. Meanwhile, intermediate encoder layers contain complementary visual details, but learning to fuse them for reconstruction can produce a latent distribution that is difficult to model. Existing fusion methods require empirical tuning of layer selection or staged optimization of fusion and decoding, increasing configuration effort or training complexity. We introduce HiRAE (Hierarchical Representation Autoencoder), which learns a hierarchical fusion framework over the full encoder hierarchy to improve reconstruction fidelity while maintaining compatibility with generative modeling. HiRAE groups encoder layers by depth and learns residual corrections to the deepest representation. Group-wise norm caps bound these corrections relative to the deep anchor, with tighter budgets for shallower groups. Our HiRAE-24 preserves the latent token count and channel dimension. On ImageNet-256, HiRAE-24 reduces reconstruction FID from 0.299 to 0.209 relative to RAEv2 while maintaining competitive guided generation quality. For text-to-image generation, HiRAE-24 improves alignment over RAEv2 on GenEval, DPG-Bench, and GenAI-Bench both before and after supervised fine-tuning. Under the same generator-training and evaluation protocol, post-fine-tuning GenEval increases from 84.86 to 87.70.
展开完整摘要收起摘要↓
Pretrained visual representations support image generation, but may not fully preserve the fine-grained details needed for faithful reconstruction. Meanwhile, intermediate encoder layers contain complementary visual details, but learning to fuse them for reconstruction can produce a latent distribution that is difficult to model. Existing fusion methods require empirical tuning of layer selection or staged optimization of fusion and decoding, increasing configuration effort or training complexity. We introduce HiRAE (Hierarchical Representation Autoencoder), which learns a hierarchical fusion framework over the full encoder hierarchy to improve reconstruction fidelity while maintaining compatibility with generative modeling. HiRAE groups encoder layers by depth and learns residual corrections to the deepest representation. Group-wise norm caps bound these corrections relative to the deep anchor, with tighter budgets for shallower groups. Our HiRAE-24 preserves the latent token count and channel dimension. On ImageNet-256, HiRAE-24 reduces reconstruction FID from 0.299 to 0.209 relative to RAEv2 while maintaining competitive guided generation quality. For text-to-image generation, HiRAE-24 improves alignment over RAEv2 on GenEval, DPG-Bench, and GenAI-Bench both before and after supervised fine-tuning. Under the same generator-training and evaluation protocol, post-fine-tuning GenEval increases from 84.86 to 87.70.
Interactive panoramic video generation aims to synthesize immersive 360\textdegree{} videos that remain visually coherent while following user-specified camera trajectories during exploration. However, progress is limited by a coupled data-and-model gap: existing panoramic video datasets are often short, weakly annotated, or lack camera trajectories, while existing camera-controlled video generation models are designed for perspective videos and do not directly support panoramic geometry. In this paper, we introduce MUGEN and Wan360 to address these limitations. MUGEN is a large-scale real-world panoramic video dataset tailored to interactive 360-degree world exploration, comprising over 1,300 hours of at least 4K panoramic videos with rich semantic and geometric annotations. Built on MUGEN, we further present Wan360, a camera-controllable interactive panoramic video generation model. Panoramic videos are commonly represented by EquiRectangular Projection (ERP), which unfolds a spherical 360-degree view into a rectangular frame with cyclic longitude seams and pole distortions. To this end, Wan360 introduces three parameter-free ERP-aware components: periodic longitude RoPE for seam-consistent positional encoding, ERP-aware padding for reducing boundary artifacts, and random roll yaw for consistent learning. For camera control, Wan360 uses a panoramic Plücker embedding that represents camera motion with ERP rays rather than perspective pinhole rays. Experiments show that MUGEN serves as a data foundation for panoramic world exploration, and that Wan360 enables high-quality, temporally coherent, camera-controllable 360-degree video generation.
展开完整摘要收起摘要↓
Interactive panoramic video generation aims to synthesize immersive 360\textdegree{} videos that remain visually coherent while following user-specified camera trajectories during exploration. However, progress is limited by a coupled data-and-model gap: existing panoramic video datasets are often short, weakly annotated, or lack camera trajectories, while existing camera-controlled video generation models are designed for perspective videos and do not directly support panoramic geometry. In this paper, we introduce MUGEN and Wan360 to address these limitations. MUGEN is a large-scale real-world panoramic video dataset tailored to interactive 360-degree world exploration, comprising over 1,300 hours of at least 4K panoramic videos with rich semantic and geometric annotations. Built on MUGEN, we further present Wan360, a camera-controllable interactive panoramic video generation model. Panoramic videos are commonly represented by EquiRectangular Projection (ERP), which unfolds a spherical 360-degree view into a rectangular frame with cyclic longitude seams and pole distortions. To this end, Wan360 introduces three parameter-free ERP-aware components: periodic longitude RoPE for seam-consistent positional encoding, ERP-aware padding for reducing boundary artifacts, and random roll yaw for consistent learning. For camera control, Wan360 uses a panoramic Plücker embedding that represents camera motion with ERP rays rather than perspective pinhole rays. Experiments show that MUGEN serves as a data foundation for panoramic world exploration, and that Wan360 enables high-quality, temporally coherent, camera-controllable 360-degree video generation.
作者Weihan Xu, Kan Jen Cheng, Koichi Saito, Jingyu Shi, Tingle Li, Yisi Liu, Liming Wang, Masato Ishii, Takashi Shibuya, Gopala Anumanchipalli, Paul Pu Liang
Adding or removing a sounding object requires coordinated changes to visual content and sound while preserving the surrounding scene. Yet paired supervision for localized non-speech audiovisual editing remains limited, as visual and acoustic edits must target the same object and isolate its sound from overlapping sources. To address this gap, we introduce AVIOBench, a dataset comprising 37.9 hours of paired audiovisual examples spanning 1{,}878 target-object names. AVIOBench links the visual presence and acoustic contribution of each target object through a shared identity and visual mask. Our automated pipeline uses visual grounding and cross-modal consistency to select target-sound removal candidates, then jointly refines the audiovisual pairs to improve perceptual quality and cross-modal consistency. Building on this dataset, we propose AVIO, which adapts a pretrained text-to audiovisual generation model through source-conditioned feature modulation to jointly learn object addition and removal. A reference-frame curriculum gradually reduces reference conditioning during training, enabling one model to perform instruction-only editing with optional visual guidance. Quantitative and qualitative evaluations demonstrate effective audiovisual object removal and addition, with optional reference guidance providing appearance and placement control for addition.
展开完整摘要收起摘要↓
Adding or removing a sounding object requires coordinated changes to visual content and sound while preserving the surrounding scene. Yet paired supervision for localized non-speech audiovisual editing remains limited, as visual and acoustic edits must target the same object and isolate its sound from overlapping sources. To address this gap, we introduce AVIOBench, a dataset comprising 37.9 hours of paired audiovisual examples spanning 1{,}878 target-object names. AVIOBench links the visual presence and acoustic contribution of each target object through a shared identity and visual mask. Our automated pipeline uses visual grounding and cross-modal consistency to select target-sound removal candidates, then jointly refines the audiovisual pairs to improve perceptual quality and cross-modal consistency. Building on this dataset, we propose AVIO, which adapts a pretrained text-to audiovisual generation model through source-conditioned feature modulation to jointly learn object addition and removal. A reference-frame curriculum gradually reduces reference conditioning during training, enabling one model to perform instruction-only editing with optional visual guidance. Quantitative and qualitative evaluations demonstrate effective audiovisual object removal and addition, with optional reference guidance providing appearance and placement control for addition.
Interpretability methods such as probes, activation patches and learned editors are designed to reveal or modify a model's current computation. World models pose a harder requirement: because their predictions become inputs to later predictions, a useful internal correction must survive after editing stops. We therefore propose RolloutFaith, a framework that measures semantic improvement both in the prediction produced at intervention time and over later autonomous predictions under fixed events, actions, noise, and information budgets. We evaluate ten fitted editors on three world models across Crafter, Cartpole, and CoinRun. We also use Reference Activation Patching, which replaces a model activation with the paired activation computed from the real observation, to measure the correction available at the chosen interface. This reference intervention improves later predictions in all nine model and task combinations and outperforms the best fitted editor in eight, yet its sustained gain decreases with horizon in five of nine combinations. Current fitted editors recover only limited and inconsistent long term effects. By restoring individual state components to their untouched values, we find that persistent effects travel through the newest generated frame in DIAMOND, recurrent memory in DreamerV3, and both in STORM. These findings suggest that training should reward future consequences. To test this hypothesis, we propose Delayed LoReFT, which optimizes the same low rank intervention through four frozen future transitions and improves sustained intervention effects to some extent.
展开完整摘要收起摘要↓
Interpretability methods such as probes, activation patches and learned editors are designed to reveal or modify a model's current computation. World models pose a harder requirement: because their predictions become inputs to later predictions, a useful internal correction must survive after editing stops. We therefore propose RolloutFaith, a framework that measures semantic improvement both in the prediction produced at intervention time and over later autonomous predictions under fixed events, actions, noise, and information budgets. We evaluate ten fitted editors on three world models across Crafter, Cartpole, and CoinRun. We also use Reference Activation Patching, which replaces a model activation with the paired activation computed from the real observation, to measure the correction available at the chosen interface. This reference intervention improves later predictions in all nine model and task combinations and outperforms the best fitted editor in eight, yet its sustained gain decreases with horizon in five of nine combinations. Current fitted editors recover only limited and inconsistent long term effects. By restoring individual state components to their untouched values, we find that persistent effects travel through the newest generated frame in DIAMOND, recurrent memory in DreamerV3, and both in STORM. These findings suggest that training should reward future consequences. To test this hypothesis, we propose Delayed LoReFT, which optimizes the same low rank intervention through four frozen future transitions and improves sustained intervention effects to some extent.
Autoregressive (AR) video diffusion enables low-latency, streamable video generation, but prediction errors often accumulate over long rollouts. Training the generator on its own rollouts exposes it to these imperfect histories. However, existing video-level distribution matching distillation (DMD) scores the whole rollout jointly. Because a chunk is evaluated together with its past and future, its correction can favor matching artifacts in the surrounding context merely to preserve temporal consistency. To provide a clearer visual-quality signal, we introduce Rollout-Marginal Distillation (RMD). RMD retains the generated history for AR prediction but scores each chunk independently against a chunk teacher, ensuring its quality correction is not compromised by an imperfect temporal context. To compensate for the lack of temporal context in independent chunk scoring, RMD subsequently applies video-level DMD to restore temporal coherence. Extensive experiments demonstrate that RMD maintains high visual quality far beyond its training horizon and outperforms video-level DMD baselines. Code and video results are available at https://cjeen.github.io/RMD
展开完整摘要收起摘要↓
Autoregressive (AR) video diffusion enables low-latency, streamable video generation, but prediction errors often accumulate over long rollouts. Training the generator on its own rollouts exposes it to these imperfect histories. However, existing video-level distribution matching distillation (DMD) scores the whole rollout jointly. Because a chunk is evaluated together with its past and future, its correction can favor matching artifacts in the surrounding context merely to preserve temporal consistency. To provide a clearer visual-quality signal, we introduce Rollout-Marginal Distillation (RMD). RMD retains the generated history for AR prediction but scores each chunk independently against a chunk teacher, ensuring its quality correction is not compromised by an imperfect temporal context. To compensate for the lack of temporal context in independent chunk scoring, RMD subsequently applies video-level DMD to restore temporal coherence. Extensive experiments demonstrate that RMD maintains high visual quality far beyond its training horizon and outperforms video-level DMD baselines. Code and video results are available at https://cjeen.github.io/RMD
World simulation is inherently multisensory, demanding synchronized visual and acoustic dynamics in real time. Yet prevailing interactive world models remain strictly silent, focusing exclusively on visual rendering and control while overlooking the acoustic dimension. We present HelixWorld, a real-time interactive audio-visual world model where visual scenes and camera-grounded spatial stereo sound co-evolve natively under user interaction. We curate a high-fidelity spatial audio-visual dataset with true stereo acoustics and metric camera poses, upon which we pre-train a bidirectional teacher conditioned on 6-DoF camera trajectories and user actions. To enable low-latency causal interaction, we distill the teacher into a few-step streaming student via an online trajectory distillation loss, sustaining drift-free joint audio-visual rollouts at 24 FPS on a single GPU. Furthermore, we formalize spatial-acoustic consistency and introduce HelixBench to evaluate whether synthesized sound fields faithfully track dynamic viewpoint motion. Extensive experiments demonstrate that HelixWorld matches state-of-the-art silent world models in visual fidelity and responsiveness, while significantly surpassing existing baselines in camera-aligned spatial-acoustic immersion.
展开完整摘要收起摘要↓
World simulation is inherently multisensory, demanding synchronized visual and acoustic dynamics in real time. Yet prevailing interactive world models remain strictly silent, focusing exclusively on visual rendering and control while overlooking the acoustic dimension. We present HelixWorld, a real-time interactive audio-visual world model where visual scenes and camera-grounded spatial stereo sound co-evolve natively under user interaction. We curate a high-fidelity spatial audio-visual dataset with true stereo acoustics and metric camera poses, upon which we pre-train a bidirectional teacher conditioned on 6-DoF camera trajectories and user actions. To enable low-latency causal interaction, we distill the teacher into a few-step streaming student via an online trajectory distillation loss, sustaining drift-free joint audio-visual rollouts at 24 FPS on a single GPU. Furthermore, we formalize spatial-acoustic consistency and introduce HelixBench to evaluate whether synthesized sound fields faithfully track dynamic viewpoint motion. Extensive experiments demonstrate that HelixWorld matches state-of-the-art silent world models in visual fidelity and responsiveness, while significantly surpassing existing baselines in camera-aligned spatial-acoustic immersion.
作者Zhirui Xing, Long Ye, Kaige Li, Ziyi Xu, Ming Meng
Co-speech gesture generation aims to synthesize natural gestures that are both temporally synchronized with speech and semantically consistent with the spoken content. Although recent methods can generate rhythmically plausible motions, they often rely heavily on acoustic prosody while underutilizing textual semantics, especially when semantic annotations are incomplete, noisy, or unavailable. Consequently, the generated gestures may follow speech rhythm while failing to express the intended semantics. To address this problem, we propose a reliability-aware semantic-rhythm control framework for co-speech gesture generation. We first learn a discrete motion prior that represents continuous gestures in a compact and structured motion-code space. We then introduce a dual-branch semantic contribution estimation mechanism consisting of a full multimodal branch and an audio-only branch. Their distributional discrepancy is formulated as conditional information gain to quantify how much textual semantics changes the predicted motion. Based on this estimate, a controllable semantic-rhythm objective selectively strengthens semantic guidance in content-relevant segments while limiting unnecessary semantic intervention in rhythm-dominant segments. Furthermore, we treat background noise as an acoustic reliability condition and introduce noise-conditioned feature modulation together with beneficial latent perturbation to improve generation robustness under realistic acoustic environments. Experiments on benchmark datasets demonstrate that the proposed framework achieves a favorable balance among semantic expressiveness, rhythmic synchronization, motion diversity, and robustness, enabling reliable and controllable co-speech gesture generation.
展开完整摘要收起摘要↓
Co-speech gesture generation aims to synthesize natural gestures that are both temporally synchronized with speech and semantically consistent with the spoken content. Although recent methods can generate rhythmically plausible motions, they often rely heavily on acoustic prosody while underutilizing textual semantics, especially when semantic annotations are incomplete, noisy, or unavailable. Consequently, the generated gestures may follow speech rhythm while failing to express the intended semantics. To address this problem, we propose a reliability-aware semantic-rhythm control framework for co-speech gesture generation. We first learn a discrete motion prior that represents continuous gestures in a compact and structured motion-code space. We then introduce a dual-branch semantic contribution estimation mechanism consisting of a full multimodal branch and an audio-only branch. Their distributional discrepancy is formulated as conditional information gain to quantify how much textual semantics changes the predicted motion. Based on this estimate, a controllable semantic-rhythm objective selectively strengthens semantic guidance in content-relevant segments while limiting unnecessary semantic intervention in rhythm-dominant segments. Furthermore, we treat background noise as an acoustic reliability condition and introduce noise-conditioned feature modulation together with beneficial latent perturbation to improve generation robustness under realistic acoustic environments. Experiments on benchmark datasets demonstrate that the proposed framework achieves a favorable balance among semantic expressiveness, rhythmic synchronization, motion diversity, and robustness, enabling reliable and controllable co-speech gesture generation.
Modern image editors excel at semantic manipulation and visual synthesis, yet remain limited in precise spatial control, motivating the development of drag-based editing. However, existing drag-based methods often struggle to balance drag accuracy with natural, plausible, and intent-aligned generation. We propose MoRe-Drag, a motion-grounded drag-based editing method. Our key insight is to treat pixel-space warping as coarse motion evidence, and to inject this evidence into the generative sampling trajectory. Specifically, MoRe-Drag performs region-aware latent recomposition over refinement, inpainting, and anchor regions, coupled with stage-adaptive conditioning that progressively shifts from motion-grounded structure formation to semantic refinement. We further support an instruction-free interface by adapting the MLLM-based text encoder for drag-aware instruction inference. Experiments on DragBench-SR and DragBench-DR show that MoRe-Drag substantially improves drag precision over strong base editors and achieves superior drag accuracy among SOTA drag-based methods, while delivering strong semantic consistency and visually realistic results. Code and dataset will be publicly released.
展开完整摘要收起摘要↓
Modern image editors excel at semantic manipulation and visual synthesis, yet remain limited in precise spatial control, motivating the development of drag-based editing. However, existing drag-based methods often struggle to balance drag accuracy with natural, plausible, and intent-aligned generation. We propose MoRe-Drag, a motion-grounded drag-based editing method. Our key insight is to treat pixel-space warping as coarse motion evidence, and to inject this evidence into the generative sampling trajectory. Specifically, MoRe-Drag performs region-aware latent recomposition over refinement, inpainting, and anchor regions, coupled with stage-adaptive conditioning that progressively shifts from motion-grounded structure formation to semantic refinement. We further support an instruction-free interface by adapting the MLLM-based text encoder for drag-aware instruction inference. Experiments on DragBench-SR and DragBench-DR show that MoRe-Drag substantially improves drag precision over strong base editors and achieves superior drag accuracy among SOTA drag-based methods, while delivering strong semantic consistency and visually realistic results. Code and dataset will be publicly released.
作者Xiangcheng Zhan, Zirui Chen, Yicheng Zhao, Ziteng Gao, Shuo Yang
World-Action Models (WAMs) couple action generation with predictions of how physical interactions unfold. However, current post-deployment learning paradigms typically improve behavior without requiring better world predictions. Especially in dexterous manipulation, small execution errors can compound in high-dimensional action spaces, hindering policy improvement and pushing interactions beyond the world model's training distribution. Motivated by this, we propose Direct Experience World-Model Optimization (DEWO), a post-deployment learning paradigm for WAMs that, alongside action imitation, refines world representations through visual experience to better condition action generation. Specifically, it identifies interaction turning points and learns from successful and failed futures to support classifier-free guidance. An additional value head estimates task progress from video representations and activates guidance when progress stalls during inference. Across five DexJoCo tasks, DEWO improves average success across all three WAM formulations. Ablations show that visual supervision from successful and failed continuations improves both prediction and control beyond action supervision alone. On four real-world tasks across Wuji and Sharpa, 3 x 3 grid evaluations show that two rounds of deployment learning increase success from 51.0% to 71.7% in cells with at least one initial success, a gain of 20.7 percentage points. These findings support continued predictive learning for improving control through deployment experience, making world modeling an active part of WAM adaptation.
展开完整摘要收起摘要↓
World-Action Models (WAMs) couple action generation with predictions of how physical interactions unfold. However, current post-deployment learning paradigms typically improve behavior without requiring better world predictions. Especially in dexterous manipulation, small execution errors can compound in high-dimensional action spaces, hindering policy improvement and pushing interactions beyond the world model's training distribution. Motivated by this, we propose Direct Experience World-Model Optimization (DEWO), a post-deployment learning paradigm for WAMs that, alongside action imitation, refines world representations through visual experience to better condition action generation. Specifically, it identifies interaction turning points and learns from successful and failed futures to support classifier-free guidance. An additional value head estimates task progress from video representations and activates guidance when progress stalls during inference. Across five DexJoCo tasks, DEWO improves average success across all three WAM formulations. Ablations show that visual supervision from successful and failed continuations improves both prediction and control beyond action supervision alone. On four real-world tasks across Wuji and Sharpa, 3 x 3 grid evaluations show that two rounds of deployment learning increase success from 51.0% to 71.7% in cells with at least one initial success, a gain of 20.7 percentage points. These findings support continued predictive learning for improving control through deployment experience, making world modeling an active part of WAM adaptation.
We introduce FracGen, a fracture-aware video generation model that produces plausible, controllable fracture dynamics from a single image of an intact object, conditioned on physics signals. To train FracGen, we build FracSim, a fracture-aware simulation framework that augments material point method (MPM) simulation with a continuum damage model, producing paired fracture videos and dense, pixel-aligned physical fields at no additional cost beyond standard rendering. FracGen leverages these maps in two ways: it is trained to jointly predict them alongside RGB video, encouraging the model to capture physical state rather than surface appearance; and it is supervised with physics-informed losses that encourage consistency among the predicted maps. As a result, FracGen captures distinct material-specific fracture behavior without expensive test-time simulation or per-scene tuning, while offering fine-grained control over where an object tears, how fast the crack propagates, and how much deformation precedes failure. We further introduce a benchmark for evaluating the physical plausibility of generated fracture video, and show through extensive experiments that FracGen outperforms existing video generation baselines in both physical and visual fidelity. Results are best viewed in our project website: https://fracgen.github.io/.
展开完整摘要收起摘要↓
We introduce FracGen, a fracture-aware video generation model that produces plausible, controllable fracture dynamics from a single image of an intact object, conditioned on physics signals. To train FracGen, we build FracSim, a fracture-aware simulation framework that augments material point method (MPM) simulation with a continuum damage model, producing paired fracture videos and dense, pixel-aligned physical fields at no additional cost beyond standard rendering. FracGen leverages these maps in two ways: it is trained to jointly predict them alongside RGB video, encouraging the model to capture physical state rather than surface appearance; and it is supervised with physics-informed losses that encourage consistency among the predicted maps. As a result, FracGen captures distinct material-specific fracture behavior without expensive test-time simulation or per-scene tuning, while offering fine-grained control over where an object tears, how fast the crack propagates, and how much deformation precedes failure. We further introduce a benchmark for evaluating the physical plausibility of generated fracture video, and show through extensive experiments that FracGen outperforms existing video generation baselines in both physical and visual fidelity. Results are best viewed in our project website: https://fracgen.github.io/.