Pith. sign in

REVIEW 2 major objections 4 minor 298 references

Scalable physical intelligence requires separating general reasoning from local execution: an embodied brain issues state-transition requests, while a harness, shared contracts, and gated learning make systems reusable.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · grok-4.5

2026-07-14 03:50 UTC pith:E7HP6AZL

load-bearing objection Solid systems roadmap that organizes the WAM/VLA mess into brain–harness–contracts; useful agenda, not a demonstrated stack. the 2 major comments →

arxiv 2607.11689 v1 pith:E7HP6AZL submitted 2026-07-13 cs.RO cs.AI

From World Action Models to Embodied Brains: A Roadmap for Open-World Physical Intelligence

Yuanzhi Liang , Xufeng Zhan , Haibin Huang , Chi Zhang , Xuelong Li This is my paper
classification cs.RO cs.AI
keywords embodied brainworld action modelsphysical intelligencevision-language-actionphysical harnesstrace cardsrobot learningopen-world agents
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

This paper argues that artificial general intelligence needs agents that can reason and act in the physical world, yet progress in action models, vision-language-action policies, and world models stays fragmented. Models bind to incompatible action spaces and prediction targets, datasets and tasks use different conventions, and runtime systems hide interfaces, so bigger models or more trajectories do not accumulate into reusable capability. The authors organize these limits into three coupled gaps—model roles and representations, objectives and standardization, and system composition—and propose a co-evolution roadmap centered on the embodied brain. That brain integrates multimodal context, compares candidate interventions, and emits an intended state transition or capability request rather than direct actuator commands. World Action Models supply prototypes for its predictive functions; a physical harness grounds the output through tools, controllers, verification, and trace logging; shared Embodiment, Task, and Trace Cards align heterogeneous data and embodiments; and closed-loop post-training admits only verified interaction as reusable experience. The resulting modular stack is offered as a path to adaptive, self-improving embodied agents.

Core claim

Physical intelligence cannot scale cumulatively while model-level reasoning, embodiment-specific control, data conventions, and runtime logic remain entangled. The paper’s claim is that an embodied brain should integrate evidence, compare interventions, and communicate intermediate intent (a state transition or capability request), while a physical harness, tool models, declared contracts, and gated closed-loop learning keep components independently testable, mutually compatible, and jointly improvable.

What carries the argument

The embodied brain—a long-term model target that reasons over multimodal context and interventions then issues state-transition or capability requests—together with its supporting WAM prediction contract (decision-relevant consequence plus horizon, frame, uncertainty, and validity) and the physical harness that grounds intent via tools, controllers, verifiers, and Trace Cards.

Load-bearing premise

Declaring intermediate intent and capability interfaces will mostly confine hardware or controller changes to adapters and tools instead of forcing the reasoning model itself to be relearned from scratch.

What would settle it

Hold the brain intent interface fixed, swap a gripper, controller, or morphology, train only the declared adapters, and measure whether task success and failure attribution recover without full brain retraining; large residual relearning falsifies the separation claim.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • General physical reasoning can transfer across grippers, controllers, and morphologies without treating any single actuator space as permanent model output.
  • Heterogeneous datasets and tasks become comparable once interventions, frames, uncertainty, and outcomes share declared semantics.
  • Failures can be attributed to intent, representation translation, tool selection, control, or environment rather than a single end-task score.
  • Verified interaction traces can be admitted to post-training only after safety, quality, and regression gates.
  • Brain, harness, and tool components can be substituted and improved independently while still contributing to a shared physical-intelligence stack.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • If Trace Cards become common, community benchmarks could score prediction quality, interface grounding, and control quality separately, reducing the current conflation of end-task success with foresight.
  • The same brain–harness split could later support multi-agent physical collaboration, where one agent’s intent is executed by another’s body through a shared capability registry.
  • Near-term video and latent world-model work would accumulate faster if forced to declare the decision-relevant consequence fields the paper proposes rather than reporting visual fidelity alone.
  • Digital agent harness patterns for tools and tracing may transfer more cleanly into robotics once spatial, temporal, and embodiment contracts mature.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

2 major / 4 minor

Summary. This roadmap paper surveys the evolution from action policies, VLA models, and world models toward World Action Models (WAMs), then diagnoses three coupled gaps that limit cumulative progress: model roles/representations, objectives/standardization, and ecosystem/systems composition. It proposes a co-evolution stack centered on an embodied brain that integrates multimodal context, compares interventions, and emits an intermediate intent (state transition or capability request) rather than actuator commands; a physical harness grounds that intent via tools, controllers, verifiers, and Trace Cards; and shared Embodiment/Task/Trace contracts plus gated closed-loop post-training make heterogeneous experience reusable. WAMs are framed as provisional prototypes for decision-relevant consequence prediction (horizon, frame, uncertainty, validity), not as a fixed architecture. The paper contributes a literature taxonomy (Tables 1–4), gap analysis (Table 5), ownership contracts (Tables 6–9), and near-term milestones for prediction contracts, brain–harness grounding, and replayable traces.

Significance. If the interface diagnosis is right, the paper offers a useful organizing frame for a fragmented robotics/embodied-AI literature: it separates general physical reasoning from local execution and turns modularity into falsifiable tests (tool substitution, cross-embodiment adaptation, decision-grounded prediction, regression-gated updates). Strengths include a citation-backed survey of predictive-control interfaces and data resources (Tables 1–3), explicit ownership boundaries (Table 6), and layered evaluation that refuses to collapse success into a single end-task score (Table 9). The work is conceptual rather than empirical; its value is agenda-setting and standardization, not a demonstrated transfer result. That is appropriate for a roadmap if the contracts remain provisional and the design hypothesis is kept as a hypothesis.

major comments (2)
  1. [§1, §4.2, §4.7] §1 and §4.2 state the load-bearing design hypothesis that intermediate intent/capability interfaces will primarily localize embodiment change to adapters and tools rather than force relearning of model-level reasoning. The paper correctly labels this a hypothesis and lists tests (tool substitution, cross-embodiment adaptation), but the near-term agenda in §4.7 still under-specifies success criteria: what residual performance loss, adapter cost, or retrain budget would falsify the claim that brain-level requests remain reusable? Without quantitative or protocol-level thresholds, the central organizational claim is hard to evaluate as more than a plausible architecture sketch.
  2. [§3.3, §4.4, Tables 2, 7–9] Tables 7–9 propose Embodiment/Task/Trace Cards and layered evaluation as the standardization response to the objective gap (§3.3). The minimum fields are sensible, but the manuscript does not show a worked mapping from any major public resource (e.g., Open X-Embodiment, DROID, BridgeData V2 in Table 2) onto those fields, nor which currently missing annotations would block replay. A short concrete mapping or gap inventory would make the contracts operational rather than aspirational and would strengthen the claim that shared semantics can begin before a universal 3D/4D representation exists.
minor comments (4)
  1. [§4.1] Figure numbering in the source text is inconsistent (e.g., “Figure 12” appears as a caption label immediately before “Figure 11” in §4.1). Please renumber figures and ensure in-text references match.
  2. [§2.6, Table 1] Table 1 footnote on π0.7 is careful, but the main-text placement still risks readers treating a VLA with optional visual subgoals as an explicit-foresight architecture. A one-sentence reminder in §2.6 would help.
  3. [References] Several arXiv-style citations lack venue/year consistency (e.g., survey and concurrent WAM papers). Standardize bibliography entries for archival clarity.
  4. [Abstract, §1] The abstract and introduction slightly differ in emphasis (AGI framing vs. physical-intelligence stack). Align the opening claim so the contribution is clearly a systems roadmap, not a solved AGI path.

Circularity Check

0 steps flagged

No circularity: conceptual roadmap defines interfaces and hypotheses without fitted predictions or definitional reductions.

full rationale

This is a survey-and-roadmap paper, not a derivation of quantitative predictions from first principles or fitted parameters. It diagnoses three literature-supported gaps (model/representation, objective/standardization, ecosystem/systems), then proposes provisional contracts (WAM prediction contract, embodied-brain intent interface, Embodiment/Task/Trace Cards, physical harness) and explicitly labels the key transfer claim as a design hypothesis to be tested via tool substitution and cross-embodiment adaptation (§1, §4.2). There are no equations equating outputs to inputs by construction, no parameters fitted to data and then re-presented as predictions, and no uniqueness theorems imported from the authors that force the roadmap. Self-citations (e.g., Uni-Inter, TeleBoost, prior Liang et al. works) appear only as illustrative examples of related systems or techniques; none is load-bearing for the central organizational claim. Defining terms (WAM, embodied brain, Trace Card) and then reasoning about their roles is normal conceptual work, not circularity. The paper is self-contained as an agenda with falsifiable near-term milestones.

Axiom & Free-Parameter Ledger

0 free parameters · 5 axioms · 4 invented entities

The paper’s load-bearing content is conceptual architecture, not fitted constants. It imports standard embodied-AI assumptions (closed-loop interaction, embodiment-relative actions, value of prediction for control) and invents named system entities and contracts whose usefulness is hypothesized rather than independently measured outside this text.

axioms (5)
  • domain assumption General-purpose agency requires closed-loop physical interaction connecting perception, prediction, decision, and action under real-world constraints.
    Stated in §1 as the motivation for physical intelligence and AGI grounding; taken from broader AI perspectives (Sutton, Silver & Sutton, LeCun, Li) rather than derived here.
  • domain assumption Diversity of embodiments/data is not the obstacle; hidden semantics of predictions, actions, and outcomes are what block accumulation.
    Core diagnostic premise in §1 and §3; organizes the three-gap analysis.
  • ad hoc to paper Separating model-level intent from harness-mediated execution can make physical reasoning reusable across tools and morphologies when interfaces remain inspectable.
    Explicit design hypothesis in §1 and §4.2; not proven, proposed as testable via substitution and adaptation experiments.
  • ad hoc to paper Decision-relevant consequence prediction (with horizon, frame, uncertainty, validity) is a better WAM contract than a mandatory output modality such as RGB video.
    Prediction contract introduced in §1 and §4.2 as the paper’s provisional position on WAM evaluation.
  • domain assumption Verified interaction traces with provenance can be admitted as reusable experience for post-training without automatically reinforcing model errors if gates are strong enough.
    Closed-loop learning section §4.5; borrows digital post-training/agent-harness patterns and applies them to physical risk settings.
invented entities (4)
  • Embodied brain no independent evidence
    purpose: Long-term model target that integrates multimodal context, compares interventions, and emits intended state transitions or capability requests rather than actuator commands.
    Central invented system role; distinguished from the full stack and from current WAMs; no independent empirical existence beyond the proposal.
  • Physical harness no independent evidence
    purpose: Runtime layer that grounds intent, resolves tools/controllers, verifies, monitors, recovers, and logs Trace Cards.
    Analogized to digital agent harnesses and loosely to a nervous system; proposed as necessary orchestration, not demonstrated as a shipped standard.
  • WAM prediction contract no independent evidence
    purpose: Interface requiring decision-relevant consequences plus horizon, reference frame, uncertainty, and validity conditions for candidate interventions.
    Paper’s provisional standardization of what WAMs should expose; not an established community standard.
  • Embodiment Card / Task Card / Trace Card no independent evidence
    purpose: Lightweight semantic records making embodiments, tasks, and interaction traces comparable and replayable across systems.
    Proposed experience contracts in §4.4; fields are specified but not validated as a field-wide schema with external adoption evidence in the paper.

pith-pipeline@v1.1.0-grok45 · 32557 in / 3318 out tokens · 32504 ms · 2026-07-14T03:50:36.187054+00:00 · methodology

0 comments
read the original abstract

Artificial general intelligence ultimately requires agents that can reason and act in the physical world. Action models, vision-language-action policies, and world models have advanced this goal, while World Action Models (WAMs) are particularly promising because they connect candidate interventions with predicted consequences. However, progress remains fragmented: models use incompatible action spaces and prediction targets, datasets and tasks follow different conventions, and runtime systems expose limited interfaces for reuse and evaluation. We review the evolution toward WAMs and organize these limitations into three coupled gaps: model roles and representations, objectives and standardization, and system composition. Building on this analysis, we propose a co-evolution roadmap for physical intelligence centered on the \emph{embodied brain}, a long-term model target for integrating multimodal context, comparing candidate interventions, and issuing state-transition or capability requests rather than direct actuator commands. WAMs provide promising prototypes for its predictive functions, while a physical harness grounds model outputs through tools, controllers, verification, and trace logging. Shared contracts align heterogeneous models, data, tasks, and embodiments, and closed-loop post-training converts verified interaction into reusable experience. Together, these components define a modular physical-intelligence stack for adaptive and self-improving embodied agents.

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

298 extracted references · 80 linked inside Pith

  1. [1]

    arXiv preprint arXiv:2511.18919 , year =

    Learning What to Trust: Bayesian Prior-Guided Optimization for Visual Generation , author =. arXiv preprint arXiv:2511.18919 , year =

    arXiv

  2. [2]

    arXiv preprint arXiv:2511.18719 , year =

    Seeing What Matters: Visual Preference Policy Optimization for Visual Generation , author =. arXiv preprint arXiv:2511.18719 , year =

    Pith/arXiv arXiv

  3. [3]

    arXiv preprint arXiv:2511.19356 , year =

    Growing with the Generator: Self-paced GRPO for Video Generation , author =. arXiv preprint arXiv:2511.19356 , year =

    arXiv

  4. [4]

    Jordan and Ion Stoica , title =

    Philipp Moritz and Robert Nishihara and Stephanie Wang and Alexey Tumanov and Richard Liaw and Eric Liang and William Paul and Michael I. Jordan and Ion Stoica , title =. CoRR , volume =. 2017 , url =. 1712.05889 , timestamp =

    Pith/arXiv arXiv 2017

  5. [5]

    ACM SIGGRAPH 2024 Conference Papers , year=

    Motionctrl: A unified and flexible motion controller for video generation , author=. ACM SIGGRAPH 2024 Conference Papers , year=

    2024

  6. [7]

    arXiv preprint arXiv:2411.06525 , year=

    I2vcontrol-camera: Precise video camera control with adjustable motion strength , author=. arXiv preprint arXiv:2411.06525 , year=

    Pith/arXiv arXiv

  7. [8]

    Advances in Neural Information Processing Systems , year=

    Light field networks: Neural scene representations with single-evaluation rendering , author=. Advances in Neural Information Processing Systems , year=

  8. [9]

    arXiv preprint arXiv:2406.02509 , year=

    Camco: Camera-controllable 3d-consistent image-to-video generation , author=. arXiv preprint arXiv:2406.02509 , year=

    Pith/arXiv arXiv

  9. [10]

    arXiv preprint arXiv:2406.10126 , year=

    Training-free camera control for video generation , author=. arXiv preprint arXiv:2406.10126 , year=

    Pith/arXiv arXiv

  10. [11]

    Proceedings of the Computer Vision and Pattern Recognition Conference , year=

    Ac3d: Analyzing and improving 3d camera control in video diffusion transformers , author=. Proceedings of the Computer Vision and Pattern Recognition Conference , year=

  11. [12]

    arXiv preprint arXiv:2407.12781 , year=

    Vd3d: Taming large video diffusion transformers for 3d camera control , author=. arXiv preprint arXiv:2407.12781 , year=

    Pith/arXiv arXiv

  12. [13]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , year=

    Snap video: Scaled spatiotemporal transformers for text-to-video synthesis , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , year=

  13. [14]

    Advances in Neural Information Processing Systems , year=

    Collaborative video diffusion: Consistent multi-video generation with camera control , author=. Advances in Neural Information Processing Systems , year=

  14. [15]

    arXiv preprint arXiv:2410.10774 , year=

    Cavia: Camera-controllable multi-view video diffusion with view-integrated attention , author=. arXiv preprint arXiv:2410.10774 , year=

    Pith/arXiv arXiv

  15. [16]

    arXiv preprint arXiv:2412.07760 , year=

    Syncammaster: Synchronizing multi-camera video generation from diverse viewpoints , author=. arXiv preprint arXiv:2412.07760 , year=

    Pith/arXiv arXiv

  16. [17]

    arXiv preprint arXiv:2503.10592 , year=

    Cameractrl ii: Dynamic scene exploration via camera-controlled video diffusion models , author=. arXiv preprint arXiv:2503.10592 , year=

    Pith/arXiv arXiv

  17. [18]

    ACM SIGGRAPH 2024 Conference Papers , year=

    Motion-i2v: Consistent and controllable image-to-video generation with explicit motion modeling , author=. ACM SIGGRAPH 2024 Conference Papers , year=

    2024

  18. [19]

    arXiv preprint arXiv:2308.08089 , year=

    Dragnuwa: Fine-grained control in video generation by integrating text, image, and trajectory , author=. arXiv preprint arXiv:2308.08089 , year=

    Pith/arXiv arXiv

  19. [20]

    Proceedings of the Computer Vision and Pattern Recognition Conference , year=

    Tora: Trajectory-oriented diffusion transformer for video generation , author=. Proceedings of the Computer Vision and Pattern Recognition Conference , year=

  20. [21]

    Proceedings of the AAAI Conference on Artificial Intelligence , volume=

    Trackgo: A flexible and efficient method for controllable video generation , author=. Proceedings of the AAAI Conference on Artificial Intelligence , volume=

  21. [22]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , year=

    Peekaboo: Interactive video generation via masked-diffusion , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , year=

  22. [23]

    arXiv preprint arXiv:2402.01566 , year=

    Boximator: Generating rich and controllable motions for video synthesis , author=. arXiv preprint arXiv:2402.01566 , year=

    Pith/arXiv arXiv

  23. [24]

    Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , year=

    Gligen: Open-set grounded text-to-image generation , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , year=

  24. [25]

    arXiv preprint arXiv:2406.16863 , year=

    Freetraj: Tuning-free trajectory control in video diffusion models , author=. arXiv preprint arXiv:2406.16863 , year=

    Pith/arXiv arXiv

  25. [26]

    arXiv preprint arXiv:2503.16421 , year=

    Magicmotion: Controllable video generation with dense-to-sparse trajectory guidance , author=. arXiv preprint arXiv:2503.16421 , year=

    arXiv

  26. [27]

    arXiv preprint arXiv:2412.07759 , year=

    3dtrajmaster: Mastering 3d trajectory for multi-entity motion in video generation , author=. arXiv preprint arXiv:2412.07759 , year=

    Pith/arXiv arXiv

  27. [28]

    Proceedings of the Computer Vision and Pattern Recognition Conference , year=

    Levitor: 3d trajectory oriented image-to-video synthesis , author=. Proceedings of the Computer Vision and Pattern Recognition Conference , year=

  28. [29]

    Proceedings of the Computer Vision and Pattern Recognition Conference , year=

    Motion prompting: Controlling video generation with motion trajectories , author=. Proceedings of the Computer Vision and Pattern Recognition Conference , year=

  29. [30]

    Proceedings of the AAAI Conference on Artificial Intelligence , year=

    Image conductor: Precision control for interactive video synthesis , author=. Proceedings of the AAAI Conference on Artificial Intelligence , year=

  30. [31]

    arXiv preprint arXiv:2505.22944 , year=

    ATI: Any Trajectory Instruction for Controllable Video Generation , author=. arXiv preprint arXiv:2505.22944 , year=

    Pith/arXiv arXiv

  31. [32]

    arXiv preprint arXiv:2501.05020 , year=

    Perception-as-Control: Fine-grained Controllable Image Animation with 3D-aware Motion Representation , author=. arXiv preprint arXiv:2501.05020 , year=

    Pith/arXiv arXiv

  32. [33]

    arXiv preprint arXiv:2502.07531 , year=

    Vidcraft3: Camera, object, and lighting control for image-to-video generation , author=. arXiv preprint arXiv:2502.07531 , year=

    Pith/arXiv arXiv

  33. [34]

    IEEE transactions on pattern analysis and machine intelligence , year=

    Real-time scene text detection with differentiable binarization and adaptive scale fusion , author=. IEEE transactions on pattern analysis and machine intelligence , year=

  34. [35]

    arXiv preprint arXiv:2507.13347 , year=

    pi3 : Permutation-Equivariant Visual Geometry Learning , author=. arXiv preprint arXiv:2507.13347 , year=

    Pith/arXiv arXiv

  35. [36]

    5-vl technical report , author=

    Qwen2. 5-vl technical report , author=. arXiv preprint arXiv:2502.13923 , year=

    Pith/arXiv arXiv

  36. [37]

    Proceedings of the Computer Vision and Pattern Recognition Conference , year=

    Segment Any Motion in Videos , author=. Proceedings of the Computer Vision and Pattern Recognition Conference , year=

  37. [39]

    arXiv preprint arXiv:2507.12462 , year=

    Spatialtrackerv2: 3d point tracking made easy , author=. arXiv preprint arXiv:2507.12462 , year=

    Pith/arXiv arXiv

  38. [40]

    Proceedings of the IEEE/CVF International Conference on Computer Vision , year=

    Free-form motion control: Controlling the 6d poses of camera and objects in video generation , author=. Proceedings of the IEEE/CVF International Conference on Computer Vision , year=

  39. [41]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , year=

    High-resolution image synthesis with latent diffusion models , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , year=

  40. [42]

    arXiv preprint arXiv:2310.05737 , year=

    Language Model Beats Diffusion--Tokenizer is Key to Visual Generation , author=. arXiv preprint arXiv:2310.05737 , year=

    Pith/arXiv arXiv

  41. [43]

    arXiv preprint arXiv:2210.02747 , year=

    Flow matching for generative modeling , author=. arXiv preprint arXiv:2210.02747 , year=

    Pith/arXiv arXiv

  42. [44]

    Medical image computing and computer-assisted intervention--MICCAI 2015: 18th international conference, Munich, Germany, October 5-9, 2015, proceedings, part III 18 , year=

    U-net: Convolutional networks for biomedical image segmentation , author=. Medical image computing and computer-assisted intervention--MICCAI 2015: 18th international conference, Munich, Germany, October 5-9, 2015, proceedings, part III 18 , year=

    2015

  43. [45]

    arXiv preprint arXiv:2503.20314 , year=

    Wan: Open and Advanced Large-Scale Video Generative Models , author=. arXiv preprint arXiv:2503.20314 , year=

    Pith/arXiv arXiv

  44. [46]

    arXiv preprint arXiv:2304.09151 , year=

    Unimax: Fairer and more effective language sampling for large-scale multilingual pretraining , author=. arXiv preprint arXiv:2304.09151 , year=

    Pith/arXiv arXiv

  45. [47]

    International conference on machine learning , year=

    Learning transferable visual models from natural language supervision , author=. International conference on machine learning , year=

  46. [48]

    Proceedings of the IEEE/CVF international conference on computer vision , pages=

    Scalable diffusion models with transformers , author=. Proceedings of the IEEE/CVF international conference on computer vision , pages=

  47. [49]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , year=

    AC3D: Analyzing and Improving 3D Camera Control in Video Diffusion Transformers , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , year=

  48. [50]

    Proceedings of the IEEE/CVF international conference on computer vision , year=

    Adding conditional control to text-to-image diffusion models , author=. Proceedings of the IEEE/CVF international conference on computer vision , year=

  49. [51]

    arXiv preprint arXiv:2412.12091 , year=

    Wonderland: Navigating 3D Scenes from a Single Image , author=. arXiv preprint arXiv:2412.12091 , year=

    Pith/arXiv arXiv

  50. [52]

    Depth Pro: Sharp Monocular Metric Depth in Less Than a Second , journal =

    Aleksei Bochkovskii and Ama\". Depth Pro: Sharp Monocular Metric Depth in Less Than a Second , journal =. 2024 , url =

    2024

  51. [53]

    Pixelwise View Selection for Unstructured Multi-View Stereo , booktitle=

    Sch\". Pixelwise View Selection for Unstructured Multi-View Stereo , booktitle=

  52. [54]

    International Conference on Learning Representations , year=

    MVSFormer++: Revealing the Devil in Transformer's Details for Multi-View Stereo , author=. International Conference on Learning Representations , year=

  53. [55]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , year=

    MVGenMaster: Scaling Multi-View Generation from Any Image via 3D Priors Enhanced Diffusion Model , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , year=

  54. [56]

    International Conference on Learning Representations , year=

    DMV3D: Denoising Multi-View Diffusion using 3D Large Reconstruction Model , author=. International Conference on Learning Representations , year=

  55. [57]

    arXiv preprint arXiv:1711.05101 , year=

    Decoupled weight decay regularization , author=. arXiv preprint arXiv:1711.05101 , year=

    Pith/arXiv arXiv

  56. [58]

    arXiv preprint arXiv:1812.01717 , year=

    Towards accurate generative models of video: A new metric & challenges , author=. arXiv preprint arXiv:1812.01717 , year=

    Pith/arXiv arXiv

  57. [59]

    , author=

    pytorch-fid: FID Score for PyTorch. , author=. https://github.com/ mseitzer/pytorch-fid , year=

  58. [60]

    arXiv preprint arXiv:2104.14806 , year=

    Godiva: Generating open-domain videos from natural descriptions , author=. arXiv preprint arXiv:2104.14806 , year=

    Pith/arXiv arXiv

  59. [61]

    arXiv preprint arXiv:2409.02048 , year=

    Viewcrafter: Taming video diffusion models for high-fidelity novel view synthesis , author=. arXiv preprint arXiv:2409.02048 , year=

    Pith/arXiv arXiv

  60. [62]

    arXiv preprint arXiv:2307.04725 , year=

    Animatediff: Animate your personalized text-to-image diffusion models without specific tuning , author=. arXiv preprint arXiv:2307.04725 , year=

    Pith/arXiv arXiv

  61. [63]

    arXiv preprint arXiv:2404.15789 , year=

    Motionmaster: Training-free camera motion transfer for video generation , author=. arXiv preprint arXiv:2404.15789 , year=

    Pith/arXiv arXiv

  62. [64]

    European Conference on Computer Vision , year=

    Dreammotion: Space-time self-similar score distillation for zero-shot video editing , author=. European Conference on Computer Vision , year=

  63. [65]

    arXiv preprint arXiv:2406.05338 , year=

    Motionclone: Training-free motion cloning for controllable video generation , author=. arXiv preprint arXiv:2406.05338 , year=

    Pith/arXiv arXiv

  64. [66]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , year=

    Space-time diffusion features for zero-shot text-driven motion transfer , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , year=

  65. [67]

    arXiv preprint arXiv:1805.09817 , year=

    Stereo magnification: Learning view synthesis using multiplane images , author=. arXiv preprint arXiv:1805.09817 , year=

    Pith/arXiv arXiv

  66. [68]

    Journal of Computing in Civil Engineering , year=

    Development of an image data set of construction machines for deep learning object detection , author=. Journal of Computing in Civil Engineering , year=

  67. [69]

    2025 , eprint=

    Understanding World or Predicting Future? A Comprehensive Survey of World Models , author=. 2025 , eprint=

    2025

  68. [70]

    arXiv preprint arXiv:2507.17744 , year=

    Yume: An Interactive World Generation Model , author=. arXiv preprint arXiv:2507.17744 , year=

    Pith/arXiv arXiv

  69. [71]

    arXiv preprint arXiv:2503.11647 , year=

    Recammaster: Camera-controlled generative rendering from a single video , author=. arXiv preprint arXiv:2503.11647 , year=

    Pith/arXiv arXiv

  70. [72]

    arXiv preprint , year=

    Generative Video Compression: Towards 0.01 author=. arXiv preprint , year=

  71. [73]

    arXiv preprint arXiv:2509.26645 , year=

    TTT3R: 3D Reconstruction as Test-Time Training , author=. arXiv preprint arXiv:2509.26645 , year=

    Pith/arXiv arXiv

  72. [74]

    2025 , eprint=

    Is Sora a World Simulator? A Comprehensive Survey on General World Models and Beyond , author=. 2025 , eprint=

    2025

  73. [75]

    2025 , eprint=

    Ctrl-World: A Controllable Generative World Model for Robot Manipulation , author=. 2025 , eprint=

    2025

  74. [76]

    2025 , eprint=

    World-Env: Leveraging World Model as a Virtual Environment for VLA Post-Training , author=. 2025 , eprint=

    2025

  75. [77]

    2025 , eprint=

    MineWorld: a Real-Time and Open-Source Interactive World Model on Minecraft , author=. 2025 , eprint=

    2025

  76. [78]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , month =

    Zuo, Sicheng and Zheng, Wenzhao and Huang, Yuanhui and Zhou, Jie and Lu, Jiwen , title =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , month =. 2025 , pages =

    2025

  77. [79]

    2025 , eprint=

    Dual-Stream Diffusion for World-Model Augmented Vision-Language-Action Model , author=. 2025 , eprint=

    2025

  78. [80]

    2025 , eprint=

    VideoVerse: How Far is Your T2V Generator from a World Model? , author=. 2025 , eprint=

    2025

  79. [81]

    2025 , eprint=

    Text2World: Benchmarking Large Language Models for Symbolic World Model Generation , author=. 2025 , eprint=

    2025

  80. [82]

    2025 , eprint=

    OccTENS: 3D Occupancy World Model via Temporal Next-Scale Prediction , author=. 2025 , eprint=

    2025

Showing first 80 references.