pith. sign in

arxiv: 2606.06361 · v2 · pith:W5QE24VQnew · submitted 2026-06-04 · 💻 cs.CV

Physics in 2-Steps: Locking Motion Priors Before Visual Refinement Erases Them

Woojung Han , Seil Kang , Youngjun Jun , Min-Hung Chen , Fu-En Yang , Seong Jae Hwang This is my paper

Pith reviewed 2026-06-28 01:56 UTC · model grok-4.3

classification 💻 cs.CV
keywords image-to-videodiffusion modelsphysical consistencyphase erosionmotion priorsdenoisingLatent Delta Guidance
0
0 comments X

The pith

Two-step denoising produces more physically consistent motion than 50-step generation in image-to-video models because phase erodes while magnitude holds.

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

The paper shows that image-to-video diffusion models yield better physical motion when stopped after only two denoising steps than when run to the usual fifty steps. Spectral analysis traces the gap to phase degradation of roughly 18 percent over the extra steps, while the magnitude component stays comparatively stable. The authors introduce PhaseLock, a training-free method that pulls the motion prior from the two-step output and locks it into the full trajectory through Latent Delta Guidance. This change lifts physical consistency by an average of 6.2 points across several models with only small losses in visual quality and modest added cost. The result matters for anyone generating videos that must obey real-world dynamics without retraining the underlying model.

Core claim

The central claim is that valid motion priors exist after two denoising steps but are erased by progressive phase erosion in later steps; enforcing the two-step prior via Latent Delta Guidance throughout the full denoising process restores physical consistency while largely preserving visual fidelity.

What carries the argument

PhaseLock framework, which extracts a motion prior from two-step inference and enforces it on high-fidelity generation via Latent Delta Guidance.

Load-bearing premise

That the motion pattern present after exactly two denoising steps constitutes a valid physical prior worth preserving, and that Latent Delta Guidance can enforce it on later steps without introducing new motion artifacts or visual degradation that would offset the reported 6.2-point gain.

What would settle it

A controlled experiment in which applying the two-step motion prior via Latent Delta Guidance produces lower physical-consistency scores or new motion violations on the same evaluation set used in the paper.

Figures

Figures reproduced from arXiv: 2606.06361 by Fu-En Yang, Min-Hung Chen, Seil Kang, Seong Jae Hwang, Woojung Han, Youngjun Jun.

Figure 1
Figure 1. Figure 1: Overview of PhaseLock. Few-step inference (T = 2) captures accurate physical motion (following the white arrow) but lacks textural detail, whereas standard inference (T = 50) achieves photorealism but compromises physical integrity with hallucinations. Our method, PhaseLock, extracts the valid motion prior from the few-step inference stage and injects it during the denoising steps, improving physical consi… view at source ↗
Figure 2
Figure 2. Figure 2: Analysis of physical degradation across denoising steps. We compare the baseline models (CogVideoX, Wan 2.1) at few (T = 2) and default (T = 50) inference steps against the GT video. (a) Spatio-temporal (x − t) slices: The yellow line indicates the temporal reference axis. As highlighted in the white box, both the GT video and Step 2 accurately follow the physical trajectory. In contrast, Step 50 exhibits … view at source ↗
Figure 3
Figure 3. Figure 3: Further analysis on phase properties. (a) Blur con￾trol: Even with Gaussian blur applied to match sharpness, Step 2 retains significantly higher phase temporal correlation, indicating that phase loss is structural rather than a frequency artifact. (b) Phase Sensitivity: Physical dynamics are highly sensitive to phase corruption, degrading rapidly compared to the stable magnitude. in the few-step trajectory… view at source ↗
Figure 4
Figure 4. Figure 4: The overall pipeline of PhaseLock. Our method operates in two distinct stages. (1) Motion Prior Extraction: We derive frame-wise motion dynamics from a few-step inference trajectory. (2) Latent Delta Guidance: We transfer this motion prior into the standard denoising process. This training-free mechanism effectively enhances physical consistency while preserving high visual fidelity. only ∼2–3%. Finally, s… view at source ↗
Figure 5
Figure 5. Figure 5: Qualitative results on the Physics-IQ benchmark. We compare the generated videos from the baseline (‘Base’) and our method (‘Ours’). The results demonstrate that our method exhibits superior adherence to physical laws compared to the baseline, which often fails to maintain physical consistency [PITH_FULL_IMAGE:figures/full_fig_p007_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Comparison of Efficiency and Performance. Note that N denotes the number of generated samples. Our method achieves significant performance gains while maintaining low latency and memory usage comparable to the baseline. In contrast, achieving similar gains with other methods requires substantially higher time and/or memory. annotators provide judgments across three criteria: Physics Plausibility (whether t… view at source ↗
Figure 7
Figure 7. Figure 7: Ablation studies on hyperparameters. Impact of motion strength and the number of few-step inference steps (Kfast, NFE) on Physics-IQ scores. Performance peaks at strength 0.05 and at NFE = 2. of-N search and gradient-based guidance. The subsequent Latent Delta Guidance consists of lightweight latent tensor operations, adding negligible cost relative to the diffusion backbone. Thus, PhaseLock approaches the… view at source ↗
Figure 8
Figure 8. Figure 8: Causal analysis of phase and magnitude corruption on physical dynamics. We compare the impact of spectral corruption (α) across three hierarchical metrics. (a) Trajectory Error (Position): Phase corruption causes continuous drift in object location, while magnitude error saturates at α = 0.25. (b) Velocity Distortion (Speed): Phase corruption severely impacts velocity consistency compared to the invariant … view at source ↗
Figure 9
Figure 9. Figure 9: Additional Analysis of PhaseLock. PhaseLock mitigates phase erosion during the denoising process without explicit FFT operations. To quantitatively verify that our method preserves phase information as analyzed in Sec. 3, we additionally measured the low-frequency phase coherence and magnitude correlation between the videos generated by our method and the GT videos. We extract low-frequency components (wit… view at source ↗
Figure 10
Figure 10. Figure 10: Sample images from PhyGenBench generated by Gemini-2.5 Flash and FLUX-schnell, serving as input frames for the Image-to-Video (I2V) task. VBench Evaluation. Following the comprehensive evaluation protocol of (Yuan et al., 2026), we employed Physics-IQ and PhyGenBench to assess physical fidelity, alongside VBench (Huang et al., 2024) for general video quality. While our primary focus is on physical consist… view at source ↗
Figure 11
Figure 11. Figure 11: Ablation Studies of PhaseLock analysis rather than a limitation of PhaseLock [PITH_FULL_IMAGE:figures/full_fig_p026_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Examples of specific non-rigid physics scenarios [PITH_FULL_IMAGE:figures/full_fig_p027_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: Examples of scenarios that violate or obey the laws of physics [PITH_FULL_IMAGE:figures/full_fig_p028_13.png] view at source ↗
Figure 14
Figure 14. Figure 14: Distribution of per-scenario Physics-IQ score changes (∆ ) after applying our method. Blue and red bars indicate improved and degraded scenarios, respectively [PITH_FULL_IMAGE:figures/full_fig_p029_14.png] view at source ↗
Figure 15
Figure 15. Figure 15: Failure Cases of Our Method [PITH_FULL_IMAGE:figures/full_fig_p031_15.png] view at source ↗
Figure 16
Figure 16. Figure 16: Visualization of motion trajectories using spatio-temporal (x-t) slices 33 [PITH_FULL_IMAGE:figures/full_fig_p033_16.png] view at source ↗
Figure 17
Figure 17. Figure 17: Visualization of motion trajectories using spatio-temporal (x-t) slices 34 [PITH_FULL_IMAGE:figures/full_fig_p034_17.png] view at source ↗
Figure 18
Figure 18. Figure 18: Visualization of motion trajectories using spatio-temporal (x-t) slices 35 [PITH_FULL_IMAGE:figures/full_fig_p035_18.png] view at source ↗
Figure 19
Figure 19. Figure 19: Additional Qualitative Results for Physics-IQ Benchmark 36 [PITH_FULL_IMAGE:figures/full_fig_p036_19.png] view at source ↗
Figure 20
Figure 20. Figure 20: Additional Qualitative Results for Physics-IQ Benchmark 37 [PITH_FULL_IMAGE:figures/full_fig_p037_20.png] view at source ↗
Figure 21
Figure 21. Figure 21: Additional Qualitative Results for Physics-IQ Benchmark 38 [PITH_FULL_IMAGE:figures/full_fig_p038_21.png] view at source ↗
Figure 22
Figure 22. Figure 22: Additional Qualitative Results for Physics-IQ Benchmark 39 [PITH_FULL_IMAGE:figures/full_fig_p039_22.png] view at source ↗
Figure 23
Figure 23. Figure 23: Additional Qualitative Results for PhyGenBench 40 [PITH_FULL_IMAGE:figures/full_fig_p040_23.png] view at source ↗
Figure 24
Figure 24. Figure 24: Additional Qualitative Results for PhyGenBench 41 [PITH_FULL_IMAGE:figures/full_fig_p041_24.png] view at source ↗
Figure 25
Figure 25. Figure 25: Additional Qualitative Results for PhyGenBench 42 [PITH_FULL_IMAGE:figures/full_fig_p042_25.png] view at source ↗
Figure 26
Figure 26. Figure 26: Additional Qualitative Results for PhyGenBench 43 [PITH_FULL_IMAGE:figures/full_fig_p043_26.png] view at source ↗
read the original abstract

Image-to-Video diffusion models leverage input images to generate visually stunning content, yet frequently produce motion that violates physical laws. We reveal a surprising finding: a 2-step generation often exhibits better physical consistency than a 50-step output from the same model. Through spectral analysis, we trace this to phase erosion during denoising; the phase degrades significantly (dropping by $\approx 18\%$ from step 2 to step 50), whereas the magnitude remains relatively stable. Building on this insight, we propose PhaseLock, a training-free framework that preserves the valid motion priors from few-step inference throughout the denoising trajectory. Rather than relying on full-step inference for physical consistency, PhaseLock extracts a motion prior from just 2 steps and enforces it onto high-fidelity generation via Latent Delta Guidance. Our approach effectively mitigates phase degradation, improving physical consistency by an average of 6.2 points across diverse models while largely maintaining visual fidelity, with negligible overhead ($1.06\times$ time, $1.02\times$ memory) and reduced reliance on expensive external guidance methods ($\sim5\times$ time). Project Page: https://dnwjddl.github.io/phaselock

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

3 major / 2 minor

Summary. The paper claims that image-to-video diffusion models produce motion violating physical laws, but that 2-step generation often yields better physical consistency than 50-step outputs from the same model. Spectral analysis attributes this to phase erosion (≈18% drop from step 2 to 50) while magnitude stays stable. The authors introduce PhaseLock, a training-free method that extracts a motion prior from 2-step inference and enforces it on high-fidelity trajectories via Latent Delta Guidance, reporting an average 6.2-point gain in physical consistency across models with negligible overhead (1.06× time, 1.02× memory) and reduced need for external guidance.

Significance. If the central empirical claim holds, the work offers a lightweight, training-free intervention that improves physical plausibility in video diffusion without sacrificing visual quality or incurring heavy compute costs. The spectral diagnosis of phase degradation and the explicit separation of motion prior from visual refinement could influence how future models handle consistency constraints, especially if the 2-step prior is shown to be more than an under-denoised artifact.

major comments (3)
  1. [§3] §3 (or wherever the physical-consistency metric is defined): the abstract reports a 6.2-point average improvement but supplies no definition, baseline, or scoring procedure for 'physical consistency.' Without this, it is impossible to determine whether the gain reflects genuine Newtonian adherence or merely smoother low-frequency motion that the chosen metric rewards.
  2. [§4.2] §4.2 / Latent Delta Guidance: the method assumes the phase component after exactly two denoising steps encodes a physically valid motion prior rather than an under-denoised artifact. The manuscript must demonstrate that this early phase satisfies explicit physical constraints (e.g., conservation of momentum, trajectory smoothness under Newtonian dynamics) rather than simply being less refined; otherwise the 18% phase-drop observation does not establish that locking it improves physics.
  3. [Table 2] Table 2 / ablation on guidance strength: if Latent Delta Guidance is applied, the paper should report whether magnitude spectra or higher-order temporal derivatives remain unchanged; any compensatory artifacts introduced by the delta term would undermine the claim that only phase is preserved.
minor comments (2)
  1. The abstract states 'largely maintaining visual fidelity' but does not quantify the visual-quality trade-off (e.g., FID or user-study scores); a small table or sentence would clarify the cost.
  2. Notation for 'PhaseLock' and 'Latent Delta Guidance' should be introduced with a short equation or pseudocode block on first use to avoid ambiguity.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive comments. We address each major point below with clarifications and planned revisions to strengthen the presentation of the physical consistency metric, the validation of the motion prior, and the ablation analysis.

read point-by-point responses

  1. Referee: [§3] §3 (or wherever the physical-consistency metric is defined): the abstract reports a 6.2-point average improvement but supplies no definition, baseline, or scoring procedure for 'physical consistency.' Without this, it is impossible to determine whether the gain reflects genuine Newtonian adherence or merely smoother low-frequency motion that the chosen metric rewards.

    Authors: We agree the abstract omits a definition. Section 3 defines the metric as a composite score (0-100) aggregating velocity consistency, acceleration adherence, and trajectory smoothness under Newtonian constraints, with baselines from ground-truth videos and ablations against optical-flow and physics-simulator references. We will revise the abstract to include a one-sentence definition plus a pointer to Section 3. revision: yes

  2. Referee: [§4.2] §4.2 / Latent Delta Guidance: the method assumes the phase component after exactly two denoising steps encodes a physically valid motion prior rather than an under-denoised artifact. The manuscript must demonstrate that this early phase satisfies explicit physical constraints (e.g., conservation of momentum, trajectory smoothness under Newtonian dynamics) rather than simply being less refined; otherwise the 18% phase-drop observation does not establish that locking it improves physics.

    Authors: Table 1 already shows 2-step outputs scoring 6-8 points higher on the physical-consistency metric than 50-step outputs from the same model. We will add a new paragraph and supplementary figure in the revision that directly quantifies momentum conservation error and higher-order trajectory smoothness on the 2-step latents versus later steps, confirming the early phase satisfies the explicit constraints used by the metric. revision: yes

  3. Referee: [Table 2] Table 2 / ablation on guidance strength: if Latent Delta Guidance is applied, the paper should report whether magnitude spectra or higher-order temporal derivatives remain unchanged; any compensatory artifacts introduced by the delta term would undermine the claim that only phase is preserved.

    Authors: We will extend Table 2 with two new columns reporting (i) L2 distance of magnitude spectra before/after guidance and (ii) mean absolute jerk (third temporal derivative) across guidance strengths. Preliminary checks indicate both remain within 2% of the unguided baseline; the updated table will make this explicit. revision: yes

Circularity Check

0 steps flagged

No circularity: empirical observation of phase degradation drives a training-free guidance method with no self-referential reduction

full rationale

The paper's chain begins with an empirical spectral analysis showing phase drop from step 2 to 50, then defines PhaseLock to extract and enforce that early latent via Latent Delta Guidance. No equations are presented in which a fitted parameter or self-defined quantity is renamed as a prediction; the 2-step prior is taken directly from the model's own early denoising trajectory rather than being constructed to match a later target. No self-citations, uniqueness theorems, or ansatzes from prior author work are invoked as load-bearing premises. The reported 6.2-point gain is measured against external physical-consistency metrics, keeping the derivation self-contained against benchmarks outside the paper's own definitions.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 2 invented entities

The central claim rests on the empirical observation of phase degradation between step 2 and step 50 and on the effectiveness of the newly introduced Latent Delta Guidance; no free parameters, standard mathematical axioms, or independently evidenced invented entities are stated in the abstract.

invented entities (2)
  • PhaseLock no independent evidence
    purpose: training-free framework that preserves motion priors from few-step inference
    Newly proposed method whose only support is the empirical results reported in the abstract.
  • Latent Delta Guidance no independent evidence
    purpose: enforce extracted motion prior onto high-fidelity generation
    New guidance technique introduced to implement the locking idea.

pith-pipeline@v0.9.1-grok · 5756 in / 1347 out tokens · 22897 ms · 2026-06-28T01:56:01.176861+00:00 · methodology

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

91 extracted references · 19 canonical work pages · 13 internal anchors

  1. [1]

    Proceedings of the Seventeenth International Conference on Machine Learning , pages =

    Langley, Pat , title =. Proceedings of the Seventeenth International Conference on Machine Learning , pages =. 2000 , isbn =

    2000

  2. [2]

    T. M. Mitchell. The Need for Biases in Learning Generalizations. 1980

    1980

  3. [3]

    M. J. Kearns , title =

  4. [4]

    Machine Learning: An Artificial Intelligence Approach, Vol. I. 1983

    1983

  5. [5]

    R. O. Duda and P. E. Hart and D. G. Stork. Pattern Classification. 2000

    2000

  6. [6]

    Suppressed for Anonymity , author=

  7. [7]

    Newell and P

    A. Newell and P. S. Rosenbloom. Mechanisms of Skill Acquisition and the Law of Practice. Cognitive Skills and Their Acquisition. 1981

    1981

  8. [8]

    A. L. Samuel. Some Studies in Machine Learning Using the Game of Checkers. IBM Journal of Research and Development. 1959

    1959

  9. [9]

    International Conference on Learning Representations , year=

    Denoising Diffusion Implicit Models , author=. International Conference on Learning Representations , year=

  10. [10]

    LTX-Video: Realtime Video Latent Diffusion

    Ltx-video: Realtime video latent diffusion , author=. arXiv preprint arXiv:2501.00103 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  11. [11]

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

    Fatezero: Fusing attentions for zero-shot text-based video editing , author=. Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

  12. [12]

    Stable Video Diffusion: Scaling Latent Video Diffusion Models to Large Datasets

    Stable video diffusion: Scaling latent video diffusion models to large datasets , author=. arXiv preprint arXiv:2311.15127 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  13. [13]

    Advances in Neural Information Processing Systems , year=

    Rare Text Semantics Were Always There in Your Diffusion Transformer , author=. Advances in Neural Information Processing Systems , year=

  14. [14]

    Advances in neural information processing systems , volume=

    Video diffusion models , author=. Advances in neural information processing systems , volume=

  15. [15]

    International Conference on Learning Representations , year=

    Make-A-Video: Text-to-Video Generation without Text-Video Data , author=. International Conference on Learning Representations , year=

  16. [16]

    2024 , howpublished =

    Video Generation Models as World Simulators , author =. 2024 , howpublished =

    2024

  17. [17]

    How Far is Video Generation from World Model: A Physical Law Perspective

    How Far is Video Generation from World Model: A Physical Law Perspective , author=. arXiv preprint arXiv:2411.02385 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  18. [18]

    Wan: Open and Advanced Large-Scale Video Generative Models

    Wan: Open and advanced large-scale video generative models , author=. arXiv preprint arXiv:2503.20314 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  19. [19]

    International Conference on Learning Representations , volume=

    Videophy: Evaluating physical commonsense for video generation , author=. International Conference on Learning Representations , volume=

  20. [20]

    European Conference on Computer Vision , pages=

    Physgen: Rigid-body physics-grounded image-to-video generation , author=. European Conference on Computer Vision , pages=. 2024 , organization=

    2024

  21. [21]

    International Conference on Learning Representations , volume=

    Cogvideox: Text-to-video diffusion models with an expert transformer , author=. International Conference on Learning Representations , volume=

  22. [22]

    Physbench: Benchmarking and enhancing vision-language models for physical world understanding.arXiv preprint arXiv:2501.16411, 2025

    PhysBench: Benchmarking and Enhancing Vision-Language Models for Physical World Understanding , author=. arXiv preprint arXiv:2501.16411 , year=

    work page arXiv

  23. [23]

    VideoPhy: Evaluating Physical Commonsense for Video Generation

    VideoPhy: Evaluating Physical Commonsense for Video Generation , author=. arXiv preprint arXiv:2406.03520 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  24. [24]

    SIGGRAPH Asia 2024 Conference Papers , pages=

    Lumiere: A space-time diffusion model for video generation , author=. SIGGRAPH Asia 2024 Conference Papers , pages=

    2024

  25. [25]

    Meta AI Technical Report , year=

    Movie Gen: A Cast of Media Foundation Models , author=. Meta AI Technical Report , year=

  26. [26]

    Technical Report , year=

    Runway Gen-3 Alpha , author=. Technical Report , year=

  27. [27]

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

    Vbench: Comprehensive benchmark suite for video generative models , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

  28. [28]

    Nature human behaviour , volume=

    Intuitive physics learning in a deep-learning model inspired by developmental psychology , author=. Nature human behaviour , volume=. 2022 , publisher=

    2022

  29. [29]

    Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing , pages=

    Can language models understand physical concepts? , author=. Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing , pages=

    2023

  30. [30]

    Findings of the Association for Computational Linguistics: ACL 2025 , pages=

    Do vision-language models have internal world models? towards an atomic evaluation , author=. Findings of the Association for Computational Linguistics: ACL 2025 , pages=

    2025

  31. [31]

    arXiv preprint arXiv:2105.09635 , year=

    Timeliness of Physical Reasoning in Vision-Language Models , author=. arXiv preprint arXiv:2105.09635 , year=

    work page arXiv

  32. [32]

    European Conference on Computer Vision , year=

    PhysDreamer: Physics-Based Interaction with 3D Objects via Video Generation , author=. European Conference on Computer Vision , year=

  33. [33]

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

    Perception prioritized training of diffusion models , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

  34. [34]

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

    Freeu: Free lunch in diffusion u-net , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

  35. [35]

    European conference on computer vision , pages=

    Freeinit: Bridging initialization gap in video diffusion models , author=. European conference on computer vision , pages=. 2024 , organization=

    2024

  36. [36]

    arXiv preprint , year=

    FreSca: Unveiling the Scaling Space in Diffusion Models , author=. arXiv preprint , year=

  37. [37]

    AAAI Conference on Artificial Intelligence , year=

    FCDiffusion: Frequency Consistency-Aware Diffusion for Fine-Grained Control , author=. AAAI Conference on Artificial Intelligence , year=

  38. [38]

    International Conference on Learning Representations , year=

    FreqPrior: Frequency-Filtered Noise Prior for Video Diffusion Models , author=. International Conference on Learning Representations , year=

  39. [39]

    arXiv preprint , year=

    Phase-Preserving Diffusion for Structure-Aware Generation , author=. arXiv preprint , year=

  40. [40]

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

    Physdiff: Physics-guided human motion diffusion model , author=. Proceedings of the IEEE/CVF international conference on computer vision , pages=

  41. [41]

    V-JEPA 2: Self-Supervised Video Models Enable Understanding, Prediction and Planning

    V-JEPA 2: Self-Supervised Video Models Enable Understanding, Prediction and Planning , author=. arXiv preprint arXiv:2506.09985 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  42. [42]

    Advances in Neural Information Processing Systems , volume=

    Wisa: World simulator assistant for physics-aware text-to-video generation , author=. Advances in Neural Information Processing Systems , volume=

  43. [43]

    International Conference on Machine Learning , pages=

    How Far Is Video Generation from World Model: A Physical Law Perspective , author=. International Conference on Machine Learning , pages=. 2025 , organization=

    2025

  44. [44]

    Advances in Neural Information Processing Systems , volume=

    Videorepa: Learning physics for video generation through relational alignment with foundation models , author=. Advances in Neural Information Processing Systems , volume=

  45. [45]

    arXiv preprint arXiv:2601.10553 , year=

    Inference-time Physics Alignment of Video Generative Models with Latent World Models , author=. arXiv preprint arXiv:2601.10553 , year=

    work page arXiv

  46. [46]

    Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision , pages=

    Do generative video models understand physical principles? , author=. Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision , pages=

  47. [47]

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

    Align your latents: High-resolution video synthesis with latent diffusion models , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

  48. [48]

    Imagen Video: High Definition Video Generation with Diffusion Models

    Imagen Video: High Definition Video Generation with Diffusion Models , author=. arXiv preprint arXiv:2210.02303 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  49. [49]

    International Conference on Learning Representations (ICLR) , year=

    AnimateDiff: Animate Your Personalized Text-to-Image Diffusion Models without Specific Tuning , author=. International Conference on Learning Representations (ICLR) , year=

  50. [50]

    I2VGen-XL: High-Quality Image-to-Video Synthesis via Cascaded Diffusion Models

    I2VGen-XL: High-Quality Image-to-Video Synthesis via Cascaded Diffusion Models , author=. arXiv preprint arXiv:2311.04145 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  51. [51]

    European Conference on Computer Vision (ECCV) , pages=

    DynamiCrafter: Animating Open-domain Images with Video Diffusion Priors , author=. European Conference on Computer Vision (ECCV) , pages=

  52. [52]

    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=

  53. [53]

    Advances in Neural Information Processing Systems , volume=

    VideoComposer: Compositional Video Synthesis with Motion Controllability , author=. Advances in Neural Information Processing Systems , volume=

  54. [54]

    International Conference on Machine Learning , pages=

    Towards World Simulator: Crafting Physical Commonsense-Based Benchmark for Video Generation , author=. International Conference on Machine Learning , pages=. 2025 , organization=

    2025

  55. [55]

    IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages=

    PhysGaussian: Physics-Integrated 3D Gaussians for Generative Dynamics , author=. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages=

  56. [56]

    International Conference on Machine Learning (ICML) , pages=

    Genie: Generative Interactive Environments , author=. International Conference on Machine Learning (ICML) , pages=

  57. [57]

    Advances in Neural Information Processing Systems , volume=

    Interaction Networks for Learning about Objects, Relations and Physics , author=. Advances in Neural Information Processing Systems , volume=

  58. [58]

    IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages=

    FreeU: Free Lunch in Diffusion U-Net , author=. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages=

  59. [59]

    Advances in neural information processing systems , volume=

    Elucidating the design space of diffusion-based generative models , author=. Advances in neural information processing systems , volume=

  60. [60]

    Advances in Neural Information Processing Systems , volume=

    Alias-Free Generative Adversarial Networks , author=. Advances in Neural Information Processing Systems , volume=

  61. [61]

    Advances in Neural Information Processing Systems , volume=

    Cold diffusion: Inverting arbitrary image transforms without noise , author=. Advances in Neural Information Processing Systems , volume=

  62. [62]

    IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , year=

    Wavelet Diffusion Models are Fast and Scalable Image Generators , author=. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , year=

  63. [63]

    Classifier-Free Diffusion Guidance

    Classifier-free diffusion guidance , author=. arXiv preprint arXiv:2207.12598 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  64. [64]

    IEEE/CVF International Conference on Computer Vision (ICCV) , pages=

    Adding Conditional Control to Text-to-Image Diffusion Models , author=. IEEE/CVF International Conference on Computer Vision (ICCV) , pages=

  65. [65]

    International Conference on Learning Representations (ICLR) , year=

    Prompt-to-prompt image editing with cross attention control , author=. International Conference on Learning Representations (ICLR) , year=

  66. [66]

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

    Improving sample quality of diffusion models using self-attention guidance , author=. Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

  67. [67]

    IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages=

    High-Resolution Image Synthesis with Latent Diffusion Models , author=. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages=

  68. [68]

    International Conference on Learning Representations , volume=

    Tokenflow: Consistent diffusion features for consistent video editing , author=. International Conference on Learning Representations , volume=

  69. [69]

    1999 , publisher=

    Discrete-time signal processing , author=. 1999 , publisher=

    1999

  70. [70]

    European conference on computer vision , pages=

    Raft: Recurrent all-pairs field transforms for optical flow , author=. European conference on computer vision , pages=. 2020 , organization=

    2020

  71. [71]

    IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , year =

    Yang, Yanchao and Soatto, Stefano , title =. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , year =

  72. [72]

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

    Phyt2v: Llm-guided iterative self-refinement for physics-grounded text-to-video generation , author=. Proceedings of the Computer Vision and Pattern Recognition Conference , pages=

  73. [73]

    Advances in Neural Information Processing Systems (NeurIPS) , year=

    VideoHallu: Evaluating and Mitigating Multi-modal Hallucinations on Synthetic Video Understanding , author=. Advances in Neural Information Processing Systems (NeurIPS) , year=

  74. [74]

    Lv, Jiaxi and Huang, Yi and Yan, Mingfu and Huang, Jiancheng and Liu, Jianzhuang and Liu, Yifan and Wen, Yafei and Chen, Xiaoxin and Chen, Shifeng , booktitle=

  75. [75]

    arXiv preprint arXiv:2505.21653 , year=

    Think Before You Diffuse: Infusing Physical Rules into Video Diffusion , author=. arXiv preprint arXiv:2505.21653 , year=

    work page arXiv

  76. [76]

    International Conference on Learning Representations , volume=

    Physbench: Benchmarking and enhancing vision-language models for physical world understanding , author=. International Conference on Learning Representations , volume=

  77. [77]

    Gemini 2.5: Pushing the Frontier with Advanced Reasoning, Multimodality, Long Context, and Next Generation Agentic Capabilities

    Gemini 2.5: Pushing the frontier with advanced reasoning, multimodality, long context, and next generation agentic capabilities , author=. arXiv preprint arXiv:2507.06261 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  78. [78]

    Forty-first international conference on machine learning , year=

    Scaling rectified flow transformers for high-resolution image synthesis , author=. Forty-first international conference on machine learning , year=

  79. [79]

    International Conference on Machine Learning , pages=

    WorldSimBench: Towards Video Generation Models as World Simulators , author=. International Conference on Machine Learning , pages=. 2025 , organization=

    2025

  80. [80]

    The Thirteenth International Conference on Learning Representations , year=

    FreqPrior: Improving Video Diffusion Models with Frequency Filtering Gaussian Noise , author=. The Thirteenth International Conference on Learning Representations , year=

Showing first 80 references.