Pith. sign in

REVIEW 5 major objections 6 minor 3 cited by

A three-stage training pipeline can turn appearance-driven video diffusion models into physics-consistent 4D world models by grounding geometry and motion in simulation and rewarding physically plausible long-horizon trajectories.

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 · deepseek-v4-flash

2026-08-02 19:06 UTC pith:JLFWSRNS

load-bearing objection Well-assembled recipe, but the 4D benchmark measures the RL reward—the real evidence is Physics-IQ, and it's under-defined. the 5 major comments →

arxiv 2603.03485 v3 pith:JLFWSRNS submitted 2026-03-03 cs.CV cs.AIcs.RO

Phys4D: Fine-Grained Physics-Consistent 4D Modeling from Video Diffusion

Haoran Lu , Shang Wu , Songling Liu , Jianshu Zhang , Maojiang Su , Guo Ye , Chenwei Xu , Lie Lu
show 5 more authors
Pranav Maneriker Fan Du Manling Li Zhaoran Wang Han Liu
This is my paper
classification cs.CV cs.AIcs.RO
keywords video diffusion4D world modelphysics consistencygeometry-motion couplingsimulation-grounded trainingreinforcement learningdepth and optical flowworldline evaluation
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.

Phys4D claims that video diffusion models trained only to match pixels do not reliably respect real-world dynamics, and that this can be fixed by injecting explicit geometry and motion supervision. The paper proposes a three-stage training paradigm: first, depth and motion heads are bootstrapped on pseudo-labeled generated and internet videos; second, a physics-grounded fine-tuning stage couples depth and optical flow through a warp-consistency loss on simulated data; third, reinforcement learning maximizes a reward defined as negative 4D Chamfer distance between generated and simulated spatiotemporal point clouds. The authors report that this recipe improves Physics-IQ scores across several open-source video diffusion backbones and reduces trajectory drift and depth-warp errors on a new 4D benchmark, while keeping visual quality competitive. A sympathetic reader would care because it offers a path from appearance-driven generators to world models that keep objects intact, move them plausibly, and maintain coherent geometry over time.

Core claim

The paper's claim is that fine-grained physical consistency in video generation can be learned, not just approximated by appearance matching. Phys4D takes a pretrained video diffusion model, adds lightweight depth and motion prediction heads, and progressively trains them: first on pseudo-labels to bootstrap 4D scene understanding, then on simulation data with a warp-based loss that enforces that depth warped by predicted motion agrees with depth at the next frame, and finally through reinforcement learning where the reward is the negative 4D Chamfer distance between the generated spatiotemporal point cloud and a simulation-derived ground-truth point cloud. Reported results show improvement

What carries the argument

The central mechanism is the coupling of geometry and motion through warp consistency and through a 4D point-cloud reward. The warp loss L_warp = ||W(D_t, F_{t->t+1}) - D_{t+1}||_1 makes predicted depth and optical flow agree across adjacent frames, operationalizing the intuitive identity that a coherent world moves together. The 4D Chamfer distance CD_4D(P_gen, P_phys) compares generated point clouds in (x, y, z, time) against simulation ground truth, turning physical plausibility into a differentiable reward. A stochastic Flow-SDE converts the deterministic denoising trajectory into a policy that can be optimized with PPO for exploration.

Load-bearing premise

The load-bearing premise is that the physics simulator's fluids, cloth, deformables, and rigid bodies behave faithfully and diversely enough that supervision and rewards computed from it teach real physical consistency rather than simulator-specific quirks.

What would settle it

Take real-world videos of falling, rolling, and pouring scenes with known 3D trajectories (e.g., captured with depth sensors or motion capture), run Phys4D on the same initial frames, and measure worldline and 4D Chamfer errors against the real observations; if these errors are no smaller than those of an appearance-only baseline, the simulation grounding is not transferring. A second check: vary simulator gravity or material parameters and test whether Phys4D's outputs follow the simulated value rather than defaulting to a generic learned prior.

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

If this is right

  • Physics-consistent video generation can be achieved without replacing the base generative model: the pipeline is architecture-agnostic and transfers to multiple open-source video diffusion backbones.
  • Simulation-generated data can provide dense, temporally aligned supervision (depth, optical flow, scene flow, trajectories) that real video at scale cannot, enabling direct supervision of geometry-motion coupling.
  • Explicit 4D world-level evaluation—per-frame geometry, temporal warp consistency, long-horizon worldlines, and novel-time interpolation—can expose physical inconsistencies that appearance-based metrics miss.
  • Each training stage contributes monotonically: physics-grounded supervised fine-tuning gives the largest gain, and simulation-grounded reinforcement learning further corrects residual physical violations.
  • A model trained this way can interpolate coherent geometry and motion at unseen timestamps, suggesting it learns a continuous (x, y, z, t) world rather than a sequence of independent frames.

Where Pith is reading between the lines

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

  • The same recipe could extend to other generative setups—image-to-4D, interactive world models, or action-conditioned generation—wherever a trajectory-level reward can be defined from simulation or pseudo ground truth.
  • A testable extension is to replace simulation ground truth with strong off-the-shelf depth and flow estimators on real video and measure whether the Physics-IQ and 4D gains persist outside the simulated distribution.
  • The novel-time interpolation diagnostic could serve as a general probe for any 4D generator (e.g., Gaussian-splatting or NeRF-based models), not just video diffusion, because it tests whether the model truly encodes continuous time.
  • The heaviest caveat is transfer: if simulator physics differ from real physics, both training and the in-distribution 4D benchmark could reward simulator-specific artifacts; confirming gains on real observed trajectories would be the decisive follow-up.

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

5 major / 6 minor

Summary. The paper proposes Phys4D, a three-stage training pipeline that converts a pretrained video diffusion model into a physics-consistent 4D world model. Stage 1 bootstraps depth and motion heads with pseudo-labels from monocular estimators; Stage 2 fine-tunes the backbone with simulation-derived supervision and a warp-based geometry–motion consistency loss; Stage 3 applies PPO with a terminal reward defined as negative 4D Chamfer distance against simulator ground truth. The authors also introduce a three-tier 4D evaluation suite (per-frame depth, temporal warp consistency, worldline/4D reconstruction) and report results on Physics-IQ plus their proposed benchmark across three open-source backbones, claiming consistent gains in fine-grained physical consistency.

Significance. If the reported results are valid, the paper makes a useful contribution: it shows a concrete recipe for injecting simulation-based geometric and dynamic supervision into open-source video diffusion models, and it demonstrates gains on an externally defined physics benchmark (Physics-IQ) across multiple architectures. The simulation pipeline with 20000+ randomized physical configurations and 1.25M videos is also a substantial resource. The proposed 4D diagnostic metrics, if made independent of the training reward, would be valuable to the community. However, the current evaluation conflates the model's learned head with the measurement instrument, and its main 4D benchmark metric is the exact objective optimized during RL, so the central claim of generalized fine-grained physical consistency is not established by the evidence as presented.

major comments (5)
  1. [Sec. 6.2, Eq. (12)] The primary 4D metric in Table 4, CD4D, is exactly the terminal reward optimized in Stage 3 (Eqs. 9–12). The test scenes in Sec. 6.2 are simulation scenes generated by the same simulator and domain-randomization distribution used for Stage 2 and Stage 3 (Sec. 4, Table 1). Table 4 therefore measures how well the model optimizes its training reward on the training distribution, not whether it has learned general physical consistency. The authors should provide an evaluation on a held-out distribution (e.g., different simulator, real-world RGB-D sequences with independent annotations, or physical reasoning benchmarks not used in training), or clearly separate the reward metric from the benchmark metric.
  2. [Sec. 6.2, Tables 3 and 5] In Tables 3 and 5, Phys4D is evaluated with its own depth and motion heads trained on simulation ground truth from the same distribution as the evaluation scenes, while baselines are evaluated with off-the-shelf estimators (DepthAnythingV2, SEA-RAFT/OTF). This conflates estimator quality and domain fit with physical consistency. For example, Table 3 reports an AbsRel of 0.2711 for 'Phys4D (Ours)' versus 0.3929 for 'WAN + DAV2', but the gap could be largely due to the estimator rather than the generative model. The authors should compare all methods with the same measurement instrument (e.g., run DAV2/SEA-RAFT on all generated videos), or train the same head architecture on all baselines, and report both variants.
  3. [Sec. 6.1, Table 2] The 'Score (%)' column is never defined. The reader cannot verify the +11.4 absolute improvement from 18.8 to 30.2 without knowing how MSE, ST-IoU, S-IoU, and WS-IoU are aggregated. If this is the official Physics-IQ aggregate, the exact formula or a precise pointer to the benchmark code is needed; otherwise the headline claim is not auditable. Additionally, no confidence intervals or significance tests are reported anywhere in the paper. Given the stochastic nature of video generation and the small gaps in several metrics (e.g., Table 5 RGB Charb. 0.0518 vs. 0.0507), multiple seeds and a significance test are necessary to support the quantitative claims.
  4. [Sec. 5.4 / Appendix F.4] The main text states that the PPO objective and training details are 'provided in the Supp F.4', but Appendix F.4 only restates the clipped PPO objective (Eq. 37) and does not provide the promised details. Missing information includes the number of denoising steps, temperature, rollout size, number of PPO epochs, learning rate, clipping parameter epsilon, the flow-SDE noise schedule sigma_t values, the moving-region threshold delta, and the temporal weight alpha (though alpha is given in discussion). Without these settings, the Stage 3 results are not reproducible and the contribution cannot be independently validated.
  5. [Sec. 4 / Sec. 6.2] The paper does not state whether the simulation test scenes used in Tables 3–5 are disjoint from the training environments. Appendix E.1.1 says 'all models are evaluated on identical test scenes', but this does not specify whether those scenes were held out from the 250,000-environment training set. If the test scenes are drawn from the same randomized configuration space as training, the 4D benchmark results are in-distribution and cannot support the generalization claim. The authors should specify the train/test split and ideally evaluate on simulation configurations outside the training ranges or on different physical solvers.
minor comments (6)
  1. [Eq. (4)] The sentence 'where ℓ denotes a combination of L1 and ˆDt, ˆFt→t+1 are ground-truth depth and optical flow respectively' is grammatically incomplete. The loss combination should be stated explicitly (e.g., L1 + perceptual + smoothness, if used).
  2. [Table 2] ST-IoU, S-IoU, and WS-IoU are not defined in the main text. Please add one-sentence definitions or cite the exact equations from Physics-IQ.
  3. [Appendix E.3.4] The text refers to 'C.3' and 'C.4' when comparing to previous sub-experiments; these should be E.3.3 and E.3.4.
  4. [Appendix E.3.3] The 'Fail Rate' threshold τ and the number of sampled pixels N are not specified in the appendix text. The definitions appear only in the repeated paragraph without numeric values; please provide the exact values used in Table 4.
  5. [Table 3] The row 'Phys4D + DAV2' is confusing: it presumably means Phys4D videos with depth estimated by DepthAnythingV2, but this is not explained in the caption or main text. Clarify whether this is an estimator-matching control and what it controls for.
  6. [Sec. 2/Related Work] Several sentences have informal phrasing (e.g., 'Detailed discussion please refer to Appendix G.1 G.2 G.3'). Also, the abstract says 'we introduce a set of 4D world consistency evaluation' — should be 'a set of 4D world consistency evaluations'.

Circularity Check

2 steps flagged

Table 4's headline 4D metric is the Stage-3 RL reward (Eq. 9) measured on the same simulator distribution, so the 4D benchmark measures the training objective rather than independently validating physics.

specific steps
  1. fitted input called prediction [Sec. 5.4 Eq. (9); Sec. 6.2 Table 4; App. E.3.2]
    "We then define the reward as the negative 4D Chamfer Distance between the two point sets: r(z0, c) = −CD4D(Pgen, Pphys). ... Metric. We measure 4D Chamfer Distance between predicted and ground-truth spatio-temporal point sets."

    Stage-3 PPO optimizes exactly this reward r = −CD4D on simulation ground truth. The headline Geometry metric in Table 4 is the same CD4D, evaluated on the same simulator-derived test distribution. Therefore the reported 'improvement' in 4D geometric consistency is the direct objective the model was trained to maximize, not an independent measure of physical understanding. The evaluation reduces to reward optimization by construction.

  2. fitted input called prediction [App. E.3.1 / E.2.1; Tables 3 and 5]
    "Baselines rely on off-the-shelf depth and flow estimators (OTF), while Phys4D directly predicts geometry and motion through learned heads. All evaluations are conducted under a fixed-camera setting."

    Phys4D's depth and motion heads are trained on ground truth from the same simulation pipeline that generates the test scenes, whereas baselines use external estimators (DepthAnythingV2, RAFT/SEA-RAFT) with no such in-distribution training. Consequently, Tables 3 and 5 partly measure how well Phys4D's own fitted heads memorize or fit the simulator distribution, conflating estimator quality and domain fit with generative physical consistency.

full rationale

The paper's central demonstration of fine-grained 4D physical consistency rests on a benchmark whose principal metric is identical to the Stage-3 RL reward. Eq. (9) defines r(z0,c) = −CD4D(Pgen,Pphys), and Sec. 6.2 / App. E.3.2 report the same 4D Chamfer Distance as the core Geometry metric on simulation ground truth. Since PPO is used to maximize this reward, Table 4 measures whether the model optimized that objective, not whether it learned general physics. The warp, worldline, and novel-time metrics in Table 4 similarly use Phys4D's own heads trained on the same simulator distribution, while baselines are evaluated with external off-the-shelf estimators, creating an estimator/domain mismatch that inflates the apparent advantage. This is genuine circularity in the 4D evaluation, warranting a score of 7. However, the paper also reports improvements on Physics-IQ (Table 2), an external benchmark that is not defined in terms of the training reward. That provides some independent, albeit partially underspecified, support for the method's physical-consistency claims. Thus the circularity is partial rather than total: the 4D benchmark reduces to the training objective, but the overall claim does not rest exclusively on that benchmark. No load-bearing self-citation or imported-uniqueness pattern was found; the simulation citations are used only as a data source.

Axiom & Free-Parameter Ledger

7 free parameters · 6 axioms · 0 invented entities

The central claim rests on two families of assumptions: (1) standard flow-matching/RL mathematics, and (2) domain assumptions that the Isaac Sim solvers are faithful physics, that off-the-shelf depth/flow labels are adequate, that the 4D Chamfer distance to simulated point clouds is a valid proxy for physical consistency, and that the score identity holds for the trained velocity. The method introduces hand-set simulator ranges, reward weights, and thresholds; these shape the training distribution and are not derived elsewhere.

free parameters (7)
  • warp loss weights λ_w, λ_rgb = not specified
    Hyperparameters in Eq. 5 controlling the warp-consistency terms; values are never given in the main text or appendix.
  • moving-region threshold δ = not specified
    Threshold selecting 'moving' points for the RL reward (App B.2); no value or sensitivity analysis reported.
  • temporal weight α = 0.03 m/frame
    Converts frame-time differences into spatial meters in Eq. 10; chosen by hand to match 'typical per-frame displacement' (App B.2).
  • simulator randomization ranges (gravity, density, Young's modulus, friction, restitution, etc.) = ranges in Table 1
    Define the training/reward distribution; chosen by hand, not derived from first principles.
  • curriculum complexity proportions = single 30%, two-body 35%, multi-object 35%
    Data-generation split, chosen by hand.
  • worldline failure-rate threshold τ and sample count N = not specified
    Evaluation-protocol parameters in App E.3.3; values are not reported.
  • Flow-SDE noise schedule σ_t = not specified
    Exploration noise magnitude in Eq. 8; no schedule is given.
axioms (6)
  • standard math Flow-matching / score-velocity identity
    Background for Eq. 2 and Eq. 34; standard in the cited literature.
  • standard math PPO and MDP formulation of denoising
    Background from Wallace et al. and Ren et al.; standard in diffusion/flow RL.
  • domain assumption Isaac Sim solvers faithfully model real physics
    Load-bearing: the entire supervision and reward pipeline assumes simulated physics transfers to real videos. Invoked in Sec. 4 and Sec. 5.4.
  • domain assumption Off-the-shelf depth/flow estimators give valid pseudo-labels for Stage 1
    Stage 1 uses DepthAnythingV2 and RAFT on generated and internet videos; if the labels are noisy, the heads inherit errors.
  • domain assumption 4D Chamfer distance to simulated point clouds is a valid proxy for physical consistency
    Used as both reward (Eq. 9) and headline metric (Table 4); no validation against human judgments or out-of-distribution physics.
  • domain assumption The score identity Eq. 34 holds for the trained velocity field
    Assumes v_θ is close to the true conditional velocity; used to build the Flow-SDE in App F.3.

pith-pipeline@v1.3.0-alltime-deepseek · 25866 in / 22594 out tokens · 199245 ms · 2026-08-02T19:06:06.222137+00:00 · methodology

0 comments
read the original abstract

Recent video diffusion models have achieved impressive capabilities as large-scale generative world models. However, these models often struggle with fine-grained physical consistency, exhibiting physically implausible dynamics over time. In this work, we present \textbf{Phys4D}, a pipeline for learning physics-consistent 4D world representations from video diffusion models. Phys4D adopts \textbf{a three-stage training paradigm} that progressively lifts appearance-driven video diffusion models into physics-consistent 4D world representations. We first bootstrap robust geometry and motion representations through large-scale pseudo-supervised pretraining, establishing a foundation for 4D scene modeling. We then perform physics-grounded supervised fine-tuning using simulation-generated data, enforcing temporally consistent 4D dynamics. Finally, we apply simulation-grounded reinforcement learning to correct residual physical violations that are difficult to capture through explicit supervision. To evaluate fine-grained physical consistency beyond appearance-based metrics, we introduce a set of \textbf{4D world consistency evaluation} that probe geometric coherence, motion stability, and long-horizon physical plausibility. Experimental results demonstrate that Phys4D substantially improves fine-grained spatiotemporal and physical consistency compared to appearance-driven baselines, while maintaining strong generative performance. Our project page is available at https://sensational-brioche-7657e7.netlify.app/

Figures

Figures reproduced from arXiv: 2603.03485 by Chenwei Xu, Fan Du, Guo Ye, Han Liu, Haoran Lu, Jianshu Zhang, Lie Lu, Manling Li, Maojiang Su, Pranav Maneriker, Shang Wu, Songling Liu, Zhaoran Wang.

Figure 1
Figure 1. Figure 1: Phys4D Our three-stage training paradigm lifts a pretrained video diffusion model into a physics-consistent 4D world model. Bottom-left: a scalable physics-simulation data collection pipeline with diverse camera view control. Top-right: Our 4D physics benchmark and diagnostics that evaluate geometric coherence, motion stability beyond appearance-based metrics. from pure 2D appearance modeling toward an exp… view at source ↗
Figure 2
Figure 2. Figure 2: Overview of Phys4D simulation framework. (a)Left A Large-scale async data collection pipeline. (b)Right Supported physical objects: rigid and articulated structures, garments, fluids, thermodynamics, deformable, inflatables, ropes, and granular materials. parameter perturbation to prevent homogeneity, resulting in over 20000 unique physical configurations [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Overview of the Phys4D training pipeline. Our three-stage paradigm progressively injects physics into a pretrained video diffusion model. Stage 1 (blue): The DiT backbone is frozen while depth and motion heads are trained on pseudo-labeled RGB videos. Stage 2 (green): The backbone is adapted via LoRA using physics simulation data with ground-truth annotations; a warp consistency loss couples depth and moti… view at source ↗
Figure 4
Figure 4. Figure 4: Qualitative comparison on Physics-IQ scenarios. We compare Wan2.2-5b baseline (middle) with Wan2.2-5b + Phys4D (right) across three physical interaction types: object placement on a rotating platform (top), ball rolling dynamics (middle), and fluid pouring (bottom). Phys4D produces more consistent object geometry, physically plausible motion, and stable temporal dynamics compared to the baseline, which exh… view at source ↗
Figure 5
Figure 5. Figure 5: Qualitative Results on 4D Experiment. From left to right: the ground-truth 4D point cloud and the generated 4D point cloud at 1/4 of the sequence; the generated 4D point cloud at 1/2 of the sequence; the novel-time generated 4D point cloud at 3/4 of the sequence; and the generated 4D point cloud at the final frame. This visualization highlights the model’s ability to maintain coherent geometry and object m… view at source ↗
Figure 6
Figure 6. Figure 6: Qualitative Result Of Phys4D Generated Scene. Appendix A: Data. We describe the data sources used throughout training, including internet videos, model-generated samples, and simulation-generated data. This section details data preprocessing, filtering strategies, and annotation procedures for depth and motion supervision, clarifying how diverse supervision signals are constructed across different stages o… view at source ↗
Figure 7
Figure 7. Figure 7: Qualitative Result Of Phys4D Generated Scene. D. Experiment D.1. Physics-IQ For evaluation, we generate 5-second videos conditioned on the switch frame following the Physics-IQ protocol. We use nucleus sampling with temperature 1.0 and top-p = 0.95 for all experiments. Phys4D achieves substantial improvements over base video diffusion models across multiple backbones. When applied to CogVideoX-5b, Phys4D i… view at source ↗

discussion (0)

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

Forward citations

Cited by 3 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. APT: Atomic Physical Transitions for Causal Video-Language Understanding

    cs.CV 2026-06 unverdicted novelty 6.0

    Introduces APT chains as ordered causal transition sequences and APT-Tune to improve VLM transition detection while preserving event-level performance.

  2. Physics-IQ Verified

    cs.CV 2026-06 unverdicted novelty 5.0

    Physics-IQ Verified refines 57.6% of samples and 34.8% of prompts from the original benchmark and produces moderate ranking shifts (Kendall's τ = 0.46) across six image-to-video models.

  3. MagicSim: A Unified Infrastructure for Executable Embodied Interaction

    cs.RO 2026-06 unverdicted novelty 5.0

    MagicSim is a unified embodied interaction infrastructure built on a deterministic batched runtime and shared MDP that supports diverse world construction, execution, task evaluation, automatic rollout generation, and...

Reference graph

Works this paper leans on

8 extracted references · 7 linked inside Pith · cited by 3 Pith papers

  1. [7]

    org/CorpusID:269282774

    URL https://api.semanticscholar. org/CorpusID:269282774. Zheng, Z., Peng, X., Yang, T., Shen, C., Li, S., Liu, H., Zhou, Y ., Li, T., and You, Y . Open-sora: Democratizing efficient video production for all. 2024a. URL https: //arxiv.org/abs/2412.20404. Zheng, Z., Peng, X., Yang, T., Shen, C., Li, S., Liu, H., Zhou, Y ., Li, T., and You, Y . Open-sora: De...

    Pith/arXiv arXiv

  2. [8]

    org/CorpusID:269502120

    URL https://api.semanticscholar. org/CorpusID:269502120. Zhou, Y ., Wang, Y ., Zhou, J., Chang, W., Guo, H., et al. Omniworld: A multi-domain and multi-modal dataset for 4d world modeling.arXiv preprint arXiv:2509.12201, 2025. 12 Phys4D: Fine-Grained Physics-Consistent 4D Modeling from Video Diffusion Appendix Overview This appendix provides supplementary...

    arXiv 2025

  3. [2002]

    org/CorpusID:110509160

    URL https://api.semanticscholar. org/CorpusID:110509160. 9 Phys4D: Fine-Grained Physics-Consistent 4D Modeling from Video Diffusion Google DeepMind. State-of-the-art video and im- age generation with veo 2 and imagen 3. https: //blog.google/technology/google-labs/ video-image-generation-update-december-2024/ , 2024. Grauman, K., Westbury, A., Byrne, D., C...

    Pith/arXiv arXiv 2024

  4. [2015]

    org/CorpusID:28559386

    URL https://api.semanticscholar. org/CorpusID:28559386. Bengio, Y ., Louradour, J., Collobert, R., and Weston, J. Curriculum learning. InProceedings of the 26th Annual International Conference on Machine Learning, pp. 41– 48, 2009. Bertiche, H., Madadi, M., Escalera, S., Black, M. J., and Mu˜noz, X. Behave: Dataset and method for track- ing human-object i...

    Pith/arXiv arXiv 2009

  5. [2020]

    org/CorpusID:214667893

    URL https://api.semanticscholar. org/CorpusID:214667893. Wallace, B., Dang, M., Rafailov, R., Zhou, L., Lou, A., Pu- rushwalkam, S., Ermon, S., Xiong, C., Joty, S., and Naik, N. Diffusion model alignment using direct preference op- timization, 2023. URL https://arxiv.org/abs/ 2311.12908. Wan, T., Wang, A., Ai, B., Wen, B., Mao, C., Xie, C.-W., Chen, D., Y...

    Pith/arXiv arXiv 2023

  6. [2022]

    org/CorpusID:252595919

    URL https://api.semanticscholar. org/CorpusID:252595919. Sohl-Dickstein, J., Weiss, E., Maheswaranathan, N., and Ganguli, S. Deep unsupervised learning using nonequi- librium thermodynamics.ICML, 2015. Song, Y . and Ermon, S. Generative modeling by estimating gradients of the data distribution.NeurIPS, 2019. Song, Y ., Sohl-Dickstein, J., Kingma, D. P., K...

    Pith/arXiv arXiv 2015

  7. [2023]

    org/CorpusID:266044186

    URL https://api.semanticscholar. org/CorpusID:266044186. Yu, H.-X., Duan, H., Herrmann, C., Freeman, W. T., and Wu, J. Wonderworld: Interactive 3d scene generation from a single image.arXiv preprint arXiv:2406.09394, 2024. URLhttps://arxiv.org/abs/2406.09394. Yuan, Y ., Song, J., Iqbal, U., Vahdat, A., and Kautz, J. Phys- diff: Physics-guided human motion...

    Pith/arXiv arXiv 2024

  8. [2024]

    org/CorpusID:269982890

    URL https://api.semanticscholar. org/CorpusID:269982890. Albergo, M. S. and Vanden-Eijnden, E. Stochastic inter- polants: A unifying framework for flows and diffusions. arXiv preprint arXiv:2303.08797, 2023. Bansal, H., Lin, Z., Xie, T., Zong, Z., Yarom, M., Bitton, Y ., Jiang, C., Sun, Y ., Chang, K.-W., and Grover, A. Videophy: Evaluating physical commo...

    Pith/arXiv arXiv 2023