Pith. sign in

REVIEW 4 major objections 4 minor 5 cited by

PERSIST replaces pixel histories with a persistent latent 3D world-frame, keeping generated video geometrically consistent over thousands of steps.

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 19:05 UTC pith:DHN5SF3Z

load-bearing objection A genuinely new architecture for persistent 3D memory in world models, but the evidence for the headline 3D-consistency claim is weaker than the FVD table suggests. the 4 major comments →

arxiv 2603.03482 v2 pith:DHN5SF3Z submitted 2026-03-03 cs.CV cs.AIcs.LG

Beyond Pixel Histories: World Models with Persistent 3D State

Samuel Garcin , Thomas Walker , Steven McDonagh , Tim Pearce , Hakan Bilen , Tianyu He , Kaixin Wang , Jiang Bian This is my paper
classification cs.CV cs.AIcs.LG
keywords world models3D representationpersistent memoryinteractive video generationflow matchingdiffusion transformerscamera modellingneural rendering
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.

PERSIST argues that interactive world models should not carry memory as a history of pixels. Instead, it maintains a persistent latent 3D world-frame centred on the agent, tracks the camera as a query into that frame, and renders each new pixel frame by projecting the 3D latents to screen space and denoising with a learned shader. Trained on a voxel game engine with ground-truth 3D states and camera poses, PERSIST produces long rollouts that stay geometrically consistent, remember revisited areas, and can be edited or re-initialised in 3D. The paper reports substantially better long-horizon quality and stability than rolling-window and key-frame-retrieval baselines, with FVD scores staying nearly flat out to 600 frames. If correct, this shifts world modelling toward worlds that persist and evolve rather than frames that are redrawn from memory.

Core claim

PERSIST decomposes world simulation into three coupled prediction problems: a world-frame model that evolves a latent voxel representation of the space around the agent, a camera model that predicts the agent's pose within that frame, and a world-to-pixel module that projects world latents through a differentiable rasterizer and denoises a pixel latent conditioned on the resulting depth-ordered stack. The camera acts as a spatial lookup key, so memory retrieval cost is independent of episode length, and geometric consistency is enforced by construction rather than learned implicitly from pixels. The paper shows that conditioning on these actively generated 3D guidance frames yields substanti

What carries the argument

The central object is the camera-centred world-frame w, a latent 3D voxel grid updated by Wθ, paired with a camera state c=⟨pos,rot,fov⟩ predicted by Cθ and a learned renderer Pθ. The projector R(c,w) rasterizes voxel features into a per-pixel, depth-ordered stack w2D via GPU depth-peeling, and Pθ — a rectified-flow causal DiT that receives more latent channels from w2D than from the pixel latent — acts as a deferred shader. Wθ is likewise a rectified-flow causal DiT with 3D spatial attention and XYZ absolute position embeddings, and can be invoked with an empty world frame to synthesise w0 from a single image. Each denoiser is trained with diffusion forcing plus 10% flat noise augmentation

Load-bearing premise

The pipeline requires ground-truth 3D world-states and camera states during training; without that privileged supervision, PERSIST cannot be trained, and the authors list this as the method's primary limitation.

What would settle it

Apply the same architecture to a dataset without ground-truth 3D annotations, using only images and estimated poses, and measure whether long-horizon FVD still beats pixel-history baselines; or ablate w2D conditioning from Pθ entirely and check whether the stability advantage disappears. Either test would show whether the gain comes from the 3D state or from the architecture itself.

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

If this is right

  • Long-horizon generation quality stays roughly flat out to 600 frames, while pixel-history baselines degrade sharply, with world frames coherent for several thousand steps.
  • Spatial memory becomes a 3D store: revisiting a previously seen region from a new viewpoint stays consistent, and retrieval cost is independent of episode length.
  • The explicit 3D state yields capabilities pixel-history models lack: full 3D world initialisation from a single image, direct w0 conditioning, mid-episode world edits, and off-screen dynamics that produce on-screen effects.
  • A learned camera model yields physically consistent motion, whereas ground-truth cameras decoupled from the world frame can cause the agent to phase through terrain.
  • World-frames can be edited and re-injected mid-episode, enabling fine-grained geometry-aware control over generated experiences.

Where Pith is reading between the lines

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

  • Because the full pipeline is differentiable, an end-to-end post-training stage on generated rollouts could directly attack the exposure-bias drift the authors observe, a direction they list as future work and the architecture seems to invite.
  • If 2D-to-3D foundation models can supply synthetic 3D annotations, the method's core assumption of ground-truth world states could be relaxed, opening real-world video and providing a decisive test of transfer outside simulators.
  • The camera-as-lookup-key framing suggests a 3D memory bank that loads spatial chunks on demand would extend the fixed-size world-frame to arbitrarily large environments without the redundancy of pixel memory banks.
  • The largest ablation penalty comes from skipping 3D upscaling, suggesting the depth and alignment quality of w2D, not the expressiveness of the 3D latent, is the binding constraint; improving projection resolution may matter more than scaling Wθ.

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

4 major / 4 minor

Summary. PERSIST is an autoregressive interactive world model that replaces pixel-history memory with a persistent latent 3D world-frame. At each step a rectified-flow DiT predicts the evolution of a voxel-based world-frame, a causal transformer predicts the camera pose, a differentiable depth-peeled projector maps the world-frame to screen-space features, and a second rectified-flow DiT renders the pixel frame conditioned on those features. The method is trained and evaluated on Luanti/Craftium trajectories with ground-truth 3D voxel grids and camera states available during training. The paper reports FVD/FID improvements over Oasis and WorldMem, a human study with higher ratings for PERSIST, and several new capabilities: single-image 3D initialization, mid-episode 3D editing, and off-screen dynamics.

Significance. If the empirical claims were fully supported, PERSIST would be a significant contribution: it provides a fixed-cost spatial memory that is not tied to episode length, enforces geometric consistency through a common 3D latent state, and enables genuinely new controls such as 3D editing and explicit world initialization. The paper is also careful in several respects: the FVD baselines in Table 1 are retrained on the same dataset and use the same VAE and flow-matching formulation, the ablations isolate the 3D upsampling, model size, and camera model, and the limitation sections are candid about the need for ground-truth 3D supervision. I found no circularity: the model is evaluated against held-out ground-truth observations, and the 3D supervision is an explicit training assumption rather than a hidden fitted quantity. However, the current evidence does not establish the central claim of improved 3D consistency and spatial memory. The main automated metric is acknowledged by the authors to be insensitive to geometric consistency, and the only direct evidence for that property comes from a human study with a domain mismatch between the compared models. These are load-bearing gaps that

major comments (4)
  1. [§6.1, Table 1 and Figure 6] No error bars, confidence intervals, or significance tests are reported for any FVD or FID number. The claim of 'substantially improved' long-horizon quality rests on differences such as 129 vs 159 vs 216 at 200 frames, but with 168 evaluation trajectories the authors can easily compute bootstrap intervals or per-seed variance. As written, the reader cannot distinguish real gains from sampling noise. The same applies to Figure 6, where FID curves are shown without uncertainty. I would require confidence intervals and, for the user study in Table 2, paired significance tests.
  2. [§6.1, Camera-GT paragraph and Figure 11] The paper states that FVD/FID favor per-frame visual quality and are known not to capture spatial and temporal coherence, and it gives a concrete example: Camera-GT achieves FVD 161/152/152, close to PERSIST's 129/141/148, yet is said to introduce physical inconsistencies that FVD does not capture. This means the headline automated metric cannot distinguish the very property the paper claims to improve. Since ground-truth world-frames W are available in this setting, the authors should report direct geometry and spatial-memory metrics, for example voxel IoU/accuracy between predicted and held-out W, depth consistency at revisited viewpoints, or a quantitative measure of geometric artifacts. Without such a metric, Table 1 does not substantiate the central claim of improved 3D consistency.
  3. [Appendix B.2 and Table 2] The human study compares PERSIST on Craftium against released Oasis and WorldMem checkpoints evaluated on MineDojo. This is a confounded comparison: the methods differ in both architecture/training data and evaluation domain. The authors attempt an adjustment with ground-truth videos from both domains, but that controls only for human perception of the two domains, not for how model artifacts interact with domain. A model trained on Minecraft can fail on Minecraft for reasons unrelated to its memory mechanism. For the human study to support the 3D-consistency and spatial-memory claims, the baselines must be retrained on the same Craftium data used for the FVD comparison, or the study must be restricted to same-domain, same-training-set conditions. As published, the only direct evidence for the paper's core qualitative claims is not a valid controlled comparison.
  4. [§5, Dataset construction; §7 Conclusion and Appendix E] The framework requires W and C to be directly obtainable from E. This is a privileged-simulator assumption that excludes most real-world video, and the authors acknowledge it. My concern is not that the assumption is hidden—it is explicitly stated—but that the abstract and introduction claim a general 'new paradigm of world model' without this scope restriction. Since the method cannot be trained without ground-truth 3D state, the empirical contribution should be framed as applying to environments that provide such state, or the paper should provide evidence that the pipeline transfers to settings where W and C are estimated rather than measured. I treat this as a scope/claim issue rather than a technical error, but it should be addressed in revision.
minor comments (4)
  1. [Table 1] WorldMem is reported only at 200 frames ('358 – –'). The missing 400/600 entries are unexplained; if rollouts became unstable or the memory bank was exhausted, that should be stated.
  2. [Figure 6] The FID curves would benefit from axis labels, a legend identifying each curve, and per-timestep error bars. The current figure is hard to read quantitatively.
  3. [References] Several references have corrupted years and page ranges, e.g., 'Hafner et al., 20 2', 'Huang et al., 20 2', and 'Po et al., 2025' missing page numbers. These should be cleaned up before publication.
  4. [Appendix B.2] The sentence 'We use 600 ground truth cameras and pixel observations' for WorldMem should clarify whether this is the same 400-frame initialization used in Table 1 or a different setting; the two numbers appear inconsistent.

Circularity Check

0 steps flagged

No significant circularity: PERSIST is a supervised empirical architecture whose 'predictions' are evaluated against held-out ground truth; no derivation step reduces to its own inputs.

full rationale

PERSIST's derivation is a supervised learning pipeline, not a mathematical derivation that re-imports its conclusions. The three components are trained with distinct supervisory signals: Wθ via rectified-flow conditional flow matching (Eq. 1) on ground-truth voxel trajectories (Eq. 4), Cθ via MSE on ground-truth camera states, and Pθ via flow matching on pixel latents conditioned on W2D produced by the fixed projection R (Eqs. 5–6). The paper's 'predictions' — future world-frames, cameras, and pixels — are evaluated against held-out ground-truth trajectories (168 held-out evaluation trajectories, Table 1), so they are not fitted values renamed as predictions. The only in-loop quantities (predicted ¯w, c, ¯o, w2D) are generated autoregressively and then compared to held-out data, which is the standard definition of a predictive evaluation. The assertion that ground-truth W and C are obtainable from E (Section 5) is a stated scope assumption, explicitly acknowledged as a limitation in Section 7 and Appendix E; it constrains applicability but does not make the central claims definitional. Likewise, the FVD insensitivity to 3D consistency (Section 6.1, Fig. 11) and the cross-domain human-study comparison (Appendix B.2) are evidence-quality concerns, not circular-reasoning steps. Self-citations (e.g., Walker et al. 2023 in related work) are contextual and not load-bearing. No step in the claimed derivation chain reduces to its own input by construction, so the circularity score is 0.

Axiom & Free-Parameter Ledger

5 free parameters · 5 axioms · 1 invented entities

The central claim depends on a privileged training signal (ground-truth 3D state), a bounded spatial memory assumption, and a set of hand-tuned coupling parameters; none of these are derived from first principles, but all are standard for learned simulators.

free parameters (5)
  • context window sizes (K_W=8, K_C=8, K_P=16) = 8/8/16
    Hand-chosen temporal context lengths for world, camera, and pixel models; they determine the memory horizon each component can rely on.
  • w2D latent channel allocation = 752 channels vs 16 for pixel latents
    The pixel denoiser gives 752 channels to the projected 3D features to bias generation toward 3D guidance; this ratio is manually set.
  • context noise levels (tau_ctx) = 0.02 (W_theta), 0.1 (P_theta)
    Noise applied to past frames at inference to mitigate exposure bias; chosen by hand.
  • noise augmentation for exposure bias = 10% flat random noise
    Added to conditioning latents during training to close the train/inference gap; magnitude is a hand-tuned hyperparameter.
  • denoising steps = 20 (default); 2/4 for speed
    Number of flow-matching steps per frame; a speed-quality trade-off.
axioms (5)
  • domain assumption Ground-truth 3D voxel states W and cameras C are obtainable from the environment for every timestep during training.
    Section 5, Dataset construction: 'we assume that W and C are directly obtainable from E.' Without this, W_theta and C_theta cannot be trained; the paper's Conclusion lists this as the primary limitation.
  • domain assumption A fixed 48^3 voxel grid centered on the agent is a sufficient proxy for the hidden state of the environment.
    Section 4 defines s_tilde=<w,c> and explicitly notes 'although not a perfect replacement for the true hidden state.' If relevant environmental state lies outside this bounded box, the model cannot represent it (Appendix E acknowledges discarded distant info).
  • domain assumption The learned world-frame and camera-prediction modules together produce projections R(c,w) that provide useful guidance for pixel generation even when both are imperfectly predicted.
    The whole pipeline relies on conditioning P_theta on w2D from predicted W and C; the Camera-GT ablation shows this coupling can fail physically (agent phasing through walls), so the assumption is non-trivial.
  • domain assumption Rectified flow training with diffusion forcing yields a generative model whose per-component coverage is sufficient to bootstrap the autoregressive loop at inference.
    Section 5.3 uses diffusion forcing plus 10% noise augmentation instead of end-to-end training; the paper attributes residual degradation to exposure bias (Appendix E).
  • standard math Rectified flow / conditional flow matching objective (Eq. 1) yields a tractable generative training signal for the voxel and pixel denoisers.
    Paper relies on Lipman et al. (2023) flow-matching theory; no new derivation is attempted.
invented entities (1)
  • Latent 3D world-frame (voxel grid) no independent evidence
    purpose: Persistent spatial memory and geometry guidance; the state that evolves over time and is projected to screen space.
    A learned representation trained with ground-truth voxel supervision; at test time it drifts (Figure 13) and is not externally validated beyond the paper's own metrics.

pith-pipeline@v1.3.0-alltime-deepseek · 21087 in / 15342 out tokens · 139220 ms · 2026-08-02T19:05:12.149614+00:00 · methodology

0 comments
read the original abstract

Interactive world models continually generate video by responding to a user's actions, enabling open-ended generation capabilities. However, existing models typically lack a 3D representation of the environment, meaning 3D consistency must be implicitly learned from data, and spatial memory is restricted to limited temporal context windows. This results in an unrealistic user experience and presents significant obstacles to downstream tasks such as training agents. To address this, we present PERSIST, a new paradigm of world model which simulates the evolution of a latent 3D scene: environment, camera, and renderer. This allows us to synthesise new frames with persistent spatial memory and consistent geometry. Both quantitative metrics and a qualitative user study show substantial improvements in spatial memory, 3D consistency, and long-horizon stability over existing methods, enabling coherent, evolving 3D worlds. We further demonstrate novel capabilities, including synthesising diverse 3D environments from a single image, as well as enabling fine-grained, geometry-aware control over generated experiences by supporting environment editing and specification directly in 3D space. Project page: https://francelico.github.io/persist.github.io

Figures

Figures reproduced from arXiv: 2603.03482 by Hakan Bilen, Jiang Bian, Kaixin Wang, Samuel Garcin, Steven McDonagh, Thomas Walker, Tianyu He, Tim Pearce.

Figure 1
Figure 1. Figure 1: Initialized with a single pixel frame, PERSIST evolves in an autoregressive loop in response to user actions . We first denoise the 3D environment centred on the agent in the form of a latent world-frame . Next, camera parameters are predicted with a feed-forward transformer. We then project the world to the camera plane to form a depth-ordered stack of world latents . Finally, pixel latents are denoised, … view at source ↗
Figure 2
Figure 2. Figure 2: PERSIST enables long-horizon spatial memory by modelling the dynamics of a 3D world-frame around the agent. Camera parameters then act as memory look-up key, fetching relevant features from the world frame via a geometric projection (here visualized as the coloured voxels). 3D environment representations. In parallel, a growing number of approaches incorporate explicit 3D represen￾tations into their genera… view at source ↗
Figure 3
Figure 3. Figure 3: PERSIST can be initialized with a single RGB frame ( , row 1), or with a single RGB and world frame ( + , row 2). We visualize the world-frames and videos produced by an autoregressive rollout of 600 timesteps. Even with a single RGB frame for initialization, PERSIST can generate cohesive and evolving worlds. learning an interactive world simulation of E consists of min￾imising the objective E[D(On, O˜n)].… view at source ↗
Figure 4
Figure 4. Figure 4: World frame w features are projected to screen-space to obtain the depth-ordered stack of features w˜2D and linear depth information d. Pixel frame prediction. Pixel frames are generated by sampling o¯t ∼ Pθ(o¯t|W2D t t−K, At t−K, O¯t−1 t−K), (6) where W2D t t−K is obtained from the projection R(C t t−K, Wt t−K). Here, Pθ acts as a learned de￾ferred shader (Thies et al., 2019) which additionally predicts i… view at source ↗
Figure 5
Figure 5. Figure 5: Video frames generated over 600 timestep episodes by PERSIST (Ours), Oasis (Decart et al., 2024) and WorldMem (Xiao et al., 2025a). core components and design choices of our method. Unless otherwise specified, our base configuration and all ablations utilize the 3D-XL denoiser. • PERSIST-S employs the smaller 3D-S denoiser to mea￾sure how an 8× reduction in spatial tokens impacts the modeling capabilities … view at source ↗
Figure 6
Figure 6. Figure 6: FID scores compared to ground truth over 600 frame episodes. PERSIST configurations remains stable, while baselines relying on pixel-histories degrade rapidly. Generation stability. Qualitative analysis of sample roll￾outs ( [PITH_FULL_IMAGE:figures/full_fig_p008_6.png] view at source ↗
Figure 8
Figure 8. Figure 8: PERSIST’s 3D state enables collision modelling with out-of-view elements (top: player moves backwards into a tree at t = 50). Since the 3D state is dynamic, it evolves even when unobserved (middle: a cave filling with water). This allows off-screen events to produce emergent on-screen effects (bottom: water flowing onto the player at t = 150). provided as a starting condition4 . This explicit 3D condi￾tion… view at source ↗
Figure 9
Figure 9. Figure 9: User Study Interface. A screenshot of our web-based platform where participants evaluate video pairs based on temporal and spatial consistency. 17 [PITH_FULL_IMAGE:figures/full_fig_p017_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Score differences during head-to-head comparisons across evaluation metrics. 18 [PITH_FULL_IMAGE:figures/full_fig_p018_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Conditioning on the ground truth camera trajectory results in physical inconsistency with the world-frame, causing the agent to phase through a wall. 21 [PITH_FULL_IMAGE:figures/full_fig_p021_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: PERSIST generates diverse but coherent initial world frames from a single RGB conditioning observation. Each row corresponds to a specific input RGB frame, while each column depicts the initial world frame sampled from Wθ during a specific generation episode. 22 [PITH_FULL_IMAGE:figures/full_fig_p022_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: Generation artifacts and subsequent recoveries over a 2000-step (83 seconds) episode generated with PERSIST. The 3D representation drifts locally, causing individual blocks to appear and disappear from the agent’s view. However we find that the 3D representation remains globally coherent and has a net stabilising effect on generation, allowing Pθ to recover from visual artifacts. This lets PERSIST generat… 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 5 Pith papers

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

  1. Dream-Cubed: Controllable Generative Modeling in Minecraft by Training on Billions of Cubes

    cs.CV 2026-04 unverdicted novelty 7.0

    Dream-Cubed releases a billion-scale voxel dataset and 3D diffusion models that generate controllable Minecraft worlds by operating directly on blocks.

  2. Closing the Loop: Training-Free Revisit Consistency for Autoregressive Generative Rendering

    cs.CV 2026-07 conditional novelty 6.0

    When a camera returns to a spot it visited long ago, loading that earlier frame into the KV cache and biasing attention with depth reprojection keeps the regenerated view consistent.

  3. Echo-Memory: A Controlled Study of Memory in Action World Models

    cs.CV 2026-06 unverdicted novelty 6.0

    A controlled study finds that block-wise state-space recurrence outperforms other memory designs for open-domain scene return in action-conditioned video models, and that standard replay metrics do not adequately meas...

  4. MilliVid: Hierarchical Latents for Long-Range Consistency in Video Generation

    cs.CV 2026-06 unverdicted novelty 5.0

    MilliVid compresses video frames into multi-scale token hierarchies and uses coarse-to-fine rollout in a diffusion model to maintain long-range geometric and object consistency on Minecraft videos.

  5. Towards Interactive Video World Modeling: Frontiers, Challenges, Benchmarks, and Future Trends

    cs.CV 2026-05 unverdicted novelty 2.0

    This survey reviews trends, challenges, benchmarks, and future directions in action-conditioned interactive world modeling for video and 3D generation.

Reference graph

Works this paper leans on

7 extracted references · 1 linked inside Pith · cited by 5 Pith papers

  1. [2]

    He, X., Peng, C., Liu, Z., Wang, B., Zhang, Y ., Cui, Q., Kang, F., Jiang, B., An, M., Ren, Y ., et al

    URLhttps://arxiv.org/abs/2509. He, X., Peng, C., Liu, Z., Wang, B., Zhang, Y ., Cui, Q., Kang, F., Jiang, B., An, M., Ren, Y ., et al. Matrix-game 2.0: An open-source, real-time, and streaming interactive world model.ArXiv preprint, abs/2508.13009, 2025. URL https://arxiv.org/abs/2508.13009. Heusel, M., Ramsauer, H., Unterthiner, T., Nessler, B., and Hoch...

    Pith/arXiv arXiv 2025

  2. [22]

    doi: 10.1109/ICCV51070.2023.00008

    IEEE, 2023. doi: 10.1109/ICCV51070.2023.00008. URL https://doi.org/10.1109/ICCV51070. 2023.00008. Yu, J., Qin, Y ., Wang, X., Wan, P., Zhang, D., and Liu, X. Gamefactory: Creating new games with generative interactive videos.ArXiv preprint, abs/2501.08325, 2025. URLhttps://arxiv.org/abs/2501.08325. Zhang, Y ., Peng, C., Wang, B., Wang, P., Zhu, Q., Kang, ...

    arXiv 2023

  3. [2017]

    URL https: //doi.org/10.1109/CVPR.2017.502

    doi: 10.1109/CVPR.2017.502. URL https: //doi.org/10.1109/CVPR.2017.502. Chang, A., Dai, A., Funkhouser, T., Halber, M., Niessner, M., Savva, M., Song, S., Zeng, A., and Zhang, Y . Matter- port3d: Learning from rgb-d data in indoor environments. International Conference on 3D Vision (3DV), 2017. Che, H., He, X., Liu, Q., Jin, C., and Chen, H. Gamegen- x: I...

    arXiv 2017

  4. [2019]

    URL https: //doi.org/10.24963/ijcai.2019/339

    doi: 10.24963/ijcai.2019/339. URL https: //doi.org/10.24963/ijcai.2019/339. Ha, D. and Schmidhuber, J. World models. 2018. doi: 10.5281/ZENODO.1207631. URL https://zenodo. org/record/1207631. Hafner, D., Yan, W., and Lillicrap, T. Training agents inside of scalable world models.ArXiv preprint, abs/ 2509, 20

    arXiv 2019

  5. [2021]

    Fan, L., Wang, G., Jiang, Y ., Mandlekar, A., Yang, Y ., Zhu, H., Tang, A., Huang, D., Zhu, Y ., and Anandkumar, A

    URL https://openreview.net/forum? id=YicbFdNTTy. Fan, L., Wang, G., Jiang, Y ., Mandlekar, A., Yang, Y ., Zhu, H., Tang, A., Huang, D., Zhu, Y ., and Anandkumar, A. Minedojo: Building open-ended embodied agents with internet-scale knowledge. In Koyejo, S., Mohamed, S., Agarwal, A., Belgrave, D., Cho, K., and Oh, A. (eds.),Advances in Neural Information Pr...

    2022

  6. [2024]

    Ge, S., Mahapatra, A., Parmar, G., Zhu, J., and Huang, J

    URL https://openreview.net/forum? id=uku9r6RROl. Ge, S., Mahapatra, A., Parmar, G., Zhu, J., and Huang, J. On the content bias in fr ´echet video distance. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2024, Seattle, WA, USA, June 16- 22, 2024, pp. 7277–7288. IEEE, 2024. doi: 10.1109/ CVPR52733.2024.00695. URL https://doi.org/ 10...

    arXiv 2024

  7. [5753]

    doi: 10.1109/CVPR.2019.00589

    Computer Vision Foundation / IEEE, 2019. doi: 10.1109/CVPR.2019.00589. URL http: //openaccess.thecvf.com/content_CVPR_ 2019/html/Zhou_On_the_Continuity_of_ Rotation_Representations_in_Neural_ Networks_CVPR_2019_paper.html. 13 World Models with Persistent 3D State A. Additional Implementation Details A.1. Additional Model Details 2D V AE.We employ a vision...

    arXiv 2019