Pith. sign in

REVIEW 5 major objections 5 minor 41 references

A single 360° photograph of a room can yield a full, freely walkable 3D scene in one feed-forward pass — no multi-view capture, no per-scene optimization, no server cluster.

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

2026-08-01 11:35 UTC pith:2F46FNZR

load-bearing objection A genuinely useful two-stage pipeline that likely works as a system, but the 'metric' claim is not actually verified and the baselines are not entirely fair. the 5 major comments →

arxiv 2607.26646 v1 pith:2F46FNZR submitted 2026-07-29 cs.CV

Genie Sim PanoWorld: An Infinite Indoor 3D World Generation Pipeline via Panoramic Scene Modeling and Simulation

Yongxin Su , Linjie Hou , Feng Wang , Jialin Tang , Zhijun Li , Qian Wang , Maoqing Yao This is my paper
classification cs.CV
keywords single-panorama 3D reconstructionpanoramic video generationfeed-forward 3D Gaussian splattingtrajectory controlshortcut modelsdiffusion modelsindoor scene generationequirectangular projection
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.

The paper is trying to establish that a single 360° photograph of a room (an equirectangular panorama) is a sufficient starting point for a full, metric-scale, freely walkable 3D scene — something usually assumed to need many overlapping photos and lengthy per-scene optimization. Its route is an explicit intermediate: the pipeline plans a collision-free walking path through the room, synthesizes a panoramic video that follows that path while inventing the content hidden behind furniture and walls, and then lifts that video into a real-time renderable 3D scene built from thousands of small Gaussian blobs. Everything is feed-forward — no per-scene optimization, no multi-view capture — and the whole run takes about 147 seconds on a single consumer GPU. A sympathetic reader would care because the input format is already everywhere: real-estate listings, virtual tours, and indoor robot training all start from one panorama, and if the claim holds, each one becomes an instant simulation asset instead of the first frame of an expensive reconstruction. The paper further reports that the pipeline beats geometry-conditioned baselines on both video fidelity and downstream reconstruction quality, and generalizes zero-shot to unseen indoor scenes.

Core claim

On the paper's own terms, the discovery is that single-panorama-to-3D need not be one monolithic mapping. The pipeline bridges generation and reconstruction with an explicit, trajectory-controllable panoramic video: a latent video diffusion model hallucinates walking-camera views, conditioned on geometry-warped frames and disocclusion masks along a navigation-mesh-planned path; a feed-forward panoramic reconstructor estimates poses and depth from that video, and a voxel-aligned Gaussian decoder fuses them into a real-time renderable 3D Gaussian scene. The full run takes about 147 seconds on one consumer GPU in four guidance-free denoising steps, beating geometry-conditioned baselines and gen

What carries the argument

The load-bearing object is the explicit, trajectory-controllable panoramic video: equirectangular frames synthesized along a navigation-mesh-planned path, turning one static view into wide-baseline multi-view evidence. Two mechanisms make it practical. In generation, long–short mixed training anchors fidelity on short warps and supervises disocclusion inpainting on long clips, and a shortcut-model self-consistency objective (student matches a two-half-step teacher's velocity) removes classifier-free guidance, cutting sampling to four steps. In reconstruction, an equirectangular visual-geometry transformer estimates poses and depth, and a voxel-aligned Gaussian head fuses RGB, depth, and ray

Load-bearing premise

The load-bearing premise is that the feed-forward pose-and-depth estimator, run on the generator's synthetic frames, is accurate enough to fuse into a metric scene; the paper's own tables (7 and 8) show that estimator's trajectory error is several times higher than classical bundle adjustment on the same frames, so any systematic drift passes straight into the reconstructed geometry.

What would settle it

Run the full pipeline from one panorama of a room whose true 3D model is known, then register the output Gaussian scene to the ground-truth model at the far end of the traversal. The paper's own numbers set the threshold: its feed-forward estimator's absolute trajectory error is roughly an order of magnitude above classical bundle adjustment on the same frames, so if that offset propagates, far-side geometry should drift by more than the voxel size used for fusion — a measurable misalignment anyone could check against the true model.

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

If this is right

  • One panorama per room — the format real-estate listings and VR tours already capture — becomes enough for a space a user can actually walk through, removing the multi-camera capture burden.
  • A complete scene costs about two and a half minutes on one consumer GPU, making on-demand world generation practical for individual users and small studios rather than multi-GPU clusters.
  • Because trajectories are planned on a collision-free navigation mesh with metric poses, the output is simulation-ready: an embodied agent can practice navigating the reconstructed room along paths that respect walls and furniture.
  • The same scene can seed a longer per-scene-optimized refinement later, so users start from the fast feed-forward result and spend extra compute only when they need more fidelity.
  • Zero-shot results on unseen indoor scenes indicate the two stages transfer without retuning, a necessary property for treating the pipeline as a general indoor world generator.

Where Pith is reading between the lines

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

  • Editorial extension: the paper's own measurements suggest the bottleneck is the pose-and-depth estimator, not the video generator — its trajectory error is roughly an order of magnitude above classical bundle adjustment on identical frames; fine-tuning that estimator on generated video with known trajectories could improve reconstruction more than further video-fidelity gains.
  • Editorial extension: the two-stage recipe — an explicitly planned, metrically grounded intermediate video between a generative model and a reconstruction model — is transferable. The paper tests it only indoors; under outdoor or unbounded settings the trajectory planner, depth scale, and disocclusion regime all change, and failures would likely surface in the reconstruction stage first.
  • Editorial extension: the shortcut-model self-consistency loss via low-rank adapters on a frozen backbone looks like a cheap route to few-step, guidance-free sampling; other video-diffusion-based 3D pipelines could adopt the same trick without retraining from scratch.
  • Editorial extension: the paper's ablations show reconstruction quality drops as more context views are fused into the voxel grid, hinting that adaptive view selection before fusion is a cheap, testable improvement over naively feeding every generated frame.

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 / 5 minor

Summary. The paper proposes Genie Sim PanoWorld, a two-stage feed-forward pipeline that turns a single equirectangular (ERP) panorama into a freely roamable 3D Gaussian scene without per-scene optimization. Stage 1 plans a collision-free SE(3) trajectory on a NavMesh derived from monocular depth, then injects geometry-warped rendering and disocclusion masks into a latent rectified-flow video diffusion model (Wan2.2-TI2V-5B + LoRA). The model is trained with long–short trajectory mixed supervision and a shortcut-model self-consistency objective, enabling 4 CFG-free denoising steps. Stage 2 applies a panoramic VGGT-style estimator (PanoVGGT) to the generated video, fuses multi-view features into a voxel grid, and decodes 3D Gaussian primitives for real-time rendering. Experiments on InteriorGS and RealSee3D compare against Matrix-3D, OmniRoam, and feed-forward reconstruction baselines, with ablations of training components, decoder modules, and pose estimators.

Significance. If the central claim holds, this is a substantial advance: an end-to-end, single-panorama-to-3D pipeline with explicit metric trajectory control, no per-scene optimization, and practical single-GPU runtime (~147 s). The paper's strengths are the clear two-stage factorization, the component-wise ablations that convincingly attribute gains to mixed training and self-consistency, the NFE reduction, and the introduction of the PanoHome dataset. However, the evidence for the 'metric 3D scene' part of the claim is incomplete: pose-error and scale-alignment gaps are not isolated, and one headline baseline is evaluated at a reduced sampling schedule. These issues prevent the stated contribution from being fully established as presented.

major comments (5)
  1. [§4.2, §4.7, Tables 7–9] All pose errors are reported after Sim(3) Umeyama alignment (Sec. 4.2), which removes global scale, rotation, and translation. The 3DGS decoder consumes PanoVGGT poses via Eqs. (8)–(9), so any scale or global drift in those poses propagates directly into the reconstructed scene. Table 7 shows VGGT ATE 0.029–0.035 on generated frames versus COLMAP 0.0035–0.022, yet no experiment isolates whether this ~10× larger pose error affects the final 3DGS output. The central 'metric 3D scene' claim is therefore not verified. Please add an end-to-end ablation using COLMAP poses as input to the same decoder on generated frames, and report metric-scale errors without Sim(3) scale alignment.
  2. [§4.3–4.4, Table 2] Matrix-3D is evaluated at 10 guided steps instead of its full 50-step schedule. This is a deliberate protocol choice (Sec. 4.3), but it reduces NFE from 100 to 20 and may handicap the baseline. Since the headline claim is that the method 'outperforms Matrix-3D at every trajectory stride,' please also report Matrix-3D at its default 50-step schedule and/or at matched NFE, and justify 10 steps as the baseline's intended configuration.
  3. [Tables 2, 4, 6, 7, 9] No error bars, multiple seeds, or significance tests are reported. Ablation deltas are small (e.g., 0.80 dB novel-view PSNR in Table 6; FVD changes of ~1.4 in Table 4), so it is unclear whether the observed differences are meaningful beyond noise. Please provide at least 3 seeds with standard deviations for the headline tables and a significance test for the main comparisons.
  4. [§4.4.3, Table 3] The RealSee3D zero-shot evaluation has no ground-truth video or trajectory; pose metrics are computed against a planned ring trajectory that derives from the same input depth used for generation. This is close to self-referential for assessing trajectory controllability and says nothing about 3D reconstruction accuracy on unseen scenes. Please clarify that zero-shot generalization is limited to appearance/consistency metrics, or add a real-scene benchmark with measured 3D geometry.
  5. [§3.2.1, Abstract/Sec. 5] The NavMesh is built from monocular Depth Anything 360 depth; if that depth has an unknown global scale, the planned SE(3) trajectory is only defined up to scale, which would contradict the 'metric' claim in the abstract and Sec. 5. The paper should either state how absolute scale is recovered or normalized, or explicitly scope the metric claim to domains with known scale (e.g., the PanoHome-trained distribution).
minor comments (5)
  1. [§4.3] Typo: 'Feed-forawrd' should be 'Feed-forward'.
  2. [§2.2.1 vs §3.2.1] Related Work discusses 'Depth Anything V3', while Method uses 'Depth Anything 360'; clarify whether these refer to the same model family and which specific version is used in the pipeline.
  3. [Tables 7 and 8] Notation is inconsistent: Table 7 uses 'VGGT (ours)' and Table 8 additionally uses 'PanoVGGT'. Define whether PanoVGGT is identical to VGGT (ours) and why both names appear.
  4. [Table 2] The 'GT track (ref.)' rows leave fidelity columns blank. Mark them clearly as pose-reference only, or fill with applicable values.
  5. [Eq. (6), §3.2.2] The text says samples are 'routed with probability ρ=0.75 to the flow-matching loss', and Eq. (6) uses b∼Bernoulli(ρ). Make the Bernoulli convention explicit (e.g., b=1 with probability 0.75) to avoid ambiguity.

Circularity Check

1 steps flagged

Minor self-referential zero-shot trajectory metric; core pipeline is independently trained and benchmarked.

specific steps
  1. self definitional [Sec. 4.2 Metrics; Sec. 4.4.3 Zero-Shot Generalization (Table 3)]
    "On RealSee3D, which has no roaming-video ground truth, we drive inference with our own collision-free ring trajectories and report FID, a loop-closure score (following OmniRoam [21]), first/last-frame LPIPS-Loop, and the same pose metrics against the planned trajectory."

    On RealSee3D, 'trajectory accuracy' (ATE/RRE/RTE) is defined as the disagreement between poses re-estimated from the generated video and the planned trajectory T. But T is the SE(3) conditioning input to the generator in Eq. (2), and the geometry-warped conditioning frames are rendered from the scene mesh along exactly that T. Thus the pose metric measures how closely the video reproduces its own conditioning trajectory, not agreement with an external scene-motion ground truth. The paper additionally reports this as zero-shot pose-drift evidence ('It also substantially reduces pose drift ... verified via COLMAP on the generated frames'), so the zero-shot controllability claim is partly self-referential. This is an evaluation-loop issue, not a derivation of the 3D-scene content, and it does

full rationale

No significant central circularity. The single-panorama-to-3D claim is not derived from fitted constants or from the authors' own prior publications. The video generator is trained on InteriorGS ground-truth tracks with flow-matching and shortcut-model losses (Eqs. 4-6); the reconstructor is trained in two stages with ground-truth camera, depth, and RGB losses (Eqs. 12-13) on RealSee3D, InteriorGS, Structured3D, and the procedurally generated PanoHome, where depth and pose are rendered from the same geometry. The reconstruction decoder is deliberately isolated from the generative stage by evaluating it on real captured videos (Table 5), which is an externally grounded benchmark rather than a self-citation. The one identified circular step is the zero-shot RealSee3D trajectory metric being measured against the planned trajectory that is itself the conditioning input; this is a self-referential evaluation of pose controllability but does not by itself establish the central scene-generation claim. Separately, all pose metrics are Sim(3)-aligned, so the 'metric' claim is not directly verified by those numbers, and VGGT pose error on generated frames is not isolated end-to-end; these are correctness/validation risks, not circular reductions. Overall, the pipeline is self-contained against external data and benchmarks, so the score is low.

Axiom & Free-Parameter Ledger

6 free parameters · 4 axioms · 1 invented entities

The central claim rests on a small set of hand-selected architecture parameters (voxel size, K, LoRA rank), a strong assumption that monocular depth yields a reliable NavMesh, and an assumption that generated video is geometrically consistent enough for VGGT-style pose recovery. The PanoHome dataset is promised but not externally available.

free parameters (6)
  • Voxel cell size epsilon in voxel fusion
    Used in Eq. 8-9 for quantizing 3D points into voxels; no value is stated, and it directly controls geometric resolution of the reconstructed scene.
  • Number of Gaussians per voxel K
    The Gaussian decoder predicts K primitives per voxel (Eq. 10); K is not reported, and it controls model capacity.
  • Self-consistency loss sampling probability rho = 0.75
    Section 3.2.2 sets rho=0.75 for routing flow-matching vs. self-consistency losses in Eq. 6.
  • Video sequence length N = 81
    Trajectory length is fixed to 81 frames (Sec. 3.2.1), which bounds the extent of exploration and therefore the claim of 'infinite' world generation.
  • Number of denoising steps at inference = 4
    The CFG-free shortcut model uses 4 Euler steps; the practicality claim depends on this choice.
  • LoRA rank and alpha = rank 48, alpha 48
    LoRA adapters are rank 48 with alpha 48; no sweep is reported, so the choice is hand-selected.
axioms (4)
  • domain assumption A navigation mesh built from monocular depth is a reliable free-space prior for collision-free trajectory planning.
    Section 3.2.1 uses Depth Anything 360 to build a point cloud, mesh, and NavMesh. The paper's own Limitations section concedes that depth errors on specular, transparent, and textureless surfaces lead to suboptimal trajectories or wall penetration.
  • domain assumption The generated panoramic video is geometrically consistent enough for a feed-forward VGGT-style pose/depth estimator to recover metric poses.
    Section 3.3.1 consumes generated frames; Table 7 shows VGGT ATE 0.029-0.035 on generated frames, about 10x worse than COLMAP, yet these estimates feed the 3DGS decoder.
  • domain assumption Shortcut-model self-consistency objectives permit CFG-free 4-step sampling without a quality penalty.
    Section 3.2.2 relies on the shortcut-model framework [5]; the paper ablates the self-consistency loss but does not independently validate the teacher two-half-step scheme on the panoramic domain.
  • domain assumption Equirectangular projection eliminates the need for per-view intrinsic estimation and improves pose accuracy.
    Section 3.3.1 claims this; it is true for ideal ERP images but depends on accurate capture and a fixed projection model.
invented entities (1)
  • PanoHome dataset no independent evidence
    purpose: Procedurally generated indoor panoramic dataset with metric depth, full 6DoF poses, and editable 3D scene assets, used to train both stages.
    PanoHome is a central contribution, but no release URL, license, or identifier is provided in the paper, so external verification is not currently possible.

pith-pipeline@v1.3.0-daily-deepseek · 19174 in / 14535 out tokens · 146913 ms · 2026-08-01T11:35:21.396936+00:00 · methodology

0 comments
read the original abstract

We address the problem of reconstructing a high-fidelity, freely navigable 3D scene from a single $360^\circ$ panorama, without per-scene optimization or multi-view capture. Existing methods either lack metric trajectory control, which hinders reliable downstream 3D reconstruction, or struggle with large disocclusions under long-range camera motion while requiring high-end multi-GPU servers.We present Genie Sim PanoWorld, a two-stage feed-forward pipeline that bridges generation and reconstruction via an explicit, trajectory-controllable panoramic video. A NavMesh-planned $\mathrm{SE}(3)$ roaming trajectory is injected into a latent video diffusion model through dense geometry-warped conditioning; long--short trajectory mixed training and a self-consistency objective based on shortcut models together yield high-fidelity video in four CFG-free denoising steps. A feed-forward panoramic reconstructor then lifts the generated video into a high-fidelity 3D Gaussian scene that supports real-time, free-viewpoint roaming and can be directly used as a simulation-ready asset for embodied AI applications. Experiments show that Genie Sim PanoWorld outperforms geometry-conditioned baselines in both panoramic video generation and downstream 3D reconstruction, while generalizing zero-shot to unseen indoor scenes.

Figures

Figures reproduced from arXiv: 2607.26646 by Feng Wang, Jialin Tang, Linjie Hou, Maoqing Yao, Qian Wang, Yongxin Su, Zhijun Li.

Figure 1
Figure 1. Figure 1: We propose Genie Sim PanoWorld, which requires only [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overview of Genie Sim PanoWorld. Stage 1 (Environment Perception & Trajectory Construction): A raw ERP panorama is processed by a depth estimator to build a dense point cloud and 3D mesh; a NavMesh is extracted from the traversable floor region and used to plan a collision-free SE(3) camera trajectory. Stage 2 (Latent Video Diffusion): For each planned pose, a geometry-warped ERP frame and its disocclusion… view at source ↗
Figure 3
Figure 3. Figure 3: Qualitative comparison on a long-range ( [PITH_FULL_IMAGE:figures/full_fig_p010_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Qualitative comparison at 480 × 960 (r=1) including OmniRoam. OmniRoam produces blurry, trajectory-inconsistent frames that diverge progressively from the intended path, while our model remains sharp and geometrically faithful. GT shown for reference; additional scenes in the supplement. outperforming Matrix-3D on all three metrics. It also sub￾stantially reduces pose drift: relative to Matrix-3D it lowers… view at source ↗
Figure 5
Figure 5. Figure 5: Zero-shot qualitative comparison on RealSee3D (no ground truth available). OmniRoam exhibits trajectory-inconsistent wall [PITH_FULL_IMAGE:figures/full_fig_p011_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Qualitative comparison of novel view synthesis on the InteriorGS dataset (60 scenes). We evaluate all methods under the 4@4 [PITH_FULL_IMAGE:figures/full_fig_p012_6.png] view at source ↗

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

41 extracted references · 3 linked inside Pith

  1. [1]

    Pano3D: A holistic benchmark and a solid baseline for 360 ◦ depth estimation

    Georgios Albanis, Nikolaos Zioulis, Petros Drakoulis, Vasileios Gkitsas, Vladimiros Sterzentsenko, Fed- erico Alvarez, Dimitrios Zarpalas, and Petros Daras. Pano3D: A holistic benchmark and a solid baseline for 360 ◦ depth estimation. InProceedings of the IEEE/CVF Conference on Computer Vision and Pat- tern Recognition, pages 3727–3737, 2021. 7, 8

    2021

  2. [2]

    Joint 2D-3D-Semantic data for indoor scene understanding.arXiv preprint arXiv:1702.01105,

    Iro Armeni, Sasha Sax, Amir R Zamir, and Silvio Savarese. Joint 2D-3D-Semantic data for indoor scene understanding.arXiv preprint arXiv:1702.01105,

    Pith/arXiv arXiv

  3. [3]

    Matterport3D: Learning from RGB-D data in indoor environments

    Angel Chang, Angela Dai, Thomas Funkhouser, Ma- ciej Halber, Matthias Niessner, Manolis Savva, Shuran Song, Andy Zeng, and Yinda Zhang. Matterport3D: Learning from RGB-D data in indoor environments. arXiv preprint arXiv:1709.06158, 2017. 7, 8

    Pith/arXiv arXiv 2017

  4. [4]

    ViewPoint: Panoramic video generation with pretrained diffusion models

    Zixun Fang, Kai Zhu, Zhiheng Liu, Yu Liu, Wei Zhai, Yang Cao, and Zheng-Jun Zha. ViewPoint: Panoramic video generation with pretrained diffusion models. In Advances in Neural Information Processing Systems, pages 12486–12508. Curran Associates, Inc., 2025. 2

    2025

  5. [5]

    One step diffusion via shortcut models

    Kevin Frans, Danijar Hafner, Sergey Levine, and Pieter Abbeel. One step diffusion via shortcut models. InInternational Conference on Learning Representa- tions (ICLR), 2025. 2, 5

    2025

  6. [6]

    Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen

    Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. LoRA: Low-Rank Adaptation of Large Language Models. InInternational Conference on Learning Representations (ICLR), 2022. 4, 8

    2022

  7. [7]

    360Loc: A dataset and benchmark for omnidirectional visual localization with cross-device queries

    Huajian Huang, Changkun Liu, Yipeng Zhu, Hui Cheng, Tristan Braud, and Sai-Kit Yeung. 360Loc: A dataset and benchmark for omnidirectional visual localization with cross-device queries. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22314–22324, 2024. 7, 8

    2024

  8. [8]

    V oyager: Long-Range and World-Consistent Video Diffusion for Explorable 3D Scene Generation.ACM Transactions on Graphics, 44(6):1–15, 2025

    Tianyu Huang, Wangguandong Zheng, Tengfei Wang, Yuhao Liu, Zhenwei Wang, Junta Wu, Jie Jiang, Hui Li, Rynson Lau, Wangmeng Zuo, and Chunchao Guo. V oyager: Long-Range and World-Consistent Video Diffusion for Explorable 3D Scene Generation.ACM Transactions on Graphics, 44(6):1–15, 2025. 2

    2025

  9. [9]

    HY-World 2.0: A Multi-Modal World Model for Re- constructing, Generating, and Simulating 3D Worlds,

    Team HY-World, Chenjie Cao, Xuhui Zuo, Zhenwei Wang, Yisu Zhang, Junta Wu, Zhenyang Liu, Yuning Gong, Yang Liu, Bo Yuan, Chao Zhang, Coopers Li, Dongyuan Guo, Fan Yang, Haiyu Zhang, Hang Cao, Jianchen Zhu, Jiaxin Lin, Jie Xiao, Jihong Zhang, Jun- lin Yu, Lei Wang, Lifu Wang, Lilin Wang, Linus, Minghui Chen, Peng He, Penghao Zhao, Qi Chen, Rui Chen, Rui Sh...

  10. [10]

    PanoWorld: A Generative Spatial World Model for Consistent Whole-House Panorama Synthesis, 2026

    Jinrang Jia, Zhenjia Li, Yijiang Hu, and Yifeng Shi. PanoWorld: A Generative Spatial World Model for Consistent Whole-House Panorama Synthesis, 2026. 2, 10

    2026

  11. [11]

    Depth Anything in 360◦: Towards scale invariance in the wild, 2025

    Hualie Jiang, Ziyang Song, Zhiqiang Lou, Rui Xu, and Minglang Tan. Depth Anything in 360◦: Towards scale invariance in the wild, 2025. 4

    2025

  12. [12]

    PanoWorld: Geometry-Consistent Panoramic Video World Modeling, 2026

    Le Jiang, Xiangyu Bai, Bishoy Galoaa, Shayda Moezzi, Caleb James Lee, Tooba Imtiaz, Edmund Yeh, Jennifer Dy, Yanzhi Wang, and Sarah Ostad- abbas. PanoWorld: Geometry-Consistent Panoramic Video World Modeling, 2026. 2

    2026

  13. [13]

    3D Gaussian Splatting for Real-Time Radiance Field Rendering

    Bernhard Kerbl, Georgios Kopanas, Thomas Leimk¨uhler, and George Drettakis. 3D Gaussian Splatting for Real-Time Radiance Field Rendering. ACM Transactions on Graphics (TOG), 42(4), 2023. 1

    2023

  14. [14]

    RealSee3D: 14 A Large-Scale Multi-View RGB-D Dataset of Indoor Scenes (Version 1.0), 2025

    Linyuan Li, Yan Wu, Xi Li, Lingli Wang, Tong Rao, Jie Zhou, Cihui Pan, and Xinchen Hui. RealSee3D: 14 A Large-Scale Multi-View RGB-D Dataset of Indoor Scenes (Version 1.0), 2025. 7, 8

    2025

  15. [15]

    CubeComposer: Spatio-temporal autoregres- sive 4K 360◦ video generation from perspective video

    Lingen Li, Guangzhi Wang, Xiaoyu Li, Zhaoyang Zhang, Qi Dou, Jinwei Gu, Tianfan Xue, and Ying Shan. CubeComposer: Spatio-temporal autoregres- sive 4K 360◦ video generation from perspective video. InProceedings of the IEEE/CVF Conference on Com- puter Vision and Pattern Recognition (CVPR), pages 32625–32635, 2026. 2

    2026

  16. [16]

    4K4DGen: Panoramic 4D generation at 4K resolution

    Renjie Li, Panwang Pan, Bangbang Yang, Dejia Xu, Shijie Zhou, zhang xuanyang, Zeming Li, Achuta Kadambi, Zhangyang Wang, Zhengzhong Tu, and Zhiwen Fan. 4K4DGen: Panoramic 4D generation at 4K resolution. InInternational Conference on Learn- ing Representations, pages 100692–100713, 2025. 2

    2025

  17. [17]

    Chen, Zhenyu Li, Guang Shi, Jiashi Feng, and Bingyi Kang

    Haotong Lin, Sili Chen, Jun Hao Liew, Donny Y . Chen, Zhenyu Li, Guang Shi, Jiashi Feng, and Bingyi Kang. Depth Anything 3: Recovering the visual space from any views.arXiv preprint arXiv:2511.10647,

    Pith/arXiv arXiv

  18. [18]

    Yaron Lipman, Ricky T. Q. Chen, Heli Ben-Hamu, Maximilian Nickel, and Matt Le. Flow matching for generative modeling. InInternational Conference on Learning Representations (ICLR), 2023. 5

    2023

  19. [19]

    Aoming Liu, Zhong Li, Zhang Chen, Nannan Li, Yi Xu, and Bryan A. Plummer. PanoFree: Tuning-Free Holistic Multi-view Image Generation with Cross- View Self-guidance. InComputer Vision – ECCV 2024, pages 146–164. Springer Nature Switzerland, Cham, 2025. 2

    2024

  20. [20]

    DynamicScaler: Seamless and scalable video generation for panoramic scenes

    Jinxiu Liu, Shaoheng Lin, Yinxiao Li, and Ming- Hsuan Yang. DynamicScaler: Seamless and scalable video generation for panoramic scenes. InProceed- ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6144–6153,

  21. [21]

    OmniRoam: World Wandering via Long- Horizon Panoramic Video Generation, 2026

    Yuheng Liu, Xin Lin, Xinke Li, Baihan Yang, Chen Wang, Kalyan Sunkavalli, Yannick Hold-Geoffroy, Hao Tan, Kai Zhang, Xiaohui Xie, Zifan Shi, and Yi- wei Hu. OmniRoam: World Wandering via Long- Horizon Panoramic Video Generation, 2026. 2, 3, 4, 8, 9, 10

    2026

  22. [22]

    GenEx: Generating an explorable world

    TaiMing Lu, Tianmin Shu, Alan Yuille, Daniel Khashabi, and Jieneng Chen. GenEx: Generating an explorable world. InInternational Conference on Learning Representations, pages 52310–52335, 2025. 2

    2025

  23. [23]

    Richter, and Vladlen Koltun

    Lars Mescheder, Wei Dong, Shiwei Li, Xuyang Bai, Marcel Santos, Peiyun Hu, Bruno Lecouat, Ming- min Zhen, Ama¨el Delaunoy, Tian Fang, Yanghai Tsin, Stephan R. Richter, and Vladlen Koltun. SHARP: Sharp monocular view synthesis in less than a second. arXiv preprint arXiv:2512.10685, 2025. 3

    arXiv 2025

  24. [24]

    Towards phys- ically executable 3D Gaussian for embodied naviga- tion, 2025

    Bingchen Miao, Rong Wei, Zhiqi Ge, Xiaoquan Sun, Shiqi Gao, Jingzhe Zhu, Renhan Wang, Siliang Tang, Jun Xiao, Rui Tang, and Juncheng Li. Towards phys- ically executable 3D Gaussian for embodied naviga- tion, 2025. 7, 8

    2025

  25. [25]

    Srinivasan, Matthew Tan- cik, Jonathan T

    Ben Mildenhall, Pratul P. Srinivasan, Matthew Tan- cik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. NeRF: Representing Scenes as Neural Radiance Fields for View Synthesis. InEuropean Conference on Computer Vision (ECCV), pages 405–421, 2020. 1

    2020

  26. [26]

    Infinite photorealistic worlds using procedu- ral generation

    Alexander Raistrick, Lahav Lipson, Zeyu Ma, Lingjie Mei, Mingzhe Wang, Yiming Zuo, Karhan Kayan, Hongyu Wen, Beining Han, Yihan Wang, Alejandro Newell, Hei Law, Ankit Goyal, Kaiyu Yang, and Jia Deng. Infinite photorealistic worlds using procedu- ral generation. InProceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition, pages 12...

    2023

  27. [27]

    Infinigen Indoors: Photorealistic indoor scenes using procedural gener- ation

    Alexander Raistrick, Lingjie Mei, Karhan Kayan, David Yan, Yiming Zuo, Beining Han, Hongyu Wen, Meenal Parakh, Stamatis Alexandropoulos, Lahav Lipson, Zeyu Ma, and Jia Deng. Infinigen Indoors: Photorealistic indoor scenes using procedural gener- ation. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 21783–...

    2024

  28. [28]

    Lyra 2.0: Explorable Generative 3D Worlds, 2026

    Tianchang Shen, Sherwin Bahmani, Kai He, Sangeetha Grama Srinivasan, Tianshi Cao, Jiawei Ren, Ruilong Li, Zian Wang, Nicholas Sharp, Zan Gojcic, Sanja Fidler, Jiahui Huang, Huan Ling, Jun Gao, and Xuanchi Ren. Lyra 2.0: Explorable Generative 3D Worlds, 2026. 2, 3

    2026

  29. [29]

    InteriorGS: A 3D Gaussian Splatting Dataset of Semantically Labeled Indoor Scenes

    SpatialVerse Research Team, Manycore Tech Inc. InteriorGS: A 3D Gaussian Splatting Dataset of Semantically Labeled Indoor Scenes. https : / / huggingface . co / datasets / spatialverse/InteriorGS, 2025. 7

    2025

  30. [30]

    Wan: Open and Advanced Large-Scale Video Generative Models, 2025

    Wan Team. Wan: Open and Advanced Large-Scale Video Generative Models, 2025. 4, 8

    2025

  31. [31]

    PanoWorld: Towards Spatial Supersensing in360 ◦ Panorama World, 2026

    Changpeng Wang, Xin Lin, Junhan Liu, Yuheng Liu, Zhen Wang, Donglian Qi, Yunfeng Yan, and Xi Chen. PanoWorld: Towards Spatial Supersensing in360 ◦ Panorama World, 2026. 2

    2026

  32. [32]

    Self-supervised learning of depth and camera motion from 360 ◦ videos

    Fu-En Wang, Hou-Ning Hu, Hsien-Tzu Cheng, Juan- Ting Lin, Shang-Ta Yang, Meng-Li Shih, Hung-Kuo Chu, and Min Sun. Self-supervised learning of depth and camera motion from 360 ◦ videos. InAsian Con- ference on Computer Vision, pages 53–68. Springer,

  33. [33]

    VGGT: Visual Geometry Grounded Trans- former, 2025

    Jianyuan Wang, Minghao Chen, Nikita Karaev, Andrea Vedaldi, Christian Rupprecht, and David Novotny. VGGT: Visual Geometry Grounded Trans- former, 2025. 3, 7, 8 15

    2025

  34. [34]

    VGGT-Ω, 2026

    Jianyuan Wang, Minghao Chen, Shangzhan Zhang, Nikita Karaev, Johannes Sch ¨onberger, Patrick La- batut, Piotr Bojanowski, David Novotny, Andrea Vedaldi, and Christian Rupprecht. VGGT-Ω, 2026. 3

    2026

  35. [35]

    Real-ESRGAN: Training Real-World Blind Super-Resolution with Pure Synthetic Data

    Xintao Wang, Liangbin Xie, Chao Dong, and Ying Shan. Real-ESRGAN: Training Real-World Blind Super-Resolution with Pure Synthetic Data. InInter- national Conference on Computer Vision Workshops (ICCVW), 2021. 5

    2021

  36. [36]

    PanoWan: Lifting diffusion video gen- eration models to 360 ◦ with latitude/longitude-aware mechanisms

    Yifei Xia, Shuchen Weng, Siqi Yang, Jingqi Liu, Chengxuan Zhu, Minggui Teng, Zijian Jia, Han Jiang, and Boxin Shi. PanoWan: Lifting diffusion video gen- eration models to 360 ◦ with latitude/longitude-aware mechanisms. InAdvances in Neural Information Pro- cessing Systems, pages 148027–148046. Curran Asso- ciates, Inc., 2025. 2

    2025

  37. [37]

    Matrix-3D: Omnidirec- tional Explorable 3D World Generation, 2025

    Zhongqi Yang, Wenhang Ge, Yuqi Li, Jiaqi Chen, Haoyuan Li, Mengyin An, Fei Kang, Hua Xue, Baixin Xu, Yuyang Yin, Eric Li, Yang Liu, Yikai Wang, Hao- Xiang Guo, and Yahui Zhou. Matrix-3D: Omnidirec- tional Explorable 3D World Generation, 2025. 2, 4, 8

    2025

  38. [38]

    PanoWorld-X: Generating Explorable Panoramic Worlds via Sphere- Aware Video Diffusion, 2025

    Yuyang Yin, HaoXiang Guo, Fangfu Liu, Mengyu Wang, Hanwen Liang, Eric Li, Yikai Wang, Xiao- jie Jin, Yao Zhao, and Yunchao Wei. PanoWorld-X: Generating Explorable Panoramic Worlds via Sphere- Aware Video Diffusion, 2025. 2

    2025

  39. [39]

    Chen, Qianyi Wu, Konstantinos N

    Cheng Zhang, Hanwen Liang, Donny Y . Chen, Qianyi Wu, Konstantinos N. Plataniotis, Camilo Cruz Gam- bardella, and Jianfei Cai. PanFlow: Decoupled Mo- tion Control for Panoramic Video Generation.Pro- ceedings of the AAAI Conference on Artificial Intelli- gence, 40(15):12385–12393, 2026. 2

    2026

  40. [40]

    World- Stereo: Bridging camera-guided video generation and scene reconstruction via 3D geometric memories

    Yisu Zhang, Chenjie Cao, Tengfei Wang, Xuhui Zuo, Junta Wu, Jianke Zhu, and Chunchao Guo. World- Stereo: Bridging camera-guided video generation and scene reconstruction via 3D geometric memories. In Proceedings of the IEEE/CVF Conference on Com- puter Vision and Pattern Recognition (CVPR), pages 40327–40339, 2026. 2, 3

    2026

  41. [41]

    Structured3D: A large photo- realistic dataset for structured 3D modeling

    Jia Zheng, Junfei Zhang, Jing Li, Rui Tang, Shenghua Gao, and Zihan Zhou. Structured3D: A large photo- realistic dataset for structured 3D modeling. InEuro- pean Conference on Computer Vision, pages 519–535. Springer, 2020. 7, 8 16

    2020