acceptodds
Under review as a conference paper at ICLR 2027

Panode: World Exploration via Incremental Panoramic Node Generation

Abstract

Generating explorable worlds from a single image is a foundational capability for interactive content creation. Existing methods mostly expand the world through perspective video generation, leading to error accumulation over long-range exploration. To address this, we present **Panode**, which recasts world exploration as sparse panoramic node generation. Instead of densely sampling frames, we build up the world from a sequence of panoramic nodes, each producing a full 360° observation. At each node, a generative model jointly completes the panoramic RGB and depth, yielding a geometry-consistent omnidirectional observation. The new RGB-D panorama is then lifted into a 3D Gaussian scene and refined via sliding-window optimization over recent nodes. For efficient exploration, we further design an active trajectory planner that chooses the next node by balancing exploration value against motion feasibility. By tightly coupling generation, reconstruction, and planning, Panode progressively expands a single image into a coherent, explorable world. Experiments demonstrate that Panode achieves strong generation quality and exploration efficiency.

Primary area:
generative models
Keywords:
3D generation, 3dgs, panorama

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_65YsOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

In this paper, the authors propose Panode, which incrementally generates a explorable 3D scenes from a initial panorama image through generating sparse panoramic node. Panoda first generate 3D Gaussians based on initial panorama image. Then, it generate a full 360 degree RGB-D image from the uncompleted rendering. New Gaussians are reconstructed based on the new generated observation and it's nearby graph node. Finally, a next step planner choose the next node for exploring. The method is combined of 3D module 1) RGB-D completion mode. 2) A 3DGS update module. 3) An trajectory planning module. In their experiments, the authors try to validate their method through RGB-D completion and world exploration tasks. The proposed method shows improvements over baselines on both RGB-D completion and world exploration tasks.

Critical strengths

The paper proposed an novel and inspired direction for 3D word generation and exploration from sparse panorama images. The whole pipeline is clean and well designed. Their results on RGB-D completion and world exploration tasks are promising. The proposed idea have potential contribution and influence to the related research area.

Critical weaknesses

The claim are well supported by the evaluation. The quantitative result are still using short trajectories from indoor scenes only. This makes the performance for long range exploration and outdoor scenes still questionable. The metrics of evaluation are 2D based (PSNR, SSIM, LPIPS) which make the accuracy of 3D reconstruction questionable as well.

Questions for authors
  1. How sensitive is Panode to the initial predicted depth?
  2. How does the method use the historical RGB-D images during training?
  3. For ablation study of adjacent-aware module in table 3, could you provide result of more general result instead only one scene?
6 comments

We highly appreciate all the reviews. We carefully read your valuable comments and find them very helpful for improving our work. The manuscript is revised based on your suggestions.

Our detailed item-by-item replies to the comments are given below.

Q1. How does Panode perform outdoors and over longer trajectories?

We use InteriorGS initially because it provides multiple, high-quality 3DGS of different scenes that can render multi-trajectory panoramic views and depth maps for quantitative evaluation. To broaden the evaluation, we collect 7 large-scale real scenes (5 outdoor and 2 indoor) from the public OverMaps360-50[1] resource and use their 3DGS scenes for evaluation. We report the long-trajectory novel-view results on these scenes in the table below :

MethodPSNR ↑SSIM ↑LPIPS ↓
OmniRoam15.0470.6020.504
Matrix-3D16.0720.6330.507
Panode (Ours)19.3850.7010.253

Figure 7 in the revised manuscript shows novel-view renderings of the Gaussian scenes. The results show that Panode still produces fidelity renderings at viewpoints far from the initial on outdoor scenes.

Although PSNR, SSIM, and LPIPS are image-based, they are widely used to evaluate generated 3D scenes because accurate renderings at unseen viewpoints require coherent appearance and geometry across views. We compare views rendered from each generated world at held-out viewpoints with the corresponding ground-truth images.

Q2. How sensitive is Panode to the initial depth estimate?

We thank the reviewer for this important suggestion. We agree that the initial depth prediction is important for our method and requires a more detailed sensitivity analysis. We evaluate 20 InteriorGS scenes with two trajectories per scene under the same 8-node setting. We test three initial depth sources: ground-truth depth, predicted depth[2], and predicted depth with 10% noise. The metrics are averages over the full trajectory.

Initial depthInitial AbsRel ↓PSNR ↑LPIPS ↓AbsRel ↓
Ground truth0.0021.310.1280.279
Predicted0.2120.140.2150.330
Predicted + 10% noise0.3118.180.3550.341

The results show that predicted depth can already support effective scene generation. Even with added noise, the subsequent filtering and completion process can still reduce part of the error. Nevertheless, we acknowledge that more accurate initial depth provides a more reliable prior and improves generation quality.

Q3. How does the method use the historical RGB-D images during training?

We apologize for the confusion. In our experiments, historical RGB-D context improves completion even when context tokens are only appended at inference. We suspect that the fine-tuning process preserves the base model[3]'s ability to process multi-image conditional inputs.

Nevertheless, to further reduce the gap between training and inference, after fine-tuning without historical context on the joint DAP, Matterport3D360, and Omni360-Scene datasets, we perform an additional 3,000 context-conditioned training steps on Matterport3D360. It's because it is difficult to obtain matching multi-view panoramas from the other two datasets. Specifically, using camera poses, we sample neighboring frames with overlapping fields of view to form image pairs. One serves as historical context, while the other uses a random mask and uses its complete RGB-D observation as the target. As in inference, we append the VAE-encoded historical context to the DiT input sequence. During training, we add noise and compute loss only on the target-frame latents, while the historical-context latents remain clean and do not receive supervision. As the experiment section describes, this stage only updates the DiT LoRA parameters and the depth-Control branch.

All results in both the original manuscript and the revised manuscript use the model from the above training procedure. These training details are added to the appendix.

Q4. Does the adjacency-aware module generalize beyond one scene?

Thank you for this suggestion. We agree that a single-scene ablation is insufficient to establish the general effectiveness of the adjacency-aware module. We extend this ablation to 20 InteriorGS scenes, with two trajectories per scene and the same 8-node setting, and also update Table 5 in the revised manuscript.

SettingPSNR ↑LPIPS ↓Holes ↓AbsRel ↓
w/o the graph18.1420.2340.1810.333
w/o the adjacency loss18.0870.2930.1780.335
Full model20.140.2150.0760.330

The results show that the full model performs best overall. The graph structure is designed to maintain spatial relationships among nodes, and removing it weakens cross-node geometric consistency. The adjacency loss further constrains local alignment between neighboring nodes and helps reduce discontinuities.

We thank the reviewer again and hope our responses address all your questions and concerns.

[1]OverTheReality. OverMaps360-50 Dataset. Hugging Face Datasets, 2026. [2]Vukasin Bozic, Isidora Slavkovic, Dominik Narnhofer, Nando Metzger, Denis Rozumny, Konrad Schindler, and Nikolai Kalischek. Unified Panoramic Geometry Estimation via Multi-View Foundation Models. Advances in Neural Information Processing Systems (NeurIPS), 2026 [3]Team HY-World. HY-World 2.0: A Multi-Modal World Model for Reconstructing, Generating, and Simulating 3D Worlds. arXiv:2604.14268, 2026.

Dear Reviewer 65Ys,

Thank you again for your time and thoughtful feedback. We would like to follow up on our responses to your concerns. We would be grateful to hear whether the clarifications and results in our rebuttal address your main concerns. If anything remains unclear or warrants further discussion, we would be happy to elaborate.

Thank you for your time and consideration.

Best regards, The Authors

Reviewer_WEmpOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents Panode, a framework for incrementally exploring and generating a 3D world from a single RGB panorama. It models world exploration as the progressive generation of sparse panoramic nodes, which are represented in RGB-D format. The method first initializes a 3D Gaussian using estimated depth. At each node, a joint RGB-D completion model then generates a 360° observation, which is integrated into a shared 3D representation through incremental optimization constrained by an adjacency graph. An active trajectory planner selects the next location to explore. Experiments show that Panode achieves better generation quality and runtime efficiency than existing methods on panoramic RGB-D completion and world exploration tasks.

Critical strengths
  1. The overall design is well motivated. Panode implements 3D scene generation from a single panorama into three components: active trajectory planning, RGB-D panorama completion for new viewpoints, and incremental Gaussian updates. Together, these components enable high-quality reconstruction with good efficiency.

  2. The technical choices within each component are interesting and intuitive . For example, RGB and depth are concatenated as image-like inputs so that they share context during generation. The trajectory planner also considers multiple factors, including completion gain and geometric context around missing regions.

Critical weaknesses
  1. Some of the paper’s claims are overstated. The introduction states, “Our key insight is that world exploration can be recast as sparse panoramic node generation,” and emphasizes its advantages over dense video frames. However, sparse-view 3D scene reconstruction has already been explored extensively. For example, RGBD2 [1] also reconstructs 3D scenes from sparse RGB-D views. The main differences are that RGBD2 uses perspective views and meshes, whereas this work uses panoramas and 3D Gaussian. I therefore do not find the overall framework sufficiently novel.

  2. The ablation study for active trajectory planning is too limited. The proposed method is compared only with random selection and a fixed forward-moving strategy. There is no ablation of the individual factors used by the planner, making it difficult to assess whether each component is necessary.

[1] Lei, Jiabao, Jiapeng Tang, and Kui Jia. "Rgbd2: Generative scene synthesis via incremental view inpainting using rgbd diffusion models." 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, 2023.

Questions for authors

With active trajectory planning, how many distinct viewpoints are required on average to reconstruct a scene? Can the current planning strategy guarantee full scene coverage? Is it possible that regions without sufficient geometric context are never selected for exploration?

6 comments

We highly appreciate all the reviews. We carefully read your valuable comments and find them very helpful for improving our work. The manuscript is revised based on your suggestions.

Our detailed item-by-item replies to the comments are given below.

Q1. How does Panode differ from sparse-view reconstruction?

We apologize for the confusion from our presentation of the contribution. We agree that sparse-view 3D reconstruction is an important and mature problem. Please allow us to clarify how Panode differs in both task and technical design.

Specifically, sparse-view reconstruction methods such as RGBD2 typically take multiple sparse RGB-D observations as input and aim to reconstruct the observation region and synthesize intermediate novel views, thereby obtaining a more complete 3D representation. In contrast, Panode starts from a single input image, progressively generates new panoramic nodes, and extends the 3D scene online. Its goal is not merely to reconstruct known regions, but to generate and complete unknown content and incrementally integrate it into a renderable 3DGS scene.

Therefore, the key distinction lies in reconstruction versus generative exploration. Sparse-view reconstruction mainly recovers a complete scene from existing multi-view observations, whereas Panode actively expands unknown regions from a single image. Moreover, Panode can continuously generate nodes according to user instructions, providing stronger interactive potential.

At the technical level, there are indeed methods such as the baseline GenRC that can complete RGB-D content. However, they rely on restrictive indoor scenes and show lower quality. Panode instead targets more general and higher-quality panoramic RGB-D completion for online scene expansion.

Q2. What does each planning component contribute?

We thank the reviewer for this important suggestion. We agree that comparisons with random selection and fixed forward motion evaluate the planner as a whole, but do not isolate the contribution of each factor. We therefore evaluate GG, CC, and RR separately on 20 InteriorGS scenes under the same 10-node budget, measuring newly visible area and path length.

PlannerNewly visible area (m²) ↑Path length (m) ↓
Full30.823.6
w/o GG21.426.9
w/o CC35.937.4
w/o RR17.830.2

Here, GG rewards supported newly visible regions, CC favors nearby reusable context, and RR discourages retraversal. The full planner covers large new area with the shortest path. Removing GG or RR substantially reduces coverage, while removing CC markedly lengthens the path and degrades generation quality by weakening the context available for completion.

Appendix Figure 14 in the revised manuscript further shows how explored area grows as active planning adds viewpoints in indoor and outdoor scenes.

Q3. How many viewpoints are needed, and can the planner cover the entire scene?

Thank you for your questions about the active planner. We agree that its coverage requires a more detailed quantitative evaluation. The viewpoint budget depends on the desired exploration range and level of detail. To quantify exploration ability, we evaluate cumulative explored area under different numbers of panoramic viewpoints in 3 InteriorGS indoor scenes and 3 OverMaps360 outdoor scenes, as shown in Figure 14 of the revised manuscript.

For typical indoor scenes, 10-20 nodes can provide relatively sufficient exploration. For outdoor scenes, the required number of nodes depends on scene scale and navigation settings.

The planner rewards exploration along the boundary between known and unknown space. As new observations provide more geometric context, a region with insufficient support can also become a candidate. Therefore, regions without sufficient geometric context are not permanently outside the exploration process. Coverage depends on reachability, scene scale, and viewpoint budget; especially for outdoor scenes, there is no fixed notion of exhaustive coverage.

We thank the reviewer again and hope our responses address your questions and concerns.

Dear Reviewer WEmp,

Thank you again for your time and thoughtful feedback. We would like to follow up on our responses to your concerns. We would be grateful to hear whether the clarifications and results in our rebuttal address your main concerns. If anything remains unclear or warrants further discussion, we would be happy to elaborate.

Thank you for your time and consideration.

Best regards, The Authors

Thank you for the response. My concerns about the ablation for active trajectory planning, the number of viewpoints, and exploration coverage have been largely addressed.

However, the response remains insufficient regarding the claims about the core framework and its distinction from RGBD2. The rebuttal primarily emphasizes the distinction between reconstruction and generation. In my view, however, the two methods share the same underlying problem formulation: at a new viewpoint, both use existing information to generatively complete unobserved regions. Indeed, the initial panorama in Panode may provide more extensive scene information than sparse perspective views. What, then, is the fundamental distinction captured by “reconstruction versus generative exploration” that the authors emphasize?

Reviewer_nfysOpenReview ↗
Rating
3
Confidence
2
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents Panode, a framework for incremental 3D world generation from a single RGB panorama. Rather than generating dense video sequences, it expands the scene through sparse panoramic nodes. The method combines joint RGB-D completion, adjacency-aware Gaussian optimization, and active trajectory planning to build an explorable 3D scene. Experiments compare its generation quality and efficiency against existing methods.

Critical strengths

1.Using sparse panoramic nodes is a sensible alternative to dense video generation, where consecutive frames contain substantial overlap. By integrating each generated RGB-D panorama into a persistent 3D Gaussian representation, the method allows the scene to be expanded and rendered from new viewpoints without repeatedly generating overlapping frames. This is a useful design choice for incremental world generation. 2.Table 2 shows that Panode achieves better image quality with substantially shorter runtime than the compared baselines. Table 1 also reports strong completion results across different missing ratios, while the adjacency ablation in Table 3 shows improved surface coverage. These results suggest a favorable trade-off between generation quality and efficiency.

Critical weaknesses

1.Limited evaluation of long-horizon 3D consistency. Table 2 evaluates image quality at six fixed positions along each reference trajectory. It is unclear whether these metrics are computed from generated panoramas or the final 3DGS renderings. More importantly, evaluation at these positions provides limited evidence of geometric consistency at unseen viewpoints or after many incremental updates. Table 3 includes a held-out-view ablation on a single scene, and Figure 12 shows qualitative results under larger viewpoint offsets, but neither provides a systematic analysis of accumulated geometric errors. Given that the goal is to construct a coherent, explorable 3D world, this is a significant gap in the evaluation. 2.There is no controlled study of node number or spacing, although sparse panoramic nodes are the main departure from dense video generation. The main comparison uses fixed trajectories, so it does not evaluate active planning; the planning ablation in Figure 8 covers only two generation calls. The joint RGB-D model also lacks controlled ablations of joint prediction and depth control. Although the paper reports generation-only runtimes, it remains unclear whether Panode's timing includes incremental 3DGS optimization and which stages are counted for the baselines. These missing comparisons make it difficult to identify which design choices account for the reported gains.

Questions for authors

1.Are the metrics in Table 2 computed from generated panoramas or final 3DGS renderings? Can the authors evaluate held-out viewpoints and quantify geometric drift over longer trajectories? 2.How do node number and spacing affect reconstruction quality, consistency, and runtime? How does active planning perform over longer exploration horizons? 3.What are the individual contributions of joint RGB-D prediction, depth control, and historical conditioning? Can the authors provide controlled ablations with metrics on the missing regions? 4.Which stages are included in the Table 2 runtimes, particularly Panode's per-node 3DGS optimization? Was Panode timed with 4 or 40 ODE steps, and what inference hardware was used?

5 comments
AuthorsOpenReview ↗

Additional Evaluation and Efficiency Analysis

We highly appreciate all the reviews. We carefully read your valuable comments and find them very helpful for improving our work.

Our detailed item-by-item replies to the comments are given below.

Q1. What does Table 2 measure, and how does quality change for held-out viewpoints along longer trajectories?

We thank the reviewer for this helpful question. We agree that Table 2 is not clear enough, and we add a novel-view 3DGS rendering comparison.

First, we clarify that Table 2 is intended to compare generation quality across methods, so the metrics are computed on the generated panoramas for all methods. Meanwhile, the results shown in the supplement (Figures 12 and 15) are 3DGS renderings at the generated node locations.

We fully agree with the reviewer's concern about reconstruction quality at held-out viewpoints. To evaluate novel-view quality, we add an experiment of 3DGS renderings from held-out trajectories. Specifically, all methods generate panoramas along the same trajectory and reconstruct a 3DGS scene. We then render each reconstructed scene from a novel trajectory. Panode directly obtains a 3DGS scene through incremental per-node optimization. For each baseline, we split its generated panoramas into perspective views and train a corresponding 3DGS scene. To further validate performance on long trajectories and real-world scenes, we use 7 large-scale real scenes (5 outdoor and 2 indoor) from the public OverMaps360[1] resource for evaluation. The table below reports the quantitative results.

MethodPSNR ↑SSIM ↑LPIPS ↓
OmniRoam15.0470.6020.504
Matrix-3D16.0720.6330.507
Panode (Ours)19.3850.7010.253

The results indicate that Panode's per-node panoramic generation and incremental 3DGS optimization better preserve cross-view appearance consistency and geometric stability.

Figure 7 in the revised manuscript shows novel-view renderings of the Gaussian scenes. The qualitative results show that Panode still produces structurally continuous and texturally stable renderings at viewpoints far from the initial view. This further validates its novel-view reconstruction advantage in long-trajectory world exploration.

Q2. How do node count and spacing affect reconstruction?

We thank the reviewer for this valuable suggestion. We agree that the original submission lacks a detailed node ablation, and we add a detailed study. We conduct experiments on 20 InteriorGS scenes, with two trajectories per scene. Along the same trajectories, we test different node counts and spacings and evaluate the 3DGS rendering quality over the full trajectory.

NodesPSNR ↑SSIM ↑LPIPS ↓
218.6810.6850.449
520.6310.7730.239
1020.5720.7130.226

The table shows that increasing the number of nodes does not significantly improve the final 3DGS rendering quality. This suggests that Panode does not rely on dense node sampling. Even with relatively sparse nodes, it can still obtain robust reconstruction results.

Figure 10 in the revised manuscript further illustrates the effect of node count on generation quality. The rendering quality tends to stabilize as the number of nodes increases, indicating that moderately sparse panoramic nodes already provide sufficient constraints for incremental 3DGS optimization.

Appendix Figure 14 in the revised manuscript separately shows that active planning continues to expand explored area as more viewpoints are added along indoor and outdoor trajectories.

AuthorsOpenReview ↗

Additional Evaluation and Efficiency Analysis

Q3. What do depth conditioning and historical context contribute to RGB-D completion?

We thank the reviewer for pointing this out. We agree that RGB-D completion warrants a more detailed ablation. To provide proper context, we evaluate the contributions of depth conditioning and historical context on the InteriorGS dataset, with one completion per scene. We use 20–40% random masks and use the 20th previous frame as historical context.

SettingPSNR ↑LPIPS ↓AbsRel ↓δ₁ ↑
Full model28.1390.0630.2350.960
w/o depth control26.0850.1620.4670.839
w/o context23.8740.3050.3820.915

The ablation shows that depth injection mainly improves depth prediction quality, while historical context contributes to both RGB and depth completion.

Q4. Which stages does Table 2 runtime include?

We thank the reviewer for pointing this out, and we sincerely apologize for the unclear runtime description. In Table 2, Panode's runtime includes per-node 3DGS optimization because generating each subsequent node depends on the scene after the previous optimization step. In contrast, the baseline runtimes include panorama generation only. If we also count the 3D reconstruction time of the baselines, Panode's efficiency advantage becomes more pronounced.

For clarity, we separately report the average generation and reconstruction time on the InteriorGS dataset. We also reconstruct 3DGS scenes from the outputs of the baselines and provide the new comparison below:

MethodGeneration time (s)Reconstruction time (s)Total time (min)
OmniRoam657.77566.7720.41
Matrix-3D1636.85539.9036.28
Ours39.35114.502.56

Q5. How many ODE steps were timed, and on what hardware?

We thank the reviewer for this important question, and we apologize for the unclear description. All inference timings use 4 ODE steps and run on a single NVIDIA RTX PRO 6000 GPU.

To analyze the effect of the number of inference steps on quality and efficiency, we add a comparison between 4 and 40 ODE steps on the InteriorGS dataset, with one completion per scene:

ODE stepsPSNR ↑LPIPS ↓AbsRel ↓δ₁ ↑Inference time (s)
428.1390.0630.2350.9604.767
4028.8660.0530.2540.96312.306

The results show that Lightning acceleration does not reduce generation quality while substantially improving inference efficiency.

We sincerely thank the reviewer again and hope our responses address all your questions and concerns.

[1]OverTheReality. OverMaps360-50 Dataset. Hugging Face Datasets, 2026.

Dear Reviewer nfys,

Thank you again for your time and thoughtful feedback. We would like to follow up on our responses to your concerns. We would be grateful to hear whether the clarifications and results in our rebuttal address your main concerns. If anything remains unclear or warrants further discussion, we would be happy to elaborate.

Thank you for your time and consideration.

Best regards,
The Authors

Comments on the paper

2 comments
AuthorsOpenReview ↗

General Response

Dear ACs and Reviewers,

Thank you for your thoughtful and constructive feedback. We carefully consider your comments and revise the manuscript accordingly. We hope that the updated manuscript meets with your approval.

We appreciate the reviewers' positive assessments of our contribution and presentation. In the revised manuscript, we add a long-trajectory novel-view rendering comparison (Table 3 and Figure 7), an RGB-D completion ablation (Table 4), and a node-count analysis (Figure 10). We update the adjacency ablation (Table 5) and add initial-depth, exploration-coverage, runtime, and ODE-step analyses in the appendix, together with details of context-conditioned training.

Our detailed point-by-point responses are provided below.

Dear authors,

Thank for engaging the discussion. I’ve sent the reviewers a reminder to respond to your rebuttal.

Best

AC

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.