- 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
- How sensitive is Panode to the initial predicted depth?
- How does the method use the historical RGB-D images during training?
- For ablation study of adjacent-aware module in table 3, could you provide result of more general result instead only one scene?
6 comments
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.
| Setting | PSNR ↑ | LPIPS ↓ | Holes ↓ | AbsRel ↓ |
|---|---|---|---|---|
| w/o the graph | 18.142 | 0.234 | 0.181 | 0.333 |
| w/o the adjacency loss | 18.087 | 0.293 | 0.178 | 0.335 |
| Full model | 20.14 | 0.215 | 0.076 | 0.330 |
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
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 :
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.
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.