acceptodds
Under review as a conference paper at ICLR 2027

SphereBridge: Single-Panorama 3D Gaussian Reconstruction for High-Resolution Real-Time Multi-Sphere Rendering

Abstract

Immersive 360-degree browsing in VR/AR and spatial photography requires turning a single panorama into a high-resolution scene representation that supports nearby views. Feed-forward 3D Gaussian Splatting (3DGS) has made nearby-view synthesis practical for perspective images, but its pinhole recipe assumes a planar local view, whereas a panorama covers the full sphere and must be reconstructed as one 360-degree scene. Tangent-crop workarounds restore pinhole compatibility at the cost of stitching seams and scale drift, and single-panorama pipelines that respect spherical geometry need over an hour of per-scene optimization. We present SphereBridge, a panorama-native pipeline that predicts a panorama-wide 3D Gaussian field from a single equirectangular panorama in one forward pass and bakes it into a Multi-Sphere Image (MSI) for high-resolution rendering. The Multi-Sphere Rendering Bridge routes the Gaussians onto concentric spherical layers, which reduces each frame to closed-form ray-sphere compositing with no per-frame Gaussian sorting, and injects the source panorama's high-frequency detail into the layer textures, which preserves detail at high output resolution. Because real captures rarely provide registered nearby views, training relies on synthetic data, and a training-free Tangent-View Scaffolding extends it to unlabeled real panoramas, using a stitched tangent-view reconstruction as a displaced input view and the real capture as supervision. Across indoor and outdoor benchmarks, SphereBridge reaches state-of-the-art single-panorama quality on LPIPS and DISTS, and its MSI renderer raises perceptual quality at high output resolution while rendering perspective views at 23.92 FPS on a single NVIDIA A800 GPU.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
novel view synthesis, 3D Gaussian splatting, multi-sphere image, single-view 3D reconstruction, real-time rendering, panorama

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_PtQNOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes SphereBridge, a method for reconstructing a 3D scene from a single panorama for high-resolution nearby-view browsing. It combines feed-forward Gaussian prediction adapted to spherical geometry with a rendering bridge that converts the Gaussians into a multi-sphere image (MSI) and injects high-resolution source textures. It also introduces Tangent-View Scaffolding to construct fine-tuning pairs from unpaired real panoramas. Experiments on six indoor and outdoor datasets show competitive perceptual quality and substantially faster MSI rendering than direct Gaussian splatting. Real-data fine-tuning improves performance on OmniPhotos but reduces performance on TartanGround, while rendering quality declines as camera displacement increases.

Critical strengths
  • The combination of feed-forward Gaussian reconstruction and MSI rendering is practical and effective. It enables fast scene construction and high-resolution nearby-view rendering, with a substantial rendering-speed improvement over direct Gaussian splatting.

  • Tangent-View Scaffolding offers a practical way to use unpaired real panoramas for fine-tuning. Using synthesized views as inputs and real captures as supervision avoids relying on synthetic target images. The Stage 2 ablation shows improved performance on OmniPhotos, supporting its usefulness for real-domain adaptation.

Critical weaknesses
  • The benefit of high-frequency texture injection is not isolated. Table 2 lacks a comparison of the same MSI with and without this injection, leaving its contribution unclear. The sensitivity of injected textures to depth and visibility errors is also not evaluated.

  • The evaluation is incomplete. Table 10 lacks external baselines and Direct GS results across displacement ranges, leaving unclear whether the method remains competitive and whether degradation comes from the predicted Gaussian scene or the MSI conversion and texture injection. Comparisons with diffusion-prior-based methods such as WorldExplorer [1] would strengthen the evaluation; the open-source HY-World 2.0 [2] is another useful reference.

[1] Manuel-Andreas Schneider et al. WorldExplorer: Towards Generating Fully Navigable 3D Scenes. SIGGRAPH Asia, 2025. [2] Team HY-World et al. HY-World 2.0: A Multi-Modal World Model for Reconstructing, Generating, and Simulating 3D Worlds. arXiv, 2026.

Questions for authors
  • Texture injection: Could the authors to report results for the same baked MSI with and without high-resolution texture injection, changing only the layer colors? Please also examine whether depth or visibility errors cause texture misalignment at novel views.

  • Camera displacement: Could the authors extend Table 10 to include Direct GS rendering of the same reconstruction and representative external baselines in each displacement interval? This would clarify whether SphereBridge remains competitive compared to baselines as displacement increases and whether MSI conversion introduces additional degradation relative to Direct GS.

  • Baseline coverage: Could the authors evaluate the panorama-to-3D reconstruction pipeline of LayerPano3D using the same input panoramas, bypassing its text-to-panorama generation stage? Table 12 compares only its layer-placement strategy, not the full reconstruction pipeline. Comparisons with WorldExplorer and the open-source HY-World 2.0, initialized from the same panoramas, would also help assess the quality–efficiency trade-off against methods that use diffusion priors.

4 comments
AuthorsOpenReview ↗

Response to Reviewer PtQN (1/2): texture injection

Thank you for the constructive review and the concrete experiments you proposed. This part answers W1 and Q1, and Part 2 answers W2, Q2, and Q3. Table numbers refer to the revised PDF (the General Response maps them to the submission).

Same baked MSI with and without injection (Q1). The variant without injection keeps the baked MSI unchanged, including its layer alphas and routing, and only skips rewriting the layer colors from the source panorama. It appears next to the full MSI in Table 2 (each output resolution), Table 5 (bridge ablation), Tables 6–11 (six datasets), Table 12 (displacement intervals), and Table 15 (360Loc up to native resolution). Cubemap PSNR / LPIPS / DISTS at the target views of Tables 6–11:

DatasetWithout injectionWith injection
MP3D24.36 / 0.279 / 0.13624.22 / 0.251 / 0.096
Replica26.53 / 0.251 / 0.13826.42 / 0.197 / 0.085
360Loc17.93 / 0.439 / 0.20018.03 / 0.351 / 0.111
TartanGround20.84 / 0.371 / 0.17820.97 / 0.283 / 0.123
PanoCity20.70 / 0.375 / 0.21320.72 / 0.237 / 0.107
OmniPhotos21.14 / 0.334 / 0.12721.10 / 0.317 / 0.094

Injection improves LPIPS and DISTS on every dataset and changes PSNR by at most 0.14 dB. At 4096×40964096 \times 4096 in Table 2, injection lowers LPIPS / DISTS from 0.485 / 0.418 to 0.423 / 0.235 at essentially the same frame rate.

Depth errors and misalignment at novel views (Q1). The six-dataset metrics above are measured at novel target views, where sharp texture misplaced by a depth error would lower the pixel metrics, yet the PSNR changes above stay within 0.14 dB. Camera displacement tests this more directly, since the parallax of a depth error grows with translation. Table 12 reports only LPIPS and DISTS, so we computed PSNR / SSIM on its displacement intervals for this response (TartanGround):

Displacement (m)(0, 0.25](0.25, 0.5](0.5, 1.0](1.0, 1.5]
Without injection21.25 / 0.67118.39 / 0.52716.49 / 0.51915.61 / 0.489
With injection21.29 / 0.69018.48 / 0.55816.43 / 0.51115.52 / 0.485

Up to 0.5 m, injection does not lower PSNR or SSIM. Beyond 0.5 m, both drop slightly, consistent with sharp texture shifting as parallax grows, while LPIPS and DISTS still improve in every interval (Table 12). Depth errors of the reconstructed surface are quantified in Appendix G (Table 16).

Visibility errors (Q1). A visibility error spreads the source visibility vpv_p of a ray over layers away from the reconstructed surface, as a foreground reconstructed with too low an opacity or the soft routing variant of Table 5 does, so part of the foreground texture lands on a wrong layer and appears as a ghost at novel views. Appendix F.6 measures this effect: M(ω)M(\omega) sums the injected weight that ray ω\omega places outside the two layers that share the reconstructed surface, which bounds how much misplaced injected texture any novel view can see from that ray. MM exceeds 0.05 on only 0.1% to 2.6% of the non-sky rays of each dataset, and these rays cluster at depth discontinuities. The highest share is on TartanGround, the dataset of the displacement test above, where injection still changes PSNR by less than 0.1 dB in every interval. Hybrid routing, whose gates block the foreground copies of soft routing (Section 3.2, Figure 6), lowers this share on TartanGround from 4.0% to 2.6%.

AuthorsOpenReview ↗

Response to Reviewer PtQN (2/2): camera displacement and diffusion-prior baselines

Camera displacement (Q2). Table 12 now evaluates the six baselines, Depth-Backproj + Bridge, direct splatting of the same SphereBridge Gaussians (Direct GS), and the MSI without and with injection in every displacement interval on TartanGround. LPIPS / DISTS:

Method(0, 0.25](0.25, 0.5](0.5, 1.0](1.0, 1.5]
PERF0.373 / 0.2010.485 / 0.2310.526 / 0.2600.561 / 0.304
Pano2Room0.444 / 0.2280.541 / 0.2530.571 / 0.2790.602 / 0.323
LayerPano3D0.421 / 0.2090.527 / 0.2490.551 / 0.2590.593 / 0.306
WorldExplorer0.584 / 0.3660.662 / 0.4290.671 / 0.4550.690 / 0.467
Depth-Backproj0.350 / 0.1870.434 / 0.2120.483 / 0.2350.520 / 0.259
Tangent-View Scaffolding0.294 / 0.1630.380 / 0.1820.423 / 0.1990.456 / 0.217
Depth-Backproj + Bridge0.278 / 0.1250.371 / 0.1630.455 / 0.2470.546 / 0.388
SphereBridge (Direct GS)0.306 / 0.1580.370 / 0.1680.412 / 0.1880.444 / 0.210
SphereBridge (w/o injection)0.357 / 0.1740.465 / 0.2180.497 / 0.2440.528 / 0.272
SphereBridge0.275 / 0.1220.387 / 0.1520.439 / 0.1880.478 / 0.220

MSI rendering ranks first on both metrics within 0.25 m and remains first or tied on DISTS up to 1.0 m. In the three farther intervals, direct splatting of the same Gaussians is ahead of the six baselines on both metrics.

The degradation has two sources. Every method degrades with displacement, Direct GS included, since a farther view exposes more content that the single source panorama never observes. The MSI conversion adds a cost on top: it discretizes the Gaussians onto 16 spheres, and the resulting parallax error grows with translation, so the MSI without injection trails Direct GS in every interval. Texture injection improves both metrics in every interval and changes PSNR by less than 0.1 dB in each interval (Part 1).

Diffusion-prior baselines (Q3). Both methods start from the same source panoramas as all other methods (Table 1, Figure 4, Tables 6–12, Appendix B.2).

  • LayerPano3D runs its full panorama-to-3D pipeline of layered decomposition, inpainting, and per-scene Gaussian optimization, with the text-to-panorama stage bypassed.
  • WorldExplorer accepts only eight horizontal 60° views, which we crop from the panorama, so its content beyond ±30° in elevation is generated. Every method is averaged over all evaluation scenes, including those on which the Gaussian optimization of WorldExplorer collapses. Even on the scenes where it does not collapse, and within the band it observes (the four side faces within ±25° of elevation), it trails SphereBridge in LPIPS and DISTS on every dataset, with LPIPS 0.402 to 0.526 against 0.164 to 0.312 and DISTS 0.184 to 0.282 against 0.070 to 0.136 (computed for this response).

On quality and efficiency, LayerPano3D takes about 20 minutes and WorldExplorer about 4 hours per scene, against 12.68 s for SphereBridge including the MSI bake (Table 1). Both trail SphereBridge in LPIPS and DISTS on all six datasets and in every displacement interval of Table 12.

We hope these results address your concerns, and we are happy to provide further details.

Reviewer_PtQNOpenReview ↗

response to authors

Thank you for the additional experiments and clarifications. The texture-injection ablations, comparisons across camera displacement ranges, and additional baselines largely address my concerns and strengthen the empirical support for the method. I am updating my recommendation to weak accept.

Thank you for reviewing our response and the revised paper. We are glad that the texture-injection ablations, the displacement comparison, and the added baselines addressed your concerns. Your questions on injection and camera motion made the attribution of each component clearer and the evaluation more complete, and we appreciate the time you spent on our submission.

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

SphereBridge targets nearby-view browsing from a single panorama. It freezes a panoramic depth encoder, predicts two depth layers, and initializes and refines ray-aligned 3D Gaussians on an equal-area spherical lattice. The predicted field is baked once into a 16-layer multi-sphere image (MSI); layer routing, spherical compositing, and injection of native-resolution source-panorama texture support fast high-resolution rendering. For real panoramas without registered nearby views, a stitched perspective reconstruction supplies a displaced pseudo input, while the original capture supervises a second training stage. The paper evaluates perceptual quality on six indoor and outdoor benchmarks and reports component ablations and 4K rendering speed on one scene.

Critical strengths
  1. A practical connection between single-panorama reconstruction and fixed-cost rendering. Spherical rays, equal-area sampling, and wrap/pole-aware padding address panorama geometry and topology. Once baked, the 16-layer MSI avoids per-frame Gaussian sorting. In Table 2 on an A800, 4096×4096 output runs at 23.92 FPS versus 1.77 FPS for direct Gaussian splatting. This is a consequential systems gain for high-resolution nearby-view browsing.
  2. Broad perceptual evaluation and meaningful component controls. On PanoCity, LPIPS/DISTS is 0.237/0.107; on OmniPhotos, it is 0.317/0.094, outperforming the other methods listed in Table 1. Removing circular padding raises directly rendered TartanGround LPIPS from 0.323 to 0.364. The authors also disclose that MP3D is not a clean win and that real-domain fine-tuning hurts synthetic-domain performance.
Critical weaknesses
  1. The high-resolution MSI comparison changes both the representation and the available texture information, and covers only one scene. Table 2 uses a single texture-rich synthetic scene. MSI receives the source panorama at its native resolution for texture injection, whereas direct splatting uses Gaussian colors predicted from a 512×1024 input. The renderer is therefore not the only difference. The experiment convincingly supports a speed advantage, but does not by itself attribute the 4K quality gain to layer routing or establish that it persists across real scenes and varied texture or depth layouts.
  2. The 3D-reconstruction claim lacks direct geometric validation. The main metrics on all six benchmarks are LPIPS and DISTS; the additional PSNR/SSIM tables also score rendered images. MP3D, Replica, and synthetic datasets supply depth or meshes that could support depth, occlusion-boundary, or geometric error measurements, but none are reported. Native-resolution texture injection may improve perceptual scores even when geometry is inaccurate. The evaluation also caps camera translation using the first percentile of source depth: on TartanGround, Appendix Table 10 shows LPIPS worsening from 0.275 at (0, 0.25] m to 0.478 at (1.0, 1.5] m. The evidence therefore more directly supports restricted nearby-view synthesis than generally accurate panoramic 3D reconstruction.
Questions for authors
  1. Can the authors report depth, occlusion-boundary, surface, or reprojection errors on datasets with geometric ground truth, broken down by camera displacement, for both the Gaussian field and baked MSI? This would directly test the reconstruction and browsable-range claims.
  2. Can the authors separate texture injection from hybrid routing under matched high-resolution inputs, and measure quality before and after baking? Please also report 4K quality, bake time, memory, and FPS across multiple indoor, outdoor, and real scenes rather than only the Table 2 example.
1 comment
AuthorsOpenReview ↗

Response to Reviewer wXaZ: attribution of the 4K quality gain

Thank you for the careful review and the precise reading of Table 2. We agree that the renderer is not the only difference. The revision separates the layered representation from texture injection and extends the comparison to real captures and to all six datasets. Table numbers refer to the revised PDF (the General Response maps them to the submission).

What produces the 4K quality gain. Table 2 now adds the same MSI baked without texture injection. It keeps the layer alphas and the routing and only skips rewriting the layer colors from the source panorama. LPIPS / DISTS / FPS:

OutputDirect GSMSI w/o injectionMSI
640×640640 \times 6400.400 / 0.211 / 111.290.402 / 0.229 / 1351.420.365 / 0.152 / 1338.25
1024×10241024 \times 10240.431 / 0.232 / 53.970.434 / 0.266 / 692.470.372 / 0.148 / 701.69
2048×20482048 \times 20480.454 / 0.264 / 9.550.459 / 0.325 / 127.550.385 / 0.166 / 127.17
4096×40964096 \times 40960.479 / 0.334 / 1.770.485 / 0.418 / 23.900.423 / 0.235 / 23.92

Without injection, the MSI renders at essentially the same rate as the full MSI but falls behind direct splatting at every resolution. The speed therefore comes from the layered representation and the 4K quality gain from texture injection, not from layer routing.

Why the texture information differs. The extra texture is what the bridge is designed to deliver. Feed-forward reconstruction at the native resolution of the panorama is impractical, since the self-attention of the DINOv3 encoder grows with the square of the number of input tokens. The bake decouples the two resolutions: reconstruction runs on the 512×1024512 \times 1024 input, and injection writes the source panorama into the layers at its native resolution once, with no cost per frame.

Real captures. Table 15 repeats the comparison on all 360Loc evaluation scenes. These real 3072×61443072 \times 6144 captures hold about 1536 pixels per 90° of view, so the comparison stops at 1536×15361536 \times 1536 output, the highest resolution with native ground truth (Appendix F.5). LPIPS / DISTS:

OutputDirect GSMSI w/o injectionMSI
640×640640 \times 6400.396 / 0.1920.439 / 0.2000.351 / 0.111
1024×10241024 \times 10240.426 / 0.1920.470 / 0.2080.383 / 0.104
1536×15361536 \times 15360.441 / 0.2000.481 / 0.2200.402 / 0.099

The MSI reaches the lowest LPIPS and DISTS at every resolution. Without injection it again falls behind direct splatting, so the attribution holds on real captures. Figure 1(c) shows direct splatting and the MSI on a real outdoor capture at 2048×40962048 \times 4096 equirectangular output.

Varied texture and depth layouts. The last group of Tables 6–11 renders the same reconstructions through the three paths on all six datasets, which span indoor and outdoor, synthetic and real scenes. Cubemap LPIPS / DISTS at 640×640640 \times 640:

DatasetDirect GSMSI w/o injectionMSI
MP3D0.280 / 0.1230.279 / 0.1360.251 / 0.096
Replica0.235 / 0.1190.251 / 0.1380.197 / 0.085
360Loc0.396 / 0.1920.439 / 0.2000.351 / 0.111
TartanGround0.323 / 0.1620.371 / 0.1780.283 / 0.123
PanoCity0.300 / 0.1720.375 / 0.2130.237 / 0.107
OmniPhotos0.316 / 0.1240.334 / 0.1270.317 / 0.094

Injection improves both metrics over the MSI without it on every dataset, and the full MSI reaches a lower DISTS than direct splatting on every dataset.

We hope this attribution addresses your question, and we are happy to provide further details.

Reviewer_u7rxOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents SphereBridge, a feed-forward framework for novel-view synthesis from a single 360° panorama. It introduces a panorama-native Gaussian reconstruction model, converts the predicted Gaussians into a multi-sphere image (MSI) for efficient high-resolution rendering, and uses tangent-view scaffolding to fine-tune on real panoramas without paired novel views.

Experiments on six benchmarks show strong perceptual quality, particularly on outdoor scenes. The MSI renderer achieves 23.92 FPS at 4096×4096 resolution on an NVIDIA A800 GPU, substantially faster than direct Gaussian splatting.

Critical strengths
  1. A practical and effective approach to single-panorama view synthesis.

The paper addresses a useful problem and combines panorama-aware reconstruction with efficient high-resolution rendering. The Gaussian-to-MSI conversion, particularly the hybrid routing and texture injection, is a promising design. At 4K resolution, it improves rendering speed from 1.77 to 23.92 FPS while also improving LPIPS and DISTS (Table 2). This is a meaningful practical result.

  1. Good experimental coverage.

The evaluation spans six indoor/outdoor and synthetic/real datasets. The method performs particularly well on outdoor scenes, and the ablations provide useful evidence for the spherical reconstruction components. The real-data fine-tuning strategy is also interesting.

Critical weaknesses
  1. The novelty and contribution of individual components are not sufficiently established.

Several building blocks are closely related to prior work: feed-forward Gaussian reconstruction (SHARP), panorama-aware geometry prediction (e.g., PanoVGGT), and multi-sphere representations (MatryODShka, MSI-NeRF). The proposed combination is useful, but it is hard to tell how much of the improvement requires the learned Gaussian representation versus a simpler depth-based MSI with the same high-resolution texture injection.

The current ablations do not fully answer this question. In particular, a matched depth-to-MSI baseline and quantitative ablations of texture injection and hybrid routing would make the core contribution much clearer.

  1. The evaluation does not fully separate perceptual quality from geometric accuracy.

The main results focus on LPIPS and DISTS, while the appendix shows that the method does not achieve the best PSNR or SSIM on any of the six datasets. This is not necessarily a problem for perceptual view synthesis, but it raises questions about reconstruction fidelity, especially since high-resolution source textures are directly injected into the MSI.

There is no direct evaluation of reconstructed geometry despite ground-truth depth being available for several datasets. Also, the evaluation focuses on nearby views, and Appendix F.2 shows substantial degradation with larger camera displacement. More geometry-focused and displacement-stratified comparisons would help establish where the method really improves reconstruction rather than appearance alone.

Questions for authors
  1. Could you compare against a simpler panoramic depth-to-MSI pipeline using the same number of layers and the same high-resolution texture injection? Also, could you provide quantitative ablations for texture injection and hybrid routing? These would help clarify which components drive the reported gains.

  2. Could you report reconstruction depth accuracy on the synthetic benchmarks and compare the methods across different camera-displacement ranges? This would help distinguish geometric improvements from perceptual improvements and better establish the usable motion range.

2 comments
AuthorsOpenReview ↗

Response to Reviewer u7rx (1/2): components and attribution

Thank you for the detailed review and for proposing concrete controls. We ran all of them. This part answers W1 and Q1, and Part 2 answers W2 and Q2. Table numbers refer to the revised PDF (the General Response maps them to the submission).

Relation to prior components. SHARP predicts Gaussians from one perspective image and is built for pinhole intrinsics and bounded fields of view, while our reconstruction handles the wrap-around border and the poles of a panorama (ablated in Table 3). PanoVGGT and DAP predict panoramic geometry, and our reconstruction builds a renderable Gaussian scene on the DAP encoder, with trained decoders that improve the geometry beyond this prior (Table 16). MatryODShka predicts an MSI from omnidirectional stereo video, and MSI-NeRF uses spherical layers as a cost volume for a NeRF over four fisheye views, so both rely on stereo or multi-view input. SphereBridge reconstructs a Gaussian field from a 512×1024512 \times 1024 copy of a single panorama and bakes it once into an MSI whose layers receive texture at the native resolution of the panorama. The controls below isolate what each part contributes.

Matched depth-to-MSI baseline (Q1). We pass the Depth-Backproj Gaussians, one per source pixel lifted from the same DAP depth at the same 512×1024512 \times 1024 input, through the unchanged bridge: 16 layers placed by the same rule, hybrid routing, and the same texture injection. Only the Gaussians differ (Depth-Backproj + Bridge, middle group of Tables 6–11). Cubemap metrics under the protocol of Table 1:

MethodMetricMP3DReplica360LocTartanGroundPanoCityOmniPhotos
Depth-Backproj + BridgeLPIPS↓0.2320.2000.4320.2870.2420.360
SphereBridgeLPIPS↓0.2510.1970.3510.2830.2370.317
Depth-Backproj + BridgeDISTS↓0.0950.0870.1150.1290.1160.108
SphereBridgeDISTS↓0.0960.0850.1110.1230.1070.094
Depth-Backproj + BridgePSNR↑23.0225.1315.8620.2119.4919.37
SphereBridgePSNR↑24.2226.4218.0320.9720.7221.10

Under the main protocol, the learned Gaussians lower the cubemap LPIPS from 0.432 to 0.351 on 360Loc and from 0.360 to 0.317 on OmniPhotos and raise the cubemap PSNR on all six datasets, while on the four synthetic datasets the two reach comparable LPIPS and DISTS. Table 12 compares the two across camera displacements on TartanGround (LPIPS / DISTS):

Displacement (m)(0, 0.25](0.25, 0.5](0.5, 1.0](1.0, 1.5]
Depth-Backproj + Bridge0.278 / 0.1250.371 / 0.1630.455 / 0.2470.546 / 0.388
SphereBridge0.275 / 0.1220.387 / 0.1520.439 / 0.1880.478 / 0.220

Depth-Backproj + Bridge stays close up to 0.5 m, where the target view exposes little content occluded from the source, and falls behind beyond that. Its Gaussians hold no content for the regions that the moving camera uncovers, whereas SphereBridge relocates surplus Gaussians into these occluded regions (Section 3.1). Figure 10 shows such disocclusions under both renderers. Part 2 shows that the reconstructed geometry also has a lower AbsRel than Depth-Backproj on all four datasets evaluated for depth.

Texture injection and hybrid routing (Q1). Table 5 ablates the bridge on TartanGround through the baked MSI. All variants share the same Gaussians, and the variant without injection keeps the same layer alphas and only skips rewriting the layer colors from the source panorama.

VariantLPIPS↓DISTS↓
Full SphereBridge (hybrid)0.2830.123
− Texture injection0.3710.178
Hard routing0.2860.128
Soft routing0.2880.129

Removing injection degrades both metrics the most, since the layers then keep only the low-frequency Gaussian colors, and injection improves LPIPS and DISTS on all six datasets as well (Tables 6–11, last group). The failures of hard and soft routing concentrate where a surface crosses a layer interface and at depth discontinuities, so full-image averaging dilutes their differences in Table 5. Hard routing leaves seams at the interfaces, and soft routing copies foreground opacity onto a nearer layer, where the copy shifts under parallax into a ghost (Section 3.2, Figure 6). Hybrid routing reaches the lowest LPIPS and DISTS of the three, and Appendix F.6 measures its local effect: the weight of injected texture that a visibility error can move off the reconstructed surface exceeds 0.05 on 4.0% of the non-sky TartanGround rays under soft routing and on 2.6% under hybrid routing.

AuthorsOpenReview ↗

Response to Reviewer u7rx (2/2): fidelity, geometry, and camera motion

PSNR and SSIM (W2). The last group of Tables 6–11 renders the same SphereBridge reconstruction three ways, which separates reconstruction fidelity from the bridge (Appendix B.2). Cubemap PSNR:

RenderingMP3DReplica360LocTartanGroundPanoCityOmniPhotos
SphereBridge (Direct GS)26.5328.8819.1222.7622.4122.29
SphereBridge (w/o injection)24.3626.5317.9320.8420.7021.14
SphereBridge24.2226.4218.0320.9720.7221.10

Direct splatting of the SphereBridge Gaussians reaches the highest cubemap PSNR and SSIM among all methods on MP3D, 360Loc, and PanoCity, and the second-highest PSNR on the other three. Baking into the 16-layer MSI lowers the PSNR by 1.1 to 2.4 dB. Texture injection then changes it by at most 0.14 dB while improving LPIPS and DISTS on all six datasets, so injected texture does not trade pixel fidelity for perceptual sharpness. The PSNR that SphereBridge loses comes from discretizing the Gaussians onto the 16 layers of the MSI.

Depth accuracy (W2, Q2). Appendix G evaluates depth as ray distance on the four datasets with dense metric ground truth (MP3D, Replica, TartanGround, PanoCity), at the source view and at the target views of the primary evaluation pairs (Table 16). SphereBridge depth, rendered directly from its predicted Gaussians, reaches the lowest AbsRel and RMSE on MP3D, TartanGround, and PanoCity at both views, and the best value in all three metrics on MP3D and PanoCity. This depth also has a lower AbsRel than Depth-Backproj on all four datasets (source / target):

AbsRel↓MP3DReplicaTartanGroundPanoCity
Depth-Backproj0.104 / 0.1150.062 / 0.0620.582 / 0.4940.168 / 0.150
SphereBridge (Direct GS)0.071 / 0.0790.055 / 0.0580.256 / 0.2460.134 / 0.122

Since Depth-Backproj lifts the same DAP depth prior that SphereBridge starts from, the trained depth and Gaussian decoders improve the geometry beyond this prior, so the improvement is not limited to appearance.

Camera displacement and usable range (W2, Q2). Table 12 now evaluates all methods on TartanGround across displacement intervals. LPIPS / DISTS of Tangent-View Scaffolding, the strongest of the six baselines, and of SphereBridge rendered by direct splatting and through the MSI:

Displacement (m)(0, 0.25](0.25, 0.5](0.5, 1.0](1.0, 1.5]
Tangent-View Scaffolding0.294 / 0.1630.380 / 0.1820.423 / 0.1990.456 / 0.217
SphereBridge (Direct GS)0.306 / 0.1580.370 / 0.1680.412 / 0.1880.444 / 0.210
SphereBridge0.275 / 0.1220.387 / 0.1520.439 / 0.1880.478 / 0.220

Quality degrades with displacement for every method, since a farther view exposes more content that the single source panorama never observes. The MSI adds a parallax error that grows with translation, because the bake discretizes the Gaussians onto 16 spheres. Its camera must also stay inside the innermost sphere, which sits at the nearest reconstructed surface, and TartanGround cameras travel close to the ground, so the usable range of the MSI is short there: MSI rendering ranks first on both metrics within 0.25 m and remains first or tied on DISTS up to 1.0 m. Since the bake leaves the Gaussian field intact, the same reconstruction can also be rendered by direct splatting, which is ahead of the six baselines on both metrics in the three intervals beyond 0.25 m.

We hope these controls make the contribution of each component clear, and we are happy to provide further details.

Comments on the paper

2 comments
AuthorsOpenReview ↗

General Response: attribution of the gains and summary of revisions

We thank the reviewers for their careful reading. All three reviews ask which part of SphereBridge produces its gains: the learned Gaussians, the MSI representation, or texture injection. The reviews also request diffusion-prior baselines, evaluation under larger camera motion, and checks of pixel fidelity and of where the injected texture lands. The revised PDF answers the first question with controls that change one factor at a time and adds the evaluations below. New and rewritten content is in blue.

Where the gains come from

  1. Speed comes from the layered representation. At 4096×40964096 \times 4096, the MSI renders at 23.92 FPS and direct splatting of the same Gaussians at 1.77 FPS. The MSI baked without texture injection renders at essentially the same rate as the full MSI (Tables 2 and 15).
  2. High-resolution quality comes from texture injection. Without injection, the MSI falls behind direct splatting of the same Gaussians at every output resolution, on the synthetic scene of Table 2 up to 4096×40964096 \times 4096 and on the real 360Loc evaluation scenes up to their native 1536×15361536 \times 1536 faces (Table 15). With injection, it reaches the lowest LPIPS and DISTS in both tables. Table 5 ablates texture injection and the three routings through the baked MSI.
  3. The learned Gaussians supply geometry and the content of occluded regions. Depth-Backproj + Bridge, a matched depth-to-MSI baseline, lifts the DAP depth prior that SphereBridge starts from into Gaussians and passes them through the same bridge. Under the main protocol it trails SphereBridge on both real-capture datasets, with cubemap LPIPS 0.432 vs. 0.351 on 360Loc and 0.360 vs. 0.317 on OmniPhotos (Tables 11 and 8). On TartanGround it stays close up to 0.5 m of camera motion, where the target view exposes little content occluded from the source, and falls behind beyond that as disocclusions grow (Table 12, Figure 10). The reconstructed depth also has a lower AbsRel than Depth-Backproj on all four datasets evaluated for depth (Table 16).

Other requested evaluations

  • Diffusion-prior baselines (PtQN). LayerPano3D, run as its panorama-to-3D pipeline without the text-to-panorama stage, and WorldExplorer start from the same source panoramas. They take about 20 minutes and about 4 hours per scene, respectively, against 12.68 s for SphereBridge, and both trail SphereBridge in LPIPS and DISTS on all six datasets (Table 1, Figure 4).
  • Larger camera motion (u7rx, PtQN). Table 12 now includes all baselines and direct splatting of the SphereBridge Gaussians in every displacement interval on TartanGround. MSI rendering ranks first on both metrics within 0.25 m. In the three intervals beyond 0.25 m, direct splatting of the same Gaussians is ahead of the six baselines on both metrics.
  • Reconstruction fidelity (u7rx). Direct splatting of the SphereBridge Gaussians reaches the first or second cubemap PSNR on every dataset. Baking them into the 16-layer MSI lowers it by 1.1 to 2.4 dB, and texture injection then changes it by at most 0.14 dB, so the PSNR that SphereBridge loses comes from discretizing the Gaussians onto the 16 layers of the MSI (Tables 6–11, Appendix B.2).
  • Placement of injected texture (PtQN). The weight of injected texture that a visibility error can move off the reconstructed surface exceeds 0.05 on only 0.1% to 2.6% of the non-sky rays of each dataset (Appendix F.6). Depth errors of the reconstructed surface are quantified in Appendix G.

Numbering. Submission Table 3 is split into Tables 3 and 4, and the new Table 5 ablates the bridge, so submission Tables 4–12 are now Tables 6–14. Existing figures and appendix sections keep their numbers.

AuthorsOpenReview ↗

Summary of the discussion and revision

Dear Area Chair and Reviewers,

We thank the Area Chair and all reviewers for their time and constructive feedback. We briefly summarize the discussion and the revised PDF (new content in blue).

  • Attribution of the gains (u7rx, wXaZ, PtQN). Controls that change one factor at a time show that speed comes from the layered MSI (23.92 vs. 1.77 FPS for direct splatting at 4096×40964096 \times 4096), high-resolution quality from texture injection (Tables 2, 5, and 15), and geometry and occluded content from the learned Gaussians. A matched depth-to-MSI baseline through the same bridge trails SphereBridge on both real-capture datasets and beyond 0.5 m of camera motion (Tables 8, 11, and 12).
  • Geometry and camera motion (u7rx, wXaZ, PtQN). The reconstructed depth has a lower AbsRel than the DAP depth prior it starts from on all four datasets with ground-truth depth (Table 16). Across displacement intervals on TartanGround, MSI rendering ranks first within 0.25 m, and direct splatting of the same Gaussians leads all six baselines beyond that (Table 12).
  • Injection robustness and baselines (PtQN). The weight of injected texture that a visibility error can move off the surface exceeds 0.05 on only 0.1% to 2.6% of the non-sky rays of each dataset (Appendix F.6). LayerPano3D and WorldExplorer take about 20 minutes and 4 hours per scene against 12.68 s for SphereBridge, and both trail it in LPIPS and DISTS on all six datasets (Table 1).

Reviewer PtQN stated that the additional experiments largely address their concerns and raised their rating, and Reviewer wXaZ indicated in our exchange that their concerns were largely addressed. We thank Reviewer wXaZ for questioning the source of the 4K quality gain, which led to the injection controls in Tables 2 and 15, and Reviewer u7rx for proposing the matched depth-to-MSI baseline and the other controls now in the revised PDF (Tables 5, 6–12, and 16).

All three reviewers recognized the practical value of SphereBridge, which reconstructs a single panorama in one forward pass and renders it at 4096×40964096 \times 4096 in real time. We thank the Area Chair and all reviewers again for their feedback.

Best regards, The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.