acceptodds
Under review as a conference paper at ICLR 2027

Squaring-the-Sphere: Distortion-Free and Seamless Panorama Generation via CubeAtlas

Abstract

Generating panoramas requires both local perspective fidelity and global spherical continuity. Existing representations face a trade-off: equirectangular projection (ERP) enables seamless panorama modeling but suffers from severe distortion and a domain gap to pre-trained image generators, whereas cubemaps preserve perspective geometry but often produce inconsistent faces and visible seams. We present Squaring-the-Sphere, a panoramic image generation framework built on a novel representation CubeAtlas, which unfolds all cubemap faces into a single compact canvas. By arranging orientation-aligned, rotated polar regions alongside the equatorial faces, this representation preserves the perspective geometry of cubemaps while organizing the cross-face relationships required for seamless panorama assembly. It further enables a pretrained diffusion transformer to model the complete panorama jointly in a single pass, without altering the backbone architecture or positional encoding. While CubeAtlas has a token count comparable to ERP, its structured layout makes many dense attention interactions unnecessary. We therefore introduce a layout-aware sparse attention mask that preserves the interactions required for panorama continuity while removing geometrically irrelevant token pairs. As a result, CubeAtlas achieves better quality than ERP modeling with lower attention computation. Furthermore, we introduce a training-free super-resolution technique and a fusion scheme that enhance fine details in distortion-free perspective space before achieving the final ERP panorama. Experiments on various benchmarks demonstrate that our method improves geometric fidelity, cross-view consistency, and high-resolution visual quality over previous methods.

Primary area:
generative models
Keywords:
Diffusion models, Panorama Generation, 3D

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_rjKoOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper investigates the trade-off between local perspective fidelity and global spherical continuity in panorama generation. Specifically, the paper notes that ERP representations lack perspective fidelity, while cubemaps produce seams, and proposes an unfolding method called CubeAtlas that claims to overcome these two limitations. Moreover, the paper proposes several other techniques, including a training-free super-resolution method and a fusion scheme, to improve visual quality.

Critical strengths
  1. The trade-off between local and global consistency is a classic problem in panorama generation.

  2. It is interesting to consider the design of representation encoding and attention mask jointly.

Critical weaknesses
  1. The design is not reasonable. In CubeAtlas encoding, the horizontally adjacent parts of the 4 up/down regions are not continuous, which still causes significant inconsistency within the 2 polar faces (Up/Down). Essentially, it sacrifices the polar faces for better continuity in middle-latitude regions.

  2. The authors claim that ERP suffers from severe distortion. However, the performance is even better in Tab.2

  • ERP has better Pole-IQ/GI than Cubemap, although Cubemap is theoretically better on pole-related metrics.
  • ERP has similar performance compared with CubeAtlas, and is even better on Pole-IQ
  1. The evaluation issues in Tab 1.
  • Only 100 images cannot reliably measure the distribution distance, like FID.
  • The advantage of the proposed method is not consistent, especially compared with HY-Pano-2.0, and it is not the best on Seam_ERP on any benchmark.
  • The geometry-free capability and SR-related ablations are not evaluated quantitatively.
  • The method is initialized from Qwen-Image, a much larger backbone than those of baselines like CubeDiff and DreamCube. The authors should report the parameter count of each method.
  • Gemini generates both the training captions and the evaluation prompts, and also serves as the judge, which may favor the proposed method, especially on TextAdh.
Questions for authors
  1. In Fig. 3, the proposed pipeline seems to rely on a perspective image to generate a panorama. Can it support text-to-panorama generation?
Reviewer_zZYZOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

The paper addresses the tension between equirectangular panorama generation, which supports global modeling but distorts polar regions, and cubemap generation, which preserves perspective geometry but can produce inconsistent faces. CubeAtlas packs four equatorial faces together with rotated half-regions from the upper and lower faces into one image-like canvas. A layout-aware attention mask restricts communication between regions while allowing the equatorial band to mediate between the polar bands. The authors fine-tune a pretrained diffusion transformer to generate the complete layout, reconstruct the polar faces by blending complementary directional estimates, and reproject the cubemap to an equirectangular panorama. A separate frozen perspective super-resolution model enhances six overlapping views for high-resolution output. The evaluation uses RealSee3D, Laval-Indoor, SUN360, and a 100-image self-collected set, with panorama baselines, boundary and VLM metrics, and a user preference study.

Critical strengths
  1. CubeAtlas places equatorial faces beside orientation-aligned polar regions, creating local canvas neighborhoods for relationships that are otherwise split across cubemap boundaries.
  2. The attention mask follows the layout's geometry. It restricts selected cross-band interactions while allowing information to pass through the equatorial band and keeping text and image conditioning globally available.
  3. WThe method separates panorama generation from high-resolution enhancement. The generated base panorama can be enhanced in perspective space and then assembled into an ERP output without changing the underlying generator.
Critical weaknesses
  1. The upper and lower polar faces are reconstructed from complementary front-back and left-right estimates using a fixed direction-dependent blend. Because the weights depend on pixel direction rather than agreement between the estimates, the rule cannot adapt where the two predictions conflict.
  2. The mask blocks direct attention between the upper and lower polar bands and limits each equatorial face to aligned polar regions. Information can pass through the equatorial band over successive layers, so communication between the two polar bands depends on this indirect route.
  3. The high-resolution stage enhances overlapping perspective views independently. Low-frequency correction and overlap blending address broad shifts and boundaries, but do not directly constrain generated high-frequency details to agree across views, so fine details may still differ across neighboring views.
  4. The conditioning image is consumed without camera calibration or explicit projection into the panorama. No direct geometric constraint ties input pixels to their corresponding output rays, so viewpoint alignment and pixel-level preservation rely on learned correlations rather than an explicit geometric guarantee.
Questions for authors
  1. When the front-back and left-right estimates disagree, what preserves structural continuity across the transition between their dominant regions?
  2. Do the six super-resolution passes share noise states or any prediction-level coupling? If not, what aligns high-frequency details in overlapping regions before blending?
  3. Does any part of training or inference explicitly constrain the input image to its corresponding directions in the output panorama, or is this alignment learned only from data?
Reviewer_89VxOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The authors propose squaring-the-sphere method for text-guided t2p generation with a new representation namely CubeAtlas, which places four equatorial cubemap faces in a central band and arranges rotated, equator-adjacent half-regions of the upper and lower faces alongside them.

This representation is interesting to me, which makes important equator–pole boundaries locally accessible within a single image canvas, allowing a pretrained image-editing DiT to jointly generate the panorama while retaining its backbone modules and positional encoding. The duplicated polar observations are subsequently assembled and combined using directional blending. Additionaly, the authors introduce a layout-aware attention mask that retains 75% of target-to-target connectivity, with globally connected text and image conditions.

The experiments including comparisons with panorama-generation systems, matched-backbone representation ablations, dense-versus-sparse attention experiments, full-versus-half polar packing, and human evaluation.

Critical strengths
  • The proposed reprensentation is generally interesting and motivated. Its arrangement makes each equatorial face locally adjacent to the relevant upper and lower neighborhoods, while avoiding the full duplication cost of retaining complete polar copies. The complementary front–back and left–right polar estimates also give a clear rationale for the directional fusion rule.

  • The design choices are well-explained by the authors. Section 4.2 explicitly keeps the backbone and other training settings fixed when comparing ERP, cubemap, and CubeAtlas. ERP and CubeAtlas also share the horizontal padding scheme, and the comparisons exclude super-resolution. These controls provide evidence that the representation contributes beyond changes in the pretrained model.

  • Sparse attention reduces relative denoising time from 1.00× to 0.84× without changing the target-token count.

  • The paper evaluates three public datasets and a self-collected set, including comparison against HY-Pano-2.0. The main comparisons and human study use base-resolution outputs without the proposed super-resolution stage, avoiding an important potential confound.

Critical weaknesses
  • Though the proposed representation is generally good, the authors fail to compare it with TanDit, which already organizes perspective tangent views into a single image grid and trains a pretrained DiT to generate the grid jointly. Its default construction uses 18 views in a 3×6 layout, followed by perspective-view super-resolution and ERP refinement. It is primarily evaluated for text-to-panorama generation.

  • Each polar face is represented twice in aggregate: the front–back half-regions reconstruct one complete estimate, and the left–right half-regions reconstruct another. Equation (2) then combines these estimates with a directional weighting rule. The evaluated cubemap faces are therefore post-fusion outputs. The reported continuity improvements may reflect several useful effects: better agreement learned by the joint generator, selection of the more reliable parts of duplicated observations, and smoothing of residual differences during fusion. Current results do not separate these effects.

  • Moreover, the calibration covers only 10 test cases and 30 outputs pooled across methods with widely different quality. Strong aggregate rank correlation does not necessarily establish reliable discrimination between the small differences among the best methods or ablation variants. Scene-level uncertainty estimates and targeted validation among similarly strong outputs would make these rankings more convincing. I suggest the authors to enlarge the evaluation set and make the results fully open.

  • Absolute inference costs and timing settings should accompany the relative speedup to make the paper more scientifically transparent.

  • The high-resolution stage is supported mainly by qualitative evidence. A dedicated quantitative or human evaluation at the final resolution would better substantiate the high-resolution claims. Separately evaluating overlap fusion and low-frequency correction would clarify their contributions.

  • It is not clear whether the proposed method will be open source or not, while closed-source method is hard to evaluated and justified.

Questions for authors

See above.

Comments on the paper

1 comment

Thanks for the authors' responses. However, my main concerns remain unresolved, particularly regarding the reliability of the evaluation metrics and the selective interpretation of results.

  1. In W1 and W2, the authors use favorable qualitative examples to dismiss unfavorable quantitative results, without explaining the discrepancies.

  2. In Tab. 2, ERP outperforms Cubemap on polar-related metrics despite Cubemap's theoretical advantage in avoiding polar distortion. Given that many metrics are defined by the paper itself, this raises concerns about the reliability of these metrics.

  3. In W3.1, reporting KID does not resolve the concern about FID's reliability with only 100 images. They are distinct metrics, and the validity of one does not establish the reliability of the other.

Therefore, I maintain my original assessment.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.