acceptodds
Under review as a conference paper at ICLR 2027

AtlasFlow: Seam-Aware Sampling and Perspective Distillation for Panorama Generation

Abstract

An equirectangular panorama is stored with an arbitrary longitude cut, and planar diffusion priors attach meaning to it: released ERP generators break there, and their samples change when the cut moves. We remove this dependence without updating the panorama backbone. AtlasFlow evaluates the frozen denoiser in two opposite longitude charts, aligns the predicted velocities, and fuses them with a symmetric partition of unity that downweights each chart near its cut. The sampler is exactly half-turn equivariant in latent space; co-rotating the chart layout extends this to every integer roll, randomized per sample at no extra cost. On Matterport3D the stored-cut error of matched DiT360 falls by 31% at no cost in aggregate quality: the complete method leads eight of ten reference and quality metrics among twelve generators and improves eight over DiT360; on ODI-SR1200 it reduces FID from 59.64 to 55.02 and FAED from 16.82 to 14.64. An 18.1K-parameter Perspective Velocity Adapter (PVA), a periodic residual distilled from the denoiser’s own square-view predictions with no teacher at inference, composes with the sampler and supplies its reference-quality and viewport gains. We further show that frame-boundary seam metrics miss a relocated seam and depend strongly on decoding: a lossless half-turn roll passes them, whereas a yaw-invariant prominence read-out detects injected off-boundary seams with AUC 0.92 where the boundary metric stays at chance.

Primary area:
generative models
Keywords:
equivariance, diffusion models, flow matching, inference-time methods, panorama generation, symmetry, frozen models

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_7aDfOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

I also served as a reviewer for what appears to be the same work submitted to WACV 2027. This version appears to be a more polished presentation of the same work rather than a substantially different contribution. Considering ICLR's dual-submission policy, I think this paper should be desk-rejected.

Critical strengths

I also served as a reviewer for what appears to be the same work submitted to WACV 2027. This version appears to be a more polished presentation of the same work rather than a substantially different contribution. Considering ICLR's dual-submission policy, I think this paper should be desk-rejected.

Critical weaknesses

I also served as a reviewer for what appears to be the same work submitted to WACV 2027. This version appears to be a more polished presentation of the same work rather than a substantially different contribution. Considering ICLR's dual-submission policy, I think this paper should be desk-rejected.

Questions for authors

I also served as a reviewer for what appears to be the same work submitted to WACV 2027. This version appears to be a more polished presentation of the same work rather than a substantially different contribution. Considering ICLR's dual-submission policy, I think this paper should be desk-rejected.

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

The paper addresses cut-dependent sampling in equirectangular panorama generation. At each denoising step, AtlasFlow evaluates the same frozen denoiser in two opposite longitude charts, maps both velocity fields back to a shared physical frame, and fuses them with a symmetric partition of unity before one scheduler update. The fixed two-chart construction has exact half-turn equivariance. Broader integer-roll equivariance requires the chart layout to move with the input, while a randomized layout gives equivariance in distribution. A separate 18,128-parameter Perspective Velocity Adapter is trained to bring ERP velocities closer to the same model's square-view predictions and is applied after chart fusion. The paper evaluates on Matterport3D-polished and ODI-SR1200.

Critical strengths
  1. AtlasFlow fuses aligned velocity fields before a shared scheduler update, so the two charts contribute to one trajectory rather than generating independent panoramas that must be reconciled later.
  2. The partition-of-unity construction gives a clear half-turn guarantee without changing the frozen denoiser's weights or positional encoding. The paper also distinguishes this fixed-layout guarantee from joint input-layout equivariance.
  3. The Perspective Velocity Adapter is a compact periodic residual module that can be composed with the sampler while keeping the underlying transformer frozen and removing the teacher branch at inference.
Critical weaknesses
  1. PVA is trained from the frozen model's one-chart velocity but is applied to the fused velocity in the complete sampler. Fusion changes the adapter's input where the chart predictions disagree, and the paper does not explain why a correction learned on the one-chart field should transfer reliably to the fused field.
  2. With a fixed two-chart layout, the exact per-sample guarantee is for a half-turn. Other integer rolls require co-transporting the chart layout, while a randomized layout gives equivariance in distribution rather than for each fixed sample. Thus the default sampler does not ensure invariance to every longitude-cut change for an individual output.
  3. The sampler makes two denoiser calls per step. The reported time is 31.1 seconds per image on one H100, compared with 15.7 seconds for the one-chart baseline, which is a substantial inference cost.
  4. The main quality comparison does not include several cited panorama methods with relevant designs, including TanDiT, SHERPA, and LF-MultiDiffusion. SphereDiff appears only in an appendix seam-diagnostic comparison on a 45-prompt subset, rather than alongside the main full-cohort quality results. The paper states that common-cohort evaluation is possible for SphereDiff and LF-MultiDiffusion, so the positioning against these methods remains incomplete.
Questions for authors
  1. Since PVA is trained on one-chart velocities and applied to the fused field, what makes its correction transferable across this input shift beyond the reported end-to-end results?
  2. Could the authors clarify why SphereDiff is limited to a 45-prompt seam diagnostic and LF-MultiDiffusion is absent from a common-cohort quality comparison that the paper says is feasible?
Reviewer_ZWxAOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper addresses a storage-layout artifact in equirectangular (ERP) panorama generation. When a panorama is stored as a rectangular image, the longitude at which the left and right edges are cut is arbitrary; however, the denoising trajectories of existing ERP generators depend on this artificial cut, so changing the longitude origin changes the generated output. Building on DiT360, the authors propose AtlasFlow, a training-free sampler that removes this cut dependence, and further train a compact PVA (Perspective Velocity Adapter) to correct the residual perspective distortion.

Critical strengths
  1. The equivariance property is a rigorous, verifiable result. The authors establish it as a formal claim rather than an empirical observation, and its benefit for panorama generation is confirmed by direct measurement.
  2. The controlled comparisons are carefully matched. Prompts, per-image seeds, resolution, number of denoising steps, guidance scale, scheduler, base checkpoint, and decoder policy are all held fixed across systems.
  3. Generalization is supported by cross-dataset and cross-backbone evidence. On ODI-SR1200 the method achieves competitive results without any dataset-specific adaptation.
  4. Seed sensitivity is examined. Multiple seeds were evaluated and the results are reported to be insensitive to seed choice.
Critical weaknesses
  1. Overlap between the PVA training data and the evaluation reference set. The images of the Matterport3D evaluation cohort (about 1k) "are present in Matterport3D-polished", the release on which PVA is optimized. This weakens the credibility of the reported gains. I recommend training an additional PVA checkpoint with the evaluation-cohort images excluded and re-running the PVA-related columns of Table 2.
  2. Compute cost is doubled while the improvement over DiT360 is limited, which reduces the practical usability of the method. Moreover, because the teacher branch is absent at inference and incurs no cost, the source of the reported 2× overhead is not clearly accounted for in the text.
  3. Novelty is weakened by prior latent-rotation panorama methods. Similar latent-rotation strategies already appear in PanoDiffusion and 360DVD. In addition, Figure 4 suggests that the improvement in boundary continuity comes mainly from AtlasFlow rather than from PVA, which further narrows the attributable contribution of the learned component.
Questions for authors
  1. Could you provide a PVA checkpoint retrained with the Matterport3D evaluation cohort (about 1k images) excluded, and report its results on the PVA-related columns of Table 2?
  2. Under a matched denoiser-call budget — for example DiT360 with twice the number of steps, or a single chart with stronger guidance — how do AtlasFlow's SeamL1/DS and overall quality metrics change? What is the actual source of the increased compute cost?
  3. What specifically distinguishes the proposed method from prior latent-rotation panorama methods such as PanoDiffusion and 360DVD? In particular, is the gain attributable to the multi-chart symmetrization itself, or to the cut-dependent weighting?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.