acceptodds
Under review as a conference paper at ICLR 2027

PanoForcing: Enforcing Memory-Scalable Full-Context for Blind Panoramic Super-Resolution

Abstract

Blind panoramic super-resolution requires realistic details that remain coherent across the full scene and faithful to the input. Such scene-wide coherence depends on full panoramic context, which is costly to preserve: processing isolated regions fragments this context, while even a one-step diffusion baseline requires over 80 GB of GPU memory for full-panorama fine-tuning. We propose PanoForcing, a one-step diffusion framework that preserves full panoramic context during the forward pass under configurable training and inference memory budgets. PanoForcing selectively constructs the training-memory-related autograd graph while maintaining full-context forward computation, and incorporates lightweight observation guidance within a single network pass to constrain generated details. Experimental results demonstrate high-quality full-context panoramic reconstruction across a broad range of controllable memory budgets. PanoForcing achieves the best WS-PSNR, LPIPS, and FID among the compared methods.

Primary area:
generative models
Keywords:
omnidirectional image, diffusion model, panorama, super resolution

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_hKBsOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary
  1. This paper tried to solve the blind panoramic image super-resolution task with limited computation resources during training.
  2. Several key components are proposed to achieve this:
    1. Instead of using ERP representation, they use the six cubemap representation. To reduce the size of the gradient map, they randomly pick some of the faces to propagate the gradients instead of all faces. This is called Spherical Subgraph Routing (SSR).
    2. To reduce the hallucination of the diffusion, the also proposed Fourier Pseudo-inverse Guidance (FPG). It avoid the intermediate features deviate too much from the input low resolution features.
    3. To reduce the size of the attention map, they also proposed Spherical Octant Attention. The global attention among all 6 faces is huge. The proposed attention mechanism first split the cubemap into a one level octree with 8 sub-cubes, then groups the regions around the cubemap corners and each sub-face only attend to faces that share corners. This significantly reduced the attention map size.
  3. They conduct experiments on different datasets including ODISR and SUN360 using Real-ESRGAN degradation.
Critical strengths
  1. I think SSR is interesting, the separation of forward and backward pass to save memory is clever. As shown in the Table S2, K=1 reduces the memory usage by around 70% compared with K=6.
  2. Table 2 shows the good efficiency of the proposed method compared with other methods. Almost half of the second best OSEDiff.
  3. The performance is the best on most metrics as shown in the Table 1. Other metrics are also comparable. Honestly, the blue in the table is really close to black... Try a more distinguishable color.
  4. Ablation study in Table 3 shows the effectiveness of both SSR and FPG.
Critical weaknesses
  1. It is a bit weird that Table 4 only shows the performance of K=1,2,3,4. What about K=5,6. If this experiment is unable to run on the available devices, how can you get the memory usage of K=5,6? It is also anti-intuitive that the K=3 is better than higher K.
  2. I am thinking if FPG is correct. For complicated degradation, the assumption in Equation 4 is hard to get, and the degradation matrix is hard to approximate. Training-free FPG with strong guidance as shown in Table S3 also shows that there might be errors in the estimated degradation.
Questions for authors
  1. The entire paper emphasizes the "full context" concept, any evidence about the necessity of the full context?
  2. If you convert the ERP to cubemap then super resolve each view, how to completly avoid seams when you stitch them together by the end? Have you analyze the seams?
  3. I wonder if the proposed modules can be generalized to other architectures,
  4. I would like to see if FPG works on real degradations.
Reviewer_VtvCOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

PanoForcing is a one-step diffusion framework for blind panoramic super-resolution under adjustable memory budgets. Its Spherical Subgraph Routing mechanism processes all six cubemap faces while selectively retaining autograd paths, and Spherical Octant Attention enables efficient interaction across faces. Fourier Pseudoinverse Guidance injects observation information at multiple network depths to improve reconstruction consistency. Experiments on ODISR and SUN 360 report competitive reconstruction quality together with substantial reductions and configurable control of training and inference memory.

Critical strengths
  1. Separating full-context forward computation from selective gradient recording is a useful approach to memory-constrained training.
  2. The framework combines panoramic interaction and observation guidance within one denoising pass.
  3. Experiments report strong reconstruction results and detailed memory measurements across multiple configurations.
Critical weaknesses
  1. The experimental comparison omits an especially relevant baseline. RealOSR is explicitly discussed as a one-step panoramic SR method with multi-depth observation guidance, and its degradation setup is followed, yet it is absent from the main comparison. This makes the incremental advantage over the closest discussed approach unclear. Panoramic evaluation is also restricted to synthetic ×4 degradations, providing limited evidence of generalization to captured panoramic inputs.

  2. The lowest-memory training claim needs qualification. Appendix A.2 reports numerical overflow at K=1K=1 with mixed precision and therefore uses FP32 for the face-selection quality study. However, Table S2’s approximately 8.7 GB configuration is measured with mixed precision. Consequently, stable training and reconstruction quality at the advertised minimum memory are not established together. Runtime measurements across memory budgets are also missing, leaving the memory–speed trade-off unclear.

  3. The connection between FPG’s theory and its learned implementation remains incomplete. Appendix C analyzes aligned linear downsampling and independent Gaussian noise, whereas the degradation pipeline includes effects outside those assumptions. Furthermore, the analysis does not establish how learned operators acting on nonlinear intermediate features enforce physical observation consistency. Comparisons against equally sized spatial conditioning modules would help establish the specific value of the Fourier pseudoinverse formulation.

  4. The central coherence claim lacks direct measurements. WS-PSNR, LPIPS, and FID assess reconstruction quality but do not directly measure cubemap boundary continuity, ERP wraparound seams, or consistency across overlapping viewports. The qualitative examples mainly demonstrate local detail restoration. Seam-focused evaluation and controlled removal of distant context would more directly test the claimed benefit of full panoramic context.

Questions for authors
  1. Can the approximately 8.7 GB configuration complete stable training, and what are the training and inference runtimes across the reported memory budgets?
  2. How does PanoForcing compare with RealOSR on reconstruction quality, seam continuity, runtime, and memory under a shared evaluation protocol?
Reviewer_NtBGOpenReview ↗
Rating
3
Confidence
2
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes PanoForcing, a one-step diffusion-based method for blind panoramic image super-resolution. It introduces Spherical Subgraph Routing to reduce training memory consumption while preserving full-context processing, and Fourier Pseudoinverse Guidance to improve observation consistency. Experiments on two datasets demonstrate competitive reconstruction performance, with the method achieving the best results on three evaluation metrics while maintaining relatively low training and inference memory costs.

Critical strengths
  1. The proposed method demonstrates clear advantages in memory efficiency and computational complexity, while achieving promising quantitative results.
  2. The visual quality of the reconstructed panoramas is also impressive, suggesting that the method can effectively balance reconstruction quality and computational efficiency.
Critical weaknesses

The generalization capability of the proposed method under blind degradation conditions remains insufficiently validated. Since the experiments primarily rely on a predefined synthetic degradation pipeline, it is unclear whether the method can generalize effectively to different degradation distributions or more realistic low-resolution panoramic inputs.

Questions for authors
  1. Could the authors provide additional experiments to better validate the generalization capability of PanoForcing under blind super-resolution settings? In particular, since the original dataset provides low-resolution images generated using its standard downsampling protocol, evaluating the proposed method on these original LR images would help assess its robustness to different degradation settings.
  2. Figure 5 compares the memory usage of PanoForcing primarily against OSEDiff. It is unclear why OSEDiff is the only baseline included in the memory analysis. Since memory scalability is a central contribution of this work, comparisons with additional representative baselines would strengthen the evidence for the claimed advantage. If such comparisons are infeasible, the authors should clarify the technical reasons and the criteria used to select OSEDiff as the memory baseline.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.