acceptodds
Under review as a conference paper at ICLR 2027

SpheRoPE: Zero-Shot Optimization-Free 360 Panorama Generation with Spherical RoPE

Abstract

We present a zero-shot, training-free and optimization-free framework for generating 360 panoramic images and videos by directly injecting spherical priors into pre-trained diffusion transformers. Existing methods either rely on costly fine-tuning on scarce panoramic data that limits generalization, or leverage multi-step optimization that incurs prohibitive inference latency. We observe that contemporary generative models natively exhibit some panoramic priors from large-scale training. However, these emergent capabilities are insufficient, as the models fundamentally fail to satisfy the rigorous topological constraints imposed by equirectangular projection (ERP). We introduce a zero-shot and optimization-free approach that resolves these constraints at inference time. Spherical RoPE replaces standard rotary position embeddings: low-frequency channels are re-parameterized as 3D Cartesian coordinates to natively encode the spherical manifold, while high-frequency channels are harmonically quantized to enforce exact periodicity. Coupled with complementary Semantic Distortion classifier-free guidance (CFG) that explicitly steers geometry, we avoid retraining and inherit the full creative breadth of state-of-the-art models. Our approach generalizes across diverse backbones and 360 generation modalities. We demonstrate this across text-to-panorama using Flux.1, Flux.2, and LTX-Video backbones, achieving competitive performance against baselines, all while remaining training-free. Our code will be publicly available.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
zero shot, 306 generation, diffusion models

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_KFg1OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents SpheRoPE, a training-free and optimization-free framework for generating 360° panoramic images and videos using pretrained diffusion transformers. It modifies the horizontal rotary positional embeddings through spectral partitioning: high-frequency channels are quantized to integer-period harmonics to enforce horizontal periodicity, while low-frequency channels use spherical Cartesian coordinates to capture spherical topology and polar convergence. The framework also introduces Semantic Distortion CFG, which uses the prediction difference between a content prompt and its geometry-augmented version to strengthen equirectangular projection characteristics.

Critical strengths
  1. It’s interesting that training-free methods can leverage the capabilities of foundation models to directly improve panoramic generation results.
  2. The paper is clearly written and easy to follow.
Critical weaknesses
  1. SHERPA [1] also modifies only the high-frequency horizontal RoPE channels, rounding the frequencies to integer-period harmonics. Under the minimum-cycle condition required by SpheRoPE, the quantization formulas for the high-frequency components of the two methods are equivalent. Furthermore, without training, it leads to a decrease in overall FAED and FID. For boundary connectivity, works such as 360DVD [2] and SphereDiffusion [3] have proposed rotating while denoising. Can modifying positional encoding provide better gains than these methods?
  2. In the comparison with other methods, the video foundation model used is LTX 2.3, while the compared methods mostly use AnimateDiff (360DVD), VideoCrafter (DynamicScaler), Wan2.1 (PanoWan), and LTX Video (SphereDiff). There is insufficient evidence to convincingly establish whether the improvements come from the proposed method or a stronger generative foundation model.
  3. Semantic Distortion CFG concatenates geometric descriptions to the same content prompt and then uses the difference between two conditional predictions to enhance ERP attributes. The application design is reasonable, but its novelty as a general method is limited, constituting a simple modification of Composable Diffusion [4] and SEGA [5] for spherical geometry.

[1] SHERPA: Seam-aware Harmonized ERP Adaptation for Open-Domain 360° Panorama Generation

[2] 360DVD: Controllable Panorama Video Generation with 360-Degree Video Diffusion Model

[3] SphereDiff: Tuning-free 360° Static and Dynamic Panorama Generation via Spherical Latent Representation

[4] Compositional Visual Generation with Composable Diffusion Models

[5] SEGA: Instructing Text-to-Image Models using Semantic Guidance

Questions for authors

See WeakNess

5 comments

We thank the reviewer for the positive evaluation and for highlighting the clarity and value of our training-free approach. Below we address the raised concerns.

Comparison with Concurrent SHERPA and Rotate-While-Denoising (W1)

We compare SHERPA's encoding without its adapters (its 14 highest-frequency width pairs snapped, the rest linear) and rotate-while-denoising (a π/2 latent roll at every step) on the same backbone and recipe: FLUX.1 with SD-CFG and circular padding, n = 1,200. We also report Pole std (pixel std along top/bottom rows, 0 for an exact pole C2, and 6.0/11.4 on real panoramas), Polar FID (FID over the eight ±60°-pitch views), and Seam-FID (FID of crops centered on the wrap):

FLUX.1 (with SD-CFG + Pad)Pole std↓Polar FID↓DS↓Seam-FID↓FAED↓FID↓KID×10²↓IS↑CS↑
Stock RoPE30.6 / 34.263.31.4253.3119.0345.902.5910.3118.49
Stock RoPE + rotation17.3 / 26.060.00.8850.9122.7448.082.839.9318.50
SHERPA (no adapters)19.1 / 27.366.80.9254.1523.2251.483.019.4718.50
SpheRoPE (Table 1)4.0 / 11.147.90.8942.7529.9445.222.6310.3218.52
  • Concurrent SHERPA [1] (trained vs. zero-shot): SHERPA is concurrent work and is training-based (trained adapters on FLUX.1), whereas SpheRoPE is zero-shot. While both snap high-frequency width channels to integer harmonics, SHERPA leaves the low-frequency band linear, which provides neither periodicity nor polar convergence (C2) in the low band, relying on trained adapters to learn the poles and global layout. Without the adapters, this encoding closes the border seam (DS 0.92) but not the poles (Pole std 19.1/27.3 vs. 4.0/11.1 for SpheRoPE, Polar FID 66.8 vs. 47.9), and snapping its lower pairs costs view quality (FID 51.48 vs. 45.22), so the poles need our low-frequency spherical band. Snapping low frequencies instead causes large OOD shifts that break global layout (Fig. 7a). Our low-frequency spherical Cartesian encoding satisfies both C1 and C2 without training (DS 0.89, Pole std 4.0/11.1, Polar FID 47.9).

  • Rotate-while-denoising [2]: Rolling the latent between steps (as in 360DVD [2], whereas SphereDiff [3] fuses perspective views rather than rotating) only relocates the seam between steps: within every forward pass, planar RoPE still treats the current borders as distant and provides zero polar convergence (C2). In Table 2, ours scores above 360DVD on all six VBench dimensions on both prompt sets, and above DynamicScaler (window shifting) on all six on SphereDiff-20, at a fraction of SphereDiff's and DynamicScaler's cost.

  • True polar convergence (C2): Rotation leaves the poles unconverged (Pole std 17.3/26.0 vs. 4.0/11.1 for SpheRoPE, close to real panoramas at 6.0/11.4). Polar views confirm it: across the eight ±60°-pitch views, SpheRoPE lowers FID from 60.0 to 47.9 and KID×10² from 2.82 to 1.69, and on zenith/nadir faces lowers KID×10² from 3.05 to 1.54 and 4.75 to 4.20. Global FAED ranks the other rows higher but, as with vanilla FLUX.1 (see our response to r3P5), barely registers unconverged poles.

  • Sharper cross-seam and view quality: While rotation blends pixel differences at the border (DS 0.88 vs. 0.89), planar attention blurs content across the wrap (Seam-FID 50.91 vs. 42.75 for SpheRoPE) and degrades per-view FID/KID/IS.

The same holds for video on LTX 2.3 (Table 2 settings), where SpheRoPE outperforms rotation on temporal stability, subject/background consistency, seam continuity (DS), and polar convergence (Pole std) across both sets:

SetMethodIQ↑CLIP↑Flicker↑Motion↑Subj.↑BG↑Seam DS↓Pole std↓
SphereDiff-20Rotation54.0027.9699.5199.5998.1397.341.097.9 / 9.2
Ours (Table 2)53.6427.9599.5899.6298.2297.711.036.1 / 6.5
Stress-20Rotation52.8929.2698.6199.3796.1595.491.5722.1 / 15.4
Ours (Table 2)51.7628.9198.9199.4696.5196.021.4816.8 / 11.0

[1] Kang et al., "SHERPA: Seam-aware Harmonized ERP Adaptation for Open-Domain 360° Panorama Generation," 2026.

[2] Wang et al., "360DVD: Controllable Panorama Video Generation with 360-Degree Video Diffusion Model," CVPR 2024.

[3] Park et al., "SphereDiff: Tuning-free 360° Static and Dynamic Panorama Generation via Spherical Latent Representation," AAAI 2026.

Video Foundation Model Comparison (W2)

To isolate our contribution on matched video backbones, we compare vanilla LTX 2.3 with our Table 2 row (identical prompts, sampler, and seeds) and evaluate our method zero-shot on Wan2.1-T2V-1.3B [4], PanoWan's exact checkpoint and prompts, plus the full method on Wan2.1-14B at 1440×720:

SetMethodIQ↑CLIP↑Flicker↑Motion↑Subj.↑BG↑Seam DS↓
SphereDiff-20Vanilla LTX 2.353.1526.6499.1799.4593.3195.426.44
Ours (Table 2)53.6427.9599.5899.6298.2297.711.03
Stress-20Vanilla LTX 2.343.6029.3496.6398.5486.5991.5410.57
Ours (Table 2)51.7628.9198.9199.4696.5196.021.48
SetMethodIQ↑CLIP↑Flicker↑Motion↑Subj.↑BG↑Seam DS↓Pole std↓
SphereDiff-20PanoWan (1.3B, LoRA)39.1027.3299.1399.4397.4096.590.843.8/7.2
Stock Wan2.1-1.3B44.7326.8198.6999.1895.7595.8912.3315.2/21.9
Ours (1.3B, cyclic)45.1926.7298.7199.1796.7595.961.1214.9/20.3
Ours (14B, full)48.4026.1898.5499.0895.0794.990.881.3/4.1
Stress-20PanoWan (1.3B, LoRA)39.4928.4297.1198.5891.6493.660.988.1/10.0
Stock Wan2.1-1.3B41.4227.0094.2396.3285.6491.8820.4734.0/34.4
Ours (1.3B, cyclic)44.0626.5995.3197.1188.8392.441.2437.8/35.1
Ours (14B, full)44.17n/a97.7798.4891.0193.130.852.6/2.3

Real ODI-SR panoramas have Pole std 6.0/11.4.

Takeaway: Against vanilla LTX 2.3, SpheRoPE improves all 6 VBench metrics on SphereDiff-20 and 5/6 on Stress-20 while closing the wrap seam (DS 6.44 → 1.03, 10.57 → 1.48). On Wan2.1-1.3B, our method closes the seam of the stock checkpoint (12.33 → 1.12, 20.47 → 1.24) and improves IQ and consistency. Panoramic video fine-tuning gives PanoWan a tighter seam and higher temporal stability and CLIP (+0.6 to +1.8) but degrades visual quality (IQ 44.73 on Stock vs. 39.10 on PanoWan), whereas we preserve visual fidelity (+6.09 and +4.57 IQ over PanoWan).

Scope on the 1.3B checkpoint: this small model relies on its lowest-frequency width channels for horizontal layout, and the spherical band collapses its generation. We therefore use a cyclic integer variant of our width encoding, which closes the seam but does not converge the poles. On Wan2.1-14B the full method works and gives the best poles on both sets and the best Stress-20 seam.

Regarding the role of the backbone: comparing zero-shot results with models trained on panoramic data is inherently unfavorable to us, so the direct comparisons are against the plain backbone (above). SpheRoPE also composes with training (LoRA fine-tuning FLUX.2 with SpheRoPE lowers FID from 25.33 to 16.63, App. B.5), and a training-free method leverages stronger backbones as soon as they are released.

Novelty of Semantic Distortion CFG (W3)

We thank the reviewer for pointing out Composable Diffusion [5] and SEGA [6], and will cite both in Sec. 2 and 3.3. While SD-CFG is an important component for leveraging pretrained ERP priors, our key novelty is Spherical RoPE (Sec. 3.2): without it, the seam degrades on both backbones (DS 0.94 → 1.22 on FLUX.2, Table 4, and 0.89 → 1.42 on FLUX.1) and on FLUX.1 the poles do not converge (Pole std 30.6/34.2 vs. 4.0/11.1, see W1). Moreover, unlike Composable Diffusion [5], which combines independent prompts ε(p_1) and ε(p_2), conditioning on p_geo alone injects conflicting scene content (Sec. 3.3). Anchoring via p_anchor = [p, p_geo] and computing ε([p, p_geo]) − ε(p) isolates the pure ERP geometric deformation for the target scene. Appending p_geo under standard CFG is the coupled special case γ = w_sem. In Table 4's setting it reaches FAED 33.33 and DS 0.99, vs. 25.40 and 0.94 with decoupled γ (see response to r3P5).

Changes in the revision:

  • SHERPA and rotate-while-denoising discussed in Sec. 2, with the SHERPA control and the image and video rotation comparisons in the appendix.
  • Vanilla LTX 2.3 and Wan2.1 rows in Table 2.
  • Composable Diffusion and SEGA cited in Sec. 2 and 3.3.

We thank the reviewer once again and hope these responses address all concerns.

[4] Team Wan, "Wan: Open and Advanced Large-Scale Video Generative Models," arXiv:2503.20314, 2025.

[5] Liu et al., "Compositional Visual Generation with Composable Diffusion Models," ECCV 2022.

[6] Brack et al., "SEGA: Instructing Text-to-Image Models using Semantic Guidance," NeurIPS 2023.

Dear Reviewer KFg1,

We thank you once again for your insightful feedback on our paper.

As the Author-Reviewer discussion period is drawing to a close, we would like to ask whether our responses have adequately addressed your concerns. Following your comments, we compared against SHERPA's encoding and rotate-while-denoising on the same backbone, for both images and video, added backbone-matched video results (vanilla LTX 2.3 and Wan2.1), and clarified the novelty of SD-CFG relative to Composable Diffusion and SEGA. We would be happy to clarify anything further. We greatly appreciate your time and consideration.

Best, Authors

Reviewer_R39cOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes SpheRoPE, a training-free and optimization-free approach to adapting pretrained diffusion transformers for 360° panorama generation. It modifies width-axis RoPE through frequency partitioning, combining cyclic linear encoding for high-frequency channels with spherical Cartesian encoding for low-frequency channels. A complementary Semantic Distortion CFG mechanism uses geometric text prompts to encourage equirectangular structure. The framework is evaluated on panoramic image and video generation, with supplementary experiments on additional backbones and compatibility with LoRA fine-tuning.

Critical strengths
  1. A practical and well-motivated approach: The method enables panorama generation through inference-time modifications without requiring task-specific training or iterative optimization. Its frequency-dependent treatment of RoPE offers a clear way to introduce spherical structure while preserving useful local spatial priors from pretrained models. The appendix examines encoding choices, guidance settings, geometric prompts, and circular VAE padding, helping explain the design decisions.

  2. Good visual quality and evidence of potential applicability: The qualitative examples show convincing panoramas with improved boundary continuity and well-preserved local detail. The image-conditioned examples also suggest that the approach can retain useful conditioning capabilities of the underlying models. Experiments on additional backbones and LoRA fine-tuning provide further evidence of the method's potential applicability.

Critical weaknesses
  1. Insufficient isolation of the proposed method's contributions

The evaluation does not sufficiently separate the contributions of the proposed method from those of the foundation model and other inference-time components. In Table 1, Ours (FLUX.1) achieves a strong FAED but has substantially worse multi-view FID, KID, and IS than Ours (FLUX.2), and underperforms several baselines on these multi-view metrics. Moreover, without any panorama-specific training, Vanilla FLUX.2 already achieves an FAED of 33.38 (Table 4), lower than the FAED scores of all competing baselines in Table 1. The video comparisons in Table 2 also use different backbones for Ours and the competing methods. These differences make it difficult to determine how much of the improvement comes from the proposed method itself.

Although Table 4 shows improvements over Vanilla FLUX.2, component attribution remains unclear because it does not clearly specify whether circular VAE padding is used in each row or the exact prompt format and content. Section 4.4 ("Component Contributions") identifies boundary seams as a weakness of Vanilla FLUX.2, while Figure 13 shows that circular VAE padding can itself remove a visible seam. Appendix B.1 explains the distinct roles of SpheRoPE and circular VAE padding, but without explicit quantitative controls for prompts and padding, it is difficult to establish how much additional improvement comes from the proposed positional encoding and guidance components.

  1. Missing evaluation of video dynamics

Table 2 omits metrics of generated dynamics, although the geometric prompt in Appendix C explicitly restricts camera and scene motion while Stress-20 requests dynamic scenes. Appendix D acknowledges the method's reliance on a static camera and minimal atmospheric motion. Without a dynamics metric alongside the temporal-stability metrics, it remains unclear whether the reported stability gains come at the cost of reduced motion. This limits the evidence for the method's effectiveness on dynamic panoramic video generation.

Questions for authors
  1. Could you clarify the exact prompts and circular VAE padding settings for each row of Table 4, and report results with and without padding for both Vanilla FLUX.2 and the full method, keeping all other settings fixed within each comparison? This would clarify the contributions of SpheRoPE and SD-CFG relative to padding.

  2. Could you add Vanilla LTX 2.3 to Table 2 and provide complete quantitative results for your existing Wan2.1 experiments, using the same Wan2.1 checkpoint as PanoWan and matched evaluation settings? These comparisons would help isolate the method's contribution from that of the backbone.

  3. Could you report VBench's Dynamic Degree or a comparable metric for all methods in Table 2 alongside the temporal-stability metrics, to assess whether the reported stability gains are accompanied by reduced motion?

5 comments

We thank the reviewer for the positive evaluation and for highlighting the practical, training-free nature of SpheRoPE. Below we address the raised concerns.

Isolating SpheRoPE, SD-CFG, and Circular VAE Padding (W1 & Q1)

In Table 4, circular VAE padding is on in all four rows. Vanilla and (+) SpheRoPE use a per-caption panoramic rewrite ("This 360-degree equirectangular panorama captures …") without SD-CFG, while (+) SD-CFG and Ours use p for ε_cond, the empty prompt for ε_uncond, and [p, p_geo] for ε_geo (App. C). Below we report padding off/on pairs (identical latents decoded with and without padding), alongside Seam-FID (FID of perspective crops centered on the wrap):

ConfigurationPadFAED↓DS↓Seam-FID↓FID↓KID×10²↓IS↑CS↑
Vanilla FLUX.2Off33.785.8532.3525.811.0514.2618.42
Vanilla FLUX.2On33.381.3731.5126.051.0614.2418.42
Ours, FLUX.2Off25.302.9431.3525.440.8612.7818.43
Ours, FLUX.2 (Table 4)On25.400.9430.9625.330.9012.7518.43

Padding changes essentially only DS in both pairs (FAED within 0.4). Without padding, the decoder border artifact dominates DS, yet SpheRoPE's matching latent content already halves it (DS 2.94 vs. 5.85). With padding enabled, Vanilla still retains a content seam (DS 1.37 vs. 1.07 for real panoramas) that SpheRoPE closes (0.94), while improving FAED by 7.98 points (33.38 → 25.40) and lowering Seam-FID and crop FID/KID.

The components act at different stages and fix different mechanisms of seam formation:

  • SpheRoPE (inside the transformer): Standard planar RoPE treats the left and right edges as distant tokens. SpheRoPE fixes this inductive bias inside self-attention so both edges attend as neighbors, making the generated latent content match across the wrap.

  • Circular VAE padding (inside the CNN decoder): Even when latent content matches, the zero-padded convolutional decoder sees only one side at image borders, leaving a thin pixel-level seam (Fig. 13). Circular padding copies the wrapped latent columns only before decoding (no positional encoding ever sees the padding). Conversely, padding alone only blends boundary pixels and cannot make the transformer generate matching structures across the boundary (Fig. 6, middle).

Backbone-Matched Video Comparisons: Vanilla LTX 2.3 and Wan2.1 (W1 & Q2)

Vanilla LTX 2.3: Following the reviewer's suggestion, we compare vanilla LTX 2.3 with our Table 2 row under identical prompts, sampler, and seeds (Seam DS = per-frame ERP DS). Ours improves all 6 VBench dimensions on SphereDiff-20 and 5/6 on Stress-20, and closes the wrap seam that vanilla LTX 2.3 leaves open:

SetMethodIQ↑CLIP↑Flicker↑Motion↑Subj.↑BG↑Seam DS↓
SphereDiff-20Vanilla LTX 2.353.1526.6499.1799.4593.3195.426.44
Ours (Table 2)53.6427.9599.5899.6298.2297.711.03
Stress-20Vanilla LTX 2.343.6029.3496.6398.5486.5991.5410.57
Ours (Table 2)51.7628.9198.9199.4696.5196.021.48

Wan2.1: we also ran our method zero-shot on PanoWan's exact checkpoint, Wan2.1-T2V-1.3B [1] (896×448, 81 frames, CFG 5, 50 steps, PanoWan's prompts), alongside the stock checkpoint, and the full method on Wan2.1-14B at 1440×720 (Table 2 protocol, all 14 perspective clips, Pole std = top/bottom-row pixel std):

SetMethodIQ↑CLIP↑Flicker↑Motion↑Subj.↑BG↑Seam DS↓Pole std↓
SphereDiff-20PanoWan (1.3B, LoRA)39.1027.3299.1399.4397.4096.590.843.8/7.2
Stock Wan2.1-1.3B44.7326.8198.6999.1895.7595.8912.3315.2/21.9
Ours (1.3B, cyclic)45.1926.7298.7199.1796.7595.961.1214.9/20.3
Ours (14B, full)48.4026.1898.5499.0895.0794.990.881.3/4.1
Stress-20PanoWan (1.3B, LoRA)39.4928.4297.1198.5891.6493.660.988.1/10.0
Stock Wan2.1-1.3B41.4227.0094.2396.3285.6491.8820.4734.0/34.4
Ours (1.3B, cyclic)44.0626.5995.3197.1188.8392.441.2437.8/35.1
Ours (14B, full)44.17n/a97.7798.4891.0193.130.852.6/2.3

Real ODI-SR panoramas have Pole std 6.0/11.4.

Wan2.1 takeaway: Over the stock 1.3B model, our method closes the seam (12.33 → 1.12, 20.47 → 1.24) and improves IQ and subject/background consistency, with slightly lower CLIP. PanoWan's video fine-tuning gives a tighter seam and slightly higher stability and CLIP (+0.6 to +1.8) but lowers IQ (44.73 on Stock vs. 39.10), so we beat it by +6.09 and +4.57 IQ.

Scope on the 1.3B checkpoint: this small model relies on its lowest-frequency width channels for horizontal layout, and the spherical band collapses its generation. We therefore use a cyclic integer variant of our width encoding, which closes the seam but does not converge the poles. On Wan2.1-14B the full method works and gives the best poles on both sets and the best Stress-20 seam.

Evaluation of Video Dynamics (W2 & Q3)

We agree that stability must be read alongside dynamics. Our Table 2 LTX recipe is conservative because its p_geo explicitly requests a static camera and scene (App. C/D). Motion level is controlled by p_geo, not restricted by SpheRoPE: on Wan2.1 (which uses the image p_geo without static clauses), our method achieves a higher VBench Dynamic Degree (DD) than PanoWan on Stress-20 (79.6 vs. 66.4). Following the reviewer's suggestion, we evaluate DD on Stress-20, whose prompts request dynamic scenes. On LTX 2.3, dropping the static-camera/scene clauses from p_geo (motion-permissive p_geo) raises Stress-20 DD from 13.2 to 51.8, on par with SphereDiff (53.2), which it beats on all 6 VBench metrics, while keeping the seam closed (Seam DS 1.09 vs. 10.57 for vanilla LTX 2.3) and leading all baselines in IQ, motion smoothness, and subject consistency:

Stress-20IQ↑CLIP↑Flicker↑Motion↑Subj.↑BG↑DD↑
360DVD (LoRA)50.1928.6298.6698.4293.3094.8239.6
PanoWan (LoRA)39.4928.4297.1198.5891.6493.6666.4
DynamicScaler (optimization)50.6532.3994.4596.1791.5893.2563.9
SphereDiff (optimization)33.8228.2096.5297.7288.5593.3453.2
Ours (LTX 2.3, zero-shot)51.7628.9198.9199.4696.5196.0213.2
Ours, motion-permissive p_geo52.2529.8197.8399.1794.2694.5451.8

Changes in the revision:

  • Padding off/on pairs with Seam-FID next to Table 4.
  • Vanilla LTX 2.3, Wan2.1, and the motion-permissive p_geo variant alongside Dynamic Degree (DD) in Table 2 (and in the appendix), with the static-camera p_geo stated.

We thank the reviewer once again.

[1] Team Wan, "Wan: Open and Advanced Large-Scale Video Generative Models," arXiv:2503.20314, 2025.

Dear Reviewer R39c,

We thank you once again for your insightful feedback on our paper.

As the Author-Reviewer discussion period is drawing to a close, we would like to ask whether our responses have adequately addressed your concerns. Following your questions, we added padding off/on pairs for vanilla FLUX.2 and our method, vanilla LTX 2.3, the full Wan2.1 results on PanoWan's checkpoint, and VBench Dynamic Degree for all Table 2 methods. We would be happy to clarify anything further. We greatly appreciate your time and consideration.

Best, Authors

Reviewer_r3P5OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes SpheRoPE, a training-free and optimization-free method for 360° panorama generation with pre-trained diffusion transformers. It makes two inference-time changes. First, it modifies the width-axis RoPE: high-frequency channels are snapped to integer harmonics so that the left and right borders connect, and low-frequency channels are replaced with spherical Cartesian coordinates so that each pole converges to a single point. Second, it adds a third CFG branch conditioned on the user prompt plus a fixed geometric prompt to steer the output toward ERP geometry. The method is applied to FLUX.1/FLUX.2 for images and LTX 2.3 for video, and is reported to achieve competitive or better results than training-based and optimization-based baselines while running much faster than multi-view fusion methods.

Critical strengths
  1. Simple and practical approach. By modifying only the positional encoding and the guidance at inference time, the method turns existing foundation models into panorama generators at no training cost. It transfers directly to new backbones and inherits their native capabilities, such as image-conditioned generation and joint audio-video generation.

  2. Frequency-aware design of the positional encoding. Treating high-frequency and low-frequency RoPE channels differently is a reasonable design choice, and the ablation in Figure 7 illustrates the failure modes of applying either encoding to all channels.

  3. Efficiency and broad evaluation. As a single-pass method, it is substantially faster than multi-view fusion approaches (about 20× faster than SphereDiff). The evaluation covers images and videos, ten image baselines, an LLM-based evaluation, a user study, and multiple hyperparameter ablations, and the code is publicly available.

Critical weaknesses
  1. The comparison with baselines is largely a comparison of backbones. In Table 4, vanilla FLUX.2 guided only by a panoramic prompt already reaches FAED 33.38, FID 26.05, KID 1.06, and IS 14.24, which is better than most baselines in Table 1. With the full method, FID barely changes (25.33) and IS decreases (12.75). When the backbone is matched, the picture is different: with FLUX.1, the proposed method obtains FID 45.22 and KID 2.63, considerably worse than DiT360 (23.28 / 0.92), a training-based method built on the same backbone, and it is better only on FAED. About half of the image baselines are based on Stable Diffusion-era models, and the video baselines (e.g., 360DVD) are also built on much weaker backbones than LTX 2.3. Therefore, the claim that a zero-shot method can match or outperform trained baselines is not sufficiently supported, as a large part of the reported advantage may come from the stronger backbone.

  2. The effect of Semantic Distortion CFG is not validated against simple alternatives. Eq. 4 adds a guidance term toward a prompt that appends geometric keywords to the user prompt, which is a direct application of multi-condition CFG. The most direct baseline, appending the same geometric prompt to the user prompt and using standard CFG only, is missing, and it is unclear whether the "panoramic prompt" used for the vanilla baseline in Table 4 is the same p_geo. Moreover, FLUX.1-dev and FLUX.2-dev are guidance-distilled models that by default do not compute an unconditional branch, so it is unclear how ε_uncond is obtained, and the stated 1.5× NFE overhead appears to be closer to 3× relative to the default inference of these models.

  3. The technical contribution beyond existing and concurrent techniques is unclear. Most components of the method have precedents. Modifying the RoPE of pre-trained DiTs at inference time, including treating individual frequency components differently, has been explored in training-free extrapolation methods (e.g., RIFLEx and DyPE); spherical or circular coordinates in RoPE have been used in prior work (e.g., IaaW); circular VAE padding is adopted from 360Anything; and Semantic Distortion CFG is a form of multi-condition guidance. Snapping only the high-frequency width channels to integer harmonics to close the seam is also used in concurrent work (SHERPA). I do not hold the concurrent work against the paper, but taken together, the contribution appears to be mainly a combination of known techniques, with the spherical encoding of low-frequency channels being the main new element. The authors should more clearly state and discuss the technical increment of their method relative to these works.

Questions for authors
  1. Backbone-matched comparisons. Could the authors report vanilla FLUX.1 results under the same protocol as Table 4, and compare against training-based methods built on the same backbone (e.g., DiT360, or other open-source FLUX-based panorama models)? For video, could the authors report vanilla LTX 2.3 with a panoramic prompt? These comparisons would show how much of the improvement comes from SpheRoPE itself rather than from the backbone.

  2. Validation and cost of Semantic Distortion CFG. Could the authors add a baseline in which p_geo is appended to the user prompt and only standard CFG is used, and clarify which prompt the vanilla baseline in Table 4 uses? For the guidance-distilled FLUX models, how is ε_uncond computed, and what is the actual inference overhead compared with the default sampling of each backbone?

5 comments

We thank the reviewer for the constructive feedback, and for pointing out our method's simplicity, frequency-aware RoPE design, and >20× speedup over SphereDiff. Below we address the raised concerns.

SpheRoPE Moves With the Backbone (W1 & Q1)

We agree that the backbone matters for quality, and this is the premise of our method. SpheRoPE is training-free, so any RoPE-based DiT becomes a panorama generator as soon as it is released, and every gain in foundation models transfers directly. The questions are whether SpheRoPE adds a consistent gain on top of each backbone and how it compares with training on the same backbone. Following the reviewer's request, we evaluate vanilla FLUX.1 and an additional backbone, Qwen-Image [1], under the Table 4 protocol (ODI-SR, n = 1,200), next to the trained methods built on FLUX.1:

Method (backbone, type)FAED↓DS↓FID↓KID×10²↓IS↑CS↑
LayerPano3D (FLUX.1, LoRA)55.710.7331.581.2210.4818.44
DiT360 (FLUX.1, LoRA)43.422.5923.280.9212.9018.52
Ours (FLUX.1, zero-shot)29.940.8945.222.6310.3218.52
Vanilla FLUX.119.998.6432.611.4612.6918.47
Vanilla FLUX.1, fisheye prompt22.0510.1534.441.5213.1818.48
Vanilla FLUX.233.381.3726.051.0614.2418.42
Ours (FLUX.2, zero-shot)25.400.9425.330.9012.7518.43
Ours (FLUX.2, +LoRA)27.440.8416.630.3912.1718.35
Vanilla Qwen-Image68.301.1421.430.7111.1318.51
Ours (Qwen-Image, zero-shot)19.630.7132.771.3912.6418.55
  • Across backbones: our method closes the seam on every backbone (DS 8.64 → 0.89 on FLUX.1, 1.37 → 0.94 on FLUX.2, 1.14 → 0.71 on Qwen-Image, and 6.44 → 1.03 and 10.57 → 1.48 on LTX 2.3 video below, vs. 1.07 for real panoramas). It also converges the poles wherever the vanilla model leaves them open: Pole std, the pixel std along the top/bottom rows (0 for an exact pole, C2), drops from 24.6/27.8 to 4.0/11.1 on FLUX.1, vs. 6.0/11.4 for real panoramas.
  • Trained FLUX.1 methods: DiT360's dataset-matched training gives better per-crop FID/KID, but its planar RoPE leaves a strong seam (DS 2.59 vs. 0.89), and we improve FAED by 13.48 points. LayerPano3D also closes the seam (0.73), at FAED 55.71 vs. 29.94.
  • Training composes with SpheRoPE: LoRA fine-tuning FLUX.2 with SpheRoPE on Matterport (App. B.5) lowers FID from 25.33 to 16.63, KID×10² from 0.90 to 0.39, and DS from 0.94 to 0.84.
  • Vanilla FLUX.1 reaches a low FAED without producing a panorama. When presented with out-of-distribution, non-panorama images, FAED can rank them above valid panoramas: re-rendering our own FLUX.1 panoramas as 150° perspective photos lowers their FAED from 29.9 to 20.8. Prompted for a "180-degree fisheye photograph", vanilla FLUX.1 scores almost the same (22.05 vs. 19.99). FAED averages features over longitude and gives the pole rows near-zero weight (Oh et al., ECCV 2022), so it misses the open seam (DS 8.64 and 10.15) and the unconverged poles (24.6/27.8 and 43.7/35.5). FID on the 8 horizon views also favors such wide single views (32.6 vs. 45.2), while at ±60° pitch and on the up/down faces ours is better (47.9 vs. 50.5, 78.5 vs. 80.9). Per-crop FID/KID also favor stock Qwen-Image, as they score each crop as a perspective photo. FAED is therefore informative only among valid panoramas (closed seam, converged poles). There SpheRoPE is better: FLUX.2 33.38 → 25.40, Qwen-Image 68.30 → 19.63, and the best closed-seam FAED on FLUX.1. App. B.2 of the uploaded revised PDF shows examples and states this scope of FAED.

Video (LTX 2.3). We compare vanilla LTX 2.3 with our Table 2 row under identical prompts, sampler, and seeds (Seam DS = per-frame ERP DS). Ours improves all 6 VBench metrics on SphereDiff-20 and 5/6 on Stress-20, closes the seam, and converges the poles (Pole std 17.9/16.1 → 6.1/6.5 on SphereDiff-20 and 38.5/24.3 → 16.8/11.0 on Stress-20):

SetMethodIQ↑CLIP↑Flicker↑Motion↑Subj.↑BG↑Seam DS↓
SphereDiff-20Vanilla LTX 2.353.1526.6499.1799.4593.3195.426.44
Ours (Table 2)53.6427.9599.5899.6298.2297.711.03
Stress-20Vanilla LTX 2.343.6029.3496.6398.5486.5991.5410.57
Ours (Table 2)51.7628.9198.9199.4696.5196.021.48

In light of the reviewer's comment, in the revision we replace the claim of outperforming trained baselines (abstract, contribution 3, Sec. 4.3) with: SpheRoPE is competitive with trained methods on the same backbone and is preferred by users over every baseline in our user study (Table 3). We also add a backbone column to Tables 1 and 2.

[1] Qwen Team, "Qwen-Image Technical Report," arXiv:2508.02324, 2025.

Validation and Cost of Semantic Distortion CFG (W2 & Q2)

In Table 4, vanilla FLUX.2 uses a per-caption panoramic rewrite ("This 360-degree equirectangular panorama captures …"), not p_geo. Appending p_geo directly to p under standard 1-way CFG forces a single scale w_sem to steer both scene content and ERP geometry simultaneously (the coupled special case γ = w_sem of Eq. 4). Decoupling (ε_geo − ε_cond) via γ enforces ERP geometry without diluting prompt adherence. In Table 4's setting, appending yields FAED 33.33, whereas decoupling (Ours, γ = 6) reaches 25.40 and also improves DS, FID and KID, at similar CS and lower IS:

Setting (Table 4, FLUX.2, SpheRoPE)FAED↓DS↓FID↓KID×10²↓IS↑CS↑
p_geo appended, standard CFG33.330.9928.711.3414.0018.46
SD-CFG, γ = 6 (Ours, Table 4)25.400.9425.330.9012.7518.43

Regarding ε_uncond and NFE: for standard-CFG backbones (Qwen-Image, Wan2.1), SD-CFG costs 1.5× NFE (3 vs. 2 NFEs/step). For guidance-distilled FLUX.1/2-dev (1 NFE/step default), evaluating ε_uncond (empty prompt at distilled guidance scale: 3.5 for FLUX.1, 4.0 for FLUX.2), ε_cond, and ε_geo is 3× NFE relative to default sampling. SpheRoPE itself adds no overhead.

We will clarify this in Sec. 3.3 and App. D. Even at 3× NFE, our runtime remains fast (1024×512 on a B200 GPU):

Method (type)s/gen↓
Diffusion360 / SMGD / DiT360 (full FT / LoRA)0.67 / 1.86 / 2.67
Ours (FLUX.1, zero-shot)4.75
PAR / UniPano (full FT / LoRA)8.05 / 8.47
Ours (FLUX.2, zero-shot)16.64
StitchDiffusion / PanFusion / Text2Light (LoRA / full FT)18.60 / 21.51 / 24.61

Technical Contribution (W3)

  • New: the low-frequency spherical Cartesian width band with its frequency split rule (C1 and C2, Sec. 3.2), and the anchored geometric guidance of SD-CFG (Sec. 3.3). Harmonic snapping of the high band is also used by concurrent SHERPA.
  • Borrowed: circular VAE padding (360Anything, App. B.1).
  • Result: zero-shot 360° generation on a frozen DiT, where RIFLEx and DyPE stay planar, and IaaW and SHERPA require training.

This statement will go into Sec. 2 of the revision. In more detail:

  • Why zero-shot RoPE modification on S² is hard: Modifying a frozen DiT's RoPE at inference time can be fragile, as naive geometric edits push attention phases out of distribution (OOD) and collapse generation (Fig. 7). Prior frequency-aware edits (RIFLEx, DyPE) only rescale 1D/planar length on an unchanged Euclidean grid. Bending a planar grid into a closed spherical manifold (S²) zero-shot required identifying that high and low frequencies fail in opposite ways: high frequencies need linear phase continuity across the seam (C1), while low frequencies govern global layout and require polar convergence (C2) without large spectral shifts.

  • vs. IaaW [2]: IaaW maps the entire domain to sphere angles and multiplicatively entangles latitude, longitude, and time in every channel (their Eq. 2), destroying the pretrained factorized RoPE and requiring full fine-tuning. SpheRoPE modifies only the low-frequency width band with bounded Cartesian coordinates whose local step rate matches the pretrained 1-unit/column step (0.043 rad RMS deviation at 1-token distance), keeping every channel in-distribution.

  • vs. Concurrent SHERPA: SHERPA requires trained adapters. It snaps half the width channels to integer harmonics and leaves the low band linear, relying on training to learn the poles (C2). Evaluating SHERPA's encoding without its trained adapters on FLUX.1 (its 14 highest-frequency width pairs snapped, the rest linear, under our SD-CFG and padding, n = 1,200) closes the border seam (DS 0.92 vs. 0.89) but leaves the poles unconverged (Pole std 19.1/27.3 vs. 4.0/11.1, Polar FID 66.8 vs. 47.9), and snapping its lower pairs to one or two cycles costs view quality (FID 51.48 vs. 45.22, Seam-FID 54.15 vs. 42.75). Snapping all channels instead pushes low frequencies OOD (Fig. 7a). Our low-frequency spherical Cartesian encoding is what satisfies C1 and C2 zero-shot.

Changes in the revision:

  • A backbone column and backbone-matched rows (vanilla FLUX.1, Qwen-Image, vanilla LTX 2.3) in Tables 1 and 2, and the softened claim in the abstract, contribution 3 and Sec. 4.3.
  • The appended-p_geo baseline in Table 4, and the NFE cost of SD-CFG in Sec. 3.3 and App. D.
  • App. B.2: why distribution metrics need a valid panorama, with the fisheye control and examples.
  • The contribution statement above in Sec. 2, with IaaW and SHERPA.

We thank the reviewer once again and hope the above responses address all concerns.

[2] Gui et al., "Image as a world: Generating interactive world from single image via panoramic video generation," NeurIPS 2025.

Dear Reviewer r3P5,

We thank you once again for your insightful feedback on our paper.

As the Author-Reviewer discussion period is drawing to a close, we would like to ask whether our responses have adequately addressed your concerns. Following your questions, we added backbone-matched comparisons (vanilla FLUX.1 and Qwen-Image next to the FLUX.1-based DiT360 and LayerPano3D, and vanilla LTX 2.3), the appended-p_geo baseline with the NFE cost of SD-CFG, and a clearer statement of our technical contribution. We would be happy to clarify anything further. We greatly appreciate your time and consideration.

Best, Authors

Comments on the paper

1 comment

Dear AC and Reviewers,

As the discussion period is coming to a close, we would like to briefly summarize the changes in the revised PDF. All additions respond to points raised in the reviews.

  • Claims (r3P5): contribution 3 and Sec. 4.3 now state that the method is competitive with trained methods, with a backbone-matched comparison in Sec. 4.3.
  • Backbone-matched comparisons (r3P5, R39c, KFg1): Tables 1 and 2 have a backbone column and plain-backbone rows (vanilla FLUX.1, FLUX.2, Qwen-Image, LTX 2.3, stock Wan2.1-1.3B). Table R2 compares with PanoWan on its own checkpoint.
  • Zero-shot alternatives (KFg1): Table R3 compares with rotate-while-denoising and SHERPA's encoding without its adapters on FLUX.1, and Table R4 with rotation on video. Both are discussed in Sec. 2.
  • SD-CFG (r3P5, KFg1): Sec. 3.3 relates it to Composable Diffusion and SEGA and states its cost (1.5× NFE, 3× for guidance-distilled FLUX). Table 4 adds the appended-prompt control.
  • Padding and prompts (R39c): the caption of Table 4 states both for every row, and Table R5 gives padding off/on pairs.
  • Video dynamics (R39c): Dynamic Degree and a motion-permissive prompt are reported in Sec. 4.3 and App. B.3.
  • Scope of FAED (r3P5): App. B.2, with Table R1 and Fig. R1.
  • Limitations (R39c): App. D now states the failure of the spherical band on Wan2.1-1.3B, the motion/stability trade-off, and the NFE cost.

Reviewers R39c and KFg1 have indicated that our responses addressed their concerns. Reviewer r3P5 indicated the same for SD-CFG and the technical contribution, and we have replied to the remaining point in that thread.

We thank the AC and all reviewers for their time and consideration.

Best,

Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.