Pith. sign in

REVIEW 2 major objections 5 minor 2 cited by

TanDiT: Tangent-Plane Diffusion Transformer for High-Quality 360{\deg} Panorama Generation

T0 review · 2 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read TanDiT generates high-quality 360° panoramas by producing a dense grid of tangent-plane perspective views in a single diffusion pass, then refining the reprojected panorama with a pretrained network for seamless, arbitrary-resolution…

desk verdict A well-ablated panorama generation system with a plausible tangent-plane grid design; the SOTA claim is real but leans on self-designed metrics and an under-checked refinement step. read the letter →

arxiv 2506.21681 v1 pith:BPRZFE5S submitted 2025-06-26 cs.CV cs.LG

Hakan Çapuk , Andrew Bond , Muhammed Burak K{i}z{i}l , Emir Göçen , Erkut Erdem , Aykut Erdem This is my paper
classification cs.CVcs.LG
keywords panoramageneration360-degreeimagestangent-planeprojectiondiffusiontransformertext-to-imageequirectangularrefinementevaluationmetricsloopconsistency
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

TanDiT claims that high-quality 360° panorama generation can be achieved by abandoning the warped equirectangular or cubemap canvas that earlier methods denoise, and instead decomposing the sphere into a grid of tangent-plane perspective views generated jointly by a single diffusion transformer in one denoising loop. The argument is that tangent planes are nearly undistorted, and that arranging them so spatial neighbors stay adjacent in the grid lets the model's attention enforce cross-view consistency directly, with an equirectangular-conditioned refinement step that later harmonizes overlaps while preserving global layout. If correct, one pipeline reuses off-the-shelf diffusion transformers and super-resolution models, reaches arbitrary resolutions up to 4K, and generalizes to out-of-domain styles and complex prompts. The paper also introduces two panorama-specific metrics, TangentIS and TangentFID, computed on tangent views rather than the distorted global projection, and releases captioned datasets and an evaluation suite to standardize comparison.

What carries the argument

The load-bearing object is the tangent-plane grid: 18 perspective views of the sphere obtained by gnomonic projection, each resized to $192\times192$ pixels and packed into a $576\times1152$ serpentine layout so that views that are neighbors on the sphere, with deliberate overlap, remain neighbors in the grid. A pretrained diffusion transformer (Stable Diffusion 3.5-Large, LoRA-fine-tuned) denoises the entire grid in one loop, so the same attention mechanism that models spatial proximity in ordinary images is what enforces coherence across views. The second mechanism is the equirectangular-conditioned refinement: the reprojected panorama's latent $z_{\mathrm{pano}}$ is perturbed with noise at timestep $\approx 800$ and denoised by the pretrained model under the same prompt, with circular padding of the horizontal edges at every step, plus latent rotation, to fix left-right loop consistency, and patched denoising of $1024\times1024$ squares for high resolution. The two metrics, TangentIS and TangentFID, compute Inception Score and FID on each of the 18 tangent views and aggregate with $\mu - 1.96\,\sigma/\sqrt{18}$ and $\mu + 1.96\,\sigma/\sqrt{18}$, respectively, so a model must be consistent across the whole sphere, poles included, to score well.

What would settle it

Fix a prompt set, then sweep the refinement timestep below roughly 700 and above roughly 900 while measuring two things directly: seam visibility near tangent-view boundaries and at the left-right edge (for example, with the paper's own Discontinuity Score), and layout preservation (for example, whether recognizable objects keep their positions between the pre- and post-refinement panoramas). If no timestep simultaneously removes the seams and preserves the layout, the refinement mechanism fails. A second test counts repeated or duplicated content across the $1024\times1024$ denoising patches, since the paper itself shows a 'two moons' artifact, and checks how often patched refinement produces such repetitions on dense prompts at 2K and 4K.

Watch

Extended reading notes

Core claim

The central claim is that a single diffusion transformer, fine-tuned to denoise a structured grid of 18 tangent-plane views, can synthesize a complete 360° scene in one pass with better coherence than multi-branch, stitching, or blending methods, and that this is the only current approach that combines tangent-plane representations, consistency-aware generation, and arbitrary resolution support in a single diffusion loop. Tangent planes are perspective views extracted by gnomonic projection, which preserves straight lines and keeps distortion low; neighboring views overlap, and the serpentine grid layout keeps sphere-adjacent views grid-adjacent so attention can propagate local context across them. At inference the generated grid is reprojected to an equirectangular panorama (with optional per-view super-resolution), its latent is noised at a high timestep of roughly 800 and re-denoised by a pretrained diffusion transformer conditioned on the text prompt, and circular padding plus patched denoising over $1024\times1024$ tiles deliver loop consistency and resolutions up to $4096\times2048$. On the evaluation side, the paper argues that standard metrics misjudge panoramas because the Inception network and CLIP expect perspective images, and it proposes TangentIS and TangentFID, which score each tangent view separately and aggregate with the lower or upper 95% confidence bound over the 18 views, so that failure in polar regions cannot be masked by strong equatorial regions.

Load-bearing premise

The load-bearing premise is that the refinement stage can wash out seams and mismatches between tangent views by adding noise at a high timestep ($\approx 800$) to the reprojected panorama's latent and denoising it with a pretrained model, without disturbing the global scene layout or the prompt's content; the paper's own appendix shows the first stage alone produces visible artifacts, so if that balance between artifact removal and layout preservation fails, the coherence and quality claims collapse.

Editorial extensions

If this is right

  • Panorama generation inherits the quality and style range of front-line text-to-image diffusion transformers, because the model denoises undistorted perspective views rather than a warped sphere.
  • A single joint denoising pass over the tangent grid replaces multi-branch or iterative stitching pipelines, which the paper argues is what removes polar distortions and seam artifacts.
  • The same trained weights produce arbitrary output resolutions: per-view super-resolution plus patched refinement reaches at least $4096\times2048$ without retraining.
  • Evaluating on tangent views rather than the equirectangular image or cubemap changes rankings among methods; the paper shows OmniFID can rate an inconsistent, unrefined first stage above the full refined pipeline, while the proposed metrics penalize exactly that.
  • The released captions, metrics, and evaluation scripts give the field a common benchmark, so future panorama generators can be compared on identical prompts and settings.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The joint grid-generation recipe is a template beyond panoramas: any output that is a set of correlated views, such as multi-view imagery, video, or light fields, could be arranged as a serpentine latent grid and generated in one diffusion pass, with the grid layout serving as the geometry that the transformer's attention consumes.
  • Because the refinement stage is model-agnostic and needs only a pretrained diffusion model plus circular padding, a direct transfer test would be to run it on other generators' outputs, for example PanFusion or Diffusion360 panoramas, and measure whether TangentIS and TangentFID improve; if it does, the refinement, not the tangent grid, is the load-bearing contribution for consistency.
  • The confidence-bound aggregation trick is separable from both metrics: the same recipe, per-view scores on tangent planes aggregated by confidence bound, could be applied to CLIP score, FAED, or future quality measures to build a family of panorama-aware metrics without redesigning each base metric.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

2 major / 5 minor

Summary. This paper introduces TanDiT, a method for text-conditioned 360° panorama generation. It decomposes a panorama into an 18-view tangent-plane grid, trains a single diffusion transformer (LoRA-fine-tuned Stable Diffusion 3.5 Large) to generate the entire grid in one denoising pass, and then applies per-view super-resolution, an equirectangular-conditioned refinement pass with a pretrained SD3 model, circular padding, and patched denoising to harmonize seams and enable high-resolution output. The paper also proposes two new evaluation metrics (TangentIS and TangentFID), a benchmark with newly captioned datasets, and a user study. The central claim is that TanDiT achieves state-of-the-art quality and consistency in 360° panorama generation.

Significance. If the results hold, TanDiT offers a practically useful unified pipeline for 360° generation, with the tangent-plane representation reducing distortion and the refinement stage improving coherence. The paper's strengths are its extensive ablations (grid ordering, number of planes, refinement timestep, super-resolution, control experiments with pretrained SD3), a user study, and a concrete release plan for code, captions, and evaluation scripts. The proposed metrics and benchmark could become a community resource, but their value depends on external validation. The main risk is that the consistency claim rests on a refinement stage whose layout-preservation behavior is not directly measured; this, together with the validation of the new metrics, is the load-bearing point that needs strengthening.

major comments (2)
  1. [Section 3.3, Appendix E.4, Appendix I] The claim in Section 3.3 that the refinement step 'wash[es] out high-frequency mismatches while preserving the scene’s overall layout' is not directly tested. Since the first stage alone is admitted to produce visible inconsistencies (Appendix B.3, Fig. 6), the consistency and state-of-the-art claims rest entirely on this refinement. Table 5 reports only aggregate metrics; no layout-preservation check (e.g., correspondence, depth, or structural similarity between the intermediate ERP and the final output) is provided. Appendix E.4 shows that t≈900 degrades panorama-specific metrics, and Appendix I documents duplicate content (two moons) from patched denoising, demonstrating that layout/semantics can change. The authors should add a direct layout-preservation evaluation and an analysis of when the balance between artifact removal and layout preservation fails, or temper the claim accordingly.
  2. [Section 4.2, Tables 2, 5, 8, 9] TangentIS and TangentFID (Section 4.2) are introduced and used both for hyperparameter selection (grid ordering in Table 8, number of planes in Table 9, refinement timestep in Table 5) and as primary evidence for state-of-the-art performance in Table 2. No external validation of these metrics is provided: there is no demonstration that they correlate with human judgments (the user study in Section 5.2 is a single overall preference, not a per-metric validation), no error bars, and no sensitivity analysis. Figure 6 is a single anecdotal example. Since the method's components were tuned on these metrics, the state-of-the-art claim is vulnerable to circularity. Please validate the metrics against human preference (e.g., using the user-study data) and report confidence intervals or multiple-seed results for all metrics.
minor comments (5)
  1. [Section 5.1, Table 2] The sentence 'our model achieves the best or second-best accuracy in most metrics' is inaccurate; on KID, IS, CS, and OmniFID the model is not in the top two. Please revise the sentence and add a sentence explaining the OmniFID result in light of the paper's critique of that metric.
  2. [Section 5.2] The phrase 'with p < 0.05 according to the Wilson interval' conflates a confidence interval with a hypothesis test; please either report a test statistic or state that the 95% Wilson intervals for all comparisons exclude 50%.
  3. [Throughout] There are several typos: 'opposite opposite' in Section 2, 'seperately' in Section 3.3, 'seperate' in Appendix E.2, 'inbold' in Table 3 caption, and 'desinging' in Section 3.2.
  4. [Table 1] The '#' column for TanDiT reads '1/2'; please clarify its meaning (one generation pass plus one refinement pass?) and define what 'Consistent' means for the full pipeline, since the first stage alone is not consistent (Appendix B.3).
  5. [Section 4.2] The upper 95% confidence bound for TangentFID is mentioned but not defined; please provide the formula, analogous to Eq. (6) for TangentIS.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation chain: TanDiT's SOTA claim is backed by external metrics and a user study; the author-proposed metrics and design tuning raise only a mild self-referential benchmarking concern, not circularity.

full rationale

The central claim that TanDiT achieves state-of-the-art panorama generation is supported by standard external metrics (FID, KID, IS, CLIP Score, FAED, OmniFID, DS) and a 38-participant 2AFC user study (Section 5.2, Appendix H), not by a fitted parameter renamed as a prediction. The proposed TangentIS and TangentFID are defined on tangent views extracted from the final equirectangular image (Section 4.2, Appendix B.3), not on the model's internal grid, so they are an evaluation instrument rather than part of the generative derivation. Some design choices, such as grid ordering and the refinement timestep, were selected using these and other panoramic metrics (Appendix E.4, E.7, Tables 5 and 8), which is normal hyperparameter selection rather than a construction-level equivalence; the paper does not claim those tuning results as independent predictions. The refinement step's premise that high-timestep noise at Thigh≈800 washes out seams while preserving layout (Eq. 3) is an empirical assumption with explicit trade-off evidence in Table 5 and qualitative figures, not a circular definition. The admitted limitations, including duplicate-content artifacts from patched denoising (Appendix I, Figure 19) and the absence of an explicit cross-view consistency mechanism (Conclusion), are correctness and robustness concerns, not evidence that the derivation reduces to its inputs. The paper contains no load-bearing self-citation or imported uniqueness theorem: its citations to prior work such as MakeAnything [24] are external and used for grid-layout inspiration. Overall, the claimed method and its evaluation remain self-contained, with no step where an output is equivalent to an input by construction.

Assumptions & free parameters 6 free parameters · 6 assumptions · 0 invented entities

The central claim depends on standard projection geometry, several domain assumptions about latent-space continuity and refinement behavior, and hand-chosen hyperparameters. The refinement timestep assumption is the most fragile; the paper's own limitations section admits the first stage alone is inconsistent, so the refinement stage must work as intended.

free parameters (6)
  • Number of tangent planes = 18
    Chosen by ablation in Appendix E.8 (Table 9); affects grid structure and output quality.
  • Refinement noise timestep (Thigh) = ~800
    Ablation in Appendix E.4 (Table 5); balances artifact removal vs. structure preservation.
  • Grid ordering = custom (polar rows top, equator middle)
    Ablation in Appendix E.7 (Table 8); custom ordering outperforms row-wise and column-wise on panoramic metrics.
  • Super-resolution scale = 2x
    Default upscale factor per tangent view; 4x used for 4K outputs (Appendix E.9, Table 10).
  • Guidance scale = 7.0
    Inference hyperparameter in Appendix C.4; no ablation reported.
  • Number of denoising steps = 28 first stage, 17 refinement
    Inference settings in Appendix C.4 and D; refinement steps are tied to the timestep choice.
assumptions (6)
  • standard math Gnomonic projection formulas map sphere points to tangent planes (Eqs. 7-9).
    Standard map projection; cited to Snyder [36].
  • domain assumption The VAE latent space preserves spatial structure so that adjacent tangent views in the grid remain locally related in latent space.
    Section 3.2 states 'Since the VAE-based latent space preserves spatial structure, the DiT can effectively capture and propagate local dependencies across tangent views.'
  • domain assumption 18 tangent views with overlap allow near-lossless reprojection to an equirectangular panorama.
    Section 4.1: 'Following [31], we create sets of 18 tangent planes... aiming to be able to perform a near-lossless projection back to the panorama.'
  • domain assumption LLava-OneVision captions accurately describe the training images.
    Section 4.1: captions are generated automatically with no human verification stated.
  • domain assumption The refinement step with noise at timestep about 800 preserves scene layout while removing seams.
    Section 3.3: 'By using a high timestep, we wash out high-frequency mismatches while preserving the scene's overall layout.' This is the load-bearing assumption.
  • domain assumption TangentIS and TangentFID are valid proxies for panoramic image quality.
    Section 4.2: proposed metrics are validated only via one qualitative figure (Figure 6) and a distortion analysis in Appendix B.3.

how reviews work

0 comments
Cite this review

Pith. "Pith review of TanDiT: Tangent-Plane Diffusion Transformer for High-Quality 360{\deg} Panorama Generation." pith.science (2026). https://pith.science/paper/BPRZFE5S

@misc{pith2026250621681,
  author       = {Pith},
  title        = {Pith review of: TanDiT: Tangent-Plane Diffusion Transformer for High-Quality 360\deg Panorama Generation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/BPRZFE5S}},
  note         = {Machine review of arXiv:2506.21681}
}
abstract

Recent advances in image generation have led to remarkable improvements in synthesizing perspective images. However, these models still struggle with panoramic image generation due to unique challenges, including varying levels of geometric distortion and the requirement for seamless loop-consistency. To address these issues while leveraging the strengths of the existing models, we introduce TanDiT, a method that synthesizes panoramic scenes by generating grids of tangent-plane images covering the entire 360$^\circ$ view. Unlike previous methods relying on multiple diffusion branches, TanDiT utilizes a unified diffusion model trained to produce these tangent-plane images simultaneously within a single denoising iteration. Furthermore, we propose a model-agnostic post-processing step specifically designed to enhance global coherence across the generated panoramas. To accurately assess panoramic image quality, we also present two specialized metrics, TangentIS and TangentFID, and provide a comprehensive benchmark comprising captioned panoramic datasets and standardized evaluation scripts. Extensive experiments demonstrate that our method generalizes effectively beyond its training data, robustly interprets detailed and complex text prompts, and seamlessly integrates with various generative models to yield high-quality, diverse panoramic images.

Figures

Figures reproduced from arXiv: 2506.21681 by the authors.

Figure 1
Figure 1. Panoramic samples generated by our TanDiT model at varying resolutions and styles. TanDiT produces visually compelling and seamlessly consistent 360◦ panoramas across various resolutions, even at 4K, and diverse visual styles. By decomposing spherical images into multiple tangent-plane (perspective) views, our model seamlessly integrates powerful off-the-shelf diffusion transformers and super-resolution methods, ena… view at source ↗
Figure 2
Figure 2. Overview of the TanDiT training pipeline. Our method starts by decomposing a 360◦ panoramic image into a structured grid of tangent-plane projections via gnomonic projection. These projections are arranged into a single coherent grid image, ensuring adjacent placement of overlapping regions for spatial consistency. Given a dense textual caption describing the scene, the model is trained to reconstruct this grid usin… view at source ↗
Figure 3
Figure 3. Overview of the TanDiT inference pipeline. At inference time, TanDiT first generates a grid of tangent views conditioned on a text prompt. These tangent views are enhanced using a super-resolution module and then reprojected to form an intermediate equirectangular panorama. To further improve global coherence and visual quality, the latent representation of this panorama is perturbed with noise and refined by a pre-… view at source ↗
Figures from the paper (17 more)
Figure 4
Figure 4. Figure 4: Qualitative comparison of panoramic image generation methods across various scene types. StitchDiffusion produces outputs that lack fine detail and fails on the out-of-domain fireworks example. PanFusion and Diffusion360 suffer from polar distortions and seam artifacts…
Figure 5
Figure 5. Figure 5: Stylized panoramic images generated by TanDiT. Examples demonstrating the model’s ability to handle a wide range of visual domains: From left to right, top to bottom: (1) traditional painting, (2) Minecraft-style rendering, (3) black-and-white charcoal sketch, and (4) …
Figure 6
Figure 6. Figure 6: Qualitative results illustrating the limitations of standard panoramic evaluation metrics. These images are taken from our model’s first-stage outputs, with no refinement stage. There are several noticeable inconsistencies, both in the equatorial and polar regions of t…
Figure 7
Figure 7. Figure 7: Generated panorama by timestep during inference of TanDiT. TanDit maps random noise to a grid of tangent planes conditioned on a given text. C.2 Model Configurations DiT Backbone. We adopt Stable Diffusion 3.5 Large (SD3) [3] as the backbone for our diffusion transform…
Figure 8
Figure 8. Figure 8: Example panoramas before and after applying the Equirectangular Conditioned Refinement step. Refinement step mitigates the artifacts and inconsistencies arising from projecting the tangent planes to a panoramic image, and enables the generation of a visually coherent p…
Figure 13
Figure 13. Figure 13: (a) Without Patched Denoising (b) With Patched Denoising [PITH_FULL_IMAGE:figures/full_fig_p019_13.png]
Figure 9
Figure 9. Figure 9: Effect of patched denoising on high-resolution panoramic generation. (a) Without patched denoising, the model struggles to maintain texture quality and consistency across the image, resulting in visible artifacts and degradation near the boundaries. (b) With patched de…
Figure 10
Figure 10. Figure 10: Effect of latent rotation and circular padding in patched denoising. Images are horizontally rotated by 90° to better visualize the loop-consistency. Original left-right loop regions are marked with red. (a) Without circular padding, there is no sufficient information…
Figure 11
Figure 11. Figure 11: Effect of timestep in the Equirectangular-Conditioned Refinement process. We visualize the effect of different noise levels used during refinement in TanDiT. Starting from the initial equirectangular panorama, we apply refinement using noise injected at various timest…
Figure 12
Figure 12. Figure 12: A comparison of panorama representations. (a) Tangent planes are arranged row-wise from top to bottom. (b) Tangent planes from the bottom row of the equirectangular projection are moved to the top row of the tangent grid. This adjustment preserves the spatial order of…
Figure 13
Figure 13. Figure 13: A 4k image generated by TanDiT. Employing the proposed patched denoising technique in the Equirectangular-Conditioned Refinement step, we are able to generate 4k images. wide variety of visually distinct aesthetics, significantly broadening TanDiT’s potential applicat…
Figure 14
Figure 14. Figure 14: Out-of-domain stylized panoramas generated by TanDiT. TanDiT, combined with the Equirectangular-Conditioned Refinement method is able to generalize to many out-of-domain styles. visualizations demonstrate that our refinement strategy, combined with circular padding, e…
Figure 15
Figure 15. Figure 15: Example of inconsistency in CubeDiff’s panorama generation. The left image shows the perspective input used to condition CubeDiff. The right image is the resulting equirectangular panorama, where a visible seam appears near the center due to misalignment and distortio…
Figure 16
Figure 16. Figure 16: Loop Consistency in the refined panoramas. Refined panorama images are rotated by 90 degrees horizontally and marked with a red rectangle to qualitatively show the left-right continuity. Utilizing circular padding in the refinement step allows for a sufficient informa…
Figure 17
Figure 17. Figure 17: Screenshot of the survey layout. The participants are asked to choose the best generation result with high relevance to the text prompt, realism, seam consistency, coverage, and overall impression. I Further Limitations In this section, we expand on the technical and …
Figure 18
Figure 18. Figure 18: Results of the pairwise user preference study comparing our method against four baselines. Each point shows the percentage of times our approach was preferred over a given baseline, along with a 95% Wilson confidence interval. Across all comparisons, our method is sig…
Figure 19
Figure 19. Figure 19: A limitation of the patched denoising approach. In this example, two patches indepen￾dently generated a moon, resulting in duplicate content in the final image. moons generated in different patches of the sky. While the panorama remains globally consistent and structu…

Discussion (0). Sign in to comment.

Forward citations

Cited by 2 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. 360Anything: Geometry-Free Lifting of Images and Videos to 360{\deg}

    cs.CV 2026-01 conditional novelty 6.0 of 10

    360Anything lifts perspective images and videos to 360° panoramas with a diffusion transformer and sequence concatenation, requiring no camera metadata at test time.

  2. Gimbal360: Canonicalizing Planar Diffusion for Spherical Panorama Completion

    cs.CV 2026-03 reject novelty 5.0 of 10

    Gimbal360 completes 360° panoramas from unposed perspective images by rigidly auto-leveling inputs and training diffusion with a Siamese shift-equivariance loss to preserve ERP seam continuity.

Reference graph

Works this paper leans on

39 extracted references · 27 canonical work pages · cited by 2 Pith papers

  1. [1]

    High- resolution image synthesis with latent diffusion models

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High- resolution image synthesis with latent diffusion models. In Proc. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022

    work page 2022

  2. [2]

    Hierarchical text-conditional image generation with CLIP latents

    Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with CLIP latents. arXiv preprint arXiv:2204.06125, 2022

    arXiv 2022

  3. [3]

    Scaling rectified flow transformers for high-resolution image synthesis

    Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, Dustin Podell, Tim Dockhorn, Zion English, Kyle Lacey, Alex Goodwin, Yannik Marek, and Robin Rombach. Scaling rectified flow transformers for high-resolution image synthesis. In Proc. International Conference on...

    work page 2024

  4. [4]

    Black Forest Labs. Flux. https://github.com/black-forest-labs/flux , 2024

    2024

  5. [5]

    MultiDiffusion: Fusing diffusion paths for controlled image generation

    Omer Bar-Tal, Lior Yariv, Yaron Lipman, and Tali Dekel. MultiDiffusion: Fusing diffusion paths for controlled image generation. In Proc. International Conference on Machine Learning (ICML), 2023

    work page 2023

  6. [6]

    Customizing 360-degree panoramas through text-to-image diffusion models

    Hai Wang, Xiaoyu Xiang, Yuchen Fan, and Jing-Hao Xue. Customizing 360-degree panoramas through text-to-image diffusion models. In Proc. IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 4921–4931, 2023

    work page 2023

  7. [7]

    Phung, Wanli Ouyang, and Jianfei Cai

    Cheng Zhang, Qianyi Wu, Camilo Cruz Gambardella, Xiaoshui Huang, Dinh Q. Phung, Wanli Ouyang, and Jianfei Cai. Taming stable diffusion for text to 360° panorama image generation. In Proc. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6347–6357, 2024

    work page 2024

  8. [8]

    Diffusion360: Seamless 360 de- gree panoramic image generation based on diffusion models

    Mengyang Feng, Jinlin Liu, Miaomiao Cui, and Xuansong Xie. Diffusion360: Seamless 360 de- gree panoramic image generation based on diffusion models. arXiv preprint arXiv:2311.13141, 2023

    arXiv 2023

Show all 39 references
  1. [9]

    360-degree panorama generation from few unregistered NFoV images

    Jiong-Qi Wang, Ziyu Chen, Jun Ling, Rong Xie, and Li Song. 360-degree panorama generation from few unregistered NFoV images. Proc. ACM International Conference on Multimedia (ACM MM), 2023

    2023

  2. [10]

    MVDiffusion: Enabling holistic multi-view image generation with correspondence-aware diffusion

    Shitao Tang, Fuyang Zhang, Jiacheng Chen, Peng Wang, and Yasutaka Furukawa. MVDiffusion: Enabling holistic multi-view image generation with correspondence-aware diffusion. In Proc. Conference on Neural Information Processing Systems (NeurIPS), 2023

    2023

  3. [11]

    Curved Diffusion: A generative model with optical geometry control

    Andrey V oynov, Amir Hertz, Moab Arar, Shlomi Fruchter, and Daniel Cohen-Or. Curved Diffusion: A generative model with optical geometry control. In Proc. European Conference on Computer Vision (ECCV), 2023

    2023

  4. [12]

    Aoming Liu, Zhong Li, Zhang Chen, Nannan Li, Yi Xu, and Bryan A. Plummer. Panofree: Tuning-free holistic multi-view image generation with cross-view self-guidance. In Proc. European Conference on Computer Vision (ECCV), 2024. 10

    2024

  5. [13]

    CubeDiff: Repurposing diffusion-based image models for panorama generation

    Nikolai Kalischek, Michael Oechsle, Fabian Manhardt, Philipp Henzler, Konrad Schindler, and Federico Tombari. CubeDiff: Repurposing diffusion-based image models for panorama generation. In Proc. International Conference on Learning Representations (ICLR), 2025

    2025

  6. [14]

    360DVD: Controllable panorama video generation with 360-degree video diffusion model

    Qian Wang, Weiqi Li, Chong Mou, Xinhua Cheng, and Jian Zhang. 360DVD: Controllable panorama video generation with 360-degree video diffusion model. In Proc. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024

    2024

  7. [15]

    4K4DGen: Panoramic 4D generation at 4K resolution

    Renjie Li, Panwang Pan, Bangbang Yang, Dejia Xu, Shijie Zhou, Xuanyang Zhang, Zeming Li, Achuta Kadambi, Zhangyang Wang, Zhengzhong Tu, and Zhiwen Fan. 4K4DGen: Panoramic 4D generation at 4K resolution. arXiv preprint arXiv:2406.13527, 2024

    2024 arXiv

  8. [16]

    SphereD- iff: Tuning-free omnidirectional panoramic image and video generation via spherical latent representation

    Minho Park, Taewoong Kang, Jooyeol Yun, Sungwon Hwang, and Jaegul Choo. SphereD- iff: Tuning-free omnidirectional panoramic image and video generation via spherical latent representation. arXiv preprint arXiv:2504.14396, 2025

    2025

  9. [17]

    Denoising diffusion probabilistic models

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In Proc. Conference on Neural Information Processing Systems (NeurIPS), 2020

    2020

  10. [18]

    Deep unsuper- vised learning using nonequilibrium thermodynamics

    Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsuper- vised learning using nonequilibrium thermodynamics. In Proc. International Conference on Machine Learning (ICML), 2015

    2015

  11. [19]

    Auto-encoding variational bayes

    Diederik P Kingma and Max Welling. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2022

    2022 arXiv

  12. [20]

    Yaron Lipman, Ricky T. Q. Chen, Heli Ben-Hamu, Maximilian Nickel, and Matt Le. Flow matching for generative modeling. In Proc. International Conference on Learning Representa- tions (ICLR), 2023

    2023

  13. [21]

    Flow straight and fast: Learning to generate and transfer data with rectified flow

    Xingchao Liu, Chengyue Gong, and Qiang Liu. Flow straight and fast: Learning to generate and transfer data with rectified flow. In Proc. International Conference on Learning Representations (ICLR), 2023

    2023

  14. [22]

    U-net25: Convolutional networks for biomedical image segmentation

    Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net25: Convolutional networks for biomedical image segmentation. In Proc. International Conference on Medical Image Computing and Computer Assisted Intervention (MICCAI), 2015

    2015

  15. [23]

    Scalable diffusion models with transformers

    William Peebles and Saining Xie. Scalable diffusion models with transformers. In Proc. International Conference on Learning Representations (ICLR), 2023

    2023

  16. [24]

    MakeAnything: Harnessing diffusion trans- formers for multi-domain procedural sequence generation

    Yiren Song, Cheng Liu, and Mike Zheng Shou. MakeAnything: Harnessing diffusion trans- formers for multi-domain procedural sequence generation. arXiv preprint arXiv:2502.01572, 2025

    2025 arXiv

  17. [25]

    Matterport3D: Learning from RGB-D data in indoor environments

    Angel Chang, Angela Dai, Thomas Funkhouser, Maciej Halber, Matthias Nießner, Manolis Savva, Shuran Song, Andy Zeng, and Yinda Zhang. Matterport3D: Learning from RGB-D data in indoor environments. In Proc. International Conference on 3D Vision (3DV), 2017

    2017

  18. [26]

    polyhaven.com. HDRIs. https://polyhaven.com/hdris, 2025. Accessed: 2025-03

    2025

  19. [27]

    NTIRE 2023 challenge on 360° omnidirectional image and video super-resolution: Datasets, methods and results

    Mingdeng Cao, Chong Mou, Fanghua Yu, Xintao Wang, Yinqiang Zheng, Jian Zhang, Chao Dong, Gen Li, Ying Shan, Radu Timofte, Xiaopeng Sun, Weiqi Li, Zhenyu Zhang, Xuhan Sheng, Bin Chen, Haoyu Ma, Ming Cheng, Shijie Zhao, Wanwan Cui, Tianyu Xu, Chunyang Li, Long Bao, Heng Sun, Hua...

    2023

  20. [28]

    LLaV A-OneVision: Easy visual task transfer

    Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Peiyuan Zhang, Yanwei Li, Ziwei Liu, and Chunyuan Li. LLaV A-OneVision: Easy visual task transfer. Transactions on Machine Learning Research, 2025. ISSN 2835-8856

    2025

  21. [29]

    Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Niko lay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, Daniel M

    Hugo Touvron, Louis Martin, Kevin R. Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Niko lay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, Daniel M. Bikel, Lukas Blecher, Cris tian Cantón Ferrer, Moya Chen, Guillem Cucurull, David Esiobu, Jude Fernandes, J...

    2023 arXiv

  22. [30]

    Learning transferable visual models from natural language supervision

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agar- wal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In Proc. Internati...

    2021

  23. [31]

    Omnifusion: 360 monocular depth estimation via geometry-aware fusion

    Yuyan Li, Yuliang Guo, Zhixin Yan, Xinyu Huang, Ye Duan, and Liu Ren. Omnifusion: 360 monocular depth estimation via geometry-aware fusion. In Proc. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022

    2022

  24. [32]

    Goodfellow, Wojciech Zaremba, Vicki Cheung, Alec Radford, and Xi Chen

    Tim Salimans, Ian J. Goodfellow, Wojciech Zaremba, Vicki Cheung, Alec Radford, and Xi Chen. Improved techniques for training GANs. In Proc. Conference on Neural Information Processing Systems (NeurIPS), 2016

    2016

  25. [33]

    GANs trained by a two time-scale update rule converge to a local nash equilibrium

    Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. GANs trained by a two time-scale update rule converge to a local nash equilibrium. In Proc. Conference on Neural Information Processing Systems (NeurIPS), 2017

    2017

  26. [34]

    Geometry fidelity for spherical images

    Anders Christensen, Nooshin Mojab, Khushman Patel, Karan Ahuja, Zeynep Akata, Ole Winther, Mar González-Franco, and Andrea Colaço. Geometry fidelity for spherical images. In Proc. European Conference on Computer Vision (ECCV), 2024

    2024

  27. [35]

    Buzuti and C.E.P

    Lucas F. Buzuti and C.E.P. Thomaz. Fréchet autoencoder distance: A new approach for evaluation of generative adversarial networks. Comput. Vis. Image Underst., 235:103768, 2023

    2023

  28. [36]

    John P. Snyder. Map Projections - A Working Manual, 1987

    1987

  29. [37]

    Visual autoregressive modeling for image super-resolution, 2025

    Yunpeng Qu, Kun Yuan, Jinhua Hao, Kai Zhao, Qizhi Xie, Ming Sun, and Chao Zhou. Visual autoregressive modeling for image super-resolution, 2025. URL https://arxiv.org/abs/ 2501.18993

    2025 arXiv

  30. [38]

    Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen

    Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models, 2021. URL https://arxiv.org/abs/2106.09685

    2021 arXiv

  31. [39]

    A panoramic image of

    Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization, 2019. URL https: //arxiv.org/abs/1711.05101. 12 Appendix A Contributions Core Contributors • Hakan Çapuk: Trained the model, developed and tested many ideas for types of train- ing/inference, developed th...

    2019 arXiv

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.