{"id":"a3b18738-ab52-49a5-a593-b86ab0725633","arxiv_id":"2603.23179","paper_version":2,"verdict":"REJECT","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":4,"one_line_summary":"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.","lead":"A computer-vision method that turns a single tilted or rolled photo into a seamless 360° panorama by first rotating the image into a gravity-aligned 'canonical' frame and then training a diffusion model to respect the wrap-around edge of equirectangular panoramas. It introduces a new dataset and reports improved fidelity and seam continuity, but the key auto-leveling module is not fully specified.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Auto-leveling cannot recover unknown focal length/FOV with a 3-DoF rotation; Sec. 3.3's W(z_ref,R_hat) lacks any intrinsic parameter while Sec. 4 trains on variable FOV, so the no-camera-parameters claim fails for general inputs.","rationale":"The paper's central claim requires a single unposed, uncalibrated perspective image to be canonically projected into ERP without camera parameters. The most load-bearing assumption is that a 3-DoF rigid rotation can represent this projection. This fails whenever the input has unknown focal length/FOV, because rotation-only warping cannot account for the image's angular extent. The paper itself introduces variable intrinsics during training but gives the solver no way to estimate or condition on them. This is an internal mismatch between the method's parameterization and its stated problem, not merely a disagreement with prior consensus. The reader's weakest_assumption identifies the same gap, so I agree. The TEG ablation contradiction in Table 1 (w/o TEG better on some metrics) is also concerning, but it is secondary: even if the empirical numbers were clean, the missing intrinsics would still undermine the claimed mechanism. A single controlled FOV-variation experiment would settle whether the concern lands. Until that is provided, rejection is appropriate, so I do not adjust the reader's verdict.","tokens_in":15483,"tokens_out":4606,"duration_ms":48538,"concrete_test":"Take a fixed set of perspective crops with ground-truth FOV θv ∈ {30°, 50°, 70°, 100°, 120°} and known extrinsics. Run the proposed auto-leveling (a) with only the 3-DoF rotation solver, and (b) with the true FOV supplied as an additional input to the solver or as a known K inside W. Compare canonicalization angular error or completion FID. If (b) is materially better, or (a) degrades monotonically as θv moves away from the training mean, the 3-DoF manifold cannot absorb intrinsics and the no-camera-parameters claim fails.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 3.3 defines the auto-leveling warp as z_canonical = W(z_ref, R_hat), a '3-DoF rigid projection' with R_hat the only estimated quantity. A perspective image with unknown intrinsic matrix K cannot be mapped to an ERP via a rotation alone: the angular patch subtended by the image is determined by K, so any fixed W implicitly assumes a fixed FOV. Section 4 explicitly trains with variable vertical FOV (θv) and computes f_y = 1/tan(θv/2), f_x = f_y/r, so the model must handle intrinsics that the 3-DoF solver cannot represent. The soft-argmin bottleneck over rotations will collapse the dense correspondence to the best rotation under the assumed camera model, leaving residual FOV-dependent distortion. The paper provides no equation for W, no intrinsic parameter in the solver, and no experiment isolating FOV variation. The qualitative in-the-wild examples are not controlled for known FOV, and Appendix D.3 even lists telephoto inputs as a failure mode. Thus the central claim—seamless, gravity-aligned completion 'without camera parameters'—is unsupported for exactly the variable-intrinsic inputs the dataset is built to simulate.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"Gimbal360 proposes a framework for completing 360° panoramas from a single unposed perspective image by aligning the input into a gravity-aligned Canonical Viewing Space via a Differentiable Auto-Leveling module, and by enforcing S^1 continuity through Topologically Equivariant Generation. The paper introduces a curated dataset, Horizon360, and reports state-of-the-art FID/KID/FAED/CLIPScore on indoor and outdoor benchmarks against three baselines. The central claims are that auto-leveling works without camera parameters at inference and that the shift-equivariance objective mathematically enforces seamless periodic generation.","tokens_in":15817,"tokens_out":8074,"duration_ms":77314,"significance":"If the results held, the paper would address a real and useful problem: adapting large planar diffusion priors to spherical panorama completion. The decomposition into canonicalization, auto-leveling, and topological equivariance is conceptually appealing, and the dataset is a potentially valuable resource. The paper includes qualitative comparisons and ablations, and the idea of constraining learned correspondence through a rigid solver is a sensible inductive bias. However, the technical formulation is underspecified, the camera-parameter-free claim is not supported by the geometry, and the empirical evidence for the topological equivariance component is internally inconsistent. The lack of equations for the core modules and the narrow comparison set prevent verification of the main contributions.","major_comments":[{"comment":"Eq. (1) defines the auto-leveling warp as z_canonical = W(z_ref, R_hat), with R_hat a 3-DoF rotation. However, a perspective image with unknown intrinsics cannot be mapped to an ERP by a rotation alone: the angular extent of the image is determined by the focal length. Section 4 explicitly samples variable vertical FOV and computes fy = 1/tan(θv/2), fx = fy/r, so the training distribution includes inputs whose intrinsic scale the 3-DoF solver cannot represent. The paper provides no equation for W, no intrinsic parameter in the solver, and no experiment that varies FOV while holding other factors fixed. Appendix D.3's admission that telephoto inputs are a failure mode is consistent with this limitation. The central claim of 'without requiring camera parameters at inference' is therefore unsupported for general inputs.","section":"Sec. 3.3 / Sec. 4"},{"comment":"The TEG ablation is not consistently favorable. On CVRG-Pano, the full model has FID 63.28 vs. 62.74 for w/o TEG, and KID 0.8236 vs. 0.8194. On Structured3D, the full model has higher KID (0.8106 vs. 0.8012). The text claims that the 'quantitative gap' demonstrates the superiority of native topological constraints, but the table shows the opposite on several metrics. This contradiction needs to be resolved, e.g., with significance testing, more metrics, or a revised interpretation.","section":"Table 1 / Sec. 5.3"},{"comment":"The proof of equivariance is tautological: it posits L_shift -> 0 and then concludes that Rollδ(ϵθ(X,t)) = ϵθ(Rollδ(X),t), which is the definition of the loss being zero. Because L_shift is a soft penalty with a finite weight (λ_shift=0.5), convergence to zero is not guaranteed, and even approximate equality on the training distribution does not imply mathematical forcing. The statements in §3.4 and the abstract that TEG 'ensures' or 'enforces' S^1 continuity should be softened to 'encourages' or 'regularizes.'","section":"Appendix C.2"},{"comment":"The Differentiable Auto-Leveling module is not specified to a reproducible level. No equation is given for the dense-correspondence-to-rotation soft-argmin solver, the reprojection error, the rigid spherical camera model, or the differentiable warp W. The cited [23] is a stereo-disparity method, not a rotation estimator; its adaptation to 3-DoF rotation is unclear. Without these formal definitions, the claimed 'structural gradient filter' property and the module's behavior cannot be verified.","section":"Sec. 3.3"},{"comment":"The comparison set is too narrow to support the state-of-the-art claim. Only HunyuanWorld, WorldGen, and DiT360 are evaluated; none is a dedicated single-image 360-degree outpainting method. The related work lists several directly relevant baselines (PanoDiffusion [49], PanoDecouple [60], CamFreeDiff [54], Matrix3D [30]) that are not included. The authors should either include these baselines or clearly justify their exclusion. As it stands, the SOTA claim is not established against the most relevant prior work.","section":"Sec. 5.2"}],"minor_comments":[{"comment":"The header 'T able 1' contains a typo.","section":"Table 1"},{"comment":"CFG scale of 30.0 is unusually high and likely a typo for 3.0; please verify.","section":"Appendix A.2"},{"comment":"The mixture-distribution parameters are inconsistent: Sec. 4 states P(Θ) = λ·N + (1-λ)·U with λ≈0.7, while Appendix B.3 gives per-parameter percentages (70/30 for pitch, 80/20 for roll, and 50/20 for FOV). Clarify the exact generative process.","section":"Sec. 4 / Appendix B.3"},{"comment":"The in-the-wild generalization results are only qualitative. Reporting quantitative rotation/FOV estimation error or a user study would strengthen the robustness claim.","section":"Sec. 5.2 / Fig. 5"},{"comment":"The FAED metric is not defined in the text; please provide a definition or citation.","section":"Sec. 5.1"}],"recommendation":"reject","confidential_remarks":"The FOV/intrinsics issue is a fundamental flaw in the proposed method as described: a 3-DoF rotation cannot canonicalize inputs with variable focal length, yet the dataset and the 'no camera parameters' claim require exactly that. Even if this were resolved, the TEG ablation contradiction and the missing equations for the solver prevent acceptance. A resubmission could be considered if the authors (1) reformulate the solver to include intrinsics (or otherwise demonstrate how unknown FOV is handled), (2) provide the missing mathematical definitions, and (3) correct the TEG evaluation and expand the comparison set."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know. The paper frames a real problem — completing 360° panoramas from unposed perspective photos — and the combination of gravity canonicalization, circular padding, and a shift-equivariance loss is a plausible engineering direction. But the load-bearing claim, that auto-leveling needs no camera parameters, is not justified. The module is described only as a 3-DoF rigid projection W(z_ref, R_hat), with no intrinsic parameter anywhere, while the dataset section trains on variable vertical FOV. A rotation alone cannot map an unknown-FOV perspective image onto the sphere. The soft-argmin bottleneck over rotations would leave the FOV mismatch in place. The paper offers no equations for the solver or the projection manifold, and the cited [23] is a stereo-disparity method, not a rotation estimator. Appendix D.3's telephoto failure is consistent with this gap.\n\nWhat is genuinely useful: the two domain gaps (projective variance and S1 severing) are clearly identified, the canonical viewing space is a sensible way to regularize the problem, and the Horizon360 dataset with a mixture-distribution sampling curriculum looks like a practical asset, although we can't inspect it. The Siamese shift-equivariance loss is a reasonable idea; the appendix 'proof' is just the definition of equivariance, so it gives no independent support, but the training objective itself is coherent.\n\nThe numbers are another problem. The TEG ablation in Table 1 is mixed: on CVRG-Pano the full model is worse than w/o TEG in FID and KID, and on Structured3D the KID is slightly worse. The text claims a clear gap. That undercuts the ablation story as written.\n\nOn balance, this is a serious paper behind a poorly written one. The soft spots are real but fixable in principle: give the solver an explicit intrinsic model or justify why the rotation manifold is sufficient under the training FOV distribution, report ablations honestly, and release code and data. As it stands, the central claim is unsupported, so I would not take the results at face value. But a good referee could do a lot here, and the problem is important enough that I'd send it to review rather than desk reject.","headline":"Useful framing and a plausible dataset, but the no-camera-parameters claim is a 3-DoF rotation short of what the problem needs, and the ablations don't support the story.","tokens_in":16282,"tokens_out":2934,"would_cite":false,"duration_ms":32061,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A single unposed perspective image can be auto-leveled and completed into a gravity-aligned, seamless 360° panorama without camera parameters.","keywords":["panorama completion","equirectangular projection","latent diffusion","auto-leveling","zero-shot camera pose","S1 periodicity","shift-equivariance","360-degree generation"],"falsifier":"Take a perspective image with a known extreme vertical field of view (e.g., 15°) and a known ground-truth panorama; if Gimbal360's output places the horizon at the wrong latitude or visibly stretches the visible content, then the auto-leveling did not recover the intrinsic scale, falsifying the no-camera-parameters claim.","tokens_in":15370,"feed_emoji":"🌐","tokens_out":3406,"duration_ms":31883,"temperature":0.7,"pith_summary":"The paper argues that diffusion-based panorama completion fails because ordinary perspective images and equirectangular panoramas differ in both projective geometry and topology. Gimbal360 bridges this gap by first warping any input into a canonical, gravity-aligned viewing space through a learned 3-DoF rigid projection, then training the generator to respect the circular seam via a shift-equivariance constraint. A newly curated dataset of rectified panoramas supplies the needed structural prior and ground-truth alignment flows. If the argument holds, diffusion models trained on flat images can be repurposed for structurally coherent 360° scene completion from a single unposed photo, with no camera metadata at inference.","feed_headline":"One unposed photo becomes a seamless 360° panorama","feed_subtitle":"Diffusion models built for flat images are taught to respect spherical geometry and the wrap-around seam — no camera data needed.","key_machinery":"The Canonical Viewing Space — a gravity-aligned spherical coordinate system where the environmental horizon is the ERP equator — is the representation that makes planar diffusion applicable to spheres. The Differentiable Auto-Leveling module is a dense correspondence predictor followed by a soft-argmin rigid-projection bottleneck; the low-rank Jacobian of the rigid manifold filters chaotic diffusion gradients and prevents non-rigid jelly-like warps. Topologically Equivariant Generation modifies the VAE's padding to circular and trains a Siamese shift-consistency loss, forcing the diffusion transformer to be S^1-equivariant. Horizon360 supplies strictly rectified panoramas and ground-truth fl","core_discovery":"Gimbal360's central claim is that explicitly standardizing geometric and topological priors lets pre-trained planar diffusion models generate seamless 360° panoramas. Projective distortion is removed by predicting a dense correspondence field and then constraining it through a soft-argmin solver to a 3-DoF rigid rotation, mapping the perspective input into a Canonical Viewing Space where the horizon coincides with the equator. Topological continuity is enforced by replacing zero-padding with circular padding in the VAE and adding a Siamese shift-equivariance loss, so the network learns that circularly shifting the input shifts the output identically. The authors report that this yields state","pith_inferences":["If auto-leveling truly recovers the 3-DoF rotation, the same module could serve as a lightweight single-image horizon and gravity estimator, independent of the completion task.","A stress test worth running: feed the model a perspective image whose vertical field of view is far outside the training mixture; the 3-DoF bottleneck cannot absorb focal-length changes, so output geometry may degrade in a measurable way.","The shift-equivariance idea could transfer to other periodic synthesis tasks, such as 360° video or seamless texture generation, where the seam is currently patched by heuristics.","Because the dataset is curated to be gravity-aligned, the method's success depends on the quality of that canonicalization; residual tilt in training data would propagate directly into the learned structural prior."],"forward_implications":["In-the-wild photos with arbitrary pitch, roll, and focal length can be completed into gravity-aligned 360° environments without user-supplied camera extrinsics.","Seam artifacts are resolved at training time rather than by post-hoc blending, so high-frequency detail survives across the wrap-around boundary.","The same canonicalization lets a model trained on rectified panoramas generalize to off-level inputs, reducing structural hallucinations like tilted horizons and curved walls.","The shift-equivariance objective provides a template for making any Euclidean generative backbone respect periodic structure.","A curated, gravity-aligned panoramic dataset enables the geometric and topological priors to be learned rather than assumed."],"fun_headline_variants":["No camera needed: one image completes a 360° panorama","Fixing diffusion for spherical vision: one photo, full panorama","Seamless 360° from one unposed photo","One image, full sphere: diffusion without camera data"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The method assumes that an unposed perspective photo can be brought into gravity alignment by a 3-DoF rotation alone, with no camera intrinsics; if the unknown focal length or field of view is not recoverable from the soft-argmin bottleneck, the canonicalization is geometrically incomplete.","fun_headline_variants_meta":{"raw":{"variants":["No camera needed: one image completes a 360° panorama","Fixing diffusion for spherical vision: one photo, full panorama","Seamless 360° from one unposed photo","One image, full sphere: diffusion without camera data"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000578,"raw_usage":{"total_tokens":2577,"prompt_tokens":776,"completion_tokens":1801,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":520,"completion_tokens_details":{"reasoning_tokens":1733}},"tokens_in":520,"tokens_out":1801,"duration_ms":12494,"temperature":1.0,"reasoning_tokens":1733,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-04T05:40:04.560673+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a perspective image with a known extreme vertical field of view (e.g., 15°) and a known ground-truth panorama; if Gimbal360's output places the horizon at the wrong latitude or visibly stretches the visible content, then the auto-leveling did not recover the intrinsic scale, falsifying the no-camera-parameters claim.","supporting_citations":[],"review_version":1}