{"id":"7f4f1e65-50dd-476c-b8cf-a48576d77842","arxiv_id":"2506.23513","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"A panorama representation and attention scheme that lets a pretrained perspective video diffusion model generate spatially consistent 360-degree videos from an input perspective clip.","lead":"ViewPoint is a method for generating 360-degree videos from a regular perspective video clip by rearranging the spherical view into a stitched grid of overlapping perspective panels. It adapts a pretrained perspective video diffusion model to the panoramic domain, improving spatial consistency and motion dynamics over prior 360-degree video generators.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The SOTA claim is not supported because the only quantitative benchmark, ODV360, is also a training dataset and no train/test split or significance test is reported.","rationale":"The reader's weakest_assumption focuses on whether the ViewPoint map remains sufficiently perspective-like for pretrained diffusion priors. That is a legitimate technical risk, and the paper's ablations (ERP vs. CP vs. ViewPoint) provide some internal evidence that the representation helps—though only on the same training distribution. The more load-bearing concern is the evaluation validity: the quantitative comparison is performed on ODV360, a training dataset, with no train/test split, sample IDs, or statistical significance. This directly threatens the central SOTA claim, because even a perfect representation would not justify 'surpassing previous methods' if the benchmark overlaps with training data. The paper also contains small technical inconsistencies (e.g., 'Rot90' in Eq. 9, diameter/radius wording in Sec. 3.2), but these are likely fixable and less decisive than the evaluation issue. I agree with the reader's conditional verdict: the paper should not be accepted as-is, but the concern is addressable with a held-out evaluation and artifact release. Hence UNCHANGED.","tokens_in":9901,"tokens_out":9884,"duration_ms":107359,"concrete_test":"Require the authors to release the exact ODV360 sample IDs used for training and evaluation, or rerun the VBench protocol on a 20% held-out split of ODV360 plus an external 360-video benchmark (e.g., a split of WEB360 or a 360DVD test set). Compute paired bootstrap 95% confidence intervals for the four metrics across all methods. If the ViewPoint margins over Imagine360 and 360DVD are not statistically significant or shrink materially, the SOTA claim should be withdrawn.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The abstract's claim of 'state-of-the-art performance and surpassing previous methods' is grounded in Table 1, which reports VBench scores on ODV360. However, Sec. 4.1 lists ODV360 as one of the four datasets used for fine-tuning. The paper never specifies which ODV360 videos, if any, were held out from training, nor does it report sample IDs or a train/test partition. Without that information, the reported margins (e.g., +0.0258 in subject consistency over 360DVD, +0.0935 in dynamic degree over Imagine360) could reflect memorization of training clips rather than generalization. The user study (Sec. 4.5) is also underspecified (50 questionnaires, 14 video sets, no disclosure of participant recruitment or interface) and cannot substitute for a valid held-out quantitative comparison. Since every quantitative comparison and the central SOTA claim rests on this evaluation, the evidence as presented does not support the headline claim.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces ViewPoint, a framework for generating 360-degree panoramic videos from a perspective video clip and text prompt. The core idea is a novel panorama representation, the ViewPoint map, which rearranges six cube faces into four overlapping pseudo-perspective subregions with gradient-based fusion, aiming to combine the local fidelity of perspective projections with global spatial continuity. The method fine-tunes a pretrained perspective video diffusion model (Wan2.1) using a Pano-Perspective attention mechanism that alternates global self-attention over the whole map and local attention over individual subregions. Experiments compare against 360DVD, Imagine360, and Follow-Your-Canvas using VBench metrics on the ODV360 dataset, together with qualitative examples and a user study, and the paper claims state-of-the-art performance.","tokens_in":10075,"tokens_out":2835,"duration_ms":31145,"significance":"If the results are validated, the ViewPoint map is a meaningful contribution: it addresses a known limitation of equirectangular and cubemap representations for generative video models, and the overlapping-fusion idea is a plausible mechanism for preserving cross-face consistency while exploiting perspective priors. The paper includes a useful ablation over ERP, cubemap variants, and the attention design, and the qualitative results show clear improvements in polar-region distortion and temporal consistency over previous approaches. However, the quantitative evidence for the central SOTA claim is currently undermined by an evaluation setup that does not separate training and test data, so the significance of the empirical contribution is not yet established.","major_comments":[{"comment":"The quantitative evaluation in Table 1 is not a valid test of generalization because ODV360 is listed in Sec. 4.1 as one of the four fine-tuning datasets. The paper never specifies a train/test partition, sample IDs, or any held-out subset. Without that information, the margins in Table 1 (e.g., +0.0258 subject consistency over 360DVD, +0.0908 dynamic degree over Imagine360) could reflect in-distribution memorization rather than a generalizable advantage. Please report exactly which ODV360 videos were used for training, which for evaluation, or re-evaluate on a dataset not seen during training, and report error bars or significance tests.","section":"4.1 and 4.3"},{"comment":"The overlapping fusion equations contain notation inconsistencies that prevent verification of the algorithm. In Eq. (9), the first branch uses `R−90(W)` while the second branch uses `Rot90(W)`; elsewhere `R90` and `R−90` are defined in Eqs. (6)-(7), and `Rot90` is never defined. Similarly, Eq. (10) first branch applies `R−90(W)` while the second applies `R90(W)` to the overlapping region of the same subregion, which is suspicious. Please unify the notation, define `Rot90`, and provide a derivation or pseudocode that makes the overlap regions and weight indexing unambiguous.","section":"3.4, Eqs. (9)-(12)"},{"comment":"The user study is underspecified and cannot substitute for the missing held-out quantitative evaluation. The paper states only that 50 valid questionnaires were collected with 14 video sets each, with no details on participant recruitment, instructions, interface, randomization, or inter-rater agreement. Figure 7 shows raw vote counts without error bars or any statistical test, so the claim that our method 'receives the highest number of votes across all four dimensions' has no demonstrated reliability.","section":"4.5"}],"minor_comments":[{"comment":"The notation in the preliminary is confusing: the sentence says 'a noisy latent xt is obtained by adding a random noise x0 ∼ N(0, I)' but x0 is already used for the clean latent. Please rename the noise variable (e.g., x1) consistently with Eq. (1).","section":"3.1, Eq. (1)"},{"comment":"The phrase 'located to the left, right, above, and below L, respectively' appears to contain a typo: for a central face F, the adjacent faces should be to the left, right, above, and below F, not L. Please correct.","section":"3.2"},{"comment":"The text says the evaluation covers 'five dimensions' but then lists only four: subject consistency, imaging quality, motion smoothness, and dynamic degree. Either add the fifth dimension or correct the count.","section":"4.3"},{"comment":"The feature shape is written as `(batch_size, channels, f rames, height, width)` with a space in 'frames', and the reshape to `height/2, width/2` is not explained. Clarify why the two spatial dimensions are halved after Pano attention.","section":"3.3"},{"comment":"The related-work discussion of panoramic video generation omits 4K4DGEN from the quantitative comparison though it is cited; if it is not comparable due to task definition, a brief sentence explaining its exclusion would help.","section":"2"}],"recommendation":"major_revision","confidential_remarks":"The core representation and attention design are promising, but the evaluation infrastructure needs substantial rework before the SOTA claim can be taken seriously. The lack of a train/test split in the main benchmark is a standard circularity problem that should have been caught; I would ask the authors to either provide a rigorous held-out evaluation or clearly state that Table 1 measures in-distribution performance only. The notation errors in the fusion equations also suggest the method description needs careful proofreading before release."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things you should know about arXiv:2506.23513. First, the ViewPoint map representation is genuinely new and clever: it rearranges six cubemap faces into four overlapping pseudo-perspective panels with a custom fusion, which is a reasonable way to keep a pretrained perspective video diffusion model happy. Second, the quantitative evidence for the headline SOTA claim is not there yet because the only benchmark, ODV360, is also a training set, and the paper gives no split, sample IDs, or error bars.\n\nWhat is actually new: the representation in Section 3.2 and the Pano-Perspective attention split in Section 3.3 are not in prior work (CubeDiff, Imagine360, 360DVD, PanFusion), and the qualitative figures suggest the method does produce coherent 360-degree video with motion, which prior methods visibly struggle with. The idea of adapting a strong pretrained video model like Wan2.1 via a geometric remapping plus minimal fine-tuning is practical and plausible.\n\nThe soft spots, in proportion. The fusion equations (9-12) are hard to follow and likely have notation slips (e.g., Rot90 vs R90, inconsistent rotation indices). That is fixable with a rewrite but as written it undermines reproducibility. More serious is the evaluation. ODV360 is listed as a training dataset in Section 4.1, and no held-out partition is described. VBench scores on training clips can reflect memorization rather than generalization. The user study (50 questionnaires, 14 video sets, no recruitment or interface details) is too underspecified to add much weight. No code or checkpoints are released, so independent verification is limited.\n\nDespite these issues, the core idea is solid and the method itself is not fitted to the evaluation metric. The representation design shows clear thinking and honest engagement with the limitations of ERP and cubemap formats. The paper is worth serious refereeing, but it needs major revision: a proper train/test split on ODV360 or another held-out dataset, statistical significance or error bars, a cleaned-up fusion math section, and ideally code or checkpoints.\n\nRecommendation: send to peer review. The central representation is novel and potentially useful to the 360-video and world-model community; it would be a waste to desk-reject on the strength of the idea alone. For a reading group, it is a maybe—useful for discussion of representation design, not for its evaluation numbers.","headline":"The ViewPoint map is a genuinely useful new 360-degree representation, but the SOTA claim leans on an in-distribution benchmark and needs a proper held-out evaluation before it can be believed.","tokens_in":10610,"tokens_out":1947,"would_cite":true,"duration_ms":21727,"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 tiled 'ViewPoint map' lets a pretrained perspective video model produce seamless 360-degree video.","keywords":["panoramic video generation","360-degree video","diffusion models","ViewPoint map","cubemap projection","equirectangular projection","video outpainting","Pano-Perspective attention"],"falsifier":"Reproject a generated ViewPoint map back to equirectangular form and inspect the top face: if the four semicircular patches show visible seams, duplicated content, or inconsistent motion across their overlaps, the claim of seamless polar continuity is falsified.","tokens_in":9689,"feed_emoji":"🌐","tokens_out":4574,"duration_ms":43312,"temperature":0.7,"pith_summary":"This paper claims that the gap between panoramic and perspective video can be closed by changing the representation rather than retraining from scratch. It introduces the ViewPoint map, a layout of four overlapping pseudo-perspective panels derived from the six faces of a cubemap, and feeds it into a pretrained perspective video diffusion model through a Pano-Perspective attention mechanism. The stated result is high-quality 360-degree video that is spatially consistent across the sphere and temporally dynamic, beating prior panoramic video methods on VBench metrics and in user studies. If true, this would make 360-degree immersive video generation practical by reusing existing perspective video generators and their rich priors.","feed_headline":"ViewPoint map turns perspective clips into seamless 360° video","feed_subtitle":"A four-panel panorama layout plus dual attention lets a pretrained video model fill the whole sphere without polar distortion.","key_machinery":"The ViewPoint map is the central object: four cube faces are each paired with triangular fragments of adjacent faces to form four square pseudo-perspective panels, arranged so that the up face is split into semicircular regions and reprojected with the scale rule $d(\\theta)=a/(\\sin\\theta+|\\cos\\theta|)$ to make overlapping seams. Pano-Perspective attention alternates a global attention block over the whole stitched map with per-panel perspective attention blocks, so the model sees both the full sphere and local detail. Overlapping gradient fusion then blends the rhombus-shaped overlaps between subregions with position-dependent weights so adjacent panels merge smoothly. Together these pieces let the model keep the in-context generation behavior of the pretrained diffusion model while learning panoramic consistency.","core_discovery":"The central claim is that panoramic video can be generated from a perspective input clip and a text prompt by reprojecting the sphere into a ViewPoint map and fine-tuning a pretrained perspective video diffusion model (Wan2.1) on it. The paper reports that this representation removes the polar distortion of equirectangular projections and the hard seams of cubemaps, that the Pano-Perspective attention keeps opposite directions aligned while preserving local detail and motion, and that the resulting videos are more spatially and temporally consistent than those of 360DVD, Imagine360, and Follow-Your-Canvas. The paper therefore presents ViewPoint as a state-of-the-art framework for panoramic video generation.","pith_inferences":["The same ViewPoint layout could be applied to single-image panorama generation with an image diffusion model, which would test whether the representation alone transfers across model families.","A stress test the paper does not run is to generate long clips and measure drift across the top and bottom faces over time, since polar reprojection is the most likely place for the learned prior to slip.","The 4x batch reshape used for perspective attention implies each panel is processed semi-independently; ablating panel size or the number of panels could reveal how much global context is actually required.","If the representation transfers cleanly, it may also benefit neighboring tasks such as 360-degree inpainting, spherical super-resolution, and world-model training by giving diffusion models a perspective-like canvas for spherical data."],"forward_implications":["Panoramic video generation inherits the motion and quality priors of a state-of-the-art perspective video generator, so high-quality 360-degree results do not require a large corpus of panoramic video.","Because the method accepts both an input video and a text prompt, it supports conditional video-to-panorama outpainting rather than only text-to-panorama generation.","The reported VBench numbers indicate that spatial consistency and high dynamics can be obtained together, a combination earlier panoramic video methods did not achieve.","The ViewPoint representation itself, rather than the diffusion backbone, is the main bridge between perspective-pretrained models and spherical video output."],"supporting_citations":[{"why":"Provides the pretrained perspective video diffusion model (Wan2.1) that ViewPoint inflates and fine-tunes, supplying the generative priors.","marker":"[27]"},{"why":"Imagine360 is the closest prior panoramic-video-from-perspective method and serves as the main qualitative and quantitative baseline.","marker":"[25]"},{"why":"360DVD is a text-driven 360-degree video diffusion baseline whose nearly static outputs are compared against.","marker":"[31]"},{"why":"Follow-Your-Canvas is the perspective video outpainting method that ViewPoint must surpass in spatial consistency.","marker":"[5]"},{"why":"VBench supplies the four evaluation metrics on which the paper reports its best scores.","marker":"[12]"},{"why":"In-context LoRA is cited as the mechanism that lets the diffusion model treat the ViewPoint map as in-context generation input.","marker":"[11]"},{"why":"Qwen-VL is used to generate descriptive captions for the three panorama datasets that lack text annotations.","marker":"[2]"},{"why":"The NTIRE omnidirectional dataset family provides training data and the ODV360 subset used for quantitative evaluation.","marker":"[3]"}],"fun_headline_variants":["Panoramic video from a single view using ViewPoint map","ViewPoint: one perspective clip becomes full 360° video","ViewPoint map: full 360° video from one perspective clip","360° video from a single perspective clip via ViewPoint map","Seamless 360° video from a single perspective clip via ViewPoint"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method assumes that after the triangular reprojection and overlapping fusion, the ViewPoint map still looks enough like ordinary perspective video that the pretrained model keeps its perspective priors instead of having to relearn basic image statistics.","fun_headline_variants_meta":{"raw":{"variants":["Panoramic video from a single view using ViewPoint map","ViewPoint: one perspective clip becomes full 360° video","ViewPoint map: full 360° video from one perspective clip","360° video from a single perspective clip via ViewPoint map","Seamless 360° video from a single perspective clip via ViewPoint"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000549,"raw_usage":{"total_tokens":2560,"prompt_tokens":821,"completion_tokens":1739,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":437,"completion_tokens_details":{"reasoning_tokens":1649}},"tokens_in":437,"tokens_out":1739,"duration_ms":11379,"temperature":1.0,"reasoning_tokens":1649,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T21:39:26.052267+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Reproject a generated ViewPoint map back to equirectangular form and inspect the top face: if the four semicircular patches show visible seams, duplicated content, or inconsistent motion across their overlaps, the claim of seamless polar continuity is falsified.","supporting_citations":[{"cited_title":"360dvd: Controllable panorama video generation with 360-degree video diffusion model","cited_arxiv_id":null,"evidence_quote":"360DVD is a text-driven 360-degree video diffusion baseline whose nearly static outputs are compared against."},{"cited_title":"VBench: Comprehensive benchmark suite for video generative models","cited_arxiv_id":null,"evidence_quote":"VBench supplies the four evaluation metrics on which the paper reports its best scores."},{"cited_title":"Ntire 2023 challenge on 360deg omnidirectional image and video super-resolution: Datasets, methods and results","cited_arxiv_id":null,"evidence_quote":"The NTIRE omnidirectional dataset family provides training data and the ODV360 subset used for quantitative evaluation."}],"review_version":1}