{"id":"e67c0c56-39c2-4393-a400-80aa7830d0c7","arxiv_id":"2501.16164","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":2,"one_line_summary":"MetaDecorator combines diffusion-based panorama decoration with a depth-prior-constrained NeRF to build personalized 3D virtual tours, claiming higher PSNR and 10x faster training on a single scene.","lead":"MetaDecorator is a framework that lets users restyle 360-degree panoramas with text prompts and reconstructs them as 3D meshes using a depth-guided NeRF variant called DP-NeRF. It claims better rendering quality and 10x faster training, but the supporting evidence is one test scene with no timing measurements.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 10x training-time claim is unsupported: no wall-clock timing is measured, and the 20k-vs-200k epoch comparison does not establish training-time or novel-view superiority, so the central quantitative claim is unverified.","rationale":"The paper's central quantitative claim, as captured in the strongest_claim, is that DP-NeRF outperforms baselines in PSNR and achieves a 10x training-time reduction. The load-bearing assumption is that the reported comparison supports those conclusions. The reader's weakest_assumption states exactly this and identifies the three concrete gaps: no wall-clock time, no novel test views, and PSNR at input image locations. I agree with these and with the REJECT verdict. My independent reading finds the same weak point plus one additional aggravating detail: even the mechanism for the speedup is not specified precisely enough to be costed, so one cannot convert the reported 20k/200k epochs into a time ratio by any means. The paper contains no equations for the DP-NeRF depth-prior occupancy grid (Section 'ENHANCEMENT: TRAINING EFFICIENTLY IMPROVEMENT FOR GREEN AI'), no ablation, no code, and no multi-scene evaluation. These are mechanical evidence gaps, not signs of misconduct. Since the headline contribution is the efficiency-accuracy tradeoff and that claim is unverified in both dimensions, the verdict REJECT is proper. No adjustment is needed.","tokens_in":5492,"tokens_out":4572,"duration_ms":33335,"concrete_test":"Run a direct timing comparison on the same hardware: train DP-NeRF, OmniNeRF, and 360FusionNeRF on the Structured3D representative scene, record wall-clock training time, and evaluate PSNR on held-out novel views rather than training views. If the measured speedup is not approximately 10x, or if the novel-view PSNR is not superior, the central claim is falsified. A minimal analytical test is to derive the per-epoch cost of DP-NeRF from the occupancy-grid specification and check whether 20k DP-NeRF epochs equal 10x fewer wall-clock seconds than 200k baseline epochs under stated assumptions.","verdict_should_be":"REJECT","load_bearing_attack":"The headline claim 'achieving a 10x reduction in training time' (Prototype Performance, Table 1) rests on comparing 20k epochs for DP-NeRF against 200k epochs for the baselines. Epoch count is not wall-clock time, and no timing measurements are reported. The proposed speedup mechanism, occupancy-grid ray pruning (Figure 3a), is described qualitatively: occupancy is initialized from depth priors and dynamically updated, with a viewpoint weight factor that 'typically falls within a small range.' The paper gives no equations for occupancy computation, no update schedule, no sample-count per ray, no resolution or density of the grid, and no convergence criterion. Without these, the per-epoch computational cost of DP-NeRF cannot be derived, so the 10x speedup is a bare assertion. In addition, PSNR is computed 'at corresponding input image locations,' i.e., training-view reconstruction, not novel-view synthesis; no held-out test views are used. The quality comparison is also a single representative scene, with no variance or statistical test. All of this means the central result fails a basic reproducibility and evidence check.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This manuscript proposes MetaDecorator, a framework for personalizing 360-degree virtual tours by decorating static panoramas with text-prompt-driven stable diffusion and ControlNet, then reconstructing the decorated scene as a 3D mesh via a proposed DP-NeRF model. DP-NeRF is described as initializing an occupancy grid from depth priors to prune empty-space ray marching and adding depth constraints to the NeRF loss. The paper also sketches future integrations of LLM-based interaction and haptic feedback. The only quantitative evaluation appears in 'Prototype Performance' (Table 1): a single representative Structured3D scene, where DP-NeRF is compared with OmniNeRF and 360FusionNeRF in terms of PSNR, with DP-NeRF trained for 20k epochs versus 200k for the baselines, leading to a claimed 10x reduction in training time. No wall-clock times, held-out views, or variance measures are reported.","tokens_in":5733,"tokens_out":4103,"duration_ms":39812,"significance":"If the central quantitative claim were substantiated, DP-NeRF would be a useful efficiency contribution to panoramic NeRF reconstruction, and the overall framework addresses a real gap in customizing virtual tours. The paper is clearly organized and the proposed pipeline (depth-prior occupancy initialization, depth-constrained NeRF, mesh extraction) is plausible. The main value of the manuscript is its systems-level integration of 2D decoration and 3D reconstruction, but the current evidence falls far short of what would be needed to support the claimed quantitative advantages. The paper does not provide code, equations, or experimental detail to allow reproduction of DP-NeRF; no user study evaluates the decoration quality; and the quantitative result is built on an apples-to-oranges comparison.","major_comments":[{"comment":"The claim of a '10x reduction in training time' is unsupported. The paper compares DP-NeRF trained for 20k epochs with OmniNeRF and 360FusionNeRF trained for 200k epochs, but it reports no wall-clock training times. Epoch count is not a proxy for training time because the per-epoch cost depends on the occupancy-grid implementation, the number of ray samples, grid resolution, and update schedule, none of which are specified. The authors must either report measured wall-clock times on identical hardware or temper the speedup claim to 'fewer epochs.'","section":"Prototype Performance, Table 1"},{"comment":"The quality comparison is not a valid novel-view evaluation. The paper states that PSNR is computed 'at corresponding input image locations,' i.e., at the training viewpoints. This measures reconstruction fidelity on the training views, not novel-view synthesis, despite the section title claiming 'NeRF Novel View Synthesis.' A proper evaluation must hold out a subset of input images or use distinct test camera poses and report PSNR on those views.","section":"Prototype Performance, Table 1"},{"comment":"The quantitative evaluation rests on a single 'representative scene' from Structured3D with no repeated runs, no standard deviations, and no statistical test. Such a comparison cannot support the general claim that DP-NeRF 'outperforms both OmniNeRF and 360FusionNeRF.' The authors should evaluate on multiple scenes and report variance across runs, or explicitly limit the claim to a single demonstrative example.","section":"Prototype Performance, Table 1"},{"comment":"The occupancy-grid mechanism is described only qualitatively. There is no equation defining occupancy, no specification of grid resolution or density, no update schedule, and no convergence criterion. The 'viewpoint weight factor' is said to 'typically fall within a small range,' but no range is given. Without these details, the mechanism cannot be reproduced, and the claimed speedup cannot be independently assessed.","section":"Enhancement: Training Efficiently Improvement for Green AI"},{"comment":"The paper asserts that DP-NeRF 'significantly increases the number of effective training points' while also pruning ray marching to empty space. This wording is ambiguous and potentially contradictory: pruning reduces the number of sampled points, while depth constraints may improve their informativeness. The authors should clarify the intended claim and provide counts of effective samples per ray or per training step.","section":"MetaDecorator Supports Metaverse Applications / Prototype Performance"}],"minor_comments":[{"comment":"The abstract contains an ungrammatical comma: 'MetaDecorator, is a framework' should read 'MetaDecorator is a framework.'","section":"Abstract"},{"comment":"The caption states 'Panels b) and d)' for Japanese style, but the figure shows panels b), c), d), and e); please check the panel assignments and the spelling of 'Japenese' (should be 'Japanese').","section":"Figure 2 caption"},{"comment":"The section title contains a typographical error: 'EFFICIENTL Y' should be 'EFFICIENTLY'.","section":"Enhancement: Training Efficiently Improvement for Green AI"},{"comment":"Reference [6] is mis-titled: the correct title of the Stable Diffusion paper is 'High-Resolution Image Synthesis with Latent Diffusion Models,' not 'Proceedings of high-resolution image synthesis with latent diffusion models.'","section":"References"},{"comment":"The phrase 'The Lidar data' should be 'LiDAR data' for correct technical spelling.","section":"MetaDecorator Framework"},{"comment":"The two subsections on LLMs and haptics are largely speculative and are not integrated with or evaluated against the proposed framework. They could be explicitly labeled as future work rather than contributions of the current system.","section":"Future Immersive Experiences"}],"recommendation":"major_revision","confidential_remarks":"The manuscript appears to be aimed at a short 'Description' venue, but the current quantitative claims, particularly the 10x training-time reduction and PSNR superiority, are not supported by the evidence presented. The fundamental issues are fixable in a major revision by reporting wall-clock times, proper held-out novel-view evaluation, and multiple scenes. The lack of any evaluation of the decoration stage itself is also a concern for a paper about a user-facing framework. If the authors cannot add these experiments, they should substantially downscope the claims to a system description."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"What you should know: this is a short 'description' piece, not a full evaluation paper. The genuinely new bit is DP-NeRF, a panoramic NeRF variant that seeds an occupancy grid from depth priors and adds depth plus RGB losses. That is a reasonable engineering idea, and the paper is honest about its lineage: it cites OmniNeRF and 360FusionNeRF as baselines, and the depth-prior pruning is a plausible extension of Instant-NGP. No citation red flags; the one self-citation is background motivation.\n\nThe soft spots are where the reader's report lands. The entire quantitative case is Table 1: one representative Structured3D scene, DP-NeRF trained for 20k epochs versus 200k for baselines, no standard deviation, no wall-clock time, and PSNR computed at 'corresponding input image locations.' Epoch count is not training time, and the paper gives no occupancy-grid update schedule, no ray sampling details, no convergence criterion, so the 10x speedup is asserted rather than derived. Likewise, PSNR at input viewpoints measures reconstruction, not novel-view synthesis, so the quality claim is weakened. The paper does note that decorated images lower PSNR because of noise in depth and pixel values, which is a fair admission.\n\nI agree with the stress-test: the central claim fails routine evidence checks. But I would not call the paper incoherent or mis-cited. It is a plausible system description overreaching on the one piece of experimental support.\n\nWho gets value from this? Someone building a similar pipeline might read it as a starting point, and the framework figure is a useful summary of a prompt-to-mesh workflow. But a researcher checking whether DP-NeRF actually trains ten times faster would find no support.\n\nFor peer review: I would not send the current version to a rigorous venue, because the headline empirical claim is unverified and the methods section is too thin to reproduce. If this is aimed at a magazine or industry-track venue, it could appear with the speedup claim softened and explicitly labeled as a prototype observation. As is, a serious referee would mostly be writing 'where is the measurement?' rather than engaging with the idea.","headline":"A readable framework description whose one quantitative claim, 10x faster training with better PSNR, rests on a single scene, no clock-time measurements, and PSNR at training-view locations.","tokens_in":6267,"tokens_out":1949,"would_cite":false,"duration_ms":20029,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims that a text prompt can restyle a static 360-degree panorama and that a depth-prior NeRF then rebuilds it as a VR-ready 3D mesh in a fraction of the usual training time.","keywords":["virtual tours","text-driven scene decoration","diffusion models","ControlNet","panoramic NeRF","depth prior","3D mesh reconstruction","metaverse applications"],"falsifier":"Run DP-NeRF, OmniNeRF, and 360FusionNeRF on several indoor panoramas (including a set where novel camera poses are held out), measure wall-clock training time on the same GPU, and compare PSNR and structural similarity on the held-out views rather than at input image locations. If DP-NeRF is not faster in wall-clock time or does not beat both baselines on novel views, the paper's central claim is not supported.","tokens_in":5308,"feed_emoji":"🏠","tokens_out":6114,"duration_ms":53666,"temperature":0.7,"pith_summary":"MetaDecorator is a two-stage framework for turning a static 360-degree panorama into a personalized virtual tour. In the first stage, text prompts and optional style images drive a diffusion model that re-styles the whole scene or replaces selected regions, while depth, edge, and segmentation cues keep the geometry and semantics intact. In the second stage, the proposed DP-NeRF reconstructs the decorated panorama as a polygonal mesh, using the same depth information to initialize an occupancy grid and to constrain training. The paper reports that DP-NeRF reaches higher PSNR than OmniNeRF and 360FusionNeRF on a representative Structured3D scene while training for 20k epochs instead of 200k, a claimed tenfold reduction in training time. If this holds, personalized, stylized virtual tours could be produced quickly and rendered on resource-limited VR devices.","feed_headline":"Depth-prior NeRF trains 10x faster on decorated panoramas","feed_subtitle":"Text-prompted restyling plus depth-guided reconstruction could make personalized virtual tours practical on VR devices.","key_machinery":"The load-bearing object is DP-NeRF (Depth Prior and Constraint Panoramic NeRF). It builds an occupancy grid whose cells are initialized from the decorated panorama's depth map, with occupancy higher near scene points and adjusted by viewpoint weight; during training and inference, ray marching only visits high-occupancy cells, and the occupancy is updated as training proceeds. RGB and depth losses are applied together at the output, which the paper says constrains geometry and suppresses artifacts. The decoration side is carried by stable diffusion plus ControlNet, which uses depth and edge maps to keep generated content aligned with the original structure. Together these mechanisms turn one text-prompted 360-degree image into a geometry-preserving, VR-ready 3D mesh.","core_discovery":"On its own terms, the central discovery is that geometry-aware image decoration and depth-prior NeRF training belong together in one pipeline. The decoration stage uses stable diffusion conditioned by ControlNet on depth, edge, and instance segmentation maps, so the generated style does not destroy the room's structure; if no region matches the prompt, the whole panorama is re-styled. The reconstruction stage then feeds the same depth map into an occupancy grid, traces rays only through occupied cells, and applies RGB plus depth losses, which the paper says both accelerates training and reduces artifacts. The reported comparison shows DP-NeRF at 20k epochs reaching PSNR 28.36 on the raw panorama, against 28.05 for 360FusionNeRF and 26.41 for OmniNeRF at 200k epochs, with decorated inputs giving 27.16 and 27.65; the paper reads this as evidence that depth priors make panoramic NeRF faster and more accurate. Finally, marching cubes plus adaptive surface refinement converts the field into a textured mesh, making the result usable in standard 3D and metaverse applications.","pith_inferences":["A natural extension the paper does not implement is to use an LLM to decompose a single high-level request into per-object or per-region prompts, making the decoration stage conversational rather than manual.","The depth maps already produced for ControlNet and DP-NeRF could be reused to assign haptic textures to segmented surfaces, turning the paper's haptics discussion into a concrete implementation path.","The reported evaluation metric rewards reconstruction of the input views, so a fair comparison on held-out viewpoints might change the relative ranking; this is a testable extension rather than a criticism.","If depth-prior occupancy accelerates training for panoramas, the same trick should transfer to perspective-image NeRFs whenever sensor depth or estimated depth is available."],"forward_implications":["Text-prompt decoration guided by depth, edges, and segmentation can restyle a 360-degree view without breaking its geometry.","A depth-prior occupancy grid can reduce the fraction of rays wasted in empty space, making single-panorama NeRF training an order of magnitude faster if the reported epoch comparison reflects wall-clock time.","Because the pipeline outputs textured polygonal meshes, the decorated scene can be imported into ordinary 3D tools and rendered on low-power VR hardware.","The same two-stage design can be applied to any virtual tour that begins as equirectangular images, not just custom-captured datasets."],"supporting_citations":[{"why":"Supplies the text-to-image diffusion backbone used in the decoration stage.","marker":"[6]"},{"why":"Provides the geometric conditioning (depth, edges) that keeps decorations aligned with scene structure.","marker":"[9]"},{"why":"The base NeRF representation that DP-NeRF extends and the raw-image baseline in Table 1.","marker":"[7]"},{"why":"Supplies the occupancy grid technique that DP-NeRF initializes with depth priors.","marker":"[12]"},{"why":"Baseline comparison and the multi-viewpoint training strategy DP-NeRF follows.","marker":"[13]"},{"why":"Baseline comparison for panoramic reconstruction quality and efficiency.","marker":"[14]"},{"why":"Provides instance segmentation used to identify regions for prompt-based decoration.","marker":"[8]"},{"why":"Supplies the adaptive surface refinement that turns the NeRF field into a high-quality mesh.","marker":"[11]"},{"why":"Generates the coarse mesh geometry from the NeRF ray-tracing results.","marker":"[10]"}],"fun_headline_variants":["Text prompts restyle 360° panoramas; depth-prior NeRF trains 10x faster","Instant VR decor: text prompts + depth priors = 10x NeRF speedup","Text-guided VR restyling + depth priors = 10x faster NeRF","Immersive VR tours: text decor + depth priors speed NeRF 10x"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The central claim rests on assuming that PSNR measured at the original input image positions on one representative scene, with 20k epochs counted as training time, is enough to prove a tenfold training-time reduction and a quality advantage over the baselines.","fun_headline_variants_meta":{"raw":{"variants":["Text prompts restyle 360° panoramas; depth-prior NeRF trains 10x faster","Instant VR decor: text prompts + depth priors = 10x NeRF speedup","Text-guided VR restyling + depth priors = 10x faster NeRF","Immersive VR tours: text decor + depth priors speed NeRF 10x"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001034,"raw_usage":{"total_tokens":4302,"prompt_tokens":845,"completion_tokens":3457,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":461,"completion_tokens_details":{"reasoning_tokens":3361}},"tokens_in":461,"tokens_out":3457,"duration_ms":22654,"temperature":1.0,"reasoning_tokens":3361,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T13:39:56.090371+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run DP-NeRF, OmniNeRF, and 360FusionNeRF on several indoor panoramas (including a set where novel camera poses are held out), measure wall-clock training time on the same GPU, and compare PSNR and structural similarity on the held-out views rather than at input image locations. If DP-NeRF is not faster in wall-clock time or does not beat both baselines on novel views, the paper's central claim is not supported.","supporting_citations":[{"cited_title":"Proceedings of high-resolution image synthesis with latent diffusion models,","cited_arxiv_id":null,"evidence_quote":"Supplies the text-to-image diffusion backbone used in the decoration stage."},{"cited_title":"Adding conditional control to text-to-image diffusion mod- els,","cited_arxiv_id":null,"evidence_quote":"Provides the geometric conditioning (depth, edges) that keeps decorations aligned with scene structure."},{"cited_title":"NeRF: Rep- resenting scenes as neural radiance fields for view synthesis,","cited_arxiv_id":null,"evidence_quote":"The base NeRF representation that DP-NeRF extends and the raw-image baseline in Table 1."},{"cited_title":"Instant neural graphics primitives with a mul- tiresolution hash encoding,","cited_arxiv_id":null,"evidence_quote":"Supplies the occupancy grid technique that DP-NeRF initializes with depth priors."},{"cited_title":"Omni-NeRF: Neural radiance field from 360 im- age captures,","cited_arxiv_id":null,"evidence_quote":"Baseline comparison and the multi-viewpoint training strategy DP-NeRF follows."},{"cited_title":"360fusion- NeRF: Panoramic neural radiance fields with joint guidance,","cited_arxiv_id":null,"evidence_quote":"Baseline comparison for panoramic reconstruction quality and efficiency."},{"cited_title":"Segment anything,","cited_arxiv_id":null,"evidence_quote":"Provides instance segmentation used to identify regions for prompt-based decoration."},{"cited_title":"Delicate textured mesh recovery from NeRF via adaptive surface refinement,","cited_arxiv_id":null,"evidence_quote":"Supplies the adaptive surface refinement that turns the NeRF field into a high-quality mesh."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Generates the coarse mesh geometry from the NeRF ray-tracing results."}],"review_version":1}