{"id":"9a3ed6da-0e69-4955-9c99-d3f12f1e3ae4","arxiv_id":"2412.11525","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"Ordering low-resolution multi-view images into video-like sequences lets off-the-shelf video super-resolution models outperform existing 3D super-resolution pipelines.","lead":"This paper turns unordered low-resolution photos into video-like sequences so that video super-resolution models can upscale them consistently for 3D reconstruction. The ordering trick improves 3D super-resolution quality on standard benchmarks without needing to fine-tune the video model.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The SOTA claim is undermined by inconsistent metric backgrounds: the paper admits it cannot verify prior baselines' background choice, and its own NeRF-SR numbers differ between Table 3 (28.46) and Table 6 (28.21), so the headline comparison may be apples-to-oranges.","rationale":"The reader identified the VSR robustness assumption as the weakest assumption. While that assumption is indeed central to the method, the paper provides direct evidence for it in the ablation of Table 4, where ALS improves over SISR across three VSR backbones. The SOTA claim, by contrast, depends on fair comparison with published baselines, and the paper explicitly leaves the comparison protocol ambiguous. The background issue is self-admitted and can shift numbers by 0.3-0.4 dB; the margin over some baselines is only about 0.5-1.0 dB (e.g., 31.41 vs 30.83 for SRGS, 31.41 vs 30.71 for CROP). A 0.4 dB adjustment to the baselines would substantially narrow or eliminate the lead. Combined with the unexplained NeRF-SR discrepancy, this makes the headline claim insufficiently supported. The concern is testable and fixable, so a CONDITIONAL verdict remains appropriate, not a rejection.","tokens_in":17684,"tokens_out":7094,"duration_ms":57316,"concrete_test":"Recompute Tables 3 and 6 under a single evaluation protocol: render all methods with both black and white backgrounds (and identical test views and downsampling), and report per-method PSNR/SSIM/LPIPS. Confirm whether Ours-ALS remains ahead of every baseline under both backgrounds, and reconcile the NeRF-SR discrepancy (28.46 vs 28.21) by identifying which number uses which background. A secondary check: rerun the three strongest baselines (CROP, SRGS, SwinIR+3DGS) with the authors' own code or with the same background setting as Ours-ALS.","verdict_should_be":"UNCHANGED","load_bearing_attack":"In Section 4.1 ('Background Impact on Metrics'), the authors state that they measure Blender metrics on a black background, that compositing on white degrades PSNR by 0.3-0.4, and that they are 'unable to determine which background' prior works used. Several baselines (ZS-SRT, CROP, FastSR-NeRF, SRGS, GaussianSR, SuperGaussian) are cited with numbers taken from their original papers, whose background protocol is unknown. If those papers used white backgrounds, their reported PSNRs would be inflated relative to the proposed method's black-background evaluation. The internal record is also inconsistent: NeRF-SR is listed as 28.46 in Table 3, but the per-object average in Table 6 is 28.21, a gap of 0.25 dB that is not explained. Since the central claim is 'state-of-the-art results,' an unfair or unreproducible comparison would invalidate that claim even if the ordering algorithm itself works. This is not a speculative concern: the authors themselves flag the ambiguity and leave it unresolved.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a 3D super-resolution pipeline that reorders low-resolution (LR) multi-view images into video-like sequences, upscales them with a pre-trained video super-resolution (VSR) model, and then trains a 3D Gaussian Splatting (3DGS) model on the upscaled images. The ordering is performed by simple greedy and adaptive-length subsequence algorithms using ORB feature similarity or camera-pose similarity, and a multi-threshold mechanism is introduced to handle variable scene density. The authors report state-of-the-art results on NeRF-Synthetic (Blender) and Mip-NeRF 360, with internal ablations showing consistent improvements from the proposed ordering across three VSR backbones (VRT, IART, PSRT) relative to single-image upsampling and the simple greedy ordering.","tokens_in":17918,"tokens_out":5766,"duration_ms":47934,"significance":"If the evaluation is sound, the paper makes a useful empirical contribution: it shows that a lightweight, training-free ordering of raw LR images can make pre-trained VSR models effective for multi-view 3D super-resolution, avoiding the expensive 3DGS rendering and fine-tuning used by prior VSR-based work (e.g., SuperGaussian). The ablations across three VSR backbones are a genuine strength, as is the use of held-out test views for the final 3D reconstruction comparison. However, the headline state-of-the-art claim depends on cross-method comparisons whose metric background protocol is explicitly acknowledged to be ambiguous, and the manuscript contains internal numerical inconsistencies in the reported baselines. These issues must be resolved before the SOTA claim can be accepted.","major_comments":[{"comment":"The paper's central SOTA claim is undermined by the unresolved background-compositing ambiguity. In Section 4.1 the authors state that Blender metrics are measured on a black background, that white-background compositing degrades PSNR by about 0.3–0.4, and that they are 'unable to determine which background' prior works used. Since several baseline numbers in Table 3 are taken directly from those prior papers and may have been evaluated on white backgrounds, the comparison may be apples-to-oranges. The authors need to provide a like-for-like evaluation: at minimum, report their method's metrics under both black and white backgrounds, re-run the available baselines under the same protocol, and clearly state the sensitivity of the ranking to the background choice.","section":"Section 4.1, Table 3"},{"comment":"There are unexplained internal numerical inconsistencies in the reported baseline results. NeRF-SR is listed as PSNR 28.46, SSIM 0.9210, LPIPS 0.0760 in Table 3, but the per-object averages in Tables 6–8 are 28.21, 0.9180, and 0.0990, respectively. Likewise, in the Mip-NeRF 360 results, Table 12 reports 3DGS-HR with SSIM 0.7710 and LPIPS 0.2802, whereas the per-scene averages in Tables 10 and 11 are 0.7962 and 0.2550. Since these are presumably the same models and the same evaluation set, the discrepancies need to be explained or corrected; otherwise the reliability of the comparison tables is called into question.","section":"Table 3 vs Tables 6–8; Table 12 vs Tables 9–11"},{"comment":"The hyperparameter specification for the adaptive-length subsequence method is not fully reproducible. Appendix B first says that for Blender the ordering uses ORB features with pose-based thresholds, and for Mip-NeRF 360 the ordering uses pose with ORB-feature thresholds, but then states 'we set the three thresholds (angle between two camera positions) to 15°, 30°, 45° on both Blender dataset. We set two thresholds (the number of candidates by distances) to 30 and 50' without clarifying which threshold set applies to which dataset or which similarity measure. The stopping condition in Algorithm 2 also uses an undefined ε. The authors should give a single coherent specification of all thresholds, how they are combined across the multi-threshold scheme, and whether they are tuned on the test set.","section":"Appendix B, Section 3.3"}],"minor_comments":[{"comment":"The text contains a typo: 'Since ASL offers' should read 'Since ALS offers' (adaptive-length subsequence).","section":"Section 4.2"},{"comment":"The affiliation 'Samsung Electorics' is misspelled; it should be 'Samsung Electronics'.","section":"Author affiliations"},{"comment":"The stopping threshold in line 5 is written as 'sim(S(i)_j, S(i)_{j+1}) < ϵ' but ϵ is never defined in the main text; it should be linked to the angle/distance thresholds described in Appendix B.","section":"Algorithm 2"},{"comment":"The phrase 'The sub-pixel loss Lsp is LR 3DGS loss' is confusing; clarify that Lsp is computed between the original LR images and bicubic-downsampled rendered images.","section":"Section 3.4"},{"comment":"The caption in Table 12 says 'Ours' while the per-scene tables and main text refer to 'Ours-ALS'; this should be made consistent.","section":"Table 12"},{"comment":"The sentence 'we set the three thresholds ... on both Blender dataset' is grammatically unclear and should be rewritten to specify that the angle thresholds are used for the Blender dataset and the candidate-distance thresholds are used for the Mip-NeRF 360 dataset, or vice versa, as appropriate.","section":"Appendix B"}],"recommendation":"major_revision","confidential_remarks":"The paper's core idea is simple and the internal ablations are convincing for the value of ordering. However, the authors explicitly acknowledge that they cannot verify the background-compositing protocol of most prior baselines, yet they still claim state-of-the-art results. Combined with the numerical discrepancies between the summary tables and the per-object/per-scene tables, this is currently not acceptable as-is. The issues are fixable within the scope of the manuscript, but the authors must supply a consistent evaluation protocol and resolve the table inconsistencies before the SOTA claim can be sustained."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick read on arXiv:2412.11525. The core finding is real: ordering unordered LR multi-view images into pseudo-video sequences by greedy nearest-neighbor (with adaptive-length subsequencing) lets off-the-shelf VSR models do 3D super-resolution without fine-tuning, and the internal ablations are consistent across three VSR backbones. The ordering gain over SISR is modest but systematic (Table 4: about 0.2-0.3 dB PSNR), and the method sidesteps the rendering-artifact problem that SuperGaussian had to fine-tune its way out of. That is a legitimate contribution and the paper is honest about its limitations, even flagging the background issue in Section 4.1.\n\nThe soft spots are mostly around the SOTA claim. The metric background ambiguity is real: the authors measure Blender metrics on black, cannot verify what prior works used, and admit a white background degrades PSNR by 0.3-0.4 dB. That is exactly the size of their margin over SRGS (31.41 vs 30.83). The internal inconsistency for NeRF-SR—28.46 in Table 3 vs 28.21 in Table 6—is a red flag for careful reporting. And they beat their own HR-3DGS \"upper bound\" on several Mip-NeRF 360 scenes (bicycle, flowers, treehill), which suggests the upper-bound protocol differs from their evaluation protocol or the bound is not actually an upper bound. These issues do not sink the ordering idea, but they do sink the \"state-of-the-art\" wording as it stands.\n\nThe per-dataset hyperparameter tuning (angle thresholds, candidate counts, loss weights, similarity-measure choice) is a minor weakness; this is a preprocessing trick that likely needs per-dataset tuning, and the paper should say so more plainly.\n\nWho is this for? Researchers working on 3D super-resolution or on repurposing video models for multi-view tasks. The ordering insight could transfer to other video-model applications. It deserves a serious referee; with the metric protocol fixed and the numbers re-checked, it could be a useful paper. I'd recommend engaging with it, but treat the SOTA numbers as provisional until the authors reconcile backgrounds and the NeRF-SR inconsistency.","headline":"A simple, well-ablated ordering trick for VSR-based 3D super-resolution, but the SOTA claim is undermined by metric-background ambiguity and an internal numbers inconsistency.","tokens_in":18470,"tokens_out":2448,"would_cite":true,"duration_ms":21702,"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":"Ordering unordered views as video lets off-the-shelf VSR beat dedicated 3D super-resolution.","keywords":["3D super-resolution","video super-resolution","multi-view image ordering","3D Gaussian Splatting","neural radiance fields","ORB features","adaptive-length subsequence","view consistency"],"falsifier":"Train or fine-tune the same VSR model on truly random permutations of the low-resolution multi-view images and compare the resulting 3D reconstruction PSNR with the proposed ORB/pose-ordered pipeline; if random orderings match or exceed the ordered ones, then the ordering is not what drives the improvement.","tokens_in":17496,"feed_emoji":"🎥","tokens_out":4804,"duration_ms":36490,"temperature":0.7,"pith_summary":"This paper claims that the order in which low-resolution multi-view images are fed to a video super-resolution (VSR) model determines whether the model can act as a 3D super-resolution engine. The authors propose two simple ordering algorithms—a greedy nearest-neighbor chain and an adaptive-length subsequence scheme—that arrange the unordered images into video-like sequences using ORB feature similarity and camera-pose similarity. Because the input frames are the original ground-truth low-resolution images rather than renderings from a low-resolution 3D model, the VSR model avoids the stripy and blob-like artifacts that earlier VSR-based approaches suffered from. With no fine-tuning of the VSR backbone, the method reaches state-of-the-art PSNR, SSIM, and LPIPS on the NeRF-synthetic Blender and Mip-NeRF 360 benchmarks, beating both single-image-super-resolution pipelines and the prior VSR-based SuperGaussian.","feed_headline":"Video models win 3D super-resolution after a simple reordering","feed_subtitle":"No fine-tuning or smooth trajectories: ordering low-res views as video yields state-of-the-art 3D reconstruction.","key_machinery":"The load-bearing mechanism is the ordering of unordered multi-view images into video-like sequences. The paper defines a good video as one in which each frame is similar to its adjacent frames, and generates such sequences with a greedy nearest-neighbor search (Algorithm 1) using either ORB feature Hamming distance or camera-pose distance as the similarity measure, plus an adaptive-length subsequence algorithm (Algorithm 2) with multi-threshold generation that produces multiple smoother subsequences from different starting images. The pre-trained VSR model (PSRT is the default backbone) then upsamples each subsequence, and the upsampled frames are aggregated and used to train 3DGS with a sub-pixel loss that regularizes inconsistent high-frequency details.","core_discovery":"The central discovery is that a pre-trained video super-resolution model trained on natural videos can outperform dedicated 3D super-resolution methods as long as the unordered multi-view images are first reordered into a sequence whose consecutive frames are mutually similar. The paper demonstrates this with a simple greedy algorithm that repeatedly links each frame to its nearest neighbor, and an adaptive-length subsequence variant that breaks chains when similarity falls below a threshold and generates multiple overlapping sequences from every starting image. These sequences are upsampled by the VSR model and aggregated, then used to train a 3D Gaussian Splatting model with a sub-pixel consistency loss. On the Blender and Mip-NeRF 360 benchmarks, this pipeline improves over all reported SISR- and render-based 3D super-resolution baselines, closing much of the gap to the ground-truth upper bound.","pith_inferences":["The same ordering principle could be applied to other multi-view tasks that feed temporal models, such as video-diffusion-based novel view synthesis or multi-frame deblurring, where a good ordering may be as important as the model itself.","Because the ordering uses only pairwise similarity, it could be computed in an online or streaming setting, enabling VSR-based super-resolution for unordered image collections captured by casual cameras without pose estimation.","The adaptive-length subsequence with multi-threshold suggests a general recipe: use strict similarity thresholds where data is dense and relax them where data is sparse; this may transfer to other sequence-modeling applications where ordering is combinatorial.","Replacing ORB with learned local features or semantic similarity might improve ordering on scenes with repetitive texture or wide baselines, a plausible testable extension."],"forward_implications":["Any pre-trained VSR model can be dropped into the pipeline without fine-tuning, making 3D super-resolution a data-ordering problem rather than a model-training problem.","The ordering algorithms generalize across VSR backbones (VRT, IART, PSRT), with consistent gains over single-image processing.","The approach avoids the distribution-shift artifacts of render-based pipelines because the VSR model sees original ground-truth images, not renderings from a low-resolution 3D model.","The sub-pixel loss stabilizes 3DGS training on upsampled images, improving view consistency and closing part of the gap to high-resolution-trained upper bounds."],"supporting_citations":[{"why":"SuperGaussian, the prior VSR-based 3D super-resolution method that renders smooth videos from a low-resolution 3D model; this paper directly addresses its artifact and fine-tuning limitations.","marker":"Shen et al. 2024"},{"why":"PSRT, the video super-resolution transformer used as the default VSR backbone in the proposed pipeline.","marker":"Shi et al. 2022"},{"why":"ORB features, the visual similarity measure used to order the multi-view images into video-like sequences.","marker":"Rublee et al. 2011"},{"why":"3D Gaussian Splatting, the 3D representation trained on the upsampled high-resolution images to produce the final 3D model.","marker":"Kerbl et al. 2023"},{"why":"NeRF and the NeRF-synthetic Blender dataset, one of the two benchmark suites used for evaluation.","marker":"Mildenhall et al. 2021"},{"why":"Mip-NeRF 360 dataset, the real-world scene-level benchmark used to evaluate generalization beyond object-centric scenes.","marker":"Barron et al. 2022"},{"why":"NeRF-SR, a SISR-based baseline and the source of the sub-pixel consistency loss used to regularize the upsampled images.","marker":"Wang et al. 2022"}],"fun_headline_variants":["Reordered low-res views as video push 3D super-resolution to SOTA","Video models excel at 3D super-resolution with simple frame linking","No fine-tuning, just sorting: video models lead 3D super-resolution tasks","Greedy view reordering lets video models surpass dedicated 3D upscaling","Frame similarity ordering makes video models top 3D super-resolution methods"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that a video super-resolution model trained on natural videos can extract useful temporal information from a sequence of static multi-view images that are linked only by visual or pose similarity, even when the implied motion is irregular or abrupt; if that premise fails, the ordering provides no benefit and the method degrades to per-image super-resolution.","fun_headline_variants_meta":{"raw":{"variants":["Reordered low-res views as video push 3D super-resolution to SOTA","Video models excel at 3D super-resolution with simple frame linking","No fine-tuning, just sorting: video models lead 3D super-resolution tasks","Greedy view reordering lets video models surpass dedicated 3D upscaling","Frame similarity ordering makes video models top 3D super-resolution methods"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000564,"raw_usage":{"total_tokens":2679,"prompt_tokens":951,"completion_tokens":1728,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":567,"completion_tokens_details":{"reasoning_tokens":1627}},"tokens_in":567,"tokens_out":1728,"duration_ms":14543,"temperature":1.0,"reasoning_tokens":1627,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T14:49:58.892869+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train or fine-tune the same VSR model on truly random permutations of the low-resolution multi-view images and compare the resulting 3D reconstruction PSNR with the proposed ORB/pose-ordered pipeline; if random orderings match or exceed the ordered ones, then the ordering is not what drives the improvement.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"PSRT, the video super-resolution transformer used as the default VSR backbone in the proposed pipeline."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"ORB features, the visual similarity measure used to order the multi-view images into video-like sequences."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"NeRF-SR, a SISR-based baseline and the source of the sub-pixel consistency loss used to regularize the upsampled images."}],"review_version":1}