{"id":"96305e60-27f8-4098-b120-f6cf12edd003","arxiv_id":"2607.17896","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":9,"one_line_summary":"A locality-aware density-control framework for 2D Gaussian image representation that densifies coherent high-error regions and merges redundant similar Gaussians, improving PSNR at fixed budgets.","lead":"This paper introduces LocoADC, a training-time add-on for 2D Gaussian Splatting image representation that adds Gaussians to coherent error regions and merges similar neighboring Gaussians. It reports consistent PSNR gains, up to 2.93 dB over GI on CLIC at equal Gaussian budgets, at the cost of some MS-SSIM loss.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Schedule confound: no baseline runs LocoADC's staged training schedule without RGD/SDGM, so the 2.93 dB gain may not be due to locality-aware components.","rationale":"The paper is well-engineered and presents consistent PSNR improvements across baselines, with useful ablations of individual components and hyperparameter sensitivity. However, the most load-bearing assumption for the central claim is that the gains are caused by the locality-aware RGD and SDGM mechanisms, not merely by the staged training schedule. The implementation details describe a distinctive 50k/20k/50k schedule, but no experiment holds that schedule fixed while toggling the proposed components. Baselines are compared using their own schedules at the same iteration count, and the component ablations also implicitly change the schedule when merging is removed. A schedule-only control is therefore essential to attribute the reported gains. Without it, the paper's headline result—2.93 dB over GI on CLIC—could be substantially a schedule artifact. The reader identified exactly this concern, and I agree. I also note the secondary issue of MS-SSIM drops at low budgets, but the schedule confound is more fundamental because it questions the cause of the primary metric gains. The proposed test directly settles the attribution; until it is run, a conditional verdict is appropriate.","tokens_in":19344,"tokens_out":8132,"duration_ms":69949,"concrete_test":"Run GI-Plus (and GI) with LocoADC's exact schedule—50k fit+densify, 20k with densification paused and no merging, 50k fine-tune with densification (ops every 5k)—but substitute the baseline's own pixel-wise densification for RGD and disable SDGM and the color consistency loss. Compare PSNR/MS-SSIM on Kodak with a 10k budget and CLIC with a 30k budget. If this schedule-only variant closes more than half of the gap between the official-schedule baseline and full LocoADC, the component attribution fails; if it stays within 0.2 dB of the official baseline, the concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim attributes PSNR gains to the locality-aware RGD and SDGM components, but the staged training schedule is never held fixed. LocoADC uses 50k iterations of fitting+densification, 20k merge-only, 50k fine-tuning with continued densification, with operations every 5k (Sec. 4.1). Baselines are run with their official schedules for the same total iterations, not with LocoADC's schedule. The ablations in Table 2 also confound schedule with component: removing merging also removes the 20k merge-only stage, so comparisons such as row 3 vs row 1 change both the component and the optimization trajectory. A schedule-only control—applying the same merge-pause/fine-tune timing with the baseline's own densification and without SDGM/color loss—is absent. If most of the 2.93 dB CLIC gain or the ~0.7 dB Kodak 10k gain over GI-Plus arises from the extra merge-only phase and fine-tuning, the locality-aware mechanisms are not the primary cause. This is load-bearing because the abstract and contributions explicitly credit RGD and SDGM for the improvements.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes LocoADC, a density-control framework for 2D Gaussian image representation, with two components: Region-wise Gaussian Densification (RGD), which allocates new Gaussians based on locally coherent distortion structures, and Similarity-Driven Gaussian Merging (SDGM), which merges redundant neighboring Gaussians under appearance/stability constraints, supported by a local color consistency regularizer. Experiments on Kodak, DIV2K×2, and CLIC show PSNR improvements over GI, EA-GI, and GI-Plus across several Gaussian budgets, including the headline 2.93 dB gain over GI on CLIC at a 30k Gaussian budget. The appendix includes component ablations, hyperparameter sensitivity, and comparisons with 3DGS sparsification strategies.","tokens_in":19710,"tokens_out":7500,"duration_ms":70158,"significance":"If the reported gains are attributable to the proposed locality-aware mechanisms, LocoADC would be a useful, plug-and-play improvement for Gaussian-based image representation, with potential implications for image compression. The paper has clear strengths: evaluation across three datasets, three baselines, multiple budgets, component ablations, a sensitivity analysis, pseudocode, and a public code link. However, the empirical attribution is currently weakened by a training-schedule confound and by unaddressed MS-SSIM trade-offs at low budgets. These issues are fixable but load-bearing for the central claim that the RGD/SDGM components themselves are responsible for the consistent improvements.","major_comments":[{"comment":"The training schedule is not controlled between LocoADC and the baselines. LocoADC uses a staged schedule (50k iterations of fitting+densification, 20k merge-only, 50k fine-tuning with continued densification; operations every 5k), whereas baseline results are obtained using their official schedules for the same total 120k iterations. Thus the reported gains, including the 2.93 dB CLIC number, may partly reflect the extra merge-only and fine-tuning phases rather than RGD/SDGM. The ablations in Table 2 also change the schedule when a component is removed (e.g., the 20k merge-only stage is present only when merging is active). Please add a schedule-only control in which each baseline, or at least GI-Plus, runs the identical staged schedule without RGD/SDGM/color loss, and report those results alongside Table 1 and Table 2.","section":"§4.1, Tables 1–2"},{"comment":"MS-SSIM drops at low Gaussian budgets are not discussed. For example, on Kodak at 5k, GI goes from 0.9632 to 0.9560 with LocoADC, and EA-GI from 0.9614 to 0.9551; at 10k on Kodak, GI drops from 0.9824 to 0.9794. The text states LocoADC 'remain[s] competitive in MS-SSIM,' which is not supported. The ablation in Table 7 shows that merging alone reduces MS-SSIM from 0.9640 (RGD only) to 0.9561, with color consistency only partially recovering it to 0.9596. The paper should explicitly report and analyze this PSNR–MS-SSIM trade-off, and qualify the 'consistent improvement' claim accordingly.","section":"Table 1, §4.4"},{"comment":"The gradient-stability filter appears internally inconsistent. The text says 'we first filter out unstable Gaussians according to their accumulated optimization gradients,' explaining that Gaussians with large gradients are under active correction and may incur large errors after merging. But the stated condition, g_i / max_k g_k < 0.5, would select low-gradient (stable) Gaussians for removal if 'filter out' is interpreted literally, which is the opposite of the stated intent. Please clarify whether the threshold is intended to remove high-gradient Gaussians (i.e., retain those with relative gradient below 0.5) or to remove low-gradient ones, and correct Eq. (11) and Algorithm 2 to match the code. This affects the reproducibility and correctness of SDGM.","section":"Eq. (11), Algorithm 2"},{"comment":"The component ablation is not a factorial design that isolates each mechanism while holding the optimization trajectory fixed. Adding RGD changes both the densification criterion and the densification schedule; adding merging changes both the sparsification operation and the 20k merge-only phase. Without a control that uses the same schedule and operation timing while toggling only the RGD/SDGM decisions, the attribution of the observed gains to the locality-aware criteria is not established. This is closely related to the schedule-confound concern above and should be addressed in the same set of experiments.","section":"§4.4, Table 2"}],"minor_comments":[{"comment":"The text refers to 'Fig. 9(a)' and 'Fig. 9(b)' when discussing the local color consistency constraint, but the displayed figure is captioned 'Figure 7: Effectiveness of local color consistency constraint.' The figure references in the appendix's Fig. 9 are also confusing because that figure shows rate-distortion/time-quality plots. Please renumber or fix the cross-references.","section":"§4.4"},{"comment":"The notation D(x) for local distortion coherence in Eq. (7)–(8) is easily confused with the densification operator D(G_t) in Eq. (4). Consider using a different symbol, e.g., C(x) or H(x), for the coherence map.","section":"§3.2, Eq. (8)"},{"comment":"For a fair comparison, the statement that all methods use 120,000 iterations is insufficient. Please specify the exact schedules, learning rates, and other hyperparameters used for each baseline, and report the number of seeds/runs and standard deviations for the main tables.","section":"§4.1"},{"comment":"The grouping of 'spatially neighboring pixels' in RGD is not formalized. Please provide the connectivity/radius criterion used to form regional subsets, as this affects reproduction and the number of Gaussians allocated per region.","section":"Appendix A, Algorithm 1"},{"comment":"The gradient-cutoff sensitivity row shows MS-SSIM varying from 0.9622 (at 0.1) to 0.9582 (at 0.9), a larger spread than PSNR. This trade-off is not discussed and should be acknowledged, especially since the default 0.5 gives the lowest MS-SSIM among tested cutoff values.","section":"Table 6"}],"recommendation":"major_revision","confidential_remarks":"The training-schedule confound is the most serious issue. If the staged schedule alone accounts for a substantial fraction of the reported gains, the central attribution to RGD/SDGM collapses. The MS-SSIM issue is secondary but still needs explicit treatment. Both can be addressed with additional experiments and analysis, so I see this as a major revision rather than a rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: the paper is a serious empirical contribution to 2D Gaussian splatting, but before I'd trust the headline gains I'd want a schedule-only control.\n\nThe new thing here is the specific combination of region-wise densification driven by local distortion coherence, similarity-driven merging with a color consistency regularizer, applied to 2D Gaussian image representation. That combination is not in the cited literature, and the paper does a solid job motivating it with the locality analysis. The experiments are broad: three datasets, three baselines, several Gaussian budgets, plus ablations and hyperparameter sensitivity. The reported PSNR gains are consistent and often large, e.g., 2.93 dB over GI on CLIC at 30k Gaussians. The ablations in Table 5 and the comparison with magnitude-only densification give some evidence that the coherence term matters.\n\nThe soft spots are real, though. The biggest is the schedule confound. LocoADC uses a staged schedule: 50k fitting+densification, 20k merge-only, 50k fine-tuning, with operations every 5k. Baselines are run with their official schedules for 120k total iterations, not with LocoADC's schedule. So the gains could come from the merge-only pause and extended fine-tuning rather than from RGD and SDGM themselves. The ablations in Table 2 don't fix this, because removing merging also removes the 20k merge-only stage. A schedule-only baseline—same staged timing with baseline densification and no SDGM/color loss—is missing. That's load-bearing, since the abstract credits RGD and SDGM for the improvements.\n\nSecond, the main text says 'consistently' but MS-SSIM drops at low budgets. For example, on Kodak 5k, GI goes from 0.9632 to 0.956 and EA-GI from 0.9614 to 0.9551. The appendix mentions merging may slightly smooth edges, but the main text doesn't flag these trade-offs.\n\nThe code is promised but not inspectable, and the hyperparameters are hand-chosen, though the sensitivity analysis shows the method isn't brittle.\n\nAll in all, the paper is worth engaging with. The idea is plausible and the empirical scope is good. But the central attribution is not yet demonstrated. A serious referee should ask for the schedule-only ablation and a more honest metric discussion. I'd send it to review, but I'd expect major revision.","headline":"Well-engineered 2DGS density control paper, but the headline gains are confounded by a non-standard training schedule that is never controlled.","tokens_in":20171,"tokens_out":2541,"would_cite":false,"duration_ms":23178,"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":"The paper argues that the bottleneck in 2D-Gaussian image representation is locality-unaware density control, and that fixing it with region-wise densification and similarity-driven merging yields consistent quality gains under equal Gaussi","keywords":["Image representation","Gaussian splatting","Density control","Region-wise densification","Gaussian merging","Local distortion coherence","2D Gaussians","PSNR"],"falsifier":"Run GI and GI-Plus through the full LocoADC schedule with RGD replaced by standard pixel-wise densification and SDGM disabled, keeping every other detail fixed (120k iterations, three stages, 5k operation intervals, same Gaussian budgets). If the PSNR gain over the published baselines largely persists, the locality-aware components are not the cause.","tokens_in":1435,"feed_emoji":"🖼️","tokens_out":2173,"duration_ms":76136,"temperature":0.7,"pith_summary":"The paper sets out to show that how a fixed budget of 2D Gaussians is placed during optimization matters as much as how many are used. It identifies two failures in existing methods: under-reconstructed areas are densified pixel-by-pixel instead of as coherent regions, and over-reconstructed areas keep redundant neighboring Gaussians with nearly identical appearance. To fix this, it proposes LocoADC, a plug-and-play density control with Region-wise Gaussian Densification (RGD) and Similarity-Driven Gaussian Merging (SDGM), plus a local color consistency constraint that makes merging safe. Across Kodak, DIV2K×2, and CLIC, the framework improves the GI, EA-GI, and GI-Plus baselines under the same Gaussian budgets, including a 2.93 dB PSNR gain over GI on CLIC at 30k Gaussians. If correct, this means better image quality at equal storage and decoding cost, since the rasterization and decoding pipeline is unchanged.","feed_headline":"Region-wise densification and merging give Gaussian images +2.93 dB","feed_subtitle":"LocoADC gives better image quality from the same Gaussian budget, lifting GI on CLIC from 32.50 to 35.43 dB.","key_machinery":"The mechanism is a dual locality model. In image space, local distortion coherence D(x) = ||mean of normalized distortion gradients over a w×w window||², multiplied by distortion E(x), forms the densification candidate set; spatially neighboring candidates are grouped into regions and one Gaussian is initialized per region. In Gaussian space, similarity-driven merging selects stable, overlapping, color-similar Gaussian pairs via Mahalanobis distance and a color threshold, ranks them by a Pair Discrepancy Score over position, color, and covariance, and merges via area-weighted position/color averaging with second-moment covariance matching under a support-area constraint. The supporting local","core_discovery":"On its own terms, the central discovery is that reconstruction errors in Gaussian image representation are not isolated pixels but locally coherent patterns, and that spatially neighboring Gaussians converge to redundant similar shapes; exploiting both forms of locality yields a better use of a fixed Gaussian capacity. Concretely, LocoADC uses the distortion map's normalized gradient field to measure local distortion coherence, weights the distortion by this coherence, selects top candidates, groups neighboring candidates into regions, and allocates one new Gaussian per region. In parallel, it filters stable Gaussians, pairs spatially overlapping neighbors with similar colors, ranks pairs by","pith_inferences":["A clean control experiment is missing in the paper: running the baselines through the exact three-stage LocoADC schedule (50k fit+densify, 20k merge-only, 50k fine-tune, operations every 5k iterations) with RGD and SDGM disabled. Until that is reported, part of the gain could be an artifact of the schedule itself.","The same locality principles could transfer to video or 3D Gaussian representations: wherever reconstruction errors are spatially coherent and neighboring primitives converge to similar appearance, region-wise densification and similarity-driven merging should free capacity; the paper lists these as future work rather than demonstrating them.","One observable side effect to watch: merging alone lowers MS-SSIM slightly (PSNR gains from 31.90 to 32.34 dB but MS-SSIM drops from 0.9614 to 0.9561), so aggressive merging trades a bit of fine-scale metric fidelity for capacity; the color-consistency constraint only partially compensates, suggesting a perceptual ceiling on how much redundancy can be safely merged."],"forward_implications":["Under the same Gaussian budget, density control that respects locality can produce materially better reconstructions: the paper reports up to +2.93 dB PSNR (GI on CLIC at 30k Gaussians) and consistent gains across all three baselines and datasets.","Because LocoADC changes only density control and not the rasterization or decoding pipeline, the quality gains translate directly to existing Gaussian-based representations without extra decoding cost.","Merging redundant similar Gaussians is more effective than pruning small ones: the paper compares against size-based and learning-based 3DGS sparsification and reports higher PSNR with merging (32.68 dB vs 30.5 and 32.19 on Kodak).","The ablation indicates both components contribute: RGD alone adds about 0.34 dB over pixel-wise densification, and coherence-aware identification beats magnitude-only region-wise densification; SDGM plus color consistency adds further gains.","After convergence the method reaches a better PSNR-time trade-off than the baselines despite higher training cost, and its rate-distortion curve approaches learned compression methods on the reported benchmarks."],"fun_headline_variants":["Locality-aware density control: +2.93 dB on CLIC","Region-wise densification + merging: +2.93 dB on CLIC","LocoADC: smarter Gaussian allocation, +2.93 dB on CLIC","Better Gaussian images from local density control: +2.93 dB","Local allocation lifts Gaussian images: +2.93 dB on CLIC"],"cache_read_input_tokens":21504,"weakest_assumption_plain":"That the PSNR improvements are caused by the locality-aware RGD/SDGM operations rather than by the altered training schedule alone; the paper does not run the baselines through the identical schedule without those operations.","fun_headline_variants_meta":{"raw":{"variants":["Locality-aware density control: +2.93 dB on CLIC","Region-wise densification + merging: +2.93 dB on CLIC","LocoADC: smarter Gaussian allocation, +2.93 dB on CLIC","Better Gaussian images from local density control: +2.93 dB","Local allocation lifts Gaussian images: +2.93 dB on CLIC"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001075,"raw_usage":{"total_tokens":4355,"prompt_tokens":784,"completion_tokens":3571,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":528,"completion_tokens_details":{"reasoning_tokens":3483}},"tokens_in":528,"tokens_out":3571,"duration_ms":24104,"temperature":1.0,"reasoning_tokens":3483,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-01T16:39:39.319722+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run GI and GI-Plus through the full LocoADC schedule with RGD replaced by standard pixel-wise densification and SDGM disabled, keeping every other detail fixed (120k iterations, three stages, 5k operation intervals, same Gaussian budgets). If the PSNR gain over the published baselines largely persists, the locality-aware components are not the cause.","supporting_citations":[],"review_version":1}