{"id":"a1689fb1-ffa3-40cd-9eb2-79d1d148c4eb","arxiv_id":"2411.18473","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"HEMGS compresses 3D Gaussian Splatting data with a single variable-rate model and a joint autoregressive/hyperprior entropy model, achieving roughly 40% smaller files than HAC at similar quality.","lead":"This paper presents HEMGS, a compression method for 3D Gaussian Splatting (3DGS) scenes that uses a hybrid entropy model to support adjustable compression rates with a single trained model and to reduce stored size by about 40% versus a strong prior baseline. It matters because 3DGS is a popular real-time renderable 3D representation whose large file sizes limit streaming and storage.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The headline 40% savings is computed at lower rendering quality (Table 1 low-rate rows), not at matched PSNR, and the equal-quality BD-rate evidence covers only two of four benchmarks; the central quantitative claim is therefore not yet supported.","rationale":"The reader's verdict is CONDITIONAL with moderate confidence, and I agree that the paper should not be accepted as-is. However, I identify a different load-bearing concern than the reader's stated weakest assumption. The reader focuses on whether the adaptive autoregressive context selection generalizes beyond the playroom ablation; that is a legitimate modeling risk, but the more direct threat to the central claim is that the headline 40% number is not computed at matched rendering quality. Table 1's low-rate HEMGS rows achieve large size reductions while showing lower PSNR/SSIM/LPIPS than HAC on three of four datasets, so 'maintaining rendering quality' is not established. The paper does provide some BDBR support, which is real evidence in the method's favor, but it is incomplete and lacks per-scene detail. The concrete test I propose would settle the question by replacing un-matched rate-point ratios with a standard equal-quality BD-rate comparison over all benchmarks. Because the reader's rationale already flags missing RD curves and error bars, there is partial overlap, but the weakest_assumption field itself points elsewhere. I would keep the verdict at CONDITIONAL: the paper is plausible and the architecture is reasonable, but the quantitative headline needs to be verified at matched quality before the central claim is accepted.","tokens_in":13604,"tokens_out":11987,"duration_ms":114572,"concrete_test":"Recompute the comparison from full per-scene rate-distortion data for all four benchmarks: run HEMGS at all lambda values (e.g., the 16 rate points shown in Figure 4) and HAC at all available rate points, then compute the Bjøntegaard BD-rate (average bitrate difference at equal PSNR) separately for every scene and report the mean and standard error across scenes. If the average BD-rate saving over HAC is below about 30% or its confidence interval includes zero, the 'about 40%' headline should be revised or removed; the same analysis should also be reported for SSIM and LPIPS.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that HEMGS achieves about a 40% average reduction in stored size while maintaining rendering quality relative to HAC. The evidence for this is not yet quality-matched. In Table 1, the low-rate HEMGS rows show size reductions of 42.8%, 45.5%, 42.2%, and 36.6% across Mip-NeRF360, Tank&Temples, DeepBlending, and Synthetic-NeRF, but the quality comparisons are mostly worse than HAC: PSNR is 27.68 vs 27.77 on Mip-NeRF360, 30.24 vs 30.34 on DeepBlending, and 33.33 vs 33.71 on Synthetic-NeRF, with SSIM and LPIPS also slightly worse in several cases. Only Tank&Temples is marginally better. The abstract's number is therefore the mean of un-matched rate-point size ratios, not an equal-distortion comparison. The BDBR table (Table 2) covers only Mip-NeRF360 and DeepBlending, without per-scene RD data or uncertainty estimates, and Figure 3 omits Synthetic-NeRF. If HEMGS is evaluated at the PSNR of HAC, or if the standard Bjøntegaard BD-rate is computed over all four datasets, the average saving could be materially different from 40%. This is a load-bearing concern because the headline number is the paper's main quantitative claim.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes HEMGS, a hybrid entropy model for compressing anchor-based 3D Gaussian Splatting (3DGS) data. The framework combines (i) a variable-rate predictor that takes a user-specified lambda as input to produce a learned quantization step, enabling multiple rate-distortion trade-offs with a single model; (ii) a hyperprior network that fuses scene-agnostic features from a pretrained PointNet++ with scene-specific features from a hash grid; and (iii) an autoregressive network with an adaptive context selection mechanism that chooses the nearest n already-encoded anchors within a large receptive field. These components are integrated into a hybrid lossy-lossless compression pipeline that first quantizes anchor locations, local features, scaling, and offsets, then entropy-codes the quantized values with arithmetic coding. Experiments on Synthetic-NeRF, Tank&Temples, Mip-NeRF360, and DeepBlending report about a 40% average size reduction over the HAC baseline while maintaining rendering quality, as well as variable-rate operation with a single trained model.","tokens_in":13985,"tokens_out":5918,"duration_ms":48321,"significance":"If the claimed results hold, HEMGS would be a meaningful advance in 3DGS storage and transmission: a single model covering multiple rate points is practically valuable, and the reported size reductions over HAC are substantial. The paper has clear strengths: it reports actual compressed file sizes from arithmetic coding rather than cross-entropy estimates; it includes ablations isolating the scene-agnostic network, the autoregressive network, and the adaptive context selection; and it demonstrates variable-rate behavior with a single model. However, the central quantitative claim is not yet fully supported by the evidence in the main text. The headline 40% savings is computed at lower rendering quality in several datasets, the BD-rate evaluation covers only two of the four benchmarks, and the adaptive context model is validated on a single scene. The significance is therefore conditional on a more complete, quality-matched evaluation.","major_comments":[{"comment":"The claim of 'about a 40% average reduction in size while maintaining rendering quality' is not directly supported by Table 1. The low-rate rows achieve the largest size reductions (e.g., 42.8% on Mip-NeRF360, 45.5% on Tank&Temples, 42.2% on DeepBlending, 36.6% on Synthetic-NeRF), but at those operating points PSNR, SSIM, or LPIPS are worse than HAC on most datasets (e.g., Mip-NeRF360 PSNR 27.68 vs 27.77; DeepBlending PSNR 30.24 vs 30.34; Synthetic-NeRF PSNR 33.33 vs 33.71). The high-rate rows are competitive or better in some cases, but the size reductions are much smaller (12.9%–33.6%). Thus the 40% average appears to be computed at unmatched quality, not at equal distortion. The authors should either report BD-rate (or BD-PSNR) over all four datasets or clearly state the distortion level at which each size reduction is measured.","section":"Section 4.3, Table 1"},{"comment":"The BDBR results are reported for only two of the four benchmarks (Mip-NeRF360 and DeepBlending). The abstract claims an average storage saving of approximately 40% across all benchmarks, but without equal-quality comparisons for Tank&Temples and Synthetic-NeRF, the headline number is not verifiable from the presented data. Moreover, the two reported BDBR values for HAC differ substantially (52.52% vs 82.85%), and no per-scene RD data or uncertainty estimates are given, so the robustness of the average is unclear.","section":"Section 4.3, Table 2"},{"comment":"The adaptive context selection algorithm, a core component of the autoregressive network, is evaluated only on the 'playroom' scene from DeepBlending. The storage difference between 'Ours' and 'Ours w/o ACS' is 0.19 MB on that single scene, and the generalization of the chosen threshold n=20 and receptive field 25x25x25 to other scenes and datasets is not demonstrated. Please provide ablations on at least one additional scene per dataset, or otherwise justify the universality of these hyperparameters.","section":"Section 3.3.3 and Table 4"},{"comment":"The rate-distortion curves omit the Synthetic-NeRF dataset, and each curve contains only two operating points per method. Since the paper claims state-of-the-art compression across four benchmarks, the RD evidence is incomplete. The variable-rate model can produce at least 16 rate points (as shown in Figure 4 for Tank&Temples); full RD curves for all four datasets with more rate points should be provided to substantiate the claim.","section":"Section 4.3, Figure 3"},{"comment":"The mapping between the four predefined lambda values (1e-3, 2e-3, 3e-3, 4e-3) and the 'low-rate' and 'high-rate' rows in Table 1 is not reported. Without this mapping, the reader cannot reproduce the results or understand how the size-quality trade-off is controlled. Please specify which lambda corresponds to each reported row and, if only two of the four lambda values are used for the main comparisons, explain why the other two are omitted.","section":"Section 4.2 and Table 1"}],"minor_comments":[{"comment":"The description of location compression says the quantized location is 'directly losslessly encoded into a bit-stream'; please clarify the entropy coding scheme used for the location (e.g., fixed-rate arithmetic coding or a learned distribution).","section":"Section 3.2"},{"comment":"Figure 2(c) marks the scene-agnostic network as 'Frozen', but the text says it introduces 'no additional storage overhead'; please state explicitly whether the PointNet++ weights are frozen and whether any fine-tuning occurs.","section":"Section 3.3.2"},{"comment":"In Eq. (2), lambda is both the rate-distortion trade-off weight and an input to the variable-rate predictor. Please use separate symbols for these two roles to avoid ambiguity.","section":"Section 3.4"},{"comment":"The column header 'Fidlity' is misspelled; it should be 'Fidelity'.","section":"Table 5"},{"comment":"The paper references supplementary materials for per-scene results, the definition of Lanchor, and other implementation details, but the arXiv version does not include a supplementary file. Please ensure the supplementary material is publicly available at the time of publication.","section":"Throughout"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is within the scope of the journal. The main revision request is to provide a proper BD-rate evaluation over all datasets and to clarify the quality-matched comparisons; without that, the headline 40% claim is not supportable. The authors should also make supplementary material available, as per-scene data and implementation details are referenced but currently missing from the arXiv version."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a genuine incremental advance in 3DGS compression, but the paper's central number is overstated. The new piece is the combination of a variable-rate predictor (lambda-conditioned quantization step), a hyperprior that mixes a frozen PointNet++ prior with scene-specific features, and an autoregressive network with adaptive context selection. That combination is not in HAC or ContextGS, and the ablations show each component buys something on the playroom scene. The single-model multi-rate feature is useful, and the complexity table is refreshingly honest.\n\nThe soft spots are mostly evaluative. The abstract and Section 4.3 claim about a 40% average reduction \"while maintaining rendering quality\" does not hold up in Table 1. The low-rate rows, which drive the average, have lower PSNR than HAC on three of four datasets and worse SSIM/LPIPS in several cases. The high-rate rows are quality-matched or slightly better but only save 13-34%. So the 40% is a rate saving at quality expense, not an equal-distortion saving. The BDBR table is partial: only two of four benchmarks, and without full RD curves or uncertainty. Figure 3 omits Synthetic-NeRF. The adaptive context selection is only ablated on one scene, and the assumption that nearest-n elements capture dependencies is plausible but not demonstrated across scene types. No code or per-scene numbers in the main paper.\n\nNone of this is fatal to the method. The architecture is sensible, the empirical work is decent, and the limitations are acknowledged. But the headline claim needs to be re-baselined: either report BD-rate over all four datasets, or report matched-PSNR size comparisons. As it stands, the paper is a solid contribution that would benefit from a revision before acceptance.\n\nI'd send it to review — it is a serious paper with enough new machinery that a referee can improve it. I wouldn't quote the 40% number in any abstract until the evaluation matches the claim.","headline":"A useful incremental advance in 3DGS compression with a plausible new architecture, but the headline 40% saving is computed at lower rendering quality and needs a matched-quality comparison before it should be quoted.","tokens_in":14479,"tokens_out":3021,"would_cite":true,"duration_ms":25597,"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":"This paper claims that a hybrid entropy model compresses 3D Gaussian Splatting scenes to about 40% smaller files while maintaining rendering quality, using a single model that covers many bitrates via a rate hyperparameter.","keywords":["3D Gaussian Splatting","compression","entropy model","variable-rate","hyperprior","autoregressive","scene-agnostic prior","lossy-lossless coding"],"falsifier":"Compress a synthetic scene containing two spatially separated clusters of identical anchors whose attributes are correlated with each other's cluster. If the nearest-n context cannot exploit that cross-cluster redundancy, the losslessly encoded size will be no smaller than a no-context baseline, showing that the adaptive context selection fails on non-local dependencies.","tokens_in":13425,"feed_emoji":"💾","tokens_out":8570,"duration_ms":67783,"temperature":0.7,"pith_summary":"3D Gaussian Splatting is a popular way to represent a scene as a cloud of tiny colored ellipsoids, but the resulting files are heavy. This paper claims that a hybrid entropy model, HEMGS, can shrink those files by about 40% compared with the current best baseline, HAC, while rendering the same image quality. The model works on anchor points—control points that summarize groups of Gaussians—and compresses each anchor attribute with a mix of lossy quantization and lossless entropy coding. A single trained model covers many compression rates, selected by a hyperparameter λ, so users can pick the storage-quality trade-off without retraining. If the claim holds, storing and streaming 3D scenes becomes much cheaper.","feed_headline":"Hybrid entropy model cuts 3D Gaussian storage by 40%","feed_subtitle":"A single model covers many rates, so users can trade size for quality without retraining.","key_machinery":"The load-bearing mechanism is HEMGS, a hybrid entropy model with three cooperating parts that run per anchor attribute. First, the variable-rate predictor takes the rate hyperparameter λ, broadcasts it to the feature dimension, and outputs a learned quantization step that controls how coarsely the attribute is quantized. Second, the hyperprior network computes a prior feature from already-compressed location and attribute data, using a frozen scene-agnostic extractor (PointNet++) plus a scene-specific hash-grid trained on the current scene. Third, the autoregressive network selects context elements adaptively: with a maximum receptive field of 25×25×25 voxels, it retains all anchors in sparse areas but only the nearest 20 in dense areas. The prior and context features are combined to estimate the Gaussian distribution of each coding element, which is then fed to an arithmetic coder. This design is what lets HEMGS do lossy rate control and lossless entropy coding in one pass.","core_discovery":"The paper's central claim is that a single neural network, combining a variable-rate predictor, a hyperprior network, and an autoregressive network, can outperform specialized 3DGS compression methods on all four benchmarks. The variable-rate predictor produces a learned quantization step from a user-specified λ, replacing the need for multiple rate-specific models. The hyperprior network fuses a scene-agnostic feature (from a frozen pre-trained 3D feature extractor, PointNet++) with a scene-specific hash-grid prior, while the autoregressive network adapts its context window to local anchor density. Together these produce accurate per-element probability estimates for the arithmetic coder, yielding roughly 40% average storage reduction over HAC at equal PSNR, and a single model that spans 16 rate points.","pith_inferences":["The adaptive context selection (all anchors in sparse regions, nearest 20 in dense ones) is a natural candidate for other sparse 3D data such as point clouds or LiDAR scans, where density varies similarly and fixed receptive fields are inefficient.","Because the scene-agnostic network is frozen and only the scene-specific hash-grid is overfit, HEMGS may compress a new scene with less per-scene optimization than methods that retrain the whole entropy model, which could speed deployment.","A worthwhile test is whether the variable-rate predictor's learned quantization step transfers to other neural compression domains (images, video, or point clouds), where rate-conditioned retraining is currently the norm; the same λ-to-step mechanism could replace that."],"forward_implications":["With one trained model, HEMGS can produce 16 rate-distortion points by varying λ, so applications can switch storage budgets without retraining or storing multiple models.","On the four benchmark datasets, HEMGS stores scenes in about 40% less space than HAC at matched PSNR; on Mip-NeRF360 it reports BDBR savings of 52.52% over HAC and 66.47% over Context-GS.","The autoregressive network contributes a 16.10% storage saving on the playroom scene (Table 3), and the scene-agnostic prior adds another 6.95% saving, showing both components pull their weight.","Progressive coding that compresses locations, then local features, then scaling and offset uses previously compressed data as priors, cutting offset storage by 34% relative to HAC (Table 5)."],"supporting_citations":[{"why":"Supplies the anchor-based 3DGS structure that HEMGS compresses.","marker":"[29]"},{"why":"Provides the HAC baseline and the adaptive quantization procedure used for lossy compression.","marker":"[3]"},{"why":"Supplies the pre-trained PointNet++ used as the scene-agnostic feature extractor.","marker":"[36]"},{"why":"Provides the joint autoregressive and hyperprior entropy model that HEMGS adapts.","marker":"[31]"},{"why":"Introduces the hyperprior entropy model for neural compression, the conceptual foundation.","marker":"[1]"},{"why":"Provides the anchor-level context model that HEMGS extends with adaptive context selection.","marker":"[39]"}],"fun_headline_variants":["Single model compresses 3D Gaussian splats 40% smaller","40% smaller 3D Gaussian splats with one model","Hybrid entropy model shrinks 3DGS files 40%","One model, 16 rates, 40% smaller 3D Gaussian storage","One neural model delivers 16 rates, 40% smaller 3DGS"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The compression gain assumes that a raster-order traversal of anchors, plus the nearest already-encoded anchors, captures the statistical dependencies between attribute values; if the important dependencies are long-range or non-local, the lossless entropy coding improvement would fail for such scenes.","fun_headline_variants_meta":{"raw":{"variants":["Single model compresses 3D Gaussian splats 40% smaller","40% smaller 3D Gaussian splats with one model","Hybrid entropy model shrinks 3DGS files 40%","One model, 16 rates, 40% smaller 3D Gaussian storage","One neural model delivers 16 rates, 40% smaller 3DGS"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001058,"raw_usage":{"total_tokens":4444,"prompt_tokens":954,"completion_tokens":3490,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":570,"completion_tokens_details":{"reasoning_tokens":3392}},"tokens_in":570,"tokens_out":3490,"duration_ms":20823,"temperature":1.0,"reasoning_tokens":3392,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T11:09:24.909341+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compress a synthetic scene containing two spatially separated clusters of identical anchors whose attributes are correlated with each other's cluster. If the nearest-n context cannot exploit that cross-cluster redundancy, the losslessly encoded size will be no smaller than a no-context baseline, showing that the adaptive context selection fails on non-local dependencies.","supporting_citations":[{"cited_title":"Scaffold-gs: Structured 3d gaussians for view-adaptive rendering","cited_arxiv_id":null,"evidence_quote":"Supplies the anchor-based 3DGS structure that HEMGS compresses."},{"cited_title":"Hac: Hash-grid assisted context for 3d gaussian splatting compression","cited_arxiv_id":null,"evidence_quote":"Provides the HAC baseline and the adaptive quantization procedure used for lossy compression."},{"cited_title":"Pointnet++: Deep hierarchical feature learning on point sets in a metric space","cited_arxiv_id":null,"evidence_quote":"Supplies the pre-trained PointNet++ used as the scene-agnostic feature extractor."},{"cited_title":"Joint autoregressive and hierarchical priors for learned im- age compression","cited_arxiv_id":null,"evidence_quote":"Provides the joint autoregressive and hyperprior entropy model that HEMGS adapts."},{"cited_title":"Variational image compression with a scale hyperprior","cited_arxiv_id":null,"evidence_quote":"Introduces the hyperprior entropy model for neural compression, the conceptual foundation."},{"cited_title":"ContextGS : Compact 3d gaussian splatting with anchor level context model","cited_arxiv_id":null,"evidence_quote":"Provides the anchor-level context model that HEMGS extends with adaptive context selection."}],"review_version":1}