{"id":"df8457be-d921-44e3-8e94-16971339375d","arxiv_id":"2505.03310","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A gated mixture of lightweight prior networks plus gradient-guided element-wise quantization cuts 3D Gaussian Splatting file sizes by about 10 to 40 percent at matched PSNR on the tested benchmarks.","lead":"A new compression method shrinks 3D Gaussian Splatting scenes by blending several small neural predictors that estimate redundancy, then giving each stored number its own quantization precision. On four benchmarks it reports file sizes roughly 10 to 40 percent smaller than previous state-of-the-art compressors at similar image quality.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Element-wise quantization matrix depends on gradients from original training views; the decoder cannot compute it, and the paper never states that it is transmitted or accounts for its bit cost, so the lossless decoding stage is under-specified.","rationale":"The reader's weakest assumption concerns whether gradient magnitude is a reliable importance proxy for the element-wise quantization matrix. I think the more load-bearing issue is upstream of that: the paper never explains how the decoder obtains this matrix. The gradient computation in Sec. 4.2 uses a first pass over camera views with ground-truth images; the decoder, by construction, has only the bitstream and the compressed locations. If Q4 is not available at the decoder, the arithmetic decoder cannot know the set of possible quantized symbols or dequantize them, so the lossless attribute compression branch in Sec. 3.2 is under-specified. If Q4 is available because it is transmitted, then its bit cost belongs in Tables 1 and 5 and in the BDBR comparisons; the current storage breakdown does not show it. The reported C2FQ gains are small (0.15 MB and 0.01 dB in Table 6), so a hidden Q4 cost could erase the claimed advantage. I keep the reader's conditional verdict: the central SOTA claim is plausible but needs this decoding-path specification and a re-audited storage accounting before acceptance. Other inconsistencies noted by the reader, such as the Table 2 versus Sec. 4.4 text mismatch and the Table 3 wording, reinforce but do not replace this concern.","tokens_in":15590,"tokens_out":12485,"duration_ms":129562,"concrete_test":"Have the authors add the exact number of bits spent on Q4, or on any gradient-derived side information, to the Table 5 storage breakdown and recompute the BDBR comparisons in Table 3; if the reported C2FQ gains in Table 6, such as the 0.15 MB gain on Tank&Temples, are smaller than the added cost, the element-wise quantization contribution is not supported. If Q4 is not transmitted, require an explicit decoding procedure that recovers \\bar{A} without it.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"Sec. 3.4 defines the quantization matrix Q4 by expanding Q2 with averaged element-wise gradients of the loss across multiple camera views, and Sec. 4.2 describes a first pass over training views to collect these gradients. At decode time only the bitstream, compressed locations, and MoP features are available; the original views and ground truth are not. The paper does not specify how Q4 is transmitted or recomputed at the decoder. Without Q4, the arithmetic decoder cannot know the symbol alphabet or dequantize \\hat{A} = Round(A*Q4)/Q4, so the lossless attribute decoding described in Sec. 3.2 is not an implementable procedure. If Q4 is instead included in the bitstream, its cost must appear in the storage breakdowns and BDBR calculations; Table 5 lists only Location, Attributes, MLPs, and 'Others' and never identifies a quantization-matrix entry. This is a missing channel in the described codec, not a question of whether gradient magnitude is a good importance proxy.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes an anchor-based 3DGS compression pipeline with two novel components: a Mixture of Priors (MoP) hyperprior and a Coarse-to-Fine Quantization (C2FQ) module. Multiple lightweight MLPs produce diverse prior features that are gated into a single MoP feature; this feature conditions the entropy model for lossless coding and guides an element-wise quantization step matrix for lossy coding. The method is evaluated on Mip-NeRF360, BungeeNeRF, DeepBlending, and Tank&Temples against a broad set of baselines, reporting state-of-the-art rate-distortion numbers in Table 1 and BDBR gains of 16.6-41.8 percent over HAC and Context-GS on the two datasets in Table 3.","tokens_in":15803,"tokens_out":7483,"duration_ms":67540,"significance":"The contribution is potentially useful: the size/quality table and BDBR comparisons, if reproducible, indicate a meaningful improvement over strong anchor-based compression baselines. Strengths of the paper are the broad benchmark coverage, the direct comparison to HAC and Context-GS, the storage breakdown in Table 5, and the parameter-size comparison in Table 4. However, the central codec description omits how the quantization matrix is made available to the decoder and whether its bit cost is counted, and the ablation narrative in Section 4.4 is not consistent with Table 2. These issues must be resolved before the state-of-the-art claim can be accepted.","major_comments":[{"comment":"The quantization matrix Q4 is computed from averaged element-wise gradients of the training loss across multiple camera views (Sec. 3.4) and is needed to dequantize \\hat A = Round(A*Q4)/Q4. The paper does not state that Q4 is transmitted as side information, nor that the decoder can recompute it. At decode time only the bitstream and compressed locations are available; the original training views and ground-truth rendering loss are not. Without Q4 the arithmetic decoder cannot determine the quantization bin boundaries or invert the quantization. If Q4 is side information, its bit cost must be included in the storage breakdown in Table 5 and in the BDBR numbers; if it is recomputed, the required reference data must be specified. As written, the lossless decoding stage in Sec. 3.2 is not an implementable procedure.","section":"Sec. 3.2/3.4 and Fig. 2/Table 5"},{"comment":"The text says that 'Ours w/o C2FQ & MoP' compared with 'Ours w/o C2FQ' yields a further decrease of 0.05 PSNR, 0.005 SSIM, and a 0.57 MB increase in storage. Table 2 shows differences of 0.17 dB, 0.003, and 0.03 MB, respectively. The additive decomposition also does not follow from the table: removing MoP alone costs 0.22 dB and 0.10 MB, removing C2FQ alone costs 0.06 dB and 0.13 MB, but the combined ablation costs 0.23 dB and 0.16 MB relative to the complete model. Please correct the narrative and discuss the interaction between the two components rather than claiming independent gains.","section":"Sec. 4.4, Table 2"},{"comment":"The statement that compared with Scaffold-GS 'our approach achieves over 88% storage savings and consistently delivers better reconstruction quality' is not supported by Table 1 for the low-rate configuration: on DeepBlending, Ours (low-rate) has PSNR 30.20 vs. Scaffold-GS 30.21 and LPIPS 0.260 vs. 0.254. Please restrict the claim to the high-rate configuration or qualify which operating point is meant.","section":"Sec. 4.3, Table 1"},{"comment":"The core assumption that gradient magnitude measures an element's importance to rate-distortion performance is not directly tested. Table 6 only removes the quantization matrix and/or vector; it does not compare the gradient-based weighting against equal weights, random weights, or a learned per-element matrix at matched storage. Please add such a control experiment and report the normalization/clamping of the averaged gradients and the contents of the scale list. Without this, the reported C2FQ gains cannot be attributed to the gradient-guided granularity rather than to the scale selection or the MoP guidance.","section":"Sec. 3.4, Table 6"}],"minor_comments":[{"comment":"BDBR is reported for only Mip-NeRF360 and DeepBlending, although the table caption and the abstract imply all four benchmarks; please state this explicitly or add the missing datasets.","section":"Sec. 4.3, Table 3"},{"comment":"Key hyperparameters are not reported: the loss weight lambda in Eq. (2), the predefined quantization step Q0, the scale list, and the MLP/gating architecture dimensions. Without these values the experiments are not reproducible.","section":"Sec. 4.2"},{"comment":"The claim that random initialization yields diverse priors is only illustrated by a weight visualization; a quantitative diversity measure or a same-capacity single-MLP baseline would be more convincing.","section":"Sec. 3.3, Fig. 4"},{"comment":"There are several minor typos, including 'interploated' (Sec. 3.3), 'lightweigth' (Sec. 4.4), 'Gubmel' (Sec. 3.4), and a missing comma in 'Different with previous 3DGS compression networks [4,51] that employ a single MLP'.","section":"Throughout"},{"comment":"Figure 3 omits Tank&Temples despite Table 1 including it; either add that dataset to the RD curves or explain why it is excluded.","section":"Fig. 3"}],"recommendation":"major_revision","confidential_remarks":"The missing decoding protocol for Q4 is the most serious technical gap: if the authors can specify how the decoder obtains the quantization matrix and account for its bit cost in Table 5 and the BDBR calculations, the central claim could become defensible. The inconsistency between Section 4.4 and Table 2 is also fixable by re-running the ablations and rewriting the text, but it currently undermines the component-level attribution. If the Q4 side-information question cannot be resolved within the manuscript, the state-of-the-art claim should be reconsidered."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a credible incremental advance in 3DGS compression. The gated mixture of lightweight MLP hyperpriors plus gradient-scaled element-wise quantization buys real BDBR gains, roughly 16-42% over HAC and Context-GS on the two datasets where BDBR is reported. The experiments are broad and the architecture is clearly motivated. But the paper has two soft spots that are big enough that I would not take the numbers at face value yet.\n\nWhat's new: replacing the single hyperprior MLP with a soft-gated ensemble of five lightweight MLPs is a sensible adaptation of MoE to 3DGS, and using the MoP feature to drive a coarse-to-fine quantization matrix is a reasonable way to get element-wise rate control without a large auxiliary network. Neither idea is conceptually novel, but the combination is new to this subfield and the results are plausible.\n\nThe first problem is a missing channel in the codec description. Section 3.4 builds the quantization matrix Q4 from per-element gradients averaged over training views. The decoder, as described, has only the bitstream, the compressed locations, and the MoP features. It cannot compute those gradients, and the paper never says Q4 is transmitted. Without Q4, arithmetic decoding can produce the quantized indexes but not the dequantized attributes, so the lossless-to-lossy decode path is under-specified. If Q4 is in the bitstream, its cost is missing from the storage breakdowns in Table 5. This is a load-bearing omission, not a stylistic quibble, and the stress-test note is right to separate it from the question of whether gradient magnitude is a good importance proxy.\n\nThe second problem is the ablation narrative. Section 4.4 says removing MoP from 'w/o C2FQ' causes a further PSNR drop of 0.05 dB and 0.57 MB size increase, but Table 2 shows 0.17 dB and 0.03 MB. That is not a rounding difference. The text also claims consistent quality improvement over Scaffold-GS, while the low-rate DeepBlending PSNR in Table 1 is slightly below Scaffold's. These are fixable but they make me want to see code and checkpoints before citing the numbers.\n\nThe hyperparameter choices (λ, Q0, scale list, expert count) are tuned without a held-out split, but that is standard practice in this literature and not a disqualifier.\n\nWho is this for? People working on 3DGS storage and streaming. The gains are practically useful if they reproduce. I would send it to review, but with a strong request to the authors to specify the decoder side of Q4, correct the ablation section, and release artifacts. Without those, the paper is not complete enough to build on.","headline":"Credible incremental gains in 3DGS compression from a gated mixture-of-priors and gradient-weighted quantization, but the decoder side of the quantization matrix is under-specified and the ablation text contradicts its own table.","tokens_in":16340,"tokens_out":5682,"would_cite":false,"duration_ms":47445,"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 claims that replacing a single hyperprior MLP with a gated mixture of lightweight prior MLPs, and quantizing each attribute element with its own gradient-guided step, yields top rate-distortion performance for 3D Gaussian…","keywords":["3D Gaussian Splatting data compression","Mixture of Priors","Coarse-to-Fine Quantization","hyperprior entropy model","anchor-based 3DGS","element-wise quantization","rate-distortion optimization","Gumbel-Softmax"],"falsifier":"Take the trained model and re-encode the test scenes with the per-element gradient weights used to build $\\mathbf{Q}_4$ randomly permuted, leaving all other parameters and the bitstream pipeline unchanged. If the size-PSNR curve barely moves, the claimed gradient-based element-wise mechanism is not the source of the C2FQ improvement; if it degrades sharply, the mechanism is confirmed.","tokens_in":15350,"feed_emoji":"📉","tokens_out":8725,"duration_ms":74196,"temperature":0.7,"pith_summary":"The paper claims that two standard bottlenecks in 3D Gaussian Splatting compression—a weak hyperprior for entropy coding and a single shared quantization step for lossy coding—can be fixed by one mechanism. It replaces the single hyperprior MLP used in earlier anchor-based codecs with a Mixture of Priors (MoP): several lightweight MLPs produce diverse prior features, and a gating network blends them into a unified MoP feature. That feature serves both as the hyperprior for arithmetic coding and as guidance for a Coarse-to-Fine Quantization (C2FQ) stage that expands a scalar quantization step into an element-wise matrix. On Mip-NeRF360, BungeeNeRF, DeepBlending, and Tanks&Temples, the reported rate-distortion results beat the HAC and Context-GS baselines, with BDBR savings of roughly 16–42% depending on dataset. If these results hold, one shared module can improve both the lossless entropy-coding stage and the lossy quantization stage of anchor-based 3DGS codecs.","feed_headline":"Mixture-of-priors cuts 3D Gaussian scene storage up to 42%","feed_subtitle":"Five lightweight prior experts and gradient-guided quantization beat HAC and Context-GS on four benchmarks.","key_machinery":"The central object is the Mixture of Priors (MoP) feature, $G = \\sum_{i=1}^n w_i \\mathbf{p}_i$, where each $\\mathbf{p}_i$ is the output of a distinct lightweight MLP applied to interpolated anchor locations and $w_i$ comes from a small gating network. This feature carries the argument because it is the shared source for both entropy modeling and quantization guidance. The second mechanism, Coarse-to-Fine Quantization (C2FQ), turns one scalar step $Q_0$ into a quantization matrix by selecting a scale via Gumbel-Softmax/Max, expanding the rescaled value into a per-anchor vector through an MLP, and then expanding that vector into a per-element matrix using averaged gradients across camera views, avoiding extra network parameters.","core_discovery":"The paper's central claim is that a hyperprior built from several diverse lightweight priors, fused by a learned gate, gives a more accurate conditional distribution for entropy coding than the single shallow MLPs used in HAC and Context-GS, and that the same feature can drive per-element quantization. Quantization proceeds coarse-to-fine: a predefined step is first rescaled by a Gumbel-Softmax-selected scale, expanded by an MLP into a per-anchor vector, then multiplied by averaged per-element gradients to form a quantization matrix $\\mathbf{Q}_4$. Attributes are quantized as $\\hat{A} = \\mathrm{Round}(A \\times \\mathbf{Q}_4)/\\mathbf{Q}_4$, so each element of every anchor receives its own step size. The paper reports that this design reduces storage by more than 97% relative to unquantized 3DGS and by over 88% relative to Scaffold-GS, while outperforming HAC and Context-GS in both size and fidelity across four benchmarks.","pith_inferences":["The paper does not isolate how much of the C2FQ gain comes from the gradient-based matrix versus the coarser scale and vector stages; its Tank&Temples ablation suggests the matrix contributes less than the vector, so the element-wise gradient weighting is the least-tested link.","If the gradient-importance premise transfers, the same quantization-matrix recipe could be attached to other hyperprior-based 3DGS codecs, or to 4D Gaussian Splatting compression, without retraining their entropy models.","The gating weights visualized across BungeeNeRF scenes hint that the MoP feature specializes per scene; a direct test would be whether scene-adaptive gating alone, with fixed experts, explains most of the entropy-model gain.","A useful stress test is to compare the five-expert MoP against a single MLP with the same total parameter count; if that match performs equally well, diversity rather than capacity is the active ingredient."],"forward_implications":["Because the MoP feature serves both the entropy model and the quantizer, the lossless and lossy branches of an anchor-based 3DGS codec can be improved by a single change rather than two separate modules.","Element-wise quantization lets the rate-distortion trade-off be tuned at the level of individual anchor attributes, which is finer than the anchor-level or attribute-level control in HAC and Context-GS.","Keeping the expansion network-free via gradients means the per-element quantization matrix adds no stored model parameters, so its storage overhead is confined to the small MoP MLPs.","The reported BDBR reductions of roughly 16–42% against HAC and Context-GS on the tested datasets imply that substantial storage savings are available without sacrificing rendering PSNR.","The same anchor-based pipeline with MoP and C2FQ can be trained end-to-end, because the Gumbel-Softmax approximation keeps the scale selection differentiable during training."],"supporting_citations":[{"why":"Supplies the 3D Gaussian Splatting scene representation whose location and attribute data are compressed.","marker":"[24]"},{"why":"Introduces the anchor representation and anchor-based rendering pipeline that the compression framework builds on.","marker":"[40]"},{"why":"The main entropy-coding baseline: hash-grid assisted context with a shallow two-layer MLP hyperprior.","marker":"[4]"},{"why":"The anchor-level context entropy model baseline that the method compares against and aims to improve.","marker":"[51]"},{"why":"Establishes the scale-hyperprior conditional entropy model used for lossless compression of latent attributes.","marker":"[2]"},{"why":"Provides the differentiable Gumbel-Softmax approximation used to select the coarse scale in C2FQ.","marker":"[23]"},{"why":"Supplies the Mip-NeRF360 benchmark used to evaluate rate-distortion performance.","marker":"[3]"},{"why":"Supplies the DeepBlending benchmark used for the main ablations and BDBR comparisons.","marker":"[19]"},{"why":"Supplies the BungeeNeRF benchmark used for gating-weight visualization and full-model comparison.","marker":"[54]"},{"why":"Supplies the Tanks&Temples benchmark used for the quantization-stage ablation.","marker":"[26]"}],"fun_headline_variants":["Mixture-of-priors shrinks 3DGS files by 97%","Per-element quantization via MoP cuts 3DGS storage 97%","Gated priors drive 97% smaller 3DGS compression","Coarse-to-fine steps guided by MoP beat HAC on 3DGS","MoP: five priors, element-wise quantization, 97% smaller"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that an element's average gradient magnitude across camera views measures how much its quantization step matters for the final rate-distortion trade-off; if that link fails, the reported C2FQ gains could come from the coarse scale selection or the per-anchor vector stage rather than from element-wise granularity.","fun_headline_variants_meta":{"raw":{"variants":["Mixture-of-priors shrinks 3DGS files by 97%","Per-element quantization via MoP cuts 3DGS storage 97%","Gated priors drive 97% smaller 3DGS compression","Coarse-to-fine steps guided by MoP beat HAC on 3DGS","MoP: five priors, element-wise quantization, 97% smaller"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000994,"raw_usage":{"total_tokens":4264,"prompt_tokens":1053,"completion_tokens":3211,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":669,"completion_tokens_details":{"reasoning_tokens":3107}},"tokens_in":669,"tokens_out":3211,"duration_ms":25375,"temperature":1.0,"reasoning_tokens":3107,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T23:54:34.105585+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take the trained model and re-encode the test scenes with the per-element gradient weights used to build $\\mathbf{Q}_4$ randomly permuted, leaving all other parameters and the bitstream pipeline unchanged. If the size-PSNR curve barely moves, the claimed gradient-based element-wise mechanism is not the source of the C2FQ improvement; if it degrades sharply, the mechanism is confirmed.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Introduces the anchor representation and anchor-based rendering pipeline that the compression framework builds on."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"The main entropy-coding baseline: hash-grid assisted context with a shallow two-layer MLP hyperprior."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"The anchor-level context entropy model baseline that the method compares against and aims to improve."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Establishes the scale-hyperprior conditional entropy model used for lossless compression of latent attributes."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the differentiable Gumbel-Softmax approximation used to select the coarse scale in C2FQ."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the Mip-NeRF360 benchmark used to evaluate rate-distortion performance."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the DeepBlending benchmark used for the main ablations and BDBR comparisons."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the BungeeNeRF benchmark used for gating-weight visualization and full-model comparison."}],"review_version":1}