{"id":"f74b8629-e1b4-42bc-917e-231cbb3aff6f","arxiv_id":"2501.03399","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"3D Gaussian Splatting scenes can be compressed to under 10 MB by encoding all gaussian attributes into tri-plane feature images and compressing them with standard HEVC video coding.","lead":"Researchers at Fraunhofer HHI show how to shrink 3D Gaussian Splatting scenes, a popular 3D representation, to under 10 megabytes by storing scene data as image-like feature planes and compressing those planes with standard video codecs such as HEVC. The work could make high-quality 3D scenes practical for mobile devices, cloud streaming, and VR headsets that already include video decoding hardware.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The entropy model p in Eq. 2 is never defined, making the training objective underspecified and the reported compression results non-reproducible; the DCT proxy is also unvalidated against actual HEVC bitrate.","rationale":"The reader's weakest assumption correctly identifies the DCT entropy proxy as a load-bearing premise, but the more fundamental gap is that the entropy objective in Eq. (2) is not fully specified because the probability model p is never defined. Without p, the loss cannot be computed, making the method non-reproducible and the reported sizes unverifiable. This is a concrete omission, not merely a concern about generalization. The reader's focus on Qstep and proxy reliability is related but secondary; even if Qstep were optimally chosen, the undefined p would prevent independent confirmation. The paper's empirical ablations provide some support for the proxy's usefulness, but they do not quantify its fidelity to the actual codec. The central claim of under-10MB sizes with negligible quality loss is plausible and may well be correct, but the missing specification and lack of proxy validation justify the conditional verdict. Therefore, I agree with the reader's conditional assessment and do not recommend changing the verdict; the concern should be addressed by releasing code or supplying the entropy model details and a proxy-bitness correlation analysis.","tokens_in":15666,"tokens_out":13489,"duration_ms":123429,"concrete_test":"Specify and implement a standard factorized entropy model for p (e.g., a learned logistic prior as in Ballé et al. 2018) and retrain the 'bonsai' scene with the exact reported hyperparameters; compare the resulting size/PSNR to the reported 6.71 MB / 31.71 dB in Table 6. Additionally, for the trained planes, extract per-channel bit counts from HM (via per-frame bitstream statistics) and compute the Spearman rank correlation between w_c·I(F(P_c)) and the actual per-channel encoded size; if the correlation is below 0.5, the proxy is not faithful and the training signal should be revised.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that the DCT entropy loss yields under-10MB sizes rests on Eq. (2), I(P)=E[-log p(˜P)], but the density model p is never specified in the main text or supplementary. The citation to Ballé et al. suggests a learned prior, yet no architecture, training procedure, or parameterization for p is given; the same omission applies to the transform-domain loss I(F(P)). Without p, the objective is undefined and the ablations in Figure 7 and Table 2 cannot be reproduced independently. Additionally, the proxy is unvalidated: it operates on per-channel 4x4 DCT coefficients with a fixed Qstep=2^8, ignoring inter-frame prediction across the 32 concatenated channels and HEVC's adaptive transform sizes and rate-distortion optimization. No quantitative evidence is provided that minimizing I(F(P_c)) correlates with actual HM bitrate at QP=1; Figure 7 only shows qualitative RD gains on four scenes. If the proxy mismatches the codec's rate allocation, the reported sizes could be scene-specific or codec-specific, and the generalization claim weakens.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a compression method for 3D Gaussian Splatting (3DGS) that replaces per-Gaussian attributes with a compact tri-plane feature representation, trains the feature planes with a DCT-domain entropy loss and channel-wise bit allocation, and then compresses the resulting 32-channel 16-bit feature planes using standard video codecs (HM, libx265). Point positions are Morton-sorted, packed into images, and losslessly coded. The authors report storage sizes below 10 MB on dataset averages for Mip-NeRF360, DeepBlending, and Tank&Temples, with quality close to the original 3DGS, and claim compression ratios up to 76x. The paper includes per-scene results, ablation studies, and an analysis of quantization step size.","tokens_in":1787,"tokens_out":1845,"duration_ms":53501,"significance":"If the results are reproducible, the paper makes a valuable contribution by showing that standard, hardware-decodable video codecs can serve as the compression engine for 3DGS attributes, potentially enabling practical deployment on mobile devices. The DCT-domain entropy modeling and channel-wise bit allocation are sensible ideas, and the paper reports detailed per-scene metrics and ablation studies. However, the current manuscript omits a definition of the entropy model p used in the training loss, which makes the central optimization objective underspecified and the results non-reproducible. The claim that all datasets are under 10 MB is also contradicted by the paper's own per-scene tables. These issues are load-bearing and must be fixed before the results can be trusted.","major_comments":[{"comment":"The entropy loss I(P) = E[-log p(P̃)] requires a probability model p of the quantized (or noised) plane coefficients. The paper never defines p: no functional form, architecture, or training procedure is given. The citation to Ballé et al. suggests a factorized density, but neither the main text nor the supplementary material specifies how p is parameterized or learned. Since the total loss in Eq. (8) includes λ_ent L_ent, where L_ent is defined by this p, the entire training objective is undefined as written. This is not an incidental omission; the ablation results in Fig. 7 and Table 2 cannot be reproduced without knowing p. The authors must either explicitly define p (e.g., a factorized logistic model with a small neural network, as in variational image compression) or use a simple fixed density (e.g., a histogram or Laplacian) and say so.","section":"Section 3.3, Eq. (2)"},{"comment":"The text claims that the method \"achieved sizes under 10MB for all datasets.\" This is contradicted by the per-scene results in Table 6: the `garden` scene requires 15.22 MB and `stump` requires 14.30 MB on Mip-NeRF360. Only the dataset averages (9.78 MB, 8.62 MB, 7.46 MB) are under 10 MB. The claim as stated is factually incorrect and should be rephrased to \"dataset averages under 10 MB\" or the method should be re-evaluated. The abstract and introduction also use \"typically <10MB,\" which is ambiguous; the discrepancy between the headline claim and the actual per-scene data must be resolved.","section":"Section 4.2, Table 6"},{"comment":"The DCT entropy loss operates on 4×4 block DCT coefficients with a quantization step Qstep=2^8, chosen as \"the best\" from an analysis on a single scene (`bonsai`). No sensitivity analysis across scenes or bitrate ranges is provided. More fundamentally, the proxy is not validated against actual HEVC bitrate: the paper shows in Fig. 7 that adding L_ent improves RD on four scenes, but it does not demonstrate that the entropy estimate correlates with the HM output size beyond those cases. Since the method's entire rate-control mechanism depends on this proxy, the authors should provide quantitative evidence (e.g., a scatter plot of L_ent vs. encoded size across scenes and λ_ent values) or at least a cross-scene validation of the Qstep choice.","section":"Section 3.3 and Appendix Table 5"},{"comment":"The feature-plane architecture is described inconsistently. The paper first defines a tri-plane P (XY, XZ, YZ), then says \"Each feature plane has 8 channels and a 512×512 resolution with a decoder MLP to predict each attribute (color, scale, rotation, opacity) separately. Therefore, a total of 32 channels are used to predict all attributes.\" A tri-plane with three planes of 8 channels would give 24 channels, not 32. The ambiguity makes it impossible to understand whether the method uses one tri-plane per attribute, four separate planes, or some other layout. Please clarify the number of planes, the channel count per plane, and how the 32 channels are derived.","section":"Section 4.1 and Section 3.1"},{"comment":"The proposed entropy loss is applied separately to each channel's 4×4 DCT coefficients, but the actual HEVC encoding operates on the 32 concatenated frames with inter-frame prediction, adaptive transform sizes, and rate-distortion optimization. The paper provides no analysis of how well the per-channel DCT entropy approximates the output bitrate under these conditions. While the empirical RD gains in Fig. 7 are encouraging, the generalization of the proxy to other scenes, other codec configurations, or other bitrate ranges is uncertain. The authors should either add a direct comparison of predicted vs. actual bitrate or acknowledge this limitation explicitly in the paper.","section":"Section 3.3 (DCT entropy modeling)"}],"minor_comments":[{"comment":"The sentence \"we found that a Qstep of 28 yields the best results\" is ambiguous; the appendix (Table 5) uses exponents (2^0, 2^2, ...), so it should read \"Qstep of 2^8\".","section":"Section 3.3"},{"comment":"The Qstep values in Table 5 are listed as \"20, 22, ...\" without superscripts, making it unclear whether these are powers of two; please format as 2^0, 2^2, etc.","section":"Supplementary Material, Table 5"},{"comment":"The paper contains numerous garbled text fragments and encoding artifacts (e.g., sequences like \"/uni00000014/uni00000019\" in Section 4.1 and corrupted reference entries). The manuscript should be thoroughly proofread and regenerated.","section":"General"},{"comment":"The x-axis is labeled as a log2 scale, but the axis ticks are not explained; please add clear axis labels or captions that describe the unit and scale.","section":"Figure 5"},{"comment":"The exact hyperparameters λ_ent and λ_1 are not stated in the main text or supplementary material, aside from \"small amount\" for L1. Please provide the numeric values or a range used in the experiments.","section":"Section 4.1"}],"recommendation":"major_revision","confidential_remarks":"The paper presents a novel and potentially practical approach, but the missing entropy-model definition and the contradiction between the headline size claim and the per-scene data are serious issues that currently prevent acceptance. The DCT-proxy validation concern is also important but may be addressable with additional experiments. I believe the paper is worth a major revision rather than rejection, assuming the authors can supply the missing definition and correct the claims."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this paper is worth a serious look. The new thing is a tri-plane prediction of all 3DGS attributes, trained progressively, with a DCT-domain entropy loss and channel-wise bit allocation that lets a standard video codec (HEVC) compress the feature planes. That combination is genuinely new: VideoRF and TeTriRF do codestream grids, and Scaffold-GS compression methods don't touch standard codecs. The experiments are conducted on three standard datasets with per-scene numbers, an ablation table, RD curves, and two codec implementations (HM and libx265). The core claim — that you can get roughly 10 MB scenes close to original 3DGS quality — is plausible and supported by the reported numbers. Credit where due: the ablation in Table 2 shows each component helping, and Fig. 7 shows the DCT loss giving a real RD benefit. The authors also correctly concede that HAC is better on DeepBlending, although the abstract's 'outperforms' overstates things.\n\nThe big soft spot is that Eq. (2) defines I(P) = E[-log p(~P)] but p is never specified. Is it a factorized Gaussian, a learned hyperprior, a fixed uniform density? No equation, no architecture, no training procedure. Without p, the training objective Lent is undefined and nobody can reproduce the ablations or the final sizes. This is a load-bearing technical omission, not a stylistic one. The stress-test note is right about this, and it needs fixing in revision (or, preferably, code release). The related DCT-proxy concern is real but less severe: Qstep=2^8 is chosen from one scene with fixed lambda, QP is set to 1, and there is no quantitative correlation plot between Lent and actual HM bitrate. Figure 7 gives indirect evidence, but the proxy could be codec-specific.\n\nAlso, the per-scene tables contradict the '<10MB' claim: flowers is 10.25 MB, garden 15.22, stump 14.30. That's a factual overstatement that needs correcting. No error bars or multiple seeds are reported, and baseline numbers are copied from respective papers rather than generated in a common harness, so small differences in the tables shouldn't be over-interpreted.\n\nBottom line: this is a useful architecture for people working on 3DGS deployment and codec-based neural representations. The central result is believable, but the missing entropy model is a real hole and the size claims need to be cleaned up. This paper deserves peer review — a good referee will catch these issues and the authors should be able to address them.","headline":"A genuinely useful tri-plane plus video-codec pipeline for 3DGS compression with plausible sub-10MB results, but the entropy model in Eq. (2) is never defined and some per-scene claims are over-stated.","tokens_in":16448,"tokens_out":4752,"would_cite":true,"duration_ms":42325,"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":"A tri-plane feature field trained on DCT entropy lets standard HEVC compress 3D Gaussian Splatting to under 10 MB, at up to 76x compression and only 0.19 dB PSNR loss.","keywords":["3D Gaussian Splatting compression","tri-plane feature planes","frequency-domain entropy modeling","DCT transform coding","video codec compression","channel-wise bit allocation","novel view synthesis","HEVC"],"falsifier":"Run the released pipeline on a scene with repetitive high-frequency textures, sweep $\\lambda_{\\mathrm{ent}}$ while holding the video QP fixed, and compare the ordering of the DCT-entropy loss with the actual HEVC bitstream sizes; any inversion, a plane that scores lower in $L_{\\mathrm{ent}}$ but encodes to more bytes, would show the entropy proxy fails. Repeating the same sweep across Qstep values and scenes would test whether the fixed $2^8$ quantizer step remains optimal.","tokens_in":15469,"feed_emoji":"🎥","tokens_out":11187,"duration_ms":96576,"temperature":0.7,"pith_summary":"This paper tries to show that the storage bottleneck of 3D Gaussian Splatting can be broken by re-encoding the scene as video. Instead of compressing millions of gaussian attribute vectors, the authors train a tri-plane feature field that predicts every attribute, then feed those planes to a standard HEVC encoder after shaping them with a DCT-domain entropy loss. On Mip-NeRF360 this reaches up to 76x compression over the original 3DGS with only 0.19 dB PSNR loss, and under 10 MB total storage on every dataset tested. If the claim holds, 3DGS scenes could be shipped as compact standard-codec bitstreams and decoded on devices with hardware video decoding, making high-quality 3D scenes practical for mobile and head-mounted use.","feed_headline":"Video codecs shrink 3D Gaussian Splatting to under 10 MB","feed_subtitle":"DCT-domain entropy training shapes the planes so an unmodified HEVC encoder packs them to <10 MB with near-original rendering quality","key_machinery":"The load-bearing object is the optimized tri-plane feature field: three axis-aligned 2D grids $P_{XY}, P_{XZ}, P_{YZ}$ with eight channels each, queried at each gaussian position and decoded by a small MLP to produce all splat attributes. What makes the compression work is a frequency-domain entropy loss applied to each plane channel, $L_{\\mathrm{ent}} = \\sum_c w_c I(F(P_c))$, where $F$ is a $4\\times4$ block-wise DCT and the entropy is estimated with uniform noise at Qstep $2^8$; this steers the planes toward signals that a DCT-based video codec can code cheaply. Channel importance scores $w_c = \\mathrm{CI}_1(P)/\\mathrm{CI}_c(P)$ allocate fewer bits to less visible channels, progressive channel masking stabilizes training, and a piecewise-projective contraction maps unbounded scenes onto the finite plane so neighboring blocks carry real spatial correlation. At deployment, the 32 channels are concatenated into 32 16-bit frames and encoded as one HEVC sequence.","core_discovery":"On its own terms, the paper's discovery is that the millions of per-gaussian attribute vectors in 3DGS can be replaced by a compact continuous field: a tri-plane of 32 16-bit channels (three axis-aligned planes with eight levels each) predicts color, scale, rotation, and opacity through a small decoder multi-layer perceptron, so the scene is compressed by coding those planes rather than the gaussians. The decisive step is to optimize the planes for the codec's frequency domain: instead of sparsifying with L1, the training loss minimizes the entropy of block-wise 4x4 DCT coefficients, and channel importance scores re-weight that loss to spend bits where they affect rendering most. The trained planes are normalized to 16-bit integer frames and fed to an unmodified HEVC encoder at QP=1; point positions are Morton-sorted, packed, and losslessly coded. The reported result is under 10 MB per scene across Mip-NeRF360, DeepBlending, and Tank&Temples, with up to 76x compression versus 3DGS and only 0.19 dB PSNR loss on Mip-NeRF360.","pith_inferences":["Because the paper leaves point positions losslessly coded and roughly constant in size, coupling this plane codec with point pruning would likely push total size well below the reported values; the paper notes this possibility but does not implement it.","The same DCT-entropy design should transfer to newer transform-based codecs such as VVC, which the paper names as motivation but does not test; larger transform blocks and better intra prediction may improve the rate-distortion curve further, or may require retuning Qstep.","The progressive channel structure suggests a natural streaming scheme: decode low-level channels first for a coarse scene, then add higher channels for detail; the paper does not build this, but the training order and channel importance scores make it feasible.","The channel-importance weights are computed once at a fixed training iteration from a sensitivity snapshot; recomputing them for the target bitrate or for a user's expected viewing directions could yield a better bit allocation than the single static weight vector."],"forward_implications":["Scenes from the tested datasets can be stored in under 10 MB, with Mip-NeRF360 compressed up to 76x against the original 3DGS at a cost of 0.19 dB PSNR.","After decoding the feature plane frames and predicting attributes once, rendering uses the original 3DGS rasterizer with no extra per-frame overhead, so render speed stays close to uncompressed 3DGS.","Because the bitstream is produced by unmodified HEVC encoders, decoding can ride on hardware video decoders, which is the paper's route to mobile and head-mounted deployment.","Rate-distortion control is best done during training through $\\lambda_{\\mathrm{ent}}$, not by raising the codec QP at encode time; the appendix shows QP-based control yields worse rate-distortion performance.","The compression transfers to a second HEVC implementation (libx265) with similar savings, so the method is not tied to one codec software."],"supporting_citations":[{"why":"Supplies the 3D Gaussian Splatting representation, the differentiable rasterizer, and the uncompressed quality baseline the method must match.","marker":"[17]"},{"why":"Supplies the k-planes factorization (Hadamard product over plane features) used to build the tri-plane attribute predictor.","marker":"[12]"},{"why":"Shows how a standard video codec can compress a radiance field encoded as 2D feature streams, the approach this paper adapts to gaussian primitives.","marker":"[39]"},{"why":"Provides the tri-plane-plus-video-codec template for radiance-field compression that the method extends to 3DGS.","marker":"[40]"},{"why":"Defines HEVC and its DCT-based transform coding, the codec the feature planes are optimized for and evaluated with.","marker":"[34]"},{"why":"Supplies the piecewise-projective contraction used to map unbounded scenes into the finite feature plane, which the paper shows improves spatial correlation for block DCT.","marker":"[31]"},{"why":"Provides the Mip-NeRF360 dataset and the unit-sphere contraction that serves as the comparison baseline for the proposed contraction.","marker":"[5]"},{"why":"Supplies Morton-order sorting, used to pack point positions into a 2D grid that can be losslessly video-coded.","marker":"[27]"},{"why":"Supplies the sensitivity-based importance scoring that the channel importance (CI) bit allocation transfers from gaussian parameters to feature planes.","marker":"[29]"},{"why":"Supplies the reference HEVC implementation whose random-access configuration with QP offsets set to zero produced the reported bitstream sizes.","marker":"[16]"}],"fun_headline_variants":["HEVC codec shrinks 3DGS scenes to under 10 MB","DCT-trained planes let HEVC squash 3DGS to <10 MB","Codec-aware planes compress 3DGS by 76x","3DGS under 10 MB via DCT entropy + HEVC","Frequency-shaped planes: 3DGS fits in 10 MB"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing assumption is that minimizing the entropy of the planes' $4\\times4$ block frequency coefficients, with one fixed quantization strength, predicts how many bytes the video codec really spends; if that match fails on other scenes, channel counts, or bitrates, the reported size gains shrink.","fun_headline_variants_meta":{"raw":{"variants":["HEVC codec shrinks 3DGS scenes to under 10 MB","DCT-trained planes let HEVC squash 3DGS to <10 MB","Codec-aware planes compress 3DGS by 76x","3DGS under 10 MB via DCT entropy + HEVC","Frequency-shaped planes: 3DGS fits in 10 MB"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000683,"raw_usage":{"total_tokens":3118,"prompt_tokens":980,"completion_tokens":2138,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":596,"completion_tokens_details":{"reasoning_tokens":2040}},"tokens_in":596,"tokens_out":2138,"duration_ms":14419,"temperature":1.0,"reasoning_tokens":2040,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T21:52:19.472527+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the released pipeline on a scene with repetitive high-frequency textures, sweep $\\lambda_{\\mathrm{ent}}$ while holding the video QP fixed, and compare the ordering of the DCT-entropy loss with the actual HEVC bitstream sizes; any inversion, a plane that scores lower in $L_{\\mathrm{ent}}$ but encodes to more bytes, would show the entropy proxy fails. Repeating the same sweep across Qstep values and scenes would test whether the fixed $2^8$ quantizer step remains optimal.","supporting_citations":[{"cited_title":"3d gaussian splatting for real-time radiance field rendering","cited_arxiv_id":null,"evidence_quote":"Supplies the 3D Gaussian Splatting representation, the differentiable rasterizer, and the uncompressed quality baseline the method must match."},{"cited_title":"K-planes: Explicit radiance fields in space, time, and appearance","cited_arxiv_id":null,"evidence_quote":"Supplies the k-planes factorization (Hadamard product over plane features) used to build the tri-plane attribute predictor."},{"cited_title":"VideoRF: Rendering Dynamic Radiance Fields as 2D Feature Video Streams","cited_arxiv_id":null,"evidence_quote":"Shows how a standard video codec can compress a radiance field encoded as 2D feature streams, the approach this paper adapts to gaussian primitives."},{"cited_title":"TeTriRF: Temporal Tri-Plane Radiance Fields for Effi - cient Free-Viewpoint Video","cited_arxiv_id":null,"evidence_quote":"Provides the tri-plane-plus-video-codec template for radiance-field compression that the method extends to 3DGS."},{"cited_title":"Overview of the high efficiency video coding (hevc) standard","cited_arxiv_id":null,"evidence_quote":"Defines HEVC and its DCT-based transform coding, the codec the feature planes are optimized for and evaluated with."},{"cited_title":"Merf: Memory-efficient radiance fields for real-time view synthesis in unbounded scenes","cited_arxiv_id":null,"evidence_quote":"Supplies the piecewise-projective contraction used to map unbounded scenes into the finite feature plane, which the paper shows improves spatial correlation for block DCT."},{"cited_title":"Mip-nerf 360: Unbounded anti-aliased neural radiance fields","cited_arxiv_id":null,"evidence_quote":"Provides the Mip-NeRF360 dataset and the unit-sphere contraction that serves as the comparison baseline for the proposed contraction."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies Morton-order sorting, used to pack point positions into a 2D grid that can be losslessly video-coded."},{"cited_title":"Accessed: 2024-09-12","cited_arxiv_id":null,"evidence_quote":"Supplies the reference HEVC implementation whose random-access configuration with QP offsets set to zero produced the reported bitstream sizes."}],"review_version":1}