{"id":"0d110279-931b-49f9-9200-5abafbfe239b","arxiv_id":"2607.25971","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"Moving 3D Gaussian scenes can be streamed in fine DASH-compatible layers using multi-resolution anchors, transformer-based prediction, and opacity-weighted Gaussian refinement.","lead":"SplatStream packs moving 3D Gaussian scenes into fine-grained quality, resolution, and frame-rate layers that clients can request adaptively. It maps these layers to MPEG-DASH sub-representations, so a phone on a slow connection and a headset on a fast one receive different amounts of the same dynamic scene.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Spatial-quality axis is confounded with fine-grained refinement: Table II varies both Gaussian count and rendering resolution despite claiming one shared representation; no control isolates resolution, so the claimed third axis is not demonstrated.","rationale":"The reader's weakest assumption focuses on the opacity-volume importance ordering lacking a baseline (random/learned). That is a valid concern about the fine-grained axis, but it does not threaten the existence of the axis—only its optimality. A more load-bearing issue is that the spatial-quality axis may not be an independent axis at all. The method section says the same representation is used for all resolutions, yet Table II's Gaussian counts differ across spatial rows, implying the layers are actually importance-ordered subsets—the same mechanism as the fine-grained refinement in Table I. If so, the framework has only two independent scalability dimensions (temporal layering and Gaussian-subset ordering), and the multi-resolution loss is just a training trick rather than a scalable coding axis. This directly undercuts the strongest claim as phrased by the reader: 'three independent scalability axes.' The proposed test—crossing rendering resolution against Gaussian count—would settle the question. I therefore keep the verdict conditional (UNCHANGED), but with this additional condition rather than only the reader's baselines/ablations. The paper's engineering contributions may still be useful, but the spatial scalability evidence must be deconfounded before the central claim is accepted.","tokens_in":9438,"tokens_out":7652,"duration_ms":77192,"concrete_test":"On the bartender sequence, decode the full 348,060-Gaussian anchor and render it at 540p, 720p, and 1080p against resolution-matched ground truth to measure the quality spread from resolution alone. Then render the 278,127-Gaussian startup prefix at 1080p. If the PSNR difference between Table II rows 1–3 is mostly reproduced by resolution-holding subset fixed, the spatial axis is real; if it tracks Gaussian count, the axis collapses into fine-grained refinement. Also report whether #G should be constant under the stated 'same representation' training.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim of 'three independent scalability axes' depends on the spatial-quality axis being distinct from the fine-grained Gaussian-refinement axis. But the evidence for the spatial axis is confounded. Section II-B states: 'we do not train separate PLY models for 540p,720p, and1080p; instead, the same intra Gaussian representation is rendered and supervised at multiple target resolutions.' Yet Table II reports total #G = 278,127 for the 'Startup I_540' unit, 292,580 for '+I_720', and 348,060 for '+I_1080'. If the representation is truly the same, the transmitted Gaussian count should not change across spatial layers. If the counts differ because the layers are importance-ordered prefixes of the same model, then the 'spatial' rows are just three points on the same fine-grained refinement curve already reported in Table I (10–100%), and the PSNR gains (30.45→33.15→34.12) are attributable to adding Gaussians, not to any resolution-specific scalability mechanism. No experiment isolates rendering resolution while holding the Gaussian set fixed. Additionally, comparing PSNR across different rendering resolutions is only meaningful against resolution-matched ground truth, which the paper does not specify. Thus the spatial-quality axis is not established as an independent scalability dimension, undermining the headline claim of three independent axes.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"SplatStream proposes a scalable streaming framework for dynamic 3D Gaussian Splatting. The system is organized into intra anchors with multi-resolution rendering supervision, a transformer-enhanced InterGS-Lite temporal prediction chain, B-layer temporal enhancement frames, and opacity-volume importance ordering for fine-grained Gaussian prefixes. These components are mapped into an MPEG-DASH-like sub-representation structure. The paper claims three independent scalability axes: spatial quality (540p/720p/1080p), frame rate (base P-frames vs. B enhancement frames), and fine-grained Gaussian refinement (10% to 100% of Gaussians). The evaluation reports operating points on three dynamic sequences, with PSNR/SSIM/LPIPS improving monotonically as refinement percentage or spatial layer is increased.","tokens_in":9733,"tokens_out":3942,"duration_ms":37847,"significance":"If the claimed scalability axes were convincingly demonstrated, SplatStream would be a useful practical contribution: it would show how dynamic Gaussian scenes can be delivered over adaptive bitrate pipelines, with fast startup and progressive refinement. The system description is internally consistent, and the mapping of scalable Gaussian bitstreams to DASH sub-representations is clearly specified. The multi-resolution intra loss (Eq. 3), the transformer predictor (Eqs. 4-15), and the opacity-volume prefix ordering (Eqs. 19-22) are all presented in sufficient detail for reproduction. However, the current evidence does not establish the central claims: there are no comparisons with prior scalable GS frameworks, no ablation of the transformer predictor, no comparison of the importance ordering against random or learned baselines, and the spatial-quality axis is confounded with Gaussian count. The paper would be significantly strengthened by controlled experiments that isolate each scalability axis and by RD comparisons at matched rate and quality against existing baselines.","major_comments":[{"comment":"The spatial-quality axis is confounded with Gaussian-count refinement. Section II-B states that the same intra Gaussian representation is rendered and supervised at multiple target resolutions, yet Table II reports Total #G = 278,127 for the I_540 startup row, 292,580 after adding I_720, and 348,060 after adding I_1080. If the representation is truly shared, the number of transmitted Gaussians should not change across spatial layers; if the layers are importance-ordered prefixes, the spatial rows are simply three points on the fine-grained refinement dimension already reported in Table I. No experiment isolates rendering resolution while holding the Gaussian set fixed, and the quality metrics are not compared against resolution-matched ground truth. The claim of three independent scalability axes is therefore not established by the current evidence.","section":"§II-B / Table II"},{"comment":"The two learned components are never tested against baselines or ablations. The transformer predictor in §II-C is claimed to reduce redundancy in inter-quality-layer and temporal prediction, but no RD comparison is reported against InterGS-Lite alone or against the KNN/bilateral-only predictor set. Similarly, the opacity-volume importance score ρ_i in Eqs. (19)-(20) is claimed to give meaningful progressive refinement, but Tables I and II only compare refinement levels against one another; there is no comparison with random ordering, reverse ordering, or the RAP estimator [19] from which the criterion is inspired. Without such controls, the monotonic PSNR/SSIM/LPIPS improvements are consistent with any ordering that adds Gaussians, and do not demonstrate that the proposed ordering is near-optimal in a rate-quality sense.","section":"§II-C / §II-E / Tables I-II"},{"comment":"The evaluation is self-referential and does not report RD performance relative to existing scalable or non-scalable GS codecs. The claimed advantage of SplatStream—adaptive streaming with low startup cost—cannot be assessed without comparing, at matched rate and quality, with at least one prior streamable GS framework [13]-[15] and with a single-layer non-scalable codec such as InterGS-Lite. In addition, the reported operating points are extremely high-rate (828-2329 Mbps for a single 4-frame GOP). The paper should contextualize these rates (e.g., versus non-scalable full-quality transmission) and discuss whether the label 'low-latency adaptive streaming' is realistic at these bitrates.","section":"§III / Table II"}],"minor_comments":[{"comment":"The term 'B-layer' is defined as P-style prediction from the nearest previous anchor, not true bidirectional prediction. This is stated in the text, but the terminology may mislead readers; consider 'enhancement layer' or explicitly coining 'P-style B-frames' in the abstract and captions.","section":"§II-D"},{"comment":"Equation (23) should state units explicitly: if B_GOP is in bytes, then R_bw = 8 B_GOP / T_GOP yields bits/sec. The current text uses 'size' ambiguously.","section":"Eq. (23)"},{"comment":"The notation I^r_a is introduced twice and slightly awkwardly: the text says 'I^r_a is the ground-truth image downsampled to resolution r' after the rendering equation. A small cleanup would improve readability.","section":"§II-B / Eq. (3)"},{"comment":"Table II uses 'S3' for the full spatial anchor without defining it in the table. Define S3 in the caption or in the text preceding the table.","section":"Table II"},{"comment":"The RD curves do not specify the evaluation viewpoint (fixed front-facing or free-viewpoint) or whether metrics are averaged over time after the first GOP. This information is needed for reproducibility.","section":"§III / Figs. 1-3"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is a plausible systems paper, but the experimental section does not yet support the three central claims. The spatial-quality axis is confounded with Gaussian count, and neither the transformer predictor nor the importance ordering is compared with baselines. I believe these are fixable within a revision, not fundamental flaws in the framework. I would also ask the editor to note that the authors have prior self-citations [13],[14],[18],[19]; the lack of independent baselines makes the evaluation particularly vulnerable to self-referentiality."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear [Colleague],\n\nThe short version: SplatStream is a sensible engineering integration of prior work on scalable 3DGS, DASH, and predictive coding, but its experiments only evaluate the system against itself. The claimed three independent scalability axes are not actually demonstrated, and the two most novel components have no ablation or baseline.\n\nWhat's new: training a single intra Gaussian representation with multi-resolution rendering loss (so one model serves 540p/720p/1080p), adding a transformer predictor to the InterGS-Lite predictor set, and organizing the bitstream into DASH sub-representations with opacity-volume prefix ordering. The system description is clear, equations are internally consistent, and the reported rate-distortion curves are monotonic. Table II is a useful demonstration of how operating points could be selected in practice, and the DASH mapping is a practical touch.\n\nThe soft spots are substantial but not fatal. There are no comparisons to any prior codec, not even the InterGS-Lite it builds on. The transformer's contribution is never ablated, so the claim that it 'reduces redundancy' is unsupported. The opacity-volume importance ordering is never compared to random ordering or the RAP estimator it is inspired by, so the fine-grained refinement gains could just be the artifact of adding more Gaussians rather than good ordering.\n\nThe stress-test note about the spatial axis is on target. The paper says the same representation is used for all resolutions, yet Table II reports different Gaussian counts for I_540, I_720, and I_1080. If those counts are actually different subsets of the same model, then the 'spatial-quality' rows are just points on the same fine-grained refinement curve, and the PSNR gains are confounded with Gaussian count. No experiment holds the Gaussian set fixed while varying rendering resolution, so the third axis is not established.\n\nThe writing is honest about limitations and the authors cite their own related work, which is fine given the context. But this is a systems paper whose value depends on the components actually working. Without baselines or ablations, the scientific claims are unproven. For an MPEG-oriented venue with a focus on standardization and engineering, that might be acceptable as a position paper; for a general peer-reviewed venue, the experiments need real work.\n\nWho is this for? People working on 3DGS streaming and MPEG-DASH integration will find the architecture worth knowing about. It deserves a serious referee, but the recommendation would be major revision, not acceptance. I'd bring it to a reading group if the topic is relevant, and I'd cite it as related work with a note about the missing validation.","headline":"A coherent 3DGS streaming integration whose headline claims are undercut by an experimental design that never compares against a baseline or isolates the three scalability axes.","tokens_in":10254,"tokens_out":2595,"would_cite":true,"duration_ms":26205,"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":"SplatStream proposes that dynamic 3D Gaussian splatting scenes can be delivered as a DASH-style bitstream with independent spatial, temporal, and fine-grained Gaussian scalability.","keywords":["3D Gaussian Splatting","dynamic scene streaming","scalable coding","MPEG-DASH","inter-frame prediction","transformer predictor","importance ordering","progressive refinement"],"falsifier":"On the three MPEG sequences, re-run the fine-grained operating points with Gaussians ordered randomly (or by opacity alone, without scale) at identical byte totals; if the resulting PSNR and LPIPS improvement from 10% to 100% is not substantially worse, the opacity-volume importance measure is not the mechanism producing the paper's monotonic refinement curves.","tokens_in":9297,"feed_emoji":"📶","tokens_out":4254,"duration_ms":38141,"temperature":0.7,"pith_summary":"Dynamic 3D Gaussian splatting produces vivid scenes but each frame is huge, so streaming it over fluctuating bandwidth is hard. SplatStream argues this can be solved by making the representation itself scalable rather than just compressing it: the same Gaussian model is supervised at several rendering resolutions to create a compact startup layer and refinement layers; a lightweight transformer predicts higher-quality Gaussians from lower layers; temporal frames are split into a low-rate base and an enhancement B-layer; and Gaussians are ordered by an opacity-volume importance score so that the most visible primitives arrive first. The reported operating points show smooth quality growth along all three axes—for one GOP, PSNR climbs from 30.45 dB at the 828 Mbps startup point to 34.12 dB at the full spatial anchor, and from 22.86 dB at 10% Gaussian refinement to 32.34 dB at 100%. If the design holds, a client can trade bandwidth for display resolution, frame rate, and detail continuously from a single encoded stream.","feed_headline":"Three scalability axes in a single 3D Gaussian stream","feed_subtitle":"Spatial quality, frame rate, and Gaussian count adapt to bandwidth from 828 Mbps startup to 2329 Mbps full detail.","key_machinery":"The load-bearing pieces are: (1) a multi-resolution rendering loss that makes the single intra Gaussian model effective at 540p, 720p, and 1080p; (2) an extended predictor set—KNN, bilateral, and a single-head transformer with geometry-aware relative position bias—whose best candidate for each Gaussian is signaled and used to form residuals; (3) a GOP structure with a base temporal layer (I, P, P, P) and a B enhancement layer coded as P-style; (4) the opacity-volume importance score rho = sigma(alpha) * exp(s_x + s_y + s_z), which ranks Gaussians so that prefixes of size 10%, 20%, ... form progressively refining renderable subsets; and (5) an MPEG-DASH manifest that records packet types, lay","core_discovery":"The central claim is that a dynamic 3DGS sequence can be organized as a layered bitstream whose three scalability axes are decoded independently but predictively. The intra anchor is trained with a multi-resolution rendering loss, so one representation supports 540p, 720p, and 1080p rendering; inter-quality-layer redundancy is removed by a transformer predictor that estimates high-layer SH and geometric attributes from the decoded lower layer; temporal redundancy is handled by a base P-frame chain with B-layer enhancement frames predicted in the reconstructed 1080p Gaussian domain; and within each frame, Gaussians sorted by rho = sigma(alpha) * exp(s_x + s_y + s_z) form nested 10%-to-100% pr","pith_inferences":["If the opacity-volume ordering is validated against shuffled or random ordering, the same prefix machinery could be reused as a generic progressive-upload format for static and dynamic 3DGS, independent of the streaming context.","The multi-resolution supervision implies that one training run can serve multiple display resolutions without per-resolution models; a content provider could publish a single asset and let the client request the needed resolution layer.","The transformer predictor's contribution could be isolated by ablating it against the KNN and bilateral branches on the same sequences; the paper's current tables do not separate the three predictors.","A natural stress test is to run the fine-grained refinement curve on a device with a fixed rendering budget; the 10%-90% prefixes may offer a way to keep frame rate stable while detail scales."],"forward_implications":["A player can start with a 540p intra packet, then add spatial layers up to 1080p before any temporal decoding, so startup latency and bandwidth can be matched to the network.","Clients on low bandwidth can request only the base temporal layer (I, P, P, P) for lower-frame-rate playback, and add B-layer frames when bandwidth allows, preserving a single encoded stream.","Within each decoded frame, rendering can use 10% to 100% of Gaussians; this gives a frame-level progressive refinement knob in addition to spatial and temporal layers.","The DASH manifest makes the operating point explicit (packet type, byte size, Gaussian count, predictors), so adaptive selection becomes a client-side table lookup rather than a codec reconfiguration.","Cumulative bits per Gaussian decrease as refinement increases (from 395 to 192), indicating that later packets spend fewer bits per Gaussian while still improving perceptual quality."],"fun_headline_variants":["Layered SplatStream scales 3D detail, quality, and frame rate","Predictive Gaussian layers stream 3D scenes to any bandwidth","Three scalability axes in one 3D bitstream: quality, rate, count","Fine-grained 3D streaming via predictive Gaussian layers"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The fine-grained scalability rests on the opacity-volume score rho = sigma(alpha) * exp(s_x + s_y + s_z) correctly ranking how much each Gaussian contributes to rendering quality; the paper never compares this ordering with random or learned alternatives, so if the ranking is wrong the monotonic refinement gains would not be due to the proposed importance measure.","fun_headline_variants_meta":{"raw":{"variants":["Layered SplatStream scales 3D detail, quality, and frame rate","Predictive Gaussian layers stream 3D scenes to any bandwidth","Three scalability axes in one 3D bitstream: quality, rate, count","Fine-grained 3D streaming via predictive Gaussian layers"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000976,"raw_usage":{"total_tokens":3964,"prompt_tokens":709,"completion_tokens":3255,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":453,"completion_tokens_details":{"reasoning_tokens":3190}},"tokens_in":453,"tokens_out":3255,"duration_ms":23658,"temperature":1.0,"reasoning_tokens":3190,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-01T00:56:25.936512+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"On the three MPEG sequences, re-run the fine-grained operating points with Gaussians ordered randomly (or by opacity alone, without scale) at identical byte totals; if the resulting PSNR and LPIPS improvement from 10% to 100% is not substantially worse, the opacity-volume importance measure is not the mechanism producing the paper's monotonic refinement curves.","supporting_citations":[],"review_version":1}