{"id":"42950536-dcd7-4bd6-aec8-3a58ae63a499","arxiv_id":"2506.09479","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"TinySplat compresses feedforward 3D Gaussian scenes by 105-199x on two-view benchmarks (about 50x on DL3DV) while keeping rendered quality close to the uncompressed model.","lead":"TinySplat is a compression pipeline that shrinks feedforward 3D Gaussian scene data to roughly one percent of its original size with no extra training. It could make 3D scene storage and streaming practical for phones and headsets.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"VABR estimates visibility only from a 3x3 ray grid over input views (Eq. 8); the 100x color compression (k=6, e.g. 27->6 SH channels for MVSplat) is therefore unvalidated for novel views outside that sampled region, and the benchmarks never exercise such views.","rationale":"Good-faith read: TinySplat is a training-free codec for feedforward 3DGS; the core engineering is sound (reversible VPT, PCA-style VABR, HEVC backend) and the reported Re10K/ACID numbers are strong if reproducible. The load-bearing point is the condition under which VABR is lossless-in-practice. Because D is chosen from input views alone, the entire color-channel reduction is justified by a directional prior that the paper itself cannot rigorously define. The reader highlighted exactly this as the weakest assumption; I agree. A test with out-of-frustum novel views would settle it. Additional concerns (DL3DV ~50x vs 100x headline, no code release, no image-codec baseline) are real but secondary; they support the existing CONDITIONAL verdict rather than changing it.","tokens_in":18165,"tokens_out":18812,"duration_ms":233217,"concrete_test":"On 200 Re10K scenes generated by MVSplat, compress with TinySplat at the reported settings, then render target views from poses sampled at increasing azimuth/elevation and translation offsets from the input camera pair, including poses whose rays to the scene lie outside the input-view 3x3 ray cones. Compare TinySplat rendered PSNR/LPIPS against the uncompressed model at each offset. If the additional loss stays <=0.2 dB up to a 30-degree pose sweep, the visibility-region assumption is adequate; if the gap grows with offset, the 100x claim must be qualified to near-input views and the VABR D-estimation needs to be made target-aware.","verdict_should_be":"UNCHANGED","load_bearing_attack":"TinySplat's headline ratio depends on the VABR premise that a scene's SH energy relevant for rendering is concentrated on a few principal components, with the relative importance of SH bases set by visibility weights lambda_m^l computed over a 'valid direction' region D. Section III.D admits D is 'challenging to rigorously define' and estimates it in Eq. 8 with only a 3x3 grid of rays per input view. The paper also asserts (Sec. III.D motivation) that novel viewpoints are 'usually not too far' from input views. For MVSplat, whose raw appearance has 27 SH coefficients, VABR keeps only k=6 dimensions (a 4.5x channel cut); if a target view is rendered from a pose whose view directions fall outside the sampled D, the discarded SH dimensions can carry visible color variation. All reported benchmarks (Re10K, ACID, DL3DV) place target views near the input views, so this regime is untested. Consequently the central claim 'over 100x compression ... with negligible loss' is not established as a general property of the 3D representation; at minimum it is conditional on near-input viewing directions. The concern is not that the transform is internally inconsistent—the forward/inverse pair in Eq. 10 is well-defined and reversible for k=d—but that the empirical generalization of D is the load-bearing assumption.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"TinySplat proposes a training-free compression framework for feedforward 3D Gaussian Splatting (3DGS) representations. Starting from per-view Gaussian feature maps generated by networks such as MVSplat and DepthSplat, the method applies (i) a View-Projection Transform (VPT) that maps geometric parameters into input camera space, (ii) a Visibility-Aware Basis Reduction (VABR) that computes scene-adaptive SH bases weighted by visibility over sampled view directions, and (iii) quantization plus HEVC encoding of each feature plane. The paper reports that on RealEstate10K and ACID, TinySplat reduces MVSplat/DepthSplat models to roughly 0.17–0.23 MB with PSNR drops of 0.04–0.08 dB, and that it surpasses the optimization-free FCGS baseline in storage and runtime. Ablations and additional DL3DV experiments are presented.","tokens_in":18446,"tokens_out":3900,"duration_ms":43402,"significance":"If the results hold, TinySplat is a practically important contribution: it is the first feedforward-compatible, optimization-free codec that exploits the peculiar pixel-aligned structure of feedforward 3DGS, with extremely low encode/decode overhead and a simple, reproducible design. The PCA-based VABR is an interesting idea, and the paper includes component ablations and rate-distortion comparisons. The main significance caveat is that the headline compression ratio is dataset-dependent: the claimed 'over 100x' is not achieved on DL3DV, and the VABR premise about valid view directions is only tested in the easy near-input-view regime.","major_comments":[{"comment":"The abstract and conclusion claim 'over 100x compression for 3D Gaussian data generated by feedforward methods', but Table III reports DL3DV results with ratios of only about 53.7x (33 MB to 0.615 MB), 51.8x (66 MB to 1.275 MB), and 52.9x (100 MB to 1.892 MB). The 100x claim is only supported by the Re10K/ACID numbers in Table I (e.g., DepthSplat Re10K: 19 MB to 0.166 MB). The claim should be qualified per dataset and per inference model, or the abstract should state that 100x is achieved specifically on Re10K/ACID.","section":"Abstract; Table III"},{"comment":"The component-wise ablation disables VPT and VABR separately, but it does not include a control with both transforms disabled and only HEVC quantization/coding. Without this baseline, the reader cannot attribute the rate-distortion improvement to VPT and VABR versus the HEVC encoding itself. Please add a no-transform baseline to Fig. 7 and report the corresponding rates and PSNR.","section":"Fig. 7; Section IV.C"},{"comment":"VABR reduces the SH color space (e.g., 27 channels to k=6) based on visibility weights computed by casting rays through a coarse 3x3 grid over each input view. The paper itself states that the valid-direction region D is 'challenging to rigorously define', and all benchmarks place target views close to the input views. If novel views move well outside the sampled directions, the discarded SH dimensions can carry visible color variation, so the claim of 'negligible loss' is conditional on near-input viewing direction. Please provide an experiment with target views substantially displaced from the input frusta, or an analysis of how the rate-distortion tradeoff degrades with view displacement.","section":"Section III.D, Eq. (8)"},{"comment":"No error bars or variance statistics are reported for the LPI/P S N R measures. The asserted quality advantage over FCGS is often small (e.g., PSNR differences of 0.1 dB or less), and without repeated experiments or per-scene standard deviations it is unclear whether these differences are meaningful. Please report standard deviations or significance tests over the test scenes.","section":"Table I; Section IV.A"}],"minor_comments":[{"comment":"Several figures contain garbled or unreadable symbols (e.g., the SH function labels in Fig. 4 and the axes in Fig. 7), likely due to PDF font issues. These should be regenerated with standard math fonts.","section":"Fig. 3; Fig. 4; Fig. 7"},{"comment":"The forward/inverse pair in Eq. (10) is not invertible when k<d, since W is d×k and W^T W is I_k while W W^T is a projection. The text should explicitly state that this is a lossy dimensionality reduction rather than a reversible transform, to avoid confusion with the invertible VPT.","section":"Section III.D, Eq. (10)"},{"comment":"LPIPS for TinySplat is sometimes slightly worse than the raw model (e.g., DepthSplat ACID: 0.1501 vs 0.1417), while PSNR is nearly unchanged. The 'negligible loss' claim should be supported by discussing all three metrics together, not only PSNR.","section":"Section IV.A, Table I"},{"comment":"The values of alpha and Q_c are hand-tuned per channel. The 'training-free' claim is accurate in the sense of no learned parameters, but the paper should acknowledge more clearly that these hyperparameters are manually configured and may not generalize across datasets without retuning.","section":"Section IV, Implementation details; Table II"},{"comment":"The paper uses official implementations of MVSplat, DepthSplat, and FCGS, but does not provide code for TinySplat itself. Releasing the implementation would substantially strengthen the reproducibility of the reported bitrates and runtime numbers.","section":"Reproducibility"}],"recommendation":"major_revision","confidential_remarks":"The stress-test concern about VABR is legitimate and should be addressed rather than dismissed. The paper would benefit from an explicit robustness test with target views farther from the input viewpoints, and from a no-transform ablation. The abstract's over-100x claim needs dataset-specific qualification as noted. I do not see any circularity issue: the adaptive PCA transform is a standard codec component and the evaluation is against external baselines."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a solid training-free codec for feedforward 3DGS, with two genuinely new components (VPT and VABR), but the headline \"over 100x\" compression is dataset-dependent and the VABR module carries a view-direction assumption that the paper itself admits is hard to pin down. Not a breakthrough, but a real engineering contribution worth engaging.\n\nWhat is new and good: the method is fully training-free, so it can be bolted onto any current feedforward 3DGS model. VPT exploits pixel-alignment by projecting centers into camera space and coding only offsets; that is a simple idea with clear gains in spatial regularity, as the distributions in Fig. 3 show. VABR is a visibility-weighted PCA on SH coefficients. The forward/inverse pair in Eq. 10 is well-defined and invertible before quantization, which means the loss is introduced only where intended. The comparison against FCGS uses the official implementation, and the runtime advantage (about 1s encode, 42ms decode) is meaningful. The citation pattern is clean.\n\nSoft spots, in proportion. The \"over 100x\" claim holds for Re10K and ACID (roughly 114x and 105x from the Table I numbers), but on DL3DV it is about 54x (33 MB to 0.615 MB). The abstract should say \"up to 100x\" or report the dataset-specific numbers. This is an overstatement, not a fabrication. The more substantive concern is VABR's visibility estimate: it uses a 3x3 grid of rays per input view to define the \"valid direction\" region D, and the paper explicitly says D is \"challenging to rigorously define\" (Section III.D). All test configurations put target views near the input views, so the regime where discarded SH dimensions matter is not exercised. This is not fatal for the stated use case—sparse-view, near-input novel-view synthesis is the actual target—but it should be tested with wider baseline angles. The ablation disables VPT and VABR separately, but there is no plain-HEVC baseline with both transforms off; that leaves the codec-only contribution unquantified. Minor: no error bars, and no code release, which makes the exact numbers hard to verify independently.\n\nBottom line: the work is coherent, honest about its main assumption, and a credible step forward for feedforward 3DGS storage and streaming. I would send it to peer review, with a request to fix the abstract, add the missing HEVC-only baseline, and ideally test VABR under more aggressive novel-view baselines. The VABR concern is real but proportionate; it does not sink the paper.","headline":"Solid training-free codec for feedforward 3DGS; the 100x claim is dataset-dependent and the VABR view-direction assumption is real but not fatal.","tokens_in":19007,"tokens_out":3848,"would_cite":true,"duration_ms":41112,"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":"TinySplat claims a training-free pipeline can shrink feedforward 3D Gaussian scene representations by over 100x while keeping rendering quality nearly unchanged.","keywords":["3D Gaussian Splatting","feedforward 3D reconstruction","Gaussian compression","view-projection transform","visibility-aware basis reduction","spherical harmonics","HEVC video coding","novel view synthesis"],"falsifier":"Render the same compressed scene from novel views placed at increasing angular offsets beyond the input camera frustum (for example, 10, 20, and 30 degrees away) and compare PSNR against the uncompressed feedforward model; if the quality gap grows sharply with angle while a baseline compressor stays flat, the 3x3 ray sampling has mis-ordered the spherical-harmonic bases.","tokens_in":17950,"feed_emoji":"🗜️","tokens_out":10169,"duration_ms":87953,"temperature":0.7,"pith_summary":"Feedforward 3D Gaussian Splatting can reconstruct a scene from a handful of images in a single network pass, but the resulting Gaussian models are far larger than the input images and existing compression schemes require per-scene optimization that does not work on these feedforward outputs. TinySplat argues that most of that bulk is removable without any retraining by eliminating three separate redundancies: geometric redundancy through a view-projection transform into camera space, perceptual redundancy through a visibility-aware reduction of the spherical-harmonic color bases, and spatial redundancy through a standard HEVC video codec. The paper reports that compressed DepthSplat and MVSplat models shrink from 19–43 MB to roughly 0.17–0.23 MB, a more than 100-fold reduction, while losing 0.07–0.08 dB PSNR on standard benchmarks. Against the state-of-the-art training-free compressor FCGS, TinySplat matches quality at 6% of the storage size, using 25% of the encoding time and 1% of the decoding time. If these numbers hold, feedforward 3DGS becomes practical for bandwidth- and storage-limited applications such as mobile AR/VR and real-time scene streaming.","feed_headline":"TinySplat shrinks feedforward 3D Gaussian scenes over 100x","feed_subtitle":"Two reversible transforms plus a video codec cut DepthSplat and MVSplat to about 1% of their size with minimal quality loss.","key_machinery":"The two load-bearing mechanisms are VPT and VABR. VPT is a reversible coordinate transformation that exploits the pixel-aligned structure of feedforward Gaussians: each Gaussian center $\\mu$ is projected into the input view's camera space via the intrinsic matrix $K$ and extrinsic rotation $R$ and translation $T$, yielding a depth value $z$ and pixel-center offsets $(x,y)$; rotations are rotated into the camera frame and scales are divided by $z$ (with focal length $f$), and all steps are invertible so rendering is unchanged. VABR is a basis reduction for spherical-harmonic color coefficients: it computes per-basis visibility weights $\\lambda_l^m = \\frac{1}{N_s}\\sum_{i=1}^{N_s}|Y_l^m(\\omega_i)|$ from $N_s$ directions $\\omega_i$ sampled through a 3x3 grid of rays per input view, weights the SH bases by these values, and extracts the principal components of the visibility-scaled coefficient covariance, retaining $k=6$ dimensions. The remaining machinery is per-channel 14-bit quantization with standard-deviation-based steps plus independent HEVC encoding of each quantized feature plane as a grayscale image.","core_discovery":"TinySplat claims that a feedforward-generated 3D Gaussian representation is a structured, image-like tensor whose redundancy can be removed by a training-free, fully reversible pipeline. The View-Projection Transform (VPT) maps Gaussian centers into the input camera space using the input camera intrinsics and extrinsics, so that positions become a depth map plus small offsets from pixel centers and scales are divided by depth; after this reversible mapping, the geometric channels are far more locally correlated and compress better. The Visibility-Aware Basis Reduction (VABR) handles color: it assigns each spherical-harmonic basis function a visibility weight by Monte Carlo sampling viewing directions through a 3x3 grid of rays per input view, scales the SH bases by those weights, and then keeps the top $k=6$ principal components of the scaled coefficient covariance matrix. The transformed features are quantized to 14-bit integers and each channel is encoded independently as a grayscale image with an off-the-shelf HEVC codec, with the inverse transforms restoring the original parameters exactly on the decoder side. On RealEstate10K and ACID with two input views, this reduces DepthSplat models from 19 MB to 0.166–0.181 MB and MVSplat models from 43 MB to 0.216–0.229 MB, with PSNR dropping no more than 0.08 dB.","pith_inferences":["A likely consequence the paper leaves implicit is that wide-baseline novel views will stress VABR, since its visibility weights are estimated from input-view rays; a direct test would measure rate-distortion as target views move beyond the input frustum.","TinySplat's entropy analysis suggests a joint optimization of the inference network's output distribution could shrink the compressed size further, potentially by an order of magnitude, by making feature maps more codec-friendly rather than using generic HEVC.","Treating each feature channel as an independent grayscale video leaves inter-channel correlations, especially among the six color basis channels, unexploited; a codec that models cross-channel dependencies should improve rate-distortion beyond what is reported.","The fixed $k=6$ color dimension may be scene-dependent; scenes with highly multi-modal view distributions might need a larger $k$, and an adaptive $k$ per scene could trade size and quality more flexibly."],"forward_implications":["Feedforward 3DGS models can be stored and transmitted at roughly one-hundredth of their original size with rendering quality within 0.1 dB of the uncompressed model.","Because compression does not require training, the same framework can wrap any pixel-aligned feedforward 3DGS method, as demonstrated with DepthSplat and MVSplat.","Encoding a scene takes about one second and decoding about 42 milliseconds, which brings compressed feedforward 3DGS close to real-time interactive use.","Color coefficients consume the largest share of the bitstream at every quality level, so future compression gains should target the appearance model.","The compression ratio remains above 50x with 4 and 6 input views on DL3DV, so adding views does not undermine the framework."],"supporting_citations":[{"why":"supplies the DepthSplat feedforward 3DGS models used as the primary compression target and quality baseline","marker":"[12]"},{"why":"supplies the MVSplat feedforward 3DGS models demonstrating the framework's generality","marker":"[35]"},{"why":"is the state-of-the-art training-free compression baseline FCGS that TinySplat compares against in size, quality, and runtime","marker":"[11]"},{"why":"introduced the pixel-aligned feedforward 3DGS formulation whose structure TinySplat exploits","marker":"[30]"},{"why":"defines the 3D Gaussian primitive representation and rasterization renderer used for evaluation","marker":"[20]"},{"why":"provides the RealEstate10K benchmark used for the main compression and rate-distortion experiments","marker":"[57]"},{"why":"provides the ACID benchmark used for the second set of compression experiments","marker":"[58]"}],"fun_headline_variants":["TinySplat compresses 3D Gaussians over 100x with two transforms","Training-free compression cuts feedforward 3DGS storage by 100x","Reversible transforms plus HEVC shrink 3D scenes to 1% size","TinySplat: feedforward 3D scenes at 1% storage, same quality","100x smaller 3D Gaussian scenes via view-projection and basis reduction"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"Everything VABR gains depends on the assumption that the viewing directions sampled by a sparse 3x3 grid of rays through each input image are representative of every viewpoint a user will actually render, so that the spherical-harmonic bases it discards are truly invisible.","fun_headline_variants_meta":{"raw":{"variants":["TinySplat compresses 3D Gaussians over 100x with two transforms","Training-free compression cuts feedforward 3DGS storage by 100x","Reversible transforms plus HEVC shrink 3D scenes to 1% size","TinySplat: feedforward 3D scenes at 1% storage, same quality","100x smaller 3D Gaussian scenes via view-projection and basis reduction"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000289,"raw_usage":{"total_tokens":1781,"prompt_tokens":1118,"completion_tokens":663,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":734,"completion_tokens_details":{"reasoning_tokens":554}},"tokens_in":734,"tokens_out":663,"duration_ms":6011,"temperature":1.0,"reasoning_tokens":554,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T04:47:42.656677+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Render the same compressed scene from novel views placed at increasing angular offsets beyond the input camera frustum (for example, 10, 20, and 30 degrees away) and compare PSNR against the uncompressed feedforward model; if the quality gap grows sharply with angle while a baseline compressor stays flat, the 3x3 ray sampling has mis-ordered the spherical-harmonic bases.","supporting_citations":[{"cited_title":"Mvsplat: Efficient 3d gaussian splatting from sparse multi-view images,","cited_arxiv_id":null,"evidence_quote":"supplies the MVSplat feedforward 3DGS models demonstrating the framework's generality"},{"cited_title":"pixelsplat: 3d gaussian splats from image pairs for scalable generalizable 3d recon- struction,","cited_arxiv_id":null,"evidence_quote":"introduced the pixel-aligned feedforward 3DGS formulation whose structure TinySplat exploits"},{"cited_title":"Infinite nature: Perpetual view generation of natural scenes from a single image,","cited_arxiv_id":null,"evidence_quote":"provides the ACID benchmark used for the second set of compression experiments"}],"review_version":1}