{"id":"abffd8f0-d6d8-4c85-a607-e187fda775c6","arxiv_id":"2605.25563","paper_version":1,"verdict":"UNVERDICTED","confidence":"LOW","novelty_score":7.0,"correctness_risk":"unknown","formal_verification":"none","parameter_count":1,"one_line_summary":"CodecSplat integrates entropy coding of structured 2D intermediate features into the feed-forward 3DGS pipeline to produce ultra-compact scene bitstreams with controllable rate-distortion.","lead":"CodecSplat encodes an intermediate 2D Gaussian-generation feature from a feed-forward 3D Gaussian splatting model into an entropy-coded bitstream. At decode time the feature is reconstructed to predict depth and Gaussian parameters, yielding scene representations one order of magnitude smaller than post-hoc compression of the final 3D primitives.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.3","headline":"No significant objection identified","rationale":"The reader's weakest_assumption correctly isolates the critical information-flow link. Because the provided abstract already states the quantitative outcome and the method description contains no obvious circularity or hidden assumption that would invalidate the headline numbers, the second-pass review does not alter the UNVERDICTED status.","tokens_in":1798,"tokens_out":247,"duration_ms":23541,"concrete_test":"Reproduce the lowest-rate CodecSplat configuration on a 10-scene subset of RealEstate10K and compare its PSNR directly against the uncompressed feed-forward backbone on identical views; a drop larger than 1.5 dB would indicate the latent reconstruction step is the dominant quality limiter.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The abstract describes a coherent pipeline that moves entropy coding to a structured 2D intermediate feature rather than post-hoc compression of irregular 3D Gaussians. Reported PSNR values at the stated bitrates are consistent with the claim of an order-of-magnitude size reduction while retaining controllable rate-distortion trade-offs. No internal contradiction, unsupported derivation, or missing control experiment is visible from the given description.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.3","summary":"The paper introduces CodecSplat, a framework that integrates entropy coding directly into the feed-forward 3D Gaussian splatting pipeline by compressing a structured 2D intermediate Gaussian-generation feature rather than the final irregular 3D primitives. The decoder reconstructs the latent feature and uses it to predict depth and Gaussian parameters that are then rendered as 3D Gaussians. Experiments report PSNR of 23.56-26.36 dB on DL3DV and 24.76-27.05 dB on RealEstate10K at per-scene bitrates of 20.00-107.77 KiB and 3.37-12.51 KiB respectively, claiming roughly an order-of-magnitude size reduction versus post-hoc compression of generated Gaussians while retaining controllable rate-distortion behavior.","tokens_in":1880,"tokens_out":482,"duration_ms":26927,"significance":"If the empirical results hold under full scrutiny of training protocols and controls, the work demonstrates a principled way to obtain compact representations for feed-forward 3D reconstruction by exploiting intermediate feature structure for compression. This could meaningfully advance practical deployment of 3D Gaussian splatting in storage- or bandwidth-constrained settings, with the reported numbers indicating substantial gains over decoupled compression baselines.","major_comments":[],"minor_comments":[{"comment":"The abstract states the method is instantiated on 'a feed-forward Gaussian splatting backbone with depth-guided multi-view feature refinement and a hierarchical learned feature codec' but provides no further architectural diagram or pseudocode; adding a high-level pipeline figure in §3 would clarify how the codec interfaces with the refinement module.","section":null},{"comment":"Bitrate and PSNR ranges are reported without accompanying per-rate-point tables or variance across scenes; including such a table (e.g., in §4) would strengthen the claim of controllable rate-distortion behavior.","section":null},{"comment":"The claim that the approach 'avoids inefficient compression over irregular 3D Gaussian primitives' is conceptually clear from the abstract, yet a short ablation comparing codec performance on the 2D feature versus on the final 3D primitives would make the efficiency argument more concrete.","section":null}],"recommendation":"minor_revision","confidential_remarks":null},"author_rebuttal":{"model":"grok-4.3","summary":"We thank the referee for the positive summary, recognition of the work's significance, and recommendation for minor revision. The report contains no major comments.","responses":[],"tokens_in":1380,"tokens_out":48,"duration_ms":11680,"standing_objections":[]},"desk_editor":{"model":"grok-4.3","letter":"CodecSplat moves compression inside the feed-forward 3D Gaussian generation step. It encodes the 2D feature map that the decoder later turns into depth and Gaussian parameters, then decodes that latent to produce the splats. The result on DL3DV and RealEstate10K is 3–108 KiB per scene at 23–27 dB PSNR, roughly ten times smaller than taking the output Gaussians and compressing them after the fact.\n\nThe concrete advance is the placement of the hierarchical learned codec on the structured 2D feature instead of the irregular 3D output. This lets the entropy model use the spatial regularity that already exists in the backbone, and the paper shows the rate-distortion curve stays controllable.\n\nThe numbers are the main evidence offered. The abstract gives clear bit-rate and PSNR ranges on two standard datasets and states the comparison baseline explicitly.\n\nThe soft spot is that the abstract supplies almost no architecture or training detail, so it is hard to tell how much of the reported gain comes from the integration itself versus the particular backbone or codec design choices. A reader would also want to see whether the same size reduction holds against other learned 3D compression methods that were not mentioned.\n\nThe work is aimed at people who need compact, renderable 3D scenes for transmission or storage inside feed-forward pipelines. Anyone already using 3DGS backbones will see the practical difference immediately.\n\nIt is worth sending to peer review. The central engineering claim is straightforward to test and the empirical gap shown is large enough to matter.","headline":"CodecSplat gets order-of-magnitude smaller scenes by entropy-coding the 2D intermediate feature inside the feed-forward 3DGS pipeline rather than compressing the final irregular primitives afterward.","tokens_in":2365,"tokens_out":405,"would_cite":false,"duration_ms":18116,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.3","headline":"CodecSplat encodes an intermediate 2D feature map to produce 3D Gaussian scenes that are roughly ten times smaller than those obtained by compressing the final primitives.","keywords":["3D Gaussian Splatting","feed-forward reconstruction","latent coding","scene compression","neural codec","rate-distortion","multi-view synthesis","compact representation"],"falsifier":"Measure PSNR when the same backbone is run once with the original uncompressed feature map and once with the feature map reconstructed from the bitstream; a large consistent drop would falsify the claim that the latent supplies sufficient information.","tokens_in":2715,"feed_emoji":"🗜️","tokens_out":758,"duration_ms":15719,"temperature":0.7,"pith_summary":"The paper shows that feed-forward 3D Gaussian splatting can be made practical for storage and transmission by folding compression directly into the generation process. Instead of first producing irregular 3D Gaussians and then compressing them, the method compresses a structured 2D feature that the decoder later turns into depth and Gaussian parameters. On standard benchmarks this yields scene representations of only a few to a hundred kilobytes while still delivering PSNR values in the mid-20 dB range. The approach also keeps rate-distortion control intact, so users can trade size for quality in a single trained model. A sympathetic reader would care because current feed-forward pipelines otherwise produce representations that are too large to store or send at scale.","feed_headline":"Latent coding shrinks 3D Gaussian scenes by 10x","feed_subtitle":"Encoding the 2D generation feature instead of final primitives yields scenes of a few to 100 KiB at 23-27 dB PSNR.","key_machinery":"Hierarchical learned feature codec that entropy-codes the 2D intermediate Gaussian-generation feature so the decoder can reconstruct it and predict depth plus Gaussian parameters.","core_discovery":"By inserting a hierarchical learned feature codec between the encoder that produces the 2D Gaussian-generation feature and the decoder that predicts depth and Gaussian parameters, CodecSplat converts the intermediate representation into an entropy-coded bitstream; the reconstructed latent then drives the same depth-guided multi-view refinement and Gaussian prediction steps used in the uncompressed backbone, yielding 23.56-26.36 dB PSNR at 20-107 KiB per scene on DL3DV and 24.76-27.05 dB at 3.37-12.51 KiB on RealEstate10K.","pith_inferences":["The same latent-coding idea could be inserted into other feed-forward 3D reconstruction pipelines that also produce intermediate 2D features.","Because the codec already produces a compact bitstream, the framework could support progressive transmission or streaming of 3D scenes without additional engineering.","If the hierarchical codec is replaced by a more modern neural compressor, further size reductions may be possible while keeping the same downstream prediction head."],"forward_implications":["Scene storage drops by roughly an order of magnitude compared with post-hoc compression of the final Gaussians.","Rate-distortion curves remain controllable because the codec operates inside the learned pipeline.","The same depth-guided multi-view refinement and Gaussian prediction steps continue to work after the latent is decoded.","The method applies to existing feed-forward backbones without requiring per-scene optimization."],"fun_headline_variants":["CodecSplat compresses feed-forward 3D Gaussians via 2D latent coding","2D feature codec yields 23-27 dB PSNR at 3-108 KiB per scene","CodecSplat encodes intermediate 2D features into compact scene bitstreams","Latent coding compresses 3D scenes to 3-108 KiB at 23-27 dB"],"cache_read_input_tokens":2112,"weakest_assumption_plain":"The entropy-coded latent feature still contains enough information for the decoder to predict accurate depth and Gaussian parameters that render at high visual quality.","fun_headline_variants_meta":{"raw":{"variants":["CodecSplat compresses feed-forward 3D Gaussians via 2D latent coding","2D feature codec yields 23-27 dB PSNR at 3-108 KiB per scene","CodecSplat encodes intermediate 2D features into compact scene bitstreams","Latent coding compresses 3D scenes to 3-108 KiB at 23-27 dB"]},"model":"grok-4.3","cost_usd":0.008462,"raw_usage":{"total_tokens":3895,"prompt_tokens":805,"num_sources_used":0,"completion_tokens":98,"cost_in_usd_ticks":84624500,"prompt_tokens_details":{"text_tokens":805,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":2992,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":805,"tokens_out":98,"duration_ms":22316,"temperature":1.0,"reasoning_tokens":2992,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-06-29T23:07:17.854821+00:00","model_set":{"reader":"grok-4.3"},"falsifier":"Measure PSNR when the same backbone is run once with the original uncompressed feature map and once with the feature map reconstructed from the bitstream; a large consistent drop would falsify the claim that the latent supplies sufficient information.","supporting_citations":[],"review_version":1}