Pith. sign in

REVIEW 3 major objections 7 minor 58 references

Low-bitrate feed-forward 3D Gaussian scenes can be decoded by keeping structure compact and regenerating missing detail under geometry control.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · grok-4.5

2026-07-31 15:46 UTC pith:PQ2ITAZF

load-bearing objection Real systems work on low-bitrate feed-forward 3DGS; dual-stream + geometry-guided one-step decode is the contribution, but “superior to existing methods” is only shown against cascaded baselines, not the joint feed-forward codecs they cite. the 3 major comments →

arxiv 2607.24403 v1 pith:PQ2ITAZF submitted 2026-07-27 cs.CV

GenSplatCodec: Feed-Forward Gaussian Splatting Compression via One-Step Diffusion

Qiang Hu , Zhenlong Wu , Lei Huang , Zihan Zheng , Xiaoyun Zhang , Wenjun Zhang This is my paper
classification cs.CV
keywords feed-forward 3D Gaussian splattingGaussian compressiongeometry-guided generative decodingone-step diffusionnovel view synthesisrate-distortiondual-stream codec
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

Feed-forward 3D Gaussian methods reconstruct scenes quickly from sparse views, but the resulting dense Gaussians are expensive to store and send. At low bitrates, ordinary deterministic compression drops high-frequency textures and view-dependent appearance, and bolting a generative model on afterward often breaks consistency across viewpoints. GenSplatCodec reframes the problem: the codec transmits a compact Gaussian structural stream plus a small coded reference appearance stream, then a one-step generative decoder rebuilds the novel view under hierarchical geometry control from the decoded Gaussians. A three-stage training path first learns detail-aware Gaussians, then rate-constrained coding, then adapts the generator to the actual coded conditions. On standard benchmarks the method reports reconstruction quality matching or beating strong feed-forward baselines and cascaded compressors while using far less storage—roughly seventeen times less than a leading unrecompressed feed-forward model at matched fidelity—and stronger multi-view consistency than generative refinement baselines.

Core claim

The authors establish that low-bitrate feed-forward Gaussian compression works better when treated as geometry-guided generative decoding than as deterministic recovery of every visual detail: a dual-stream codec that preserves compact structure and a lightweight reference appearance cue, decoded by a one-step multi-view generator under hierarchical geometry control, yields superior rate-distortion and cross-view consistency without per-scene optimization.

What carries the argument

Geometry-guided one-step generative decoding: decoded low-rate Gaussians supply a query RGB anchor and query/reference geometry maps that are turned into multi-level latent control signals; those signals, plus the coded reference latent, drive a shared one-step multi-view U-Net so appearance transfer stays tied to the transmitted structure.

Load-bearing premise

A single coded reference appearance latent plus geometry maps from the low-rate Gaussians are enough for one generative step to restore high-frequency and view-dependent detail that stays faithful to the real scene for arbitrary query poses, including weakly observed regions.

What would settle it

On held-out sparse-view scenes, measure whether novel views in poorly observed regions stay geometrically and photometrically consistent with ground truth and with each other when the reference stream is removed or replaced by an unrelated view; a large drop in cross-view SIFT/CLIP consistency or a rise in structured error relative to the paper’s reported tables would falsify the claim.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Low-bitrate immersive streaming can ship feed-forward Gaussian scenes at a fraction of prior storage while keeping novel-view quality usable.
  • Codec design for explicit 3D representations can allocate most bits to structure and treat high-frequency appearance as a generative residual under geometric constraints.
  • One-step geometry-conditioned generation can replace multi-step diffusion post-processors when cross-view consistency is required.
  • Rate can be steered mainly by Gaussian compaction and attribute entropy coding, with only a small fixed overhead for the reference appearance stream.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • If hierarchical geometry control is the real stabilizer, similar dual-stream one-step decoding may transfer to other explicit scene formats (meshes, point radiance) under bandwidth limits.
  • Failure cases will likely concentrate where the selected reference view shares little visibility with the query, suggesting adaptive multi-reference or visibility-aware bit allocation as a natural next test.
  • The three-stage freeze-then-adapt schedule implies that joint end-to-end training of predictor, entropy model, and generator remains brittle for this class of codecs.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 7 minor

Summary. The paper proposes GenSplatCodec, a feed-forward 3D Gaussian splatting codec that pairs a compact Gaussian structural stream (dual-branch prediction, RD-guided pruning, G-PCC geometry coding plus a hyperprior/spatial/channel-context attribute entropy model) with a small reference-appearance stream (a quantized, entropy-coded VAE latent of the input view nearest the query pose). Decoding is performed by a one-step multi-view diffusion U-Net conditioned on the decoded Gaussian renderings and depth maps through a hierarchical geometry-control module, trained in three stages (feed-forward pretraining, entropy-constrained coding, codec-aware generative adaptation). Experiments on DL3DV and RealEstate10K at 224² and 518², with 2/4/6 input views, report large rate savings over uncompressed feed-forward reconstructors (≈17× vs YoNoSplat at matched PSNR) and roughly 3× over cascaded reconstructor+compressor pipelines (YoNoSplat+SOGS), together with multi-view consistency metrics, cumulative ablations, bit-allocation analysis, and runtime.

Significance. If the results hold, this is a timely and well-motivated contribution: generative decoding for low-bitrate 3DGS transmission is an active direction, and doing it with codec-derived conditions (rather than image-domain post-processing) in a single diffusion step is a clean formulation with practical appeal (1.58 s/1.80 s encode/decode, 0.12 s generative step). The experimental package is unusually complete for this venue: rate accounting is honest — both streams are included in reported bitrates (Eq. 23, §IV-A.3), the reference stream's overhead is quantified (≤3.06%, Table VI), and non-differentiable components excluded from the rate loss are explicitly disclosed (§III-D, Stage 2). The sparsity sweep (Table II), 518² evaluation (Table IV), consistency evaluation against MVSplat360/LatentSplat (Table III), and cumulative ablation isolating each design decision (Table V) are all the right experiments. The gap is that the comparison set does not include the methods the paper itself identifies as the direct prior art, so the headline claim "superior RD performance over existing methods" is currently supported only against uncompressed reconstructors and cascaded baselines.

major comments (3)
  1. [§II-C and §IV-A.2 / Table I, Fig. 3] The related-work section explicitly names TinySplat [9], CodecSplat [10], and SparseSplat [45] as the methods that 'incorporate representation design and learned coding into feed-forward 3DGS pipelines' — i.e., the direct prior art for exactly this problem — yet none appears in any table or RD curve. All quantitative comparisons are against (a) uncompressed feed-forward reconstructors, where large rate savings are expected of any competent codec, and (b) cascaded pipelines that attach compressors designed for per-scene-optimized 3DGS (FCGS, SOGS) to feed-forward outputs. The abstract's central claim ('consistently achieves superior rate-distortion performance over existing methods') is therefore unsupported precisely where it matters most. TinySplat (TCSVT) and SparseSplat (CVPR) are published and should be benchmarked at matched operating points; CodecSplat is contemporaneous but at min
  2. [Abstract / §I (contribution bullets) and §IV-B vs. Table I] The headline '≈17× lower storage than YoNoSplat' compares the codec against an uncompressed representation. The codec-vs-codec comparison in the same table is materially smaller: YoNoSplat+SOGS reaches 20.04 dB at 1.199 MB vs. Ours-Mid 20.02 dB at 0.376 MB — a real but more modest ≈3.2× advantage (and Ours-Low actually trails YoNoSplat+SOGS in PSNR by 0.83 dB at 5.8× smaller size, which is an RD trade-off, not dominance). The abstract and intro should state the advantage against the strongest compressed baseline, not only against the uncompressed reference. Relatedly, the cascaded baselines are each run at a single operating point (FCGS with λ=0.0001, SOGS default config, §IV-A.2), so the RD curves in Fig. 3 compare three tuned operating points of the proposed method against one point per baseline; please sweep baseline rate parameters (or justify why they cannot be swept) so the curve c
  3. [§III-C and §IV (Tables I, III, V)] At the low-rate operating points the decoder is synthesizing a substantial fraction of the final image content from a one-step diffusion prior conditioned on a heavily pruned/quantized Gaussian set plus one reference latent. Table V shows the deterministic decoded Gaussians at 0.577 MB reach only 17.87 dB / 0.478 LPIPS before generation, and the 'Naive One-Step SD' row improves LPIPS while decreasing PSNR — i.e., the generative step demonstrably can trade fidelity for plausibility. The claim that hierarchical geometry control keeps generation 'perceptually faithful' for arbitrary query poses (abstract, §III-C) currently rests on dataset-average metrics and the consistency suite in Table III. Two additions would make this load-bearing claim credible: (i) an analysis of reconstruction quality as a function of query-to-reference camera distance (the reference view is selected by proximity,
minor comments (7)
  1. [§III-C, Eq. (15)] The pretrained generative prior underlying the one-step U-Net is never identified (base model, distillation procedure if any, and what the 'scene-independent conditioning embedding' e_p actually is — a null text prompt?). This matters for reproducibility and for assessing how much of the low-bitrate quality comes from prior training data. Please specify the backbone, its pretraining, and which parameters are adapted in Stage 3.
  2. [§III-B, Eq. (8) and §IV-A] The reference-view selection protocol is under-specified: is I_r restricted to the input context views, or can it be any view? Since selection depends on the query pose, the transmitted bitrate is query-dependent; please clarify how per-scene sizes in Tables I–VI are aggregated over query views, and how many query views per scene are evaluated.
  3. [Table I] YoNoSplat's size is reported as exactly 6.508 MB on both DL3DV and RealEstate10K; identical sizes across datasets with different scene statistics are surprising (a fixed primitive count would explain it, but then DepthSplat's differing sizes need the same explanation). Please verify and state how 'size' is computed for the uncompressed baselines.
  4. [§IV-C (Runtime Analysis)] No hardware is specified for the 1.58 s/1.80 s encode/decode figures, and baseline runtimes (particularly the cascaded pipelines, some of which involve per-scene steps) are not reported. Please add the device and a runtime comparison.
  5. [§IV-A.3] FID on a 224² test split of a few hundred scenes is a high-variance estimate; please state the number of images used per FID computation and consider reporting it as secondary evidence rather than alongside PSNR in the headline discussion.
  6. [§III-B, Eq. (6)] The RD-priority proxy p_RD = v_j/(b^a_j + ε) ignores the geometry-coding share of each primitive's cost entirely (justified by G-PCC's non-decomposability). Since B_µ is 21–29% of the bitstream (Table VI), a brief sensitivity check on the retention ratio or an ablation of the proxy would strengthen this heuristic.
  7. [Throughout] Typographical: missing spaces in 'proposeGenSplatCodec' (§I) and similar concatenations; 'naive one-step SD' in Table V is used before any one-step diffusion model is defined; Table I mixes bold/underline inconsistently in the RE10K block (e.g., SSIM row). G-PCC quantization resolution for Gaussian centers is never stated.

Circularity Check

0 steps flagged

No significant circularity: empirical codec paper with external benchmarks; RD claims are measured, not definitional.

full rationale

GenSplatCodec is a systems/ML paper whose central claim is superior measured rate-distortion on DL3DV and RealEstate10K versus named baselines. Rate is the sum of actual transmitted bitstreams (R_G + R_R); distortion and perceptual scores compare decoded/generated images to held-out ground truth. The three-stage losses (rendering MSE/LPIPS/opacity, fidelity+rate, fidelity+LPIPS) are standard training objectives, not identities that force the reported test metrics by construction. Dual-stream coding, RD-guided pruning, geometry-attribute coding, and geometry-guided one-step decoding are architectural choices evaluated empirically; none reduces a claimed prediction to a fitted input or a self-defined quantity. Author self-citations appear only as related prior work (e.g., 4D/NeRF compression), not as load-bearing uniqueness theorems that forbid alternatives or smuggle the result. Missing direct comparison to CodecSplat/TinySplat/SparseSplat is a baseline-completeness issue, not circularity. Derivation chain is self-contained against external data.

Axiom & Free-Parameter Ledger

5 free parameters · 6 axioms · 2 invented entities

The central RD claim rests on standard neural rendering/compression assumptions plus several paper-specific modeling choices: that compact Gaussians should carry structure while a generative decoder restores discarded appearance; that one coded reference view’s VAE latent plus rendered depth controls suffice; and that staged optimization with particular losses and rate proxies yields a fair codec. Free parameters are the usual ML knobs (rate weights, loss weights, retention ratios, operating points). No new physical entities are postulated.

free parameters (5)
  • λ_rate (attribute rate penalty / operating points) = e.g. 0.002, 0.050, 0.100, 1.000
    Controls bitrate vs fidelity in Stage 2 and defines Low/Mid/High points; Table VI sweeps 0.002–1.000.
  • Loss weights λ_lpips, λ_op, λ_mse, λ_ssim, λ_2
    Balance pixel, perceptual, opacity sparsity, and generative objectives across stages (§III-D); values not fully enumerated.
  • Gaussian retention ratio / RD-guided pruning threshold
    Determines how many primitives survive compaction via p_RD ranking (§III-B); directly sets structural rate.
  • Reference-latent quantization and byte-plane coding settings
    Set the small BR overhead and appearance cue quality (§III-B, Eq. 8).
  • Geometry controller and one-step U-Net adaptation hyperparameters
    Stage-3 optimization knobs for hierarchical control injection and cross-view mixing (§III-C/D).
axioms (6)
  • domain assumption At low bitrate, deterministic recovery of full high-frequency/view-dependent appearance from compact Gaussians is inadequate; structure-preserving coding plus generative detail synthesis is the right objective.
    Stated in abstract and §I as the reformulation premise motivating the whole codec.
  • ad hoc to paper Decoded Gaussian geometry (RGB anchor + query/reference depth) plus one selected reference VAE latent provide sufficient conditions for view-consistent novel-view synthesis without raw side images.
    Core dual-stream design in §III-A/B/C; reference chosen by camera proximity.
  • ad hoc to paper Attribute coding cost is a usable proxy for per-primitive RD value when center cost cannot be decomposed (p_RD = v_j / (b^a_j + ε)).
    Explicit proxy choice in §III-B for greedy compaction.
  • domain assumption G-PCC for centers, Morton-ordered hierarchical entropy coding for attributes, and standard VAE latent coding for the reference stream are appropriate coding backends.
    Adopted in §III-B as engineering substrate; geometry rate excluded from differentiable L_rate.
  • ad hoc to paper Three-stage freeze-and-adapt training is stable enough that Stage-3 generative results reflect the true codec pathway rather than teacher-forced uncompressed cues.
    §III-D argues staged optimization avoids unstable joint training.
  • domain assumption PSNR/SSIM/LPIPS/FID/MUSIQ/MANIQA and listed consistency metrics on DL3DV/RE10K official splits are adequate to claim superior RD and view consistency.
    Evaluation protocol in §IV-A; standard in the area but imperfect for generative faithfulness.
invented entities (2)
  • GenSplatCodec dual-stream geometry-guided one-step generative codec no independent evidence
    purpose: Unify feed-forward Gaussian prediction, compact structural+reference coding, and one-step generative novel-view decoding.
    The named system is the paper’s proposed artifact—an engineered architecture, not a new physical entity. Independent evidence is only the paper’s own benchmarks unless code/models are released.
  • Hierarchical geometry control module C_φ for multi-level latent injection no independent evidence
    purpose: Convert rendered depth maps into latent-aligned controls that constrain cross-view appearance transfer.
    Introduced in §III-C as a new module within the decoder; validated only via ablations in this paper.

pith-pipeline@v1.2.0-grok45-kimik3 · 24106 in / 4121 out tokens · 85939 ms · 2026-07-31T15:46:39.917990+00:00 · methodology

0 comments
read the original abstract

Feed-forward 3D Gaussian Splatting (3DGS) enables scalable scene reconstruction without per-scene optimization, yet produces dense Gaussians that are costly to store and transmit. Existing feed-forward Gaussian compression methods formulate decoding as deterministic representation recovery, which becomes inadequate at low bitrates when high-frequency textures and view-dependent appearance are discarded. Although generative models offer a promising alternative, using them as standalone post-processing decouples generation from the transmitted scene structure, thereby compromising cross-view consistency. To address these limitations, we propose GenSplatCodec, a unified feed-forward Gaussian codec that reformulates low-bitrate Gaussian compression as geometry-guided generative decoding. We present a detail-aware feed-forward Gaussian coding scheme within a dual-stream formulation, where the resulting compact Gaussian structural stream is complemented by a lightweight reference appearance stream. We further introduce a geometry-guided one-step generative decoding approach that jointly exploits decoded structural and appearance cues through hierarchical geometry control to reconstruct high-fidelity and view-consistent novel views. Finally, we develop a three-stage optimization strategy that stabilizes the learning of the unified codec and adapts the generative decoder to codec-derived structural and appearance cues. Extensive experiments across multiple datasets demonstrate that GenSplatCodec consistently achieves superior rate-distortion (RD) performance over existing methods.

Figures

Figures reproduced from arXiv: 2607.24403 by Lei Huang, Qiang Hu, Wenjun Zhang, Xiaoyun Zhang, Zhenlong Wu, Zihan Zheng.

Figure 1
Figure 1. Figure 1: Overview of GenSplatCodec. Given sparse input views, the feed-forward encoder produces a Gaussian structural bitstream [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Architecture of the dual-stream coding scheme. (a) The Gaussian [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Rate-distortion curves on the DL3DV and RealEstate10K datasets. The RD curves demonstrate the superiority of GenSplatCodec over existing Gaussian [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Qualitative comparison of novel-view reconstruction results. Existing Gaussian compression methods suffer from noticeable texture degradation [PITH_FULL_IMAGE:figures/full_fig_p010_4.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

58 extracted references · 7 linked inside Pith

  1. [1]

    Nerf: Representing scenes as neural radiance fields for view synthesis,

    B. Mildenhall, P. P. Srinivasan, M. Tancik, J. T. Barron, R. Ramamoorthi, and R. Ng, “Nerf: Representing scenes as neural radiance fields for view synthesis,”Communications of the ACM, vol. 65, no. 1, pp. 99–106, 2021

    2021

  2. [2]

    3d gaussian splatting for real-time radiance field rendering

    B. Kerbl, G. Kopanas, T. Leimk ¨uhler, and G. Drettakis, “3d gaussian splatting for real-time radiance field rendering.”ACM Trans. Graph., vol. 42, no. 4, pp. 139–1, 2023

    2023

  3. [3]

    Dust3r: Geometric 3d vision made easy,

    S. Wang, V . Leroy, Y . Cabon, B. Chidlovskii, and J. Revaud, “Dust3r: Geometric 3d vision made easy,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2024, pp. 20 697–20 709

    2024

  4. [4]

    Vggt: Visual geometry grounded transformer,

    J. Wang, M. Chen, N. Karaev, A. Vedaldi, C. Rupprecht, and D. Novotny, “Vggt: Visual geometry grounded transformer,” inProceedings of the Computer Vision and Pattern Recognition Conference, 2025, pp. 5294– 5306

    2025

  5. [5]

    pixelnerf: Neural radiance fields from one or few images,

    A. Yu, V . Ye, M. Tancik, and A. Kanazawa, “pixelnerf: Neural radiance fields from one or few images,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2021, pp. 4578– 4587

    2021

  6. [6]

    pixelsplat: 3d gaussian splats from image pairs for scalable generalizable 3d re- construction,

    D. Charatan, S. L. Li, A. Tagliasacchi, and V . Sitzmann, “pixelsplat: 3d gaussian splats from image pairs for scalable generalizable 3d re- construction,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2024, pp. 19 457–19 467

    2024

  7. [7]

    Mvsplat: Efficient 3d gaussian splatting from sparse multi-view images,

    Y . Chen, H. Xu, C. Zheng, B. Zhuang, M. Pollefeys, A. Geiger, T.- J. Cham, and J. Cai, “Mvsplat: Efficient 3d gaussian splatting from sparse multi-view images,” inEuropean conference on computer vision. Springer, 2024, pp. 370–386

    2024

  8. [8]

    Depthsplat: Connecting gaussian splatting and depth,

    H. Xu, S. Peng, F. Wang, H. Blum, D. Barath, A. Geiger, and M. Polle- feys, “Depthsplat: Connecting gaussian splatting and depth,” inProceed- ings of the Computer Vision and Pattern Recognition Conference, 2025, pp. 16 453–16 463

    2025

  9. [9]

    Tinysplat: Feedforward approach for generating compact 3d scene representation,

    Z. Song, J. Fu, J. Zhang, X. Lu, C. Jia, S. Ma, and W. Gao, “Tinysplat: Feedforward approach for generating compact 3d scene representation,” IEEE Transactions on Circuits and Systems for Video Technology, 2026

    2026

  10. [10]

    Codecsplat: Ultra-compact latent coding for feed-forward 3d gaussian splatting,

    P. Yu, R. Jiang, Q. Zhang, D. Li, J. Wang, and Y . Guo, “Codecsplat: Ultra-compact latent coding for feed-forward 3d gaussian splatting,” arXiv preprint arXiv:2605.25563, 2026. JOURNAL OF LATEX CLASS FILES, VOL. 18, NO. 9, SEPTEMBER 2020 12

    Pith/arXiv arXiv 2026

  11. [11]

    Yonosplat: You only need one model for feedforward 3d gaussian splatting,

    B. Ye, B. Chen, H. Xu, D. Barath, and M. Pollefeys, “Yonosplat: You only need one model for feedforward 3d gaussian splatting,”arXiv preprint arXiv:2511.07321, 2025

    arXiv 2025

  12. [12]

    Mip-nerf: A multiscale representation for anti- aliasing neural radiance fields,

    J. T. Barron, B. Mildenhall, M. Tancik, P. Hedman, R. Martin-Brualla, and P. P. Srinivasan, “Mip-nerf: A multiscale representation for anti- aliasing neural radiance fields,” inProceedings of the IEEE/CVF inter- national conference on computer vision, 2021, pp. 5855–5864

    2021

  13. [13]

    Instant neural graphics primitives with a multiresolution hash encoding,

    T. M ¨uller, A. Evans, C. Schied, and A. Keller, “Instant neural graphics primitives with a multiresolution hash encoding,”ACM transactions on graphics (TOG), vol. 41, no. 4, pp. 1–15, 2022

    2022

  14. [14]

    Jointrf: End-to-end joint optimization for dynamic neural radiance field representation and compression,

    Z. Zheng, H. Zhong, Q. Hu, X. Zhang, L. Song, Y . Zhang, and Y . Wang, “Jointrf: End-to-end joint optimization for dynamic neural radiance field representation and compression,” in2024 IEEE International Conference on Image Processing (ICIP), 2024, pp. 3292–3298

    2024

  15. [15]

    Mip-nerf 360: Unbounded anti-aliased neural radiance fields,

    J. T. Barron, B. Mildenhall, D. Verbin, P. P. Srinivasan, and P. Hedman, “Mip-nerf 360: Unbounded anti-aliased neural radiance fields,” inPro- ceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2022, pp. 5470–5479

    2022

  16. [16]

    Hpc: Hierarchical progressive coding framework for volumetric video,

    Z. Zheng, H. Zhong, Q. Hu, X. Zhang, L. Song, Y . Zhang, and Y . Wang, “Hpc: Hierarchical progressive coding framework for volumetric video,” inACM MM, ser. MM ’24. New York, NY , USA: Association for Computing Machinery, 2024, p. 7937–7946

    2024

  17. [17]

    2d gaussian splatting for geometrically accurate radiance fields,

    B. Huang, Z. Yu, A. Chen, A. Geiger, and S. Gao, “2d gaussian splatting for geometrically accurate radiance fields,” inACM SIGGRAPH 2024 conference papers, 2024, pp. 1–11

    2024

  18. [18]

    Mip-splatting: Alias-free 3d gaussian splatting,

    Z. Yu, A. Chen, B. Huang, T. Sattler, and A. Geiger, “Mip-splatting: Alias-free 3d gaussian splatting,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2024, pp. 19 447–19 456

    2024

  19. [19]

    Vastgaussian: Vast 3d gaussians for large scene reconstruction,

    J. Lin, Z. Li, X. Tang, J. Liu, S. Liu, J. Liu, Y . Lu, X. Wu, S. Xu, Y . Yan et al., “Vastgaussian: Vast 3d gaussians for large scene reconstruction,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 5166–5175

    2024

  20. [20]

    4d gaussian splatting for real-time dynamic scene rendering,

    G. Wu, T. Yi, J. Fang, L. Xie, X. Zhang, W. Wei, W. Liu, Q. Tian, and X. Wang, “4d gaussian splatting for real-time dynamic scene rendering,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2024, pp. 20 310–20 320

    2024

  21. [21]

    Octree-gs: To- wards consistent real-time rendering with lod-structured 3d gaussians,

    K. Ren, L. Jiang, T. Lu, M. Yu, L. Xu, Z. Ni, and B. Dai, “Octree-gs: To- wards consistent real-time rendering with lod-structured 3d gaussians,” arXiv preprint arXiv:2403.17898, 2024

    Pith/arXiv arXiv 2024

  22. [22]

    Sugar: Surface-aligned gaussian splatting for efficient 3d mesh reconstruction and high-quality mesh rendering,

    A. Gu ´edon and V . Lepetit, “Sugar: Surface-aligned gaussian splatting for efficient 3d mesh reconstruction and high-quality mesh rendering,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2024, pp. 5354–5363

    2024

  23. [23]

    Mvsnerf: Fast generalizable radiance field reconstruction from multi-view stereo,

    A. Chen, Z. Xu, F. Zhao, X. Zhang, F. Xiang, J. Yu, and H. Su, “Mvsnerf: Fast generalizable radiance field reconstruction from multi-view stereo,” inProceedings of the IEEE/CVF international conference on computer vision, 2021, pp. 14 124–14 133

    2021

  24. [24]

    Splatter image: Ultra- fast single-view 3d reconstruction,

    S. Szymanowicz, C. Rupprecht, and A. Vedaldi, “Splatter image: Ultra- fast single-view 3d reconstruction,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2024, pp. 10 208–10 217

    2024

  25. [25]

    Gs-lrm: Large reconstruction model for 3d gaussian splatting,

    K. Zhang, S. Bi, H. Tan, Y . Xiangli, N. Zhao, K. Sunkavalli, and Z. Xu, “Gs-lrm: Large reconstruction model for 3d gaussian splatting,” inEuropean Conference on Computer Vision. Springer, 2024, pp. 1–19

    2024

  26. [26]

    Lgm: Large multi-view gaussian model for high-resolution 3d content creation,

    J. Tang, Z. Chen, X. Chen, T. Wang, G. Zeng, and Z. Liu, “Lgm: Large multi-view gaussian model for high-resolution 3d content creation,” in European Conference on Computer Vision. Springer, 2024, pp. 1–18

    2024

  27. [27]

    Pf3plat: Pose-free feed-forward 3d gaussian splatting,

    S. Hong, J. Jung, H. Shin, J. Han, J. Yang, C. Luo, and S. Kim, “Pf3plat: Pose-free feed-forward 3d gaussian splatting,”arXiv preprint arXiv:2410.22128, 2024

    Pith/arXiv arXiv 2024

  28. [28]

    No pose, no problem: Surprisingly simple 3d gaussian splats from sparse unposed images,

    B. Ye, S. Liu, H. Xu, X. Li, M. Pollefeys, M.-H. Yang, and S. Peng, “No pose, no problem: Surprisingly simple 3d gaussian splats from sparse unposed images,” inInternational Conference on Learning Representa- tions, vol. 2025, 2025, pp. 54 009–54 033

    2025

  29. [29]

    Anysplat: Feed-forward 3d gaussian splatting from unconstrained views,

    L. Jiang, Y . Mao, L. Xu, T. Lu, K. Ren, Y . Jin, X. Xu, M. Yu, J. Pang, F. Zhaoet al., “Anysplat: Feed-forward 3d gaussian splatting from unconstrained views,”ACM Transactions on Graphics (TOG), vol. 44, no. 6, pp. 1–16, 2025

    2025

  30. [30]

    Prosplat: Improved feed-forward 3d gaussian splatting for wide-baseline sparse views,

    X. Lu, J. Fu, J. Zhang, Z. Song, C. Jia, and S. Ma, “Prosplat: Improved feed-forward 3d gaussian splatting for wide-baseline sparse views,” arXiv preprint arXiv:2506.07670, 2025

    Pith/arXiv arXiv 2025

  31. [31]

    Diffsplat: Repurposing image diffusion models for scalable gaussian splat generation,

    C. Lin, P. Pan, B. Yang, Z. Li, and Y . Mu, “Diffsplat: Repurposing image diffusion models for scalable gaussian splat generation,”arXiv preprint arXiv:2501.16764, 2025

    Pith/arXiv arXiv 2025

  32. [32]

    Mvsplat360: Feed-forward 360 scene synthesis from sparse views,

    Y . Chen, C. Zheng, H. Xu, B. Zhuang, A. Vedaldi, T.-J. Cham, and J. Cai, “Mvsplat360: Feed-forward 360 scene synthesis from sparse views,”Advances in Neural Information Processing Systems, vol. 37, pp. 107 064–107 086, 2024

    2024

  33. [33]

    One-shot refiner: Boosting feed-forward novel view synthesis via one-step diffusion,

    Y . Dong, Q. Zhang, M. Jiang, Z. Wu, Q. Fan, Y . Feng, H. Zhang, H. Bao, and G. Zhang, “One-shot refiner: Boosting feed-forward novel view synthesis via one-step diffusion,”arXiv preprint arXiv:2601.14161, 2026

    arXiv 2026

  34. [34]

    Lightgaussian: Unbounded 3d gaussian compression with 15x reduction and 200+ fps,

    Z. Fan, K. Wang, K. Wen, Z. Zhu, D. Xu, and Z. Wang, “Lightgaussian: Unbounded 3d gaussian compression with 15x reduction and 200+ fps,”Advances in neural information processing systems, vol. 37, pp. 140 138–140 158, 2024

    2024

  35. [35]

    Compact 3d gaussian representation for radiance field,

    J. C. Lee, D. Rho, X. Sun, J. H. Ko, and E. Park, “Compact 3d gaussian representation for radiance field,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 21 719–21 728

    2024

  36. [36]

    Efficientgs: Streamlining gaussian splatting for large-scale high-resolution scene representation,

    W. Liu, T. Guan, B. Zhu, L. Xu, Z. Song, D. Li, Y . Wang, and W. Yang, “Efficientgs: Streamlining gaussian splatting for large-scale high-resolution scene representation,”IEEE MultiMedia, 2025

    2025

  37. [37]

    4dgcpro: efficient hierarchical 4d gaussian compression for progressive volumetric video streaming,

    Z. Zheng, Z. Wu, H. Zhong, Y . Tian, N. Cao, L. Xu, J. Yao, X. Zhang, Q. Hu, and W. Zhang, “4dgcpro: efficient hierarchical 4d gaussian compression for progressive volumetric video streaming,”Advances in Neural Information Processing Systems, vol. 38, pp. 100 860–100 885, 2026

    2026

  38. [38]

    Compgs: Smaller and faster gaussian splatting with vector quantization,

    K. Navaneet, K. Pourahmadi Meibodi, S. Abbasi Koohpayegani, and H. Pirsiavash, “Compgs: Smaller and faster gaussian splatting with vector quantization,” inEuropean Conference on Computer Vision. Springer, 2024, pp. 330–349

    2024

  39. [39]

    Compressed 3d gaussian splatting for accelerated novel view synthesis,

    S. Niedermayr, J. Stumpfegger, and R. Westermann, “Compressed 3d gaussian splatting for accelerated novel view synthesis,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recogni- tion, 2024, pp. 10 349–10 358

    2024

  40. [40]

    Hac++: Towards 100x compression of 3d gaussian splatting,

    Y . Chen, Q. Wu, W. Lin, M. Harandi, and J. Cai, “Hac++: Towards 100x compression of 3d gaussian splatting,”IEEE Transactions on Pattern Analysis and Machine Intelligence, 2025

    2025

  41. [41]

    4dgc: Rate-aware 4d gaussian compression for efficient streamable free-viewpoint video,

    Q. Hu, Z. Zheng, H. Zhong, S. Fu, L. Song, X. Zhang, G. Zhai, and Y . Wang, “4dgc: Rate-aware 4d gaussian compression for efficient streamable free-viewpoint video,” inProceedings of the Computer Vision and Pattern Recognition Conference, 2025, pp. 875–885

    2025

  42. [42]

    Vrvvc: Variable-rate nerf-based volumetric video compres- sion,

    Q. Hu, H. Zhong, Z. Zheng, X. Zhang, Z. Cheng, L. Song, G. Zhai, and Y . Wang, “Vrvvc: Variable-rate nerf-based volumetric video compres- sion,” inProceedings of the AAAI Conference on Artificial Intelligence, vol. 39, no. 4, 2025, pp. 3563–3571

    2025

  43. [43]

    Hac: Hash-grid assisted context for 3d gaussian splatting compression,

    Y . Chen, Q. Wu, W. Lin, M. Harandi, and J. Cai, “Hac: Hash-grid assisted context for 3d gaussian splatting compression,” inEuropean Conference on Computer Vision. Springer, 2024, pp. 422–438

    2024

  44. [44]

    Contextgs: Compact 3d gaussian splatting with anchor level context model,

    Y . Wang, Z. Li, L. Guo, W. Yang, A. C. Kot, and B. Wen, “Contextgs: Compact 3d gaussian splatting with anchor level context model,”Ad- vances in neural information processing systems, vol. 37, pp. 51 532– 51 551, 2024

    2024

  45. [45]

    Sparsesplat: Towards applica- ble feed-forward 3d gaussian splatting with pixel-unaligned prediction,

    Z. Zhang, X. Meng, K. Wu, and W. Ding, “Sparsesplat: Towards applica- ble feed-forward 3d gaussian splatting with pixel-unaligned prediction,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2026, pp. 5049–5058

    2026

  46. [46]

    Dl3dv-10k: A large-scale scene dataset for deep learning-based 3d vision,

    L. Ling, Y . Sheng, Z. Tu, W. Zhao, C. Xin, K. Wan, L. Yu, Q. Guo, Z. Yu, Y . Luet al., “Dl3dv-10k: A large-scale scene dataset for deep learning-based 3d vision,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 22 160–22 169

    2024

  47. [47]

    Stereo magnification: Learning view synthesis using multiplane images,

    T. Zhou, R. Tucker, J. Flynn, G. Fyffe, and N. Snavely, “Stereo magnification: Learning view synthesis using multiplane images,”arXiv preprint arXiv:1805.09817, 2018

    Pith/arXiv arXiv 2018

  48. [48]

    Fast feedfor- ward 3d gaussian splatting compression,

    Y . Chen, Q. Wu, M. Li, W. Lin, M. Harandi, and J. Cai, “Fast feedfor- ward 3d gaussian splatting compression,” inInternational Conference on Learning Representations, vol. 2025, 2025, pp. 74 859–74 872

    2025

  49. [49]

    Compact 3d scene representation via self-organizing gaussian grids,

    W. Morgenstern, F. Barthel, A. Hilsmann, and P. Eisert, “Compact 3d scene representation via self-organizing gaussian grids,” inEuropean conference on computer vision. Springer, 2024, pp. 18–34

    2024

  50. [50]

    Image quality assessment: from error visibility to structural similarity,

    Z. Wang, A. C. Bovik, H. R. Sheikh, and E. P. Simoncelli, “Image quality assessment: from error visibility to structural similarity,”IEEE transactions on image processing, vol. 13, no. 4, pp. 600–612, 2004

    2004

  51. [51]

    The unreasonable effectiveness of deep features as a perceptual metric,

    R. Zhang, P. Isola, A. A. Efros, E. Shechtman, and O. Wang, “The unreasonable effectiveness of deep features as a perceptual metric,” in Proceedings of the IEEE conference on computer vision and pattern recognition, 2018, pp. 586–595

    2018

  52. [52]

    Gans trained by a two time-scale update rule converge to a local nash equilibrium,

    M. Heusel, H. Ramsauer, T. Unterthiner, B. Nessler, and S. Hochreiter, “Gans trained by a two time-scale update rule converge to a local nash equilibrium,”Advances in neural information processing systems, vol. 30, 2017. JOURNAL OF LATEX CLASS FILES, VOL. 18, NO. 9, SEPTEMBER 2020 13

    2017

  53. [53]

    Musiq: Multi- scale image quality transformer,

    J. Ke, Q. Wang, Y . Wang, P. Milanfar, and F. Yang, “Musiq: Multi- scale image quality transformer,” inProceedings of the IEEE/CVF international conference on computer vision, 2021, pp. 5148–5157

    2021

  54. [54]

    Maniqa: Multi-dimension attention network for no-reference image quality assessment,

    S. Yang, T. Wu, S. Shi, S. Lao, Y . Gong, M. Cao, J. Wang, and Y . Yang, “Maniqa: Multi-dimension attention network for no-reference image quality assessment,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2022, pp. 1191–1200

    2022

  55. [55]

    latentsplat: Autoencoding variational gaussians for fast generalizable 3d reconstruc- tion,

    C. Wewer, K. Raj, E. Ilg, B. Schiele, and J. E. Lenssen, “latentsplat: Autoencoding variational gaussians for fast generalizable 3d reconstruc- tion,” inEuropean conference on computer vision. Springer, 2024, pp. 456–473

    2024

  56. [56]

    Vbench: Comprehensive benchmark suite for video generative models,

    Z. Huang, Y . He, J. Yu, F. Zhang, C. Si, Y . Jiang, Y . Zhang, T. Wu, Q. Jin, N. Chanpaisitet al., “Vbench: Comprehensive benchmark suite for video generative models,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 21 807–21 818

    2024

  57. [57]

    Vbench-2.0: Advancing video generation benchmark suite for intrinsic faithfulness,

    D. Zheng, Z. Huang, H. Liu, K. Zou, Y . He, F. Zhang, L. Gu, Y . Zhang, J. He, W.-S. Zhenget al., “Vbench-2.0: Advancing video generation benchmark suite for intrinsic faithfulness,”arXiv preprint arXiv:2503.21755, 2025

    Pith/arXiv arXiv 2025

  58. [58]

    Vbench++: Comprehensive and versatile benchmark suite for video generative models,

    Z. Huang, F. Zhang, X. Xu, Y . He, J. Yu, Z. Dong, Q. Ma, N. Chan- paisit, C. Si, Y . Jianget al., “Vbench++: Comprehensive and versatile benchmark suite for video generative models,”IEEE Transactions on Pattern Analysis and Machine Intelligence, 2025

    2025