Pith. sign in

REVIEW 2 major objections 7 minor 25 references

Denoising Diffusion Probabilistic Model for Point Cloud Compression at Low Bit-Rates

T0 review · 2 major / 7 minor · reviewed 2026-08-15 · deepseek-v4-flash

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

Pith's one-line read A conditional diffusion decoder, fed a quantized PointNet latent code, compresses point cloud geometry at 0.06–0.25 bits per point, beating G-PCC, Draco, D-PCC, and COT-PCC on ShapeNet and ModelNet40.

desk verdict A plausible generative low-rate point cloud compression pipeline with a real new result, but the SOTA claim is under-supported by learned baselines run outside their design range and by missing variance; worth serious refereeing. read the letter →

arxiv 2505.13316 v1 pith:EJ3KJS3F submitted 2025-05-19 cs.CV cs.AIcs.LG

Gabriele Spadaro , Alberto Presta , Jhony H. Giraldo , Marco Grangetto , Wei Hu , Giuseppe Valenzise , Attilio Fiandrotti , Enzo Tartaglione This is my paper · ORCID
classification cs.CVcs.AIcs.LG
keywords pointcloudcompressiondenoisingdiffusionprobabilisticmodellowbit-ratelearnablevectorquantizationNetencodergenerativerate-distortiongeometry
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

DDPM-PCC recasts point cloud compression at low bit rates as conditional generation. A PointNet encoder turns each point cloud into a compact latent vector, a learnable vector quantizer reduces that vector to a handful of codebook indices, and a denoising diffusion decoder generates the point cloud from those indices. The paper reports that, on ShapeNet and ModelNet40 with 2,048-point clouds, this pipeline achieves the best geometry reconstruction among G-PCC, Draco, D-PCC, and COT-PCC at roughly 0.06–0.25 bits per point, across Chamfer distance, point-to-plane PSNR, and Earth mover's distance. It also claims to be the first to use a diffusion probabilistic model for geometry-only point cloud compression. If correct, the pay-off is that bandwidth-constrained 3D applications could send a small generative condition instead of the geometry itself.

What carries the argument

The load-bearing machinery is the conditional denoising diffusion decoder paired with a learnable vector quantizer. The PointNet encoder produces z∈$R^{256}$; z is split into C equal chunks, each chunk is mapped to its nearest vector in a codebook of N=128 learned entries (a VQ-VAE-style stop-gradient quantization), and the quantized code z_hat conditions the reverse diffusion process pθ(x(t-1)|x(t), z_hat). Training minimizes the sum of the noise-prediction diffusion loss and the vector-quantization loss, so the codebook is shaped for the generation task. Varying C changes the rate while keeping the same architecture; each C has its own jointly trained codebook.

What would settle it

Count the decoder and codebook in the bitstream—either as a one-time transmission amortized over a batch or as side information—and re-train D-PCC and COT-PCC with rate-distortion weights aimed at 0.05–0.3 bits per point; if DDPM-PCC no longer dominates all three metrics at matched bit rates, the central claim fails.

Watch

Extended reading notes

Core claim

The central claim, stated on the paper's own terms, is that a diffusion-based decoder makes point cloud geometry compressible to a few codebook indices without collapsing to a coarse voxelized shape. With a codebook of size N=128, the rate is C log2 N bits: C chunks, 7 bits per index, so the reported rates of about 0.06–0.25 bits per point correspond to small values of C. At those rates the method's rate-distortion curves sit below the other codecs on Chamfer distance and Earth mover's distance and above them on point-to-plane PSNR, on both datasets. The paper also notes a ceiling: because the decoder is generative, extra rate beyond what identifies the target sample does not improve quality beyond the model's own reconstruction fidelity.

Load-bearing premise

The comparison assumes that the full cost of compression is the transmitted codebook indices, with the trained diffusion decoder and codebook treated as free shared infrastructure, and that the re-trained learned baselines fairly represent low-rate performance.

Editorial extensions

If this is right

  • Low-bit-rate point cloud delivery can be generative: the decoder synthesizes plausible geometry, so the bit budget pays for the conditioning code, not the coordinates.
  • Rate selection reduces to choosing the number of chunks C; no rate-distortion trade-off parameter λ needs to be tuned per operating point.
  • Reconstruction quality saturates at the generative model's ceiling, so adding bits beyond the code needed to identify the sample yields little or no geometry gain.
  • All three distortion measures (Chamfer distance, point-to-plane PSNR, Earth mover's distance) move in the same direction at low rates, so the reported gain is not an artifact of a single metric.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The stated rate C log2 N ignores any cost for transmitting or storing the trained diffusion decoder and codebook; a user decoding a single cloud would pay that cost, while a user decoding many clouds would amortize it, so the comparison favors the method most for repeated use.
  • If codebook indices are not uniformly distributed, entropy-coding them could lower the true bit rate below C log2 N, potentially widening the gap; measuring the entropy of transmitted indices is a direct testable extension.
  • The authors' suggested side information about input geometry is a natural next step: a hybrid that sends a few explicit coordinates alongside the code could push quality beyond the generative ceiling at higher rates.
  • Re-training D-PCC and COT-PCC with rate-distortion weights tuned for the 0.05–0.3 bits-per-point range is the key stress test; if either closes the gap, the headline claim reflects baseline calibration as much as the method itself.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

2 major / 7 minor

Summary. The manuscript proposes DDPM-PCC, a generative point-cloud geometry codec for very low bit rates. A PointNet encoder maps an input point cloud to a 256-dimensional latent vector; the latent is split into C chunks, each chunk is quantized by nearest-neighbor lookup in a learned codebook of size N=128, and the transmitted bit rate is computed as C*log2(N) bits. The decoder is a conditional denoising diffusion probabilistic model that reconstructs the point cloud from Gaussian noise conditioned on the quantized latent. Experiments on ShapeNet and ModelNet40 with 2048-point clouds compare the method against G-PCC, Draco, D-PCC, and COT-PCC in terms of Chamfer distance, point-to-plane PSNR, and Earth mover's distance versus bits per point, and the authors claim the best rate-distortion performance at low bit rates. Ablations compare the learnable vector quantizer to a fully factorized entropy model and PointNet to PointNet++ encoders. The code is publicly released.

Significance. If the empirical claims are verified, the paper would demonstrate a practical paradigm for very low-rate point cloud geometry compression, which is relevant for bandwidth-constrained applications. The proposed pipeline is simple, the code is released, and the evaluation uses three distortion metrics on two standard datasets, which are good practices. The main value is empirical: the claim of low-rate superiority over learned baselines. However, the evidence in the current form is not sufficient to establish that claim because the learned baselines are evaluated outside their designed operating range and the stochastic decoder is assessed without any uncertainty quantification. The technical contribution is also incremental relative to the prior diffusion model [21], on which the decoder is directly based. The significance is therefore conditional on a substantially strengthened experimental section.

major comments (2)
  1. [Sec. IV-A, IV-B, Fig. 3] The central claim of low-rate superiority over learned baselines is not supported by the baseline protocol. The authors retrain D-PCC [11] and COT-PCC [10] with lambda values intended to force low bit rates and then state in Sec. IV-B that these methods 'were designed to operate at higher bit-rates (typically in the range of 1 to 6 bpp)' and 'do not converge to satisfactory results' in the tested range. Modifying the lambda in a rate-distortion loss is not a controlled way to adapt a codec outside its intended operating range; it can produce collapsed or undertrained latents. Therefore the D-PCC and COT-PCC points in Fig. 3 may be strawmen, and the abstract's and Sec. IV-B's claim of 'best geometry reconstruction at a very low bit rate' is not established. The comparison should include learned baselines that are designed for, or properly adapted to, the low-rate regime, or the paper should restrict its claims to a range where the baselines are representative.
  2. [Fig. 3 and Sec. III-A] The diffusion decoder is stochastic: for a fixed quantized latent, the reverse process described by Eq. (4)-(7) samples from a Gaussian at each of T=200 denoising steps, so repeated decoding of the same latent produces different reconstructions. Fig. 3 reports single RD curves with no error bars, confidence intervals, or per-test statistics. Consequently the statements that the method is 'significantly better' and 'best according to all quality metrics' are not supported statistically; the large reported gaps could partly be stochastic variation, especially at the lowest rates. Please report means and standard deviations over multiple decoded samples per test point, and preferably over multiple training runs, and state the number of test point clouds used for each plotted point.
minor comments (7)
  1. [Sec. III-B] The sentence 'We choose C as an integer multiple of d' is mathematically backwards; the intended meaning is that d is an integer multiple of C. There is also a recurring typo 'chucks' instead of 'chunks' in the same section.
  2. [Sec. III-B] The rate formula C*log2(N) assumes fixed-length coding of the codebook indices. Please clarify whether any entropy coding is applied to the index sequence; if not, state that this is an upper bound on the rate. It would also help to state explicitly that the codebook and the decoder weights are assumed to be available at the decoder as shared side information, since this is a common but implicit convention.
  3. [Sec. IV-A and Fig. 3] The exact rate points used for each RD curve are not listed. For DDPM-PCC, the reader has to infer from the set C in {4,8,16,32,64,128,256}; for the baselines, the quantization parameters are not reported. A table with numerical bpp and metric values for each method would make the comparison reproducible and easier to read than the log-scale plots alone.
  4. [Sec. IV-A] The evaluation protocol for G-PCC and Draco is under-specified: it is not stated how the different rate operating points were selected (e.g., which quantization step sizes or compression levels), nor whether the same zero-mean unit-variance normalization was applied to the input point clouds for all codecs. Please provide these details.
  5. [Sec. IV-C and Fig. 5] The fully factorized (FF) baseline is described only as 'a more traditional end-to-end compression approach [3]'. Please specify the entropy model, the lambda values used, whether the same PointNet encoder and diffusion decoder are used, and whether the FF baseline is trained with the same diffusion loss. In addition, the dashed line labeled 'w/out Compression' is not defined in the text or caption; it appears to be the unquantized upper bound but should be stated explicitly.
  6. [Sec. IV-C and Table I] The table header contains typos: 'Encodings and Decondings times' should be 'Encoding and decoding times'. Also, the CD values in Table I are reported without variance; given that the main result also lacks variance, please add error bars or at least standard deviations for the ablations.
  7. [Sec. I and Sec. V] The statement that 'no attempts have been made to employ such methods specifically for point cloud data' should be verified against the recent literature and scoped carefully; the field is active and the 'first' claim should be supported by an explicit related-work search. Also, the Limitations section does not mention that the decoder generates a fixed number of points (2048 in the experiments), so the method's behavior for variable-size point clouds is not addressed.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the paper's claims are empirical RD comparisons; the diffusion backbone cited from [21] is component reuse, not a load-bearing self-citation.

full rationale

The paper does not derive its headline result from its assumptions by construction. DDPM-PCC is an empirical architecture (Sec. III): a PointNet encoder E_phi produces z, a VQ codebook quantizes z into C indices, the rate is C*log2(N), and a conditional DDPM decodes. Distortion is measured externally as Chamfer distance, point-to-plane PSNR, and EMD on ShapeNet/ModelNet40 test splits (Sec. IV-B). No equation identifies a fitted parameter with the reported prediction; the codebook and encoder are learned normally, and the rate formula is a conventional index count. The only overlapping-author citation is [21] (Luo and Hu; Hu is a co-author here), used as the point-cloud diffusion backbone and training schedule: 'Leveraging the point cloud auto-encoder proposed by [21]' and 'Following the same training procedure adopted in [21]'. This is component reuse of an externally published model, not a self-citation chain that forces the conclusion. The concern that D-PCC and COT-PCC were retrained outside their designed rate range (Sec. IV-A/B: 'they were designed to operate at higher bit-rates (typically in the range of 1 to 6 bpp)') is an experimental fairness/validity issue, not circularity: it weakens the baseline comparison but does not make the proposed method's curve an algebraic consequence of its inputs. The ablation in Sec. IV-C comparing VQ with fully-factorized coding and with the uncompressed generative lower bound is likewise an empirical measurement. Finding: no significant circularity.

Assumptions & free parameters 4 free parameters · 4 assumptions · 0 invented entities

The method has no fitted constants in a derivation; the result is an empirical RD comparison. The load-bearing elements are hand-set hyperparameters (N=128, d=256, C values), the diffusion backbone inherited from [21], and the assumption that the generative decoder's ceiling is high enough for the low-rate regime.

free parameters (4)
  • codebook size N = 128
    Hand-set; log2(N)=7 bits per chunk, directly sets the rate staircase.
  • latent dimension d = 256
    Encoder output dimension, inherited from [21]; determines the maximum chunk count C and thus the highest rate.
  • chunk counts C = {4,8,16,32,64,128,256}
    Rate control knob; each C requires a separately trained codebook and model, so each RD point is a distinct trained model.
  • diffusion steps T and noise schedule = T=200, beta 1e-4 to 0.05
    Taken from [21]; affects generation quality ceiling and decode time.
assumptions (4)
  • domain assumption The diffusion forward and reverse processes are applied independently to each point of the cloud
    Adopted from [21] in Sec. II; treats points as independent samples and ignores inter-point structure that could matter for sharp reconstruction.
  • domain assumption The conditional reverse process p_theta(x(t-1)|x(t), z) is a sufficient decoder for the compression task
    Sec. III-A; the entire reconstruction quality is inherited from this generative model, as acknowledged in Sec. IV-B.
  • domain assumption CD, point-to-plane PSNR, and EMD are valid fidelity measures for this low-rate regime
    Sec. IV-A evaluation follows [10,11]; these metrics may not capture perceptual quality or local density preservation.
  • standard math The DDPM noise-prediction loss (Eq. 6) trains a valid conditional generator
    Standard result from Ho et al. [18]; the paper uses it without modification.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Denoising Diffusion Probabilistic Model for Point Cloud Compression at Low Bit-Rates." pith.science (2026). https://pith.science/paper/EJ3KJS3F

@misc{pith2026250513316,
  author       = {Pith},
  title        = {Pith review of: Denoising Diffusion Probabilistic Model for Point Cloud Compression at Low Bit-Rates},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/EJ3KJS3F}},
  note         = {Machine review of arXiv:2505.13316}
}
read the original abstract

Efficient compression of low-bit-rate point clouds is critical for bandwidth-constrained applications. However, existing techniques mainly focus on high-fidelity reconstruction, requiring many bits for compression. This paper proposes a "Denoising Diffusion Probabilistic Model" (DDPM) architecture for point cloud compression (DDPM-PCC) at low bit-rates. A PointNet encoder produces the condition vector for the generation, which is then quantized via a learnable vector quantizer. This configuration allows to achieve a low bitrates while preserving quality. Experiments on ShapeNet and ModelNet40 show improved rate-distortion at low rates compared to standardized and state-of-the-art approaches. We publicly released the code at https://github.com/EIDOSLAB/DDPM-PCC.

Figures

Figures reproduced from arXiv: 2505.13316 by the authors.

Figure 1
Figure 1. Low bit-rates PCC and compression artifacts. [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Overview of the proposed method. Here for simplicity, [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Quantitative results on ShapeNet (first row) and ModelNet40 (second row). The y-axes are plotted in logarithmic scale. [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Qualitative results on ModelNet (first two rows) and ShapeNet (last two rows). We compare the reconstruction results [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]
Figure 5
Figure 5. Figure 5: Comparison between a learnable vector quantize ap [PITH_FULL_IMAGE:figures/full_fig_p006_5.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

25 extracted references · 23 canonical work pages

  1. [21]

    Diffusion probabilistic models for 3d point cloud generation,

    Shitong Luo and Wei Hu, “Diffusion probabilistic models for 3d point cloud generation,” in CVPR, 2021

    work page 2021

  2. [11]

    Density-preserving deep point cloud compression,

    Yun He, Xinlin Ren, Danhang Tang, Yinda Zhang, Xiangyang Xue, and Yanwei Fu, “Density-preserving deep point cloud compression,” in CVPR, 2022

    work page 2022

  3. [10]

    Point cloud compression via constrained optimal transport,

    Zezeng Li, Weimin Wang, Ziliang Wang, and Na Lei, “Point cloud compression via constrained optimal transport,” in ICME, 2024

    work page 2024

  4. [1]

    An overview of ongoing point cloud compression standardization activities: video-based (v-pcc) and geometry-based (g- pcc),

    D Graziosi, O Nakagami, S Kuma, A Zaghetto, T Suzuki, and A Tabatabai, “An overview of ongoing point cloud compression standardization activities: video-based (v-pcc) and geometry-based (g- pcc),” APSIPA Transactions on Signal and Information Processing , 2020

    work page 2020

  5. [2]

    Google/draco: a library for compressing and de- compressing 3d geometric meshes and point clouds,

    Frank Galligan, Michael Hemmer, Ondrej Stava, Fan Zhang, and Jamieson Brettle, “Google/draco: a library for compressing and de- compressing 3d geometric meshes and point clouds,” https://github.com/ google/draco, 2018

    work page 2018

  6. [3]

    End-to-end optimized image compression,

    Johannes Ball ´e et al., “End-to-end optimized image compression,” in ICLR, 2017

    work page 2017

  7. [4]

    Learned image compression with mixed transformer-cnn architectures,

    Jinming Liu, Heming Sun, and Jiro Katto, “Learned image compression with mixed transformer-cnn architectures,” in CVPR, 2023

    work page 2023

  8. [5]

    Learning convolutional transforms for lossy point cloud geometry compression,

    Maurice Quach, Giuseppe Valenzise, and Frederic Dufaux, “Learning convolutional transforms for lossy point cloud geometry compression,” in ICIP. IEEE, 2019

    work page 2019

Show all 25 references
  1. [6]

    Deep autoencoder-based lossy geometry compression for point clouds,

    Wei Yan, Shan Liu, Thomas H Li, Zhu Li, Ge Li, et al., “Deep autoencoder-based lossy geometry compression for point clouds,” arXiv preprint arXiv:1905.03691, 2019

    1905 arXiv

  2. [7]

    Pointnet: Deep learning on point sets for 3d classification and segmen- tation,

    Charles Ruizhongtai Qi, Hao Su, Kaichun Mo, and Leonidas J. Guibas, “Pointnet: Deep learning on point sets for 3d classification and segmen- tation,” in CVPR, 2017

    2017

  3. [8]

    Multi-space point geometry compression with progressive relation- aware transformer,

    Wenxiang Shen, BaoYe Zhang, Hao Xu, XiaoHan Li, and Jun Wu, “Multi-space point geometry compression with progressive relation- aware transformer,” IEEE TMM , 2024

    2024

  4. [9]

    Pointnet++: Deep hierarchical feature learning on point sets in a metric space,

    Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas, “Pointnet++: Deep hierarchical feature learning on point sets in a metric space,” NeurIPS, 2017

    2017

  5. [12]

    V oxelcontext-net: An octree based framework for point cloud compression,

    Zizheng Que, Guo Lu, and Dong Xu, “V oxelcontext-net: An octree based framework for point cloud compression,” in CVPR, 2021

    2021

  6. [13]

    Octattention: Octree-based large-scale contexts model for point cloud compression,

    Chunyang Fu, Ge Li, Rui Song, Wei Gao, and Shan Liu, “Octattention: Octree-based large-scale contexts model for point cloud compression,” in AAAI, 2022

    2022

  7. [14]

    Geometric 3d point cloud compression,

    Vicente Morell, Sergio Orts, Miguel Cazorla, and Jose Garcia- Rodriguez, “Geometric 3d point cloud compression,” Pattern Recogni- tion Letters , 2014

    2014

  8. [15]

    Improved deep point cloud geometry compression,

    Maurice Quach, Giuseppe Valenzise, and Frederic Dufaux, “Improved deep point cloud geometry compression,” in IEEE MMS-W) . IEEE, 2020

    2020

  9. [16]

    Lossy point cloud geometry compression via end-to-end learning,

    Jianqiang Wang, Hao Zhu, Haojie Liu, and Zhan Ma, “Lossy point cloud geometry compression via end-to-end learning,” IEEE TCSVT , 2021

    2021

  10. [17]

    Multiscale point cloud geometry compression,

    Jianqiang Wang, Dandan Ding, Zhu Li, and Zhan Ma, “Multiscale point cloud geometry compression,” in DCC, 2021

    2021

  11. [18]

    Denoising diffusion probabilistic models,

    Jonathan Ho, Ajay Jain, and Pieter Abbeel, “Denoising diffusion probabilistic models,” in NeurIPS, 2020

    2020

  12. [19]

    Lossy image compression with conditional diffusion models,

    Ruihan Yang and Stephan Mandt, “Lossy image compression with conditional diffusion models,” NeurIPS, 2023

    2023

  13. [20]

    Towards image compression with perfect realism at ultra- low bitrates,

    Marlene Careil, Matthew J Muckley, Jakob Verbeek, and St ´ephane Lathuili`ere, “Towards image compression with perfect realism at ultra- low bitrates,” in ICLR, 2023

    2023

  14. [22]

    Neural discrete representation learning,

    A ¨aron van den Oord, Oriol Vinyals, and Koray Kavukcuoglu, “Neural discrete representation learning,” in NeurIPS, 2017

    2017

  15. [23]

    ShapeNet: An Information-Rich 3D Model Repository,

    Angel X. Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, Jianxiong Xiao, Li Yi, and Fisher Yu, “ShapeNet: An Information-Rich 3D Model Repository,” Tech. Rep. arXiv:1512.03012 [cs.GR], Stanfor...

    2015 arXiv

  16. [24]

    3d shapenets for 2.5d object recognition and next-best-view prediction,

    Zhirong Wu, Shuran Song, Aditya Khosla, Xiaoou Tang, and Jianxiong Xiao, “3d shapenets for 2.5d object recognition and next-best-view prediction,” in CVPR, 2015

    2015

  17. [25]

    Adam: A method for stochastic optimization,

    Diederik P Kingma and Jimmy Ba, “Adam: A method for stochastic optimization,” in ICLR, 2015. 6

    2015

Pith tools

Reviewed August 15, 2026 · model on record in the stance chip above.