{"id":"c9889c25-05e1-47b4-8ed7-b95250b66af5","arxiv_id":"2505.13316","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A diffusion model conditioned on a vector-quantized PointNet latent reconstructs point clouds at 0.06-0.25 bits per point with better reported quality than G-PCC, Draco, D-PCC, and COT-PCC on ShapeNet and ModelNet40.","lead":"The paper compresses 3D point clouds at very low bit rates by encoding a shape into a small code, then using a diffusion model to regenerate the cloud from that code. It reports large quality gains over standard codecs like G-PCC and Draco below 0.25 bits per point on ShapeNet and ModelNet40, which could matter for bandwidth-limited 3D streaming.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The low-bit-rate SOTA comparison is undercut by learned baselines evaluated outside their designed rate range; retraining with a low-rate lambda is not a controlled test, and no variance is reported.","rationale":"Good-faith reading: the paper proposes a clear pipeline and releases code; the central claim is plausible and no mathematical contradiction appears. The load-bearing condition is that the baselines are representative at the tested rates. The authors' own concession in Sec. IV-B makes this condition insecure: D-PCC and COT-PCC are pushed to low rates with a lambda sweep, and their failure to converge is reported as evidence of superiority. That is a protocol problem rather than an internal inconsistency. The reader identified the same baseline-regime concern, but I disagree that the shared decoder and codebook cost is a serious flaw; it is standard in learned compression. The missing error bars matter because the decoder is generative and stochastic, so a single RD curve cannot establish dominance. With code available, the retraining check is feasible. Since no fundamental error was found and the concern is addressable by re-evaluation, I keep the existing CONDITIONAL verdict, meaning no change from the reader's recommendation.","tokens_in":8100,"tokens_out":6472,"duration_ms":68846,"concrete_test":"Re-run the comparison at matched low bpp with (1) D-PCC and COT-PCC trained from their official recipes but with lambda chosen by sweeping to target 0.06-0.25 bpp, full convergence, and 3 seeds each, and (2) DDPM-PCC evaluated over 3 or more sampling seeds, reporting mean plus/minus standard deviation for Chamfer distance, point-to-plane PSNR, and EMD. If the retrained baselines close the gap or overlap within error bars, the claim of best geometry reconstruction is not supported; if DDPM-PCC remains separated by more than the variance, the concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is the low-rate dominance shown in Fig. 3. For that claim to hold, G-PCC, Draco, D-PCC, and COT-PCC must be representative at 0.06-0.25 bpp. The weakest point is the learned baselines. In Sec. IV-A the authors re-train D-PCC and COT-PCC using lambda values to force low rates, and Sec. IV-B concedes that these methods were designed to operate at higher bit-rates (typically 1 to 6 bpp) and do not converge to satisfactory results in the tested range. Simply increasing lambda in a rate-distortion loss does not exercise the model's designed operating range; it can produce collapsed or undertrained latents. Thus the plotted D-PCC and COT-PCC points may be strawmen, making the headline claim an artifact of the baseline protocol rather than of the method. A supporting weakness is that DDPM-PCC's diffusion decoder is stochastic, but Fig. 3 reports no error bars, so we cannot tell whether the large reported gap is significant. I do not regard excluding the shared decoder and codebook from the bitrate as unfair: learned codecs standardly assume pre-installed decoder parameters, so the reader's second premise is less convincing than the baseline-regime concern.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","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.","tokens_in":8255,"tokens_out":10602,"duration_ms":111776,"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":[{"comment":"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.","section":"Sec. IV-A, IV-B, Fig. 3"},{"comment":"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.","section":"Fig. 3 and Sec. III-A"}],"minor_comments":[{"comment":"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.","section":"Sec. III-B"},{"comment":"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.","section":"Sec. III-B"},{"comment":"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.","section":"Sec. IV-A and Fig. 3"},{"comment":"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.","section":"Sec. IV-A"},{"comment":"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.","section":"Sec. IV-C and Fig. 5"},{"comment":"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.","section":"Sec. IV-C and Table I"},{"comment":"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.","section":"Sec. I and Sec. V"}],"recommendation":"major_revision","confidential_remarks":"The decoder is directly taken from prior work [21], which shares an author with this submission, and the contribution section does not explicitly separate reused components from new ones. I do not see evidence of misconduct, but the authors should be asked to state clearly which parts of the pipeline are inherited from [21] and to frame the novelty accordingly. The paper's central empirical claim depends on the baseline protocol and the lack of variance reporting; both are fixable with additional experiments, so I would not reject on that basis alone."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague — this paper is a genuine attempt at generative point cloud compression and deserves a referee's time, not a desk reject. The new piece is the combination: a PointNet encoder produces a 256-d latent, a learnable VQ codebook turns it into C×log2(128) bits, and a conditional DDPM decoder reconstructs the point cloud. Evaluated at roughly 0.06–0.25 bpp on ShapeNet and ModelNet40, it reports better Chamfer distance, point-to-plane PSNR, and EMD than G-PCC, Draco, D-PCC, and COT-PCC. The method is clearly described, the ablation against a fully-factorized learned codec is informative, and the code is public. The authors also state plainly that their quality is capped by the generative model, which is the right caveat.\n\nThe weak spots are experimental, not mathematical. The headline comparison rests on D-PCC and COT-PCC re-trained with lambda values meant to force low rates. The paper itself says those methods were designed for 1–6 bpp and do not converge in the tested range. Tuning one scalar in an RD loss is not a controlled way to move a learned codec outside its operating regime, so those two curves in Fig. 3 may be strawmen. I do not regard this as fatal, because G-PCC and Draco are honest baselines and beating them at low rate is already a meaningful result, but the 'state-of-the-art' claim needs better baselines or a careful discussion of what low-rate retraining does to those models. The second issue is that the decoder is stochastic and Fig. 3 has no error bars, so the size of the reported gains is uncertain. The bitrate definition that excludes shared decoder/codebook cost is standard practice for learned codecs; I would not flag it as a flaw.\n\nThe self-citation of the Lou–Hu diffusion model is component reuse, not circular argument; that prior work generated point clouds without quantization or compression measurements. The math is consistent as far as I checked, and the authors' limitation statements match what the experiments show.\n\nThis is not a breakthrough yet, but it is a well-scoped contribution to low-bitrate point-cloud geometry coding. A serious referee should push for variance reporting, fairer low-rate learned baselines, and decoder-side cost or size analysis. I recommend sending it to peer review.","headline":"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.","tokens_in":8917,"tokens_out":1989,"would_cite":true,"duration_ms":20516,"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":"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.","keywords":["point cloud compression","denoising diffusion probabilistic model","low bit-rate compression","learnable vector quantization","PointNet encoder","generative compression","rate-distortion","geometry compression"],"falsifier":"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.","tokens_in":7808,"feed_emoji":"🧊","tokens_out":8395,"duration_ms":79524,"temperature":0.7,"pith_summary":"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.","feed_headline":"A diffusion decoder beats standard codecs below 0.25 bits per point","feed_subtitle":"A few codebook indices, fed to a diffusion decoder, beat G-PCC and Draco at very low bpp.","key_machinery":"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.","core_discovery":"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.","pith_inferences":["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."],"forward_implications":["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."],"supporting_citations":[{"why":"Supplies the DDPM noise-prediction objective that the conditional decoder adapts.","marker":"[18]"},{"why":"Provides the diffusion point-cloud generation backbone and training schedule the method builds on.","marker":"[21]"},{"why":"PointNet encoder that produces the latent condition vector.","marker":"[7]"},{"why":"Introduces the learnable vector quantizer and its stop-gradient loss for latent compression.","marker":"[22]"},{"why":"G-PCC standard codec baseline in the rate-distortion comparison.","marker":"[1]"},{"why":"Draco baseline in the rate-distortion comparison.","marker":"[2]"},{"why":"D-PCC learned baseline re-trained for the low-rate comparison.","marker":"[11]"},{"why":"COT-PCC learned baseline re-trained for the low-rate comparison.","marker":"[10]"}],"fun_headline_variants":["Diffusion decoder compresses point clouds to a few codebook indices","At 0.06–0.25 bpp, diffusion beats G-PCC and Draco on geometry","Generative compression: Point clouds from 7-bit codebook entries","Low-bit-rate point clouds: Diffusion model beats standard codecs","Few bits per point: Diffusion model outperforms traditional codecs"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"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.","fun_headline_variants_meta":{"raw":{"variants":["Diffusion decoder compresses point clouds to a few codebook indices","At 0.06–0.25 bpp, diffusion beats G-PCC and Draco on geometry","Generative compression: Point clouds from 7-bit codebook entries","Low-bit-rate point clouds: Diffusion model beats standard codecs","Few bits per point: Diffusion model outperforms traditional codecs"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000196,"raw_usage":{"total_tokens":1306,"prompt_tokens":839,"completion_tokens":467,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":455,"completion_tokens_details":{"reasoning_tokens":370}},"tokens_in":455,"tokens_out":467,"duration_ms":4475,"temperature":1.0,"reasoning_tokens":370,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T20:15:53.077628+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"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.","supporting_citations":[{"cited_title":"Denoising diffusion probabilistic models,","cited_arxiv_id":null,"evidence_quote":"Supplies the DDPM noise-prediction objective that the conditional decoder adapts."},{"cited_title":"Diffusion probabilistic models for 3d point cloud generation,","cited_arxiv_id":null,"evidence_quote":"Provides the diffusion point-cloud generation backbone and training schedule the method builds on."},{"cited_title":"Pointnet: Deep learning on point sets for 3d classification and segmen- tation,","cited_arxiv_id":null,"evidence_quote":"PointNet encoder that produces the latent condition vector."},{"cited_title":"Neural discrete representation learning,","cited_arxiv_id":null,"evidence_quote":"Introduces the learnable vector quantizer and its stop-gradient loss for latent compression."},{"cited_title":"An overview of ongoing point cloud compression standardization activities: video-based (v-pcc) and geometry-based (g- pcc),","cited_arxiv_id":null,"evidence_quote":"G-PCC standard codec baseline in the rate-distortion comparison."},{"cited_title":"Google/draco: a library for compressing and de- compressing 3d geometric meshes and point clouds,","cited_arxiv_id":null,"evidence_quote":"Draco baseline in the rate-distortion comparison."},{"cited_title":"Density-preserving deep point cloud compression,","cited_arxiv_id":null,"evidence_quote":"D-PCC learned baseline re-trained for the low-rate comparison."},{"cited_title":"Point cloud compression via constrained optimal transport,","cited_arxiv_id":null,"evidence_quote":"COT-PCC learned baseline re-trained for the low-rate comparison."}],"review_version":1}