{"id":"558fbb19-4696-4634-9efd-22ade516c601","arxiv_id":"2506.17983","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"LVPNet reports lower bits-per-pixel than prior learned lossless codecs by conditioning pixel predictions on a global multi-scale latent variable with a quantization compensation module.","lead":"A new neural compression method, LVPNet, uses a global latent description of the whole image to predict pixel probabilities, then encodes those probabilities losslessly. On chest X-ray and other image sets it reports lower bits-per-pixel than previous learned lossless codecs while keeping decode fast.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The reported 2.65 BPP on Chest X-ray may be a test-set artifact: the loss in Eq. (5) omits the latent bitstream rate, so total BPP is not what is optimized.","rationale":"The reader's weakest assumption is exactly the missing rate term for z and the unverified stability of the z bitstream. This is the most load-bearing issue because the headline gain (2.65 vs. 2.74 BPP on Chest X-ray, and similar margins on other datasets) is small, and the training objective in Eq. (5) does not include the cost of storing z. The paper's Sec. 2.1 claim that the z bitstream is stable is not supported by any measurement; latent entropy depends on feature magnitudes and Huffman coding, not just the sampling ratio. There are additional concerns (entropy coding not shown to be lossless, r tuned on the test set, no code release), but they reinforce the same central uncertainty: the reported total BPP may not be reproducible or may not reflect the method as described. I agree with the reader's identification of the weakest assumption and recommend keeping the CONDITIONAL verdict, pending code release and a separated rate measurement.","tokens_in":7715,"tokens_out":1481,"duration_ms":12710,"concrete_test":"Recompute the total BPP with an explicit measurement: for a held-out Chest X-ray split, run the released code, record rate(z) = Huffman-coded bytes of z and rate(p_theta) = arithmetic-coded bytes of the pixel predictions separately, and verify that their sum equals the reported 2.65 BPP. Then retrain or re-run with r in {0.05, 0.10, 0.15, 0.20, 0.30} on a validation split and check whether the BPP-vs-r curve matches Fig. 4; if the optimal r on validation differs from 0.15, the headline number is a test-set artifact.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim is that LVPNet achieves new state-of-the-art lossless compression (e.g., 2.65 BPP on Chest X-ray vs. 2.74 for ArIB-BPS). However, Eq. (5) trains only the cross-entropy of the pixel predictions p_theta(x|z); it contains no rate term for the quantized latent z. The paper asserts (Sec. 2.1) that the z bitstream 'remains relatively stable due to the fixed sampling ratio r', but the entropy of z is determined by the feature magnitudes produced by GMSM and by the Huffman coder, not by the sampling ratio alone. If the entropy of z varies across images or datasets, then total BPP = rate(z) + rate(p_theta(x|z)) is not controlled by the optimized loss, and tuning r=0.15 on the test set (Fig. 4) can overfit the loss to the reported numbers. Moreover, no measured rate(z) is reported anywhere, so the claim that z is stable is unverified. The assertion of 'lossless' reconstruction also lacks a verification step: no code or bit-exact check is provided, and the entropy coder's correctness is not demonstrated. Because the reported gain over ArIB-BPS is small (about 0.09 BPP), an unaccounted or varying latent rate could erase the advantage.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes LVPNet, a learned lossless compression framework for medical images. A global multi-scale sensing module (GMSM) extracts a downsampled latent from the full image, quantizes it, entropy-codes it, and a quantization compensation module (QCM) plus a pixel-prediction network produce pixel-wise probability estimates that are arithmetic-coded. The authors report BPP values on Chest X-ray, CIFAR10, ImageNet32, and ImageNet64 for both 'dataset' and 'single-image' settings, claiming state-of-the-art compression efficiency, e.g., 2.65 BPP on Chest X-ray versus 2.74 for ArIB-BPS, with competitive inference time. Ablations on Chest X-ray support the contributions of GMSM and QCM.","tokens_in":8009,"tokens_out":2773,"duration_ms":23709,"significance":"If the reported BPP numbers are reproducible and the pipeline truly encodes losslessly, LVPNet would be a meaningful new data point in learned lossless image compression: it combines a global latent with prediction-driven pixel coding and demonstrates a consistent margin over ArIB-BPS across four benchmarks. The manuscript ships a public code repository, which strengthens the potential for verification. However, the significance is tempered by the fact that the core evaluation is a set of BPP tables with no lossless bitstream verification, no reported latent-rate breakdown, and limited baseline variance information, so the strength of the empirical claim is currently uncertain.","major_comments":[{"comment":"The training loss minimizes only the negative log-likelihood of pixels under p_theta(x|z). It contains no term for the entropy of the quantized latent z, yet the total bitstream is rate(z) + rate(p_theta(x|z)). The paper assumes in Sec. 2.1 that the z bitstream 'remains relatively stable due to the fixed sampling ratio r', but the entropy of z is determined by the feature magnitudes produced by GMSM and by the Huffman coder, not by the sample count alone. Since no measured rate(z) is reported anywhere, the claimed total BPP values (e.g., 2.65 for Chest X-ray in Table 1) are not shown to be controlled by the optimized loss, and a varying latent rate could erase the margin over ArIB-BPS.","section":"Sec. 2.5, Eq. (5)"},{"comment":"The sampling rate r is selected on the Chest X-ray test set. Figure 4 varies r and reports the resulting BPP on Chest X-ray, and the same r=0.15 is then used for the other datasets. This is a test-set-tuned hyperparameter, so the reported gains on Chest X-ray are optimistic, and the paper provides no evidence that r=0.15 transfers to CIFAR10, ImageNet32, or ImageNet64. A validation-set sweep or a sensitivity table on the other datasets is needed for the cross-dataset claim.","section":"Sec. 3.3, Fig. 4"},{"comment":"There is no verification that the full pipeline is lossless or that the claimed bitstream sizes are achievable. The paper states that Huffman coding is used for z and arithmetic coding for predicted probabilities, but no bit-exact decode test, code output verification, or measured table of actual coded lengths is reported. Since 'lossless' is the central claim, at least a reported encode-decode round-trip check (ideally on all test images) should accompany the BPP tables.","section":"Sec. 2.3 and Sec. 3.3"},{"comment":"Baseline variance is absent: only LVPNet has error bars in Tables 1-3, and it is unclear how many runs or training seeds they cover. Given that the headline gain over ArIB-BPS is about 0.09 BPP on Chest X-ray (2.65 vs. 2.74), it is important to know the standard deviation of the baseline and whether the baselines were re-run in the same experimental protocol or taken from their original papers.","section":"Table 1"}],"minor_comments":[{"comment":"Reference [4] appears as '?' in the text; this citation is missing or broken.","section":"Sec. 1"},{"comment":"The equations use both p_theta(x|z) and x-hat notation inconsistently: Eqs. (4)-(6) write p_theta(hat{x}_ij | Q(G(x))) but the text around them refers to true pixel values x_ij. Please clarify the notation for the target random variable.","section":"Sec. 2.5"},{"comment":"The definition of the delta-marked 'initial bit count' for iVPF, LBB, and iFlow is given only in the table caption. The meaning of 'summing the dataset's compression performance and the initial bit count' should be explained in the body text, since it affects the interpretability of those entries.","section":"Table 2"},{"comment":"The straight-through gradient in Eq. (3) uses the notation partial L/partial z but the loss in Eq. (5) is not actually a function of the latent rate; this should be made explicit so readers do not infer that the model optimizes rate(z).","section":"Sec. 2.3"},{"comment":"The training protocol omits details about dataset splits, number of training epochs, data augmentation, and how the 'single-image' models are trained. A few sentences on these details would help reproducibility.","section":"Sec. 3.1"}],"recommendation":"major_revision","confidential_remarks":"The strengths of the paper are its clear architecture narrative and the provision of code, but the main empirical claim is not yet fully supported: the trained objective does not include the latent rate, no actual lossless encode/decode verification is reported, and the sampling-rate hyperparameter is tuned on the test set. The requested additions (latent-rate measurement, round-trip verification, validation-based hyperparameter selection or sensitivity analysis, baseline variance) are load-bearing for the BPP claims and are within the scope of a revision. I do not see grounds for rejection; the issues are fixable with additional experiments rather than conceptual ones."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nQuick take on LVPNet: it's a modest but plausible step beyond ArIB-BPS. The two new pieces—GMSM's multi-scale feature aggregation and QCM's quantization-error compensation—are simple, clearly motivated, and the ablation shows each buys a real gain. The headline result (2.65 vs 2.74 BPP on Chest X-ray, smaller margins on natural images) is consistent across datasets, and inference speed is competitive. If the numbers hold, this is useful for storage-heavy medical imaging workflows.\n\nWhat's genuinely new is the combination of those modules inside the ArIB-BPS prediction-driven latent framework. The paper doesn't oversell it; it frames itself as an architecture tweak, and that's fair.\n\nWhere I'd push back: the training loss (Eq. 5) only minimizes the cross-entropy of pixel predictions. There is no term for the bits spent on the latent z. The paper asserts that z's bitstream stays \"relatively stable\" because the sampling ratio r is fixed, but that's not demonstrated—entropy depends on the learned feature distribution, not just the sample count. They never report rate(z), and since total BPP must include those bits, an unmeasured latent rate could easily erase the 0.09 BPP advantage. The tuning of r on the Chest X-ray test set (Fig. 4) is a mild form of test-set selection; it's one hyperparameter, but it should be flagged. I'd also want an actual lossless round-trip verification and baseline error bars. The paper mentions code but the full text doesn't give a working link.\n\nThese are not fatal flaws—the architecture is coherent and the results are internally consistent—but they are exactly what a referee should probe. The fix is straightforward: report the latent bitrate separately, run a sensitivity sweep on r with a validation split, and state explicitly that a bit-exact reconstruction was confirmed.\n\nWho gains from this: anyone doing learned lossless compression on medical images, especially as a baseline for future work. It deserves a serious referee; the concerns are addressable with minor additions.\n\nRecommendation: send it to peer review. I wouldn't cite it in my own work unless I were actively benchmarking in this exact line, but it's a legitimate empirical contribution.","headline":"Solid incremental extension of ArIB-BPS with two sensible modules and plausible BPP gains, but missing a measured latent-bitrate term and a lossless round-trip check.","tokens_in":8526,"tokens_out":3365,"would_cite":false,"duration_ms":36556,"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":"LVPNet claims that global latent variables, extracted by a multi-scale sensing module and refined by quantization compensation, push lossless medical image compression to 2.65 bits per pixel on Chest X-ray, beating prior learned codecs…","keywords":["lossless image compression","medical imaging","latent variable model","pixel prediction","quantization compensation","global multi-scale sensing","entropy coding","chest X-ray"],"falsifier":"Compute the actual Huffman-coded length of $z$ per image at $r=0.15$ on a held-out set of Chest X-ray images with varying pathology and image size, and add it to the probability bitstream; if this latent length varies by more than about 0.05 BPP across the set, the fixed-sampling-ratio assumption fails to control total bitrate, and the reported BPP depends on the entropy of $z$ rather than on the trained cross-entropy loss.","tokens_in":7539,"feed_emoji":"🩻","tokens_out":6668,"duration_ms":63606,"temperature":0.7,"pith_summary":"LVPNet is a learned lossless codec for medical images that predicts each pixel's probability from a compact set of global latent variables, then entropy-codes those probabilities. The paper claims this design beats prior autoregressive-and-latent models such as ArIB-BPS on Chest X-ray (2.65 vs 2.74 bits per pixel for dataset compression) and also improves on natural image benchmarks, while keeping encode/decode time competitive. Its two named mechanisms are a Global Multi-scale Sensing Module (GMSM), which aggregates multi-level features into the latent code, and a Quantization Compensation Module (QCM), which models quantization error so the predicted probabilities are more accurate. The contribution is a single framework for both whole-dataset and single-image lossless compression, and the reason to care is that medical-image storage is volume-heavy and lossless requirements rule out lossy shortcuts.","feed_headline":"Learned codec cuts medical image lossless rate to 2.65 BPP","feed_subtitle":"Global latent prediction plus quantization compensation beats prior learned codecs on X-ray and natural images.","key_machinery":"Global Multi-scale Sensing Module (GMSM): a CNN-based encoder that keeps a constant feature volume across hierarchical layers via skip connections $f^{t+1} = \\mathrm{Conv}(f^t) + \\mathrm{Pool}(f^t)$, then downsamples with $1\\times1$ convolutions to a compressed coefficient map $y$ at a sampling ratio $r$. $y$ is floor-quantized to $z$ with a straight-through gradient approximation, Huffman-coded into the stored latent bitstream, and dequantized; the Quantization Compensation Module (QCM) adds non-negative residual corrections to the dequantized features so the final pixel-probability model sees features closer to the original coefficients. Together these modules increase the information content of the latent variables and reduce the cross-entropy term that dominates the total rate.","core_discovery":"In the paper's own terms, LVPNet establishes that lossless compression of medical images can be improved by replacing sub-image autoregression with prediction conditioned on global latent variables. The quantized latent $z$ is entropy-coded once, and the bulk of the bitstream is the arithmetic-coded pixel probabilities $p_\\theta(x|z)$ produced by a probability prediction module; a fixed sampling ratio $r$ makes the latent contribution to total bitrate approximately constant, so optimizing the pixel cross-entropy loss alone is claimed sufficient to lower total bits per pixel. On Chest X-ray the method reports 2.65 BPP dataset-level and 2.72 BPP single-image, below the previous best learned codecs, with encode/decode times at 143.63 and 184.06 ms/sample respectively.","pith_inferences":["A testable extension is to measure the Huffman-coded length of $z$ across images and datasets; if the latent bitstream varies by more than a small fraction of a BPP at fixed $r$, the paper's assumption that total bitrate is controlled by the pixel-prediction loss is violated.","The sampling ratio $r=0.15$ is chosen on Chest X-ray (Fig. 4) and may not transfer to higher-entropy images; adapting $r$ per image based on a latent-entropy estimate would be a natural follow-up.","Because the loss is pure cross-entropy with no prior on $z$, coupling LVPNet with a learned entropy model on $z$ could push total BPP further; the paper does not report this.","The single-image numbers for iVPF, LBB, and iFlow include large initial-bit overheads marked with $\\Delta$ in Table 2, so a fair comparison of true single-image usefulness depends on whether model parameters are counted; LVPNet's own single-image numbers assume a shared trained model."],"forward_implications":["If correct, LVPNet sets a new state of the art on Chest X-ray lossless compression, with dataset BPP 2.65 versus 2.74 for ArIB-BPS and 2.76 for LBB.","The framework works for both whole-dataset and single-image compression, so an imaging site could use one codec for both archive-level and per-study storage.","Because the pixel predictor is conditioned on global latents rather than a raster-scan autoregressive context, decoding is parallel across pixels, which keeps inference competitive with the fastest prior method.","Quantization compensation is a drop-in ingredient: adding QCM to a plain CNN sampler reduces Chest X-ray BPP from 2.833 to 2.757 in the paper's ablation, suggesting similar gains for other entropy-coded latent codecs.","The method extends beyond medical images: reported gains on CIFAR10, ImageNet32, and ImageNet64 are improvements over prior learned codecs, so the design appears to generalize beyond radiology."],"supporting_citations":[{"why":"Baseline autoregressive-latent codec whose even sub-image latent split LVPNet aims to fix; its reported numbers are the main comparison throughout.","marker":"[16]"},{"why":"Flow-based bits-back baseline that defines the prior state of the art on dataset compression.","marker":"[23]"},{"why":"Invertible-flow codec used as the rate and speed comparison in Table 3.","marker":"[25]"},{"why":"Residual skip connections that GMSM adapts to aggregate multi-level features and keep feature volume constant.","marker":"[39]"},{"why":"Huffman coding used to entropy-encode the quantized latent variables.","marker":"[40]"},{"why":"Arithmetic coding used to encode the predicted pixel probabilities into the main bitstream.","marker":"[41]"},{"why":"ChestX-ray8 dataset that supplies the medical-image benchmark and drives the choice of sampling ratio.","marker":"[44]"},{"why":"Traditional lossless codec included as a non-learned baseline in Tables 1 and 2.","marker":"[42]"}],"fun_headline_variants":["Global latent prediction cuts lossless medical bitrate to 2.65 BPP","LVPNet predicts pixels globally, cutting lossless medical BPP to 2.65","Quantization compensation key to LVPNet's 2.65 BPP lossless medical win","New lossless codec for medical images: 2.65 BPP with global latent model"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method assumes the storage cost of the quantized latent variables stays roughly constant across images because the sampling ratio is fixed, so the training loss only needs to minimize pixel-prediction error.","fun_headline_variants_meta":{"raw":{"variants":["Global latent prediction cuts lossless medical bitrate to 2.65 BPP","LVPNet predicts pixels globally, cutting lossless medical BPP to 2.65","Quantization compensation key to LVPNet's 2.65 BPP lossless medical win","New lossless codec for medical images: 2.65 BPP with global latent model"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000649,"raw_usage":{"total_tokens":2967,"prompt_tokens":923,"completion_tokens":2044,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":539,"completion_tokens_details":{"reasoning_tokens":1961}},"tokens_in":539,"tokens_out":2044,"duration_ms":15619,"temperature":1.0,"reasoning_tokens":1961,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T23:23:32.621877+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compute the actual Huffman-coded length of $z$ per image at $r=0.15$ on a held-out set of Chest X-ray images with varying pathology and image size, and add it to the probability bitstream; if this latent length varies by more than about 0.05 BPP across the set, the fixed-sampling-ratio assumption fails to control total bitrate, and the reported BPP depends on the entropy of $z$ rather than on the trained cross-entropy loss.","supporting_citations":[{"cited_title":"Learned Lossless Image Compression based on Bit Plane Slic- ing","cited_arxiv_id":null,"evidence_quote":"Baseline autoregressive-latent codec whose even sub-image latent split LVPNet aims to fix; its reported numbers are the main comparison throughout."},{"cited_title":"Compression with flows via local bits-back coding","cited_arxiv_id":null,"evidence_quote":"Flow-based bits-back baseline that defines the prior state of the art on dataset compression."},{"cited_title":"iflow: Numerically invertible flows for efficient lossless com- pression via a uniform coder","cited_arxiv_id":null,"evidence_quote":"Invertible-flow codec used as the rate and speed comparison in Table 3."},{"cited_title":"Deep residual learning for image recognition","cited_arxiv_id":null,"evidence_quote":"Residual skip connections that GMSM adapts to aggregate multi-level features and keep feature volume constant."},{"cited_title":"A method for the construction of minimum-redundancy codes","cited_arxiv_id":null,"evidence_quote":"Huffman coding used to entropy-encode the quantized latent variables."},{"cited_title":"Synthesis of noiseless compression codes,","cited_arxiv_id":null,"evidence_quote":"Arithmetic coding used to encode the predicted pixel probabilities into the main bitstream."},{"cited_title":"Chestx-ray8: Hospital-scale chest x-ray database and benchmarks on weakly-supervised classification and localization of common thorax diseases","cited_arxiv_id":null,"evidence_quote":"ChestX-ray8 dataset that supplies the medical-image benchmark and drives the choice of sampling ratio."},{"cited_title":"FLIF: Free lossless image format based on MA- NIAC compression","cited_arxiv_id":null,"evidence_quote":"Traditional lossless codec included as a non-learned baseline in Tables 1 and 2."}],"review_version":1}