{"id":"da45dd2e-716d-4fff-bc54-9a924a1198d2","arxiv_id":"2507.19125","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A hierarchical, multi-scale context model with progressive cross-attention context fusion improves rate-distortion performance in learned image compression.","lead":"This paper introduces a new entropy-coding context model for learned image compression that codes image latents at three spatial scales from coarse to fine, and progressively fuses context across coding steps with cross-attention. The method reports rate-distortion gains over the VTM-22.0 baseline and over several recent learned codecs at similar or lower compute.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The SOTA claim for HPCM-Base rests on a 0.16–0.26 pp BD-Rate margin over MLIC++ with no uncertainty quantification; a single-seed comparison does not support a state-of-the-art readout.","rationale":"I read the paper in good faith and found the architecture and ablations internally consistent; the hierarchical coding schedule and progressive context fusion are supported by the provided ablations in Tables 2 and 3, and the paper includes a useful supplementary analysis of coding step allocations, shared parameters, and stages. My concern is not about the soundness of the construction but about the strength of the evidence for the strongest empirical claim. Table 1 shows HPCM-Base ahead of MLIC++ by only tenths of a percentage point, yet the abstract asserts state-of-the-art performance without reporting variance or significance. The reader's condition already asks that the SOTA claim be tempered and that code be released, which addresses part of this concern, but the specific quantitative fragility of the Base model comparison is not highlighted in the reader's weakest assumption. Because the claim remains plausible and the Large model has a larger margin, I do not think the verdict should change from CONDITIONAL; the reader's condition should be augmented to require uncertainty quantification for the Base-model SOTA claim. I set agreement_with_reader to partial because the reader identified the hand-designed coding schedule as the weakest assumption, whereas my strongest concern is the unreported run-to-run variability in the headline BD-Rate comparison.","tokens_in":16443,"tokens_out":13152,"duration_ms":146008,"concrete_test":"Retrain HPCM-Base three times from different random seeds using the released code, and also run MLIC++ with the same evaluation protocol; compute BD-Rate vs VTM-22.0 on Kodak, CLIC Pro Valid, and Tecnick with per-dataset means and 95% confidence intervals (or bootstrap across seeds). If the HPCM-Base interval overlaps the MLIC++ interval on any dataset, revise the abstract and Section 4.2 to say 'comparable rate-distortion performance at lower complexity' instead of 'state-of-the-art'.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central quantitative claim is the state-of-the-art BD-Rate reported in Section 4.2 and Table 1. The weakest load-bearing point is that the entire comparison is single-run and no variance is reported. HPCM-Base beats MLIC++ by 0.16 pp on Kodak (-15.31 vs -15.15), by 0.18 pp on CLIC (-14.23 vs -14.05), and by 0.26 pp on Tecnick (-18.16 vs -17.90). These margins are much smaller than the typical seed-to-seed spread of learned image codecs, which is often well above 1 pp in BD-Rate. The Base model is the one that supports the complexity-balance claim (918 kMACs/pixel vs 1282 for MLIC++), and the abstract's 'state-of-the-art' wording applies to both models. If the true performance of HPCM-Base is within noise of MLIC++, the SOTA claim for the Base model is unsupported; if it is slightly worse, the headline readout would need to change. HPCM-Large has a more comfortable margin, roughly 4 pp over MLIC++, so the overall SOTA claim may survive, but the Base-model balance claim is the fragile part. The same absence of uncertainty also affects the CHARM and DCVC-DC intra baselines in Table 1, though those are secondary to the main comparison.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes HPCM, a learned image compression method built around a hierarchical progressive context model. The latent tensor is partitioned into three scales and coded sequentially from the smallest scale to the largest, so that long-range dependencies are modeled efficiently at low resolution before short-range detail is added. A progressive context fusion module uses cross-attention to accumulate context across coding steps and scales, and structural improvements plus a shared-parameter design reduce complexity. On Kodak, CLIC Pro Valid, and Tecnick, the authors report BD-Rate savings over VTM-22.0 of 15.31%, 14.23%, and 18.16% for HPCM-Base and 19.19%, 18.37%, and 22.20% for HPCM-Large, claiming state-of-the-art rate-distortion performance and a better complexity trade-off than existing learned codecs. The code is released, and the paper includes extensive ablations of the coding schedule, fusion mechanism, step allocation, number of stages, and parameter sharing.","tokens_in":16744,"tokens_out":6462,"duration_ms":61365,"significance":"If the reported results hold, HPCM is a meaningful contribution to learned image compression: it reduces kMACs/pixel by roughly 28% relative to MLIC++ while matching or slightly improving BD-Rate in the Base configuration, and the Large configuration improves BD-Rate by ~4 percentage points at comparable complexity. The paper is strong on experimental discipline: entropy-model ablations are run under a fixed transform, the coding-step allocation is tested in four configurations, shared-parameter effects are quantified, and the code is public. The main fragility is that the Base model's state-of-the-art claim rests on BD-Rate margins of 0.16–0.26 percentage points over MLIC++, which is below typical seed-to-seed variation in learned codecs; without uncertainty quantification, that specific claim is not fully supported. The Large model's margin is much larger and supports the core contribution.","major_comments":[{"comment":"The state-of-the-art claim for HPCM-Base is not robust to training stochasticity. The BD-Rate advantage over MLIC++ is 0.16 percentage points on Kodak, 0.18 on CLIC Pro Valid, and 0.26 on Tecnick, yet the paper reports no standard deviations, confidence intervals, or multiple-seed results. In learned image compression, single-seed differences smaller than about one BD-Rate point are often within run-to-run variation. The abstract attributes state-of-the-art performance to 'our method' generally, and the Base model's complexity-balance contribution depends on this small gap. The authors should either provide uncertainty quantification (e.g., results from three training seeds with mean and standard deviation) or restrict the state-of-the-art claim to the Large model, which has a substantially larger margin.","section":"Section 4.2, Table 1"},{"comment":"The hierarchical coding schedule is entirely hand-designed: three scales, eight channel-group-specific partition patterns, and the fixed step allocation (2,3,6). The ablations in Table 2 show that varying the allocation changes BD-Rate by -2.55% to +2.39%, so this choice is load-bearing for the claimed complexity-performance trade-off. However, all evaluations are conducted on 512x768 and 1200x1200 images, and there is no evidence about how the schedule transfers to higher resolutions (e.g., 4K) or to different content types. The paper should either test the schedule at higher resolutions or, if that is infeasible, discuss the expected sensitivity of the design choices to input resolution.","section":"Section 3.2, Tables 2 and D"},{"comment":"The comparison set used to support the state-of-the-art claim is a selection of methods with public code, but several recent methods cited in the related work are not evaluated, including GroupedMixer [23] and the method of Kim et al. [20]. If any of these methods achieves better BD-Rate or a better complexity trade-off, the abstract's claim may not hold. The authors should either include these methods in Table 1 (where code is available) or provide a clear justification for their exclusion.","section":"Section 4.2, Related Work"}],"minor_comments":[{"comment":"The headers of Tables 2 and 3 contain a typo: 'Abaltion' should be 'Ablation'.","section":"Tables 2 and 3"},{"comment":"In the caption and image labels of Figure 9, 'HCPM-Large' and 'HCPM-Base' should be 'HPCM-Large' and 'HPCM-Base'.","section":"Figure 9"},{"comment":"The training dataset is referred to as 'Flicker2W'; this should likely be 'Flickr2W' (the standard Flickr 2W dataset).","section":"Section 4.1"},{"comment":"The eight channel-group-specific partition patterns are only fully described in the supplementary figure; a brief textual summary in the main text would help readers who do not consult the supplement.","section":"Section 3.2"},{"comment":"The MambaVC row has missing entries for CLIC Pro Valid and Tecnick; this is acceptable, but a footnote explaining that the results are unavailable (e.g., code or model not released for those datasets) would improve clarity.","section":"Table 1"},{"comment":"The panels in Figure F are dense; enlarging the panels or annotating the step numbers more visibly would make the alternative coding schedules easier to follow.","section":"Supplementary Figure F"}],"recommendation":"major_revision","confidential_remarks":"The paper is technically sound and the ablations are exemplary, but the state-of-the-art claim for the Base model is the main risk: the margin over MLIC++ is smaller than typical seed variance, and the abstract claims SOTA for the method as a whole. The authors need to add uncertainty quantification or soften the claim. The paper's scope fits eess.IV, and the self-citations to the authors' prior work are not excessive. I would support acceptance after the SOTA-claim issue is resolved."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Read this one if you care about learned image compression. The core idea is a hierarchical coding schedule: split latents into three spatial scales, code the smallest first (long-range context) then progressively up-sample and code larger scales with more steps. On top of that they add a cross-attention fusion that accumulates context from previous coding steps. The combination is new as far as I can tell; prior work had multi-scale orders and cross-attention separately, but not this package. The paper is well executed: ablations cover the schedule, fusion, step counts, stage count, shared params; comparisons are against strong recent baselines with released code, and they report kMACs/pixel and coding times, not just BD-Rate. The Large model shows a solid ~4pp BD-Rate gain over MLIC++ on Kodak/CLIC/Tecnick. That's a meaningful result.\n\nThe soft spot is exactly what the stress-test note says. HPCM-Base beats MLIC++ by 0.16, 0.18 and 0.26 pp on the three datasets, with no confidence intervals or multiple seeds. Those margins are well within the seed-to-seed spread typical for learned codecs. And the Base model is the one carrying the complexity-balance story. So the abstract's 'state-of-the-art rate-distortion performance' is too strong for the Base model as reported. The Large model likely supports the SOTA claim, but the Base one does not. The paper also hand-tunes the coding schedule (2,3,6) and tests only a few alternatives; there's no content-adaptive allocation and no evidence the schedule transfers to higher resolutions. That's a minor concern given the ablations, but worth asking.\n\nI don't see a load-bearing flaw. The architecture is sound, the experiments are mostly clean, and the code is promised on GitHub. The fix is cheap: temper the claims, run a few seeds, report variance, and release the code. If the code reproduces, this is a citeable contribution. I'd send it to peer review with those revision requests.","headline":"Solid architecture paper with a credible RD gain and clean ablations, but the Base model's SOTA claim is a single-seed, sub-0.3pp margin over MLIC++ with no variance reported.","tokens_in":17232,"tokens_out":2073,"would_cite":true,"duration_ms":21130,"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":"This paper establishes that coding latents coarse-to-fine with context carried across scales gives learned image compression state-of-the-art rate-distortion performance at reduced complexity.","keywords":["learned image compression","entropy coding","context modeling","hierarchical coding schedule","progressive context fusion","autoregressive model","rate-distortion optimization","long-range dependency"],"falsifier":"Train a variant that replaces the fixed hierarchical schedule with a content-adaptive ordering, or with the same progressive fusion but a single-scale context model, keeping everything else identical; if BD-Rate stays within about 0.5% of HPCM-Base, the hierarchical schedule is not the source of the gain. Alternatively, evaluate the released model on 4K or screen-content images, since the schedule was tuned and validated only at 512x768 and 1200x1200; a clear complexity or rate-distortion regression there would show the schedule does not transfer.","tokens_in":16254,"feed_emoji":"🖼️","tokens_out":5820,"duration_ms":53974,"temperature":0.7,"pith_summary":"The paper tries to establish that the expensive part of learned image compression, modeling the probability of every latent symbol so entropy coding uses as few bits as possible, can be made cheaper and more accurate if the latents are coded in a coarse-to-fine order. Instead of paying for global attention over the full latent map, the proposed HPCM codes a 4x-downsampled sub-latent first, then progressively fills in finer scales. A cross-attention fusion step carries context accumulated from earlier coding steps into the next step. If the claim holds, learned codecs can get the gains of long-range context modeling without the quadratic complexity that usually comes with it.","feed_headline":"Hierarchical context coding beats VTM-22.0 by 19.19%","feed_subtitle":"Coding coarse latents first captures long-range redundancy without the cost of global attention.","key_machinery":"Three-scale coarse-to-fine latent coding with progressive context fusion. The quantized latents $\\hat{\\mathbf{y}}$ are split into $\\hat{\\mathbf{y}}^{S_1}$ (a 4x downsampled version), $\\hat{\\mathbf{y}}^{S_2}$, and $\\hat{\\mathbf{y}}^{S_3}$, coded with 2, 3, and 6 steps respectively. After each scale is coded, it is filled back into the next scale, so the context network sees long-range structure before fine detail. The progressive context fusion module maintains a context state $C_i$ per coding step, updates it by cross-attention with the entropy-parameter state $\\psi_i$, and propagates it across scales; this carries diverse context from past coding steps into the current estimate of the generalized Gaussian entropy parameters.","core_discovery":"The central claim is that a hierarchical progressive context model (HPCM) gives state-of-the-art rate-distortion performance while using substantially less computation than recent transformer-based context models. The method partitions the quantized latent representation into three scales and codes them sequentially from the smallest to the largest, so long-range dependencies are captured cheaply at the coarse scale and short-range details are refined at the fine scale. Across eight channel groups the partition pattern varies, mixing spatial and channel context. A progressive context fusion module accumulates the entropy-model state from previous coding steps through cross-attention, then propagates it across scales by filling coarse context into the finer scale and merging it with the hyperprior. On Kodak, HPCM-Large reports 19.19% BD-Rate savings over VTM-22.0 and HPCM-Base reports 15.31%, outperforming MLIC++ and FLIC; the paper attributes the gain specifically to the hierarchical schedule and the progressive fusion.","pith_inferences":["If the coding schedule were content-adaptive, choosing the number of scales and per-scale step counts from the image statistics, the same fusion mechanism might push the gains further, since the fixed (2,3,6) allocation is only one operating point and the ablations show diminishing returns at the coarse scale.","The coarse-to-fine order already produces an embedded bitstream, so HPCM could be extended to progressive or quality-scalable transmission, where a coarse reconstruction is decoded first and the finer scales refine it.","The cross-attention fusion's attention maps concentrate on high-bitrate texture regions, suggesting a testable loop where the fusion module predicts where future coding steps should be spent.","The hierarchical schedule resembles scalable wavelet coding, so coupling HPCM with a wavelet-like transform might let the codec allocate context modeling effort across frequency bands as well as scales."],"forward_implications":["Long-range context modeling in learned codecs no longer needs full-resolution global attention: the same dependency coverage comes from first coding a coarse sub-latent, with complexity around 918 to 1261 kMACs/pixel.","A codec that codes coarse structure first will allocate compute where it pays off: the ablation reports that raising the fine-scale steps from (2,3,6) to (2,3,12) buys about 2.55% more BD-Rate savings at much higher cost, while adding coarse-scale steps buys almost nothing.","Sharing context-model weights across coding steps keeps the parameter count near 68.5M, while non-shared weights add about 121M parameters for essentially identical performance.","The entropy-model state itself is a useful context carrier: removing fusion entirely costs 4.71% in BD-Rate, while reusing only the previous state recovers most of the loss, showing the cross-attention fusion closes the remaining gap.","The released model makes the hierarchical progressive context design directly reproducible and testable on other datasets and resolutions."],"supporting_citations":[{"why":"Supplies the hyperprior branch that HPCM conditions on and from which the initial context is down-scaled.","marker":"[3]"},{"why":"The joint autoregressive and hierarchical priors baseline; HPCM replaces its raster-scan spatial context with a hierarchical schedule.","marker":"[40]"},{"why":"The channel-wise autoregressive entropy model CHARM, re-trained with HPCM's transform as a comparison baseline.","marker":"[39]"},{"why":"Source of the spatial-channel group partition and quadtree coding strategy that HPCM adapts to its three-scale schedule.","marker":"[26]"},{"why":"MLIC++ is the strongest learned-codec baseline HPCM-Base must match at lower kMACs/pixel and HPCM-Large exceeds.","marker":"[19]"},{"why":"FLIC is the frequency-aware transformer baseline that HPCM beats; it is also the complexity anchor in the BD-rate versus kMACs/pixel comparison.","marker":"[24]"},{"why":"TCM is a mixed Transformer-CNN codec used as a comparison point for both rate-distortion and complexity.","marker":"[35]"},{"why":"Entroformer motivates long-range transformer context modeling, the line of work HPCM makes cheaper.","marker":"[41]"},{"why":"M2T provides the masking-based octree decoding order that the fine-scale coding schedule follows.","marker":"[38]"}],"fun_headline_variants":["Hierarchical scales cut compute, boost image compression 19%","Progressive context model slashes bitrate 19% on Kodak","Coarse-to-fine context coding outperforms VTM-22.0","HPCM large wins 19.19% over VTM at lower complexity","Multiscale progressive context model beats VTM by 19%"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The central assumption is that a hand-designed schedule of three scales, eight channel-group partition patterns, and fixed step counts (2, 3, 6) is the right way to spend compute; if a simpler or content-adaptive schedule yielded the same rate-distortion gains, the contribution would reduce to the fusion module.","fun_headline_variants_meta":{"raw":{"variants":["Hierarchical scales cut compute, boost image compression 19%","Progressive context model slashes bitrate 19% on Kodak","Coarse-to-fine context coding outperforms VTM-22.0","HPCM large wins 19.19% over VTM at lower complexity","Multiscale progressive context model beats VTM by 19%"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000527,"raw_usage":{"total_tokens":2509,"prompt_tokens":877,"completion_tokens":1632,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":493,"completion_tokens_details":{"reasoning_tokens":1537}},"tokens_in":493,"tokens_out":1632,"duration_ms":13017,"temperature":1.0,"reasoning_tokens":1537,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T17:59:39.761885+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train a variant that replaces the fixed hierarchical schedule with a content-adaptive ordering, or with the same progressive fusion but a single-scale context model, keeping everything else identical; if BD-Rate stays within about 0.5% of HPCM-Base, the hierarchical schedule is not the source of the gain. Alternatively, evaluate the released model on 4K or screen-content images, since the schedule was tuned and validated only at 512x768 and 1200x1200; a clear complexity or rate-distortion regression there would show the schedule does not transfer.","supporting_citations":[{"cited_title":"MLIC++: Linear complex- ity multi-reference entropy modeling for learned image com- pression","cited_arxiv_id":null,"evidence_quote":"MLIC++ is the strongest learned-codec baseline HPCM-Base must match at lower kMACs/pixel and HPCM-Large exceeds."},{"cited_title":"Variational image compres- sion with a scale hyperprior","cited_arxiv_id":null,"evidence_quote":"Supplies the hyperprior branch that HPCM conditions on and from which the initial context is down-scaled."},{"cited_title":"Joint autoregressive and hierarchical priors for learned image compression","cited_arxiv_id":null,"evidence_quote":"The joint autoregressive and hierarchical priors baseline; HPCM replaces its raster-scan spatial context with a hierarchical schedule."},{"cited_title":"Channel-wise autoregres- sive entropy models for learned image compression","cited_arxiv_id":null,"evidence_quote":"The channel-wise autoregressive entropy model CHARM, re-trained with HPCM's transform as a comparison baseline."},{"cited_title":"Neural video compression with diverse contexts","cited_arxiv_id":null,"evidence_quote":"Source of the spatial-channel group partition and quadtree coding strategy that HPCM adapts to its three-scale schedule."},{"cited_title":"Frequency-aware transformer for learned image compression","cited_arxiv_id":null,"evidence_quote":"FLIC is the frequency-aware transformer baseline that HPCM beats; it is also the complexity anchor in the BD-rate versus kMACs/pixel comparison."},{"cited_title":"Learned image compression with mixed Transformer-CNN architectures","cited_arxiv_id":null,"evidence_quote":"TCM is a mixed Transformer-CNN codec used as a comparison point for both rate-distortion and complexity."},{"cited_title":"Entroformer: A Transformer-based entropy model for learned image compression","cited_arxiv_id":null,"evidence_quote":"Entroformer motivates long-range transformer context modeling, the line of work HPCM makes cheaper."},{"cited_title":"M2T: Masking Transformers twice for faster decoding","cited_arxiv_id":null,"evidence_quote":"M2T provides the masking-based octree decoding order that the fine-scale coding schedule follows."}],"review_version":2}