{"id":"48ad1b5a-6549-4db4-9697-514eca5ceb0d","arxiv_id":"2506.15228","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"ABC learns the structure of a neural image compression codec jointly with a rate-distortion-complexity objective, making the codec computationally scalable across the encoder, decoder, and autoregressive context model.","lead":"ABC is a system that lets neural image compression codecs trade computing effort for quality on the fly, adapting to the device, the image content, and the downstream vision task. It learns which parts of the network and which parallel coding stages to switch on or off.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Uncounted MACs of the VGG-based data-adaptive controller may inflate low-complexity gains; end-to-end MACs must be reported.","rationale":"The reader's weakest assumption concerns the bit-cost of transmitting the generative structure G, citing Section V-C. That concern is legitimate but likely low-impact: the intra-node structure is generated from a low-dimensional C x 2 x 2 logit tensor expanded spatially, and the inter-node structures are a handful of categorical channel-width choices, so even at low bit-rates the side information is plausibly a few hundred bits, well under 1% of a 0.2 BPP 512x512 Kodak bitstream. The paper should still measure it, but it is unlikely to overturn the central claim. A more load-bearing gap is the computational overhead of the adaptive control module. The abstract promises computational scalability and complexity control span across all NIC components, yet the module that enables data adaptivity is a VGG-based network whose MACs are not included in the complexity loss (Eq. 33), not reported in the MAC experiments, and explicitly excluded from the speed comparisons (Section VI-B2). If this controller adds tens to hundreds of kMACs/pixel, then the low-complexity data-adaptive operating points in Figures 4 and 6 are not end-to-end scalable and the reported span is overstated. A single instrumentation pass would settle this. Because the required evidence is absent but the framework is otherwise coherent, the appropriate verdict is CONDITIONAL: accept only after the control-module MACs are quantified and included in all complexity comparisons. This is consistent with the reader's CONDITIONAL verdict, though the stress-test identifies a different primary condition.","tokens_in":26084,"tokens_out":12288,"duration_ms":120220,"concrete_test":"Instrument the ABC+DA model with ptflops and count total encode MACs per pixel at each of the 8 complexity levels (ABC-L0 to ABC-L7) on the Kodak evaluation resolution, explicitly including the VGG-based data-adaptive controller g_c, the merge module, and the control MLP. Report the fraction of total MACs contributed by the control module at ABC-L7. Then re-plot Figures 4a and 6a with end-to-end MACs (including the controller) on the x-axis for ABC+DA, against ABC w/o DA, SlimCAE, and ELFIC. If the control module adds less than 5% of total MACs at every level, the concern is retired. If it adds more than 20% at ABC-L7 or shifts the ABC+DA curves across another method's curve, the claims of full computational scalability and data-adaptivity advantage require qualification.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim is that ABC achieves full computational scalability and broader complexity control span while maintaining competitive compression performance. That claim requires the reported MAC figures to reflect the total system compute at each level. Section V-C introduces an Adaptive Control Module, and for data adaptivity it says 'we utilize a VGG-based network g_c to predict from input data x.' The complexity loss in Equation 33 includes C(G_inter) for the four inter-node structures and notes C(G_intra) is constant, but there is no term for the control module. Appendix B-B3 states 'We utilize ptflops to tally the MAC figures' without listing which modules are included. If g_c runs at or near input resolution, its per-pixel MACs can be comparable to or exceed the cheapest ABC levels (for example, ABC-L7 is well below the 444 kMACs/pixel of ABC-L0), so excluding it would shift the ABC+DA curves in Figures 4 and 6 to the right at the low-complexity end, narrowing the claimed control span and weakening the data-adaptivity advantage. The paper itself notes 'the added computational overhead from the adaptive control module leads to marginally reduced performance at specific complexity thresholds' (Section VI-B1), but this overhead is never quantified in MACs or time. This is a missing measurement, not a demonstrated failure: it can be resolved by instrumenting the full encode path.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes ABC, a framework for computational scalable multi-task neural image compression based on Bayesian network (BayesNet) structure learning. The BayesNet structure G is decomposed into an inter-node bipartite graph that selects among slimmable channel-width edges for the analysis, synthesis, and hyper transforms, and an intra-node multipartite graph that learns topological partitions for a dynamic masked convolution acting as a parallel autoregressive context model. An adaptive control module conditions G on device budget, input content, and downstream task. Experiments compare BD-Rate against kMACs/pixel on Kodak and CLIC, task-specific accuracy on ImageNet and MS-COCO, and processing speed against WebP, BPG, and NIC baselines; ablations cover one- vs two-stage optimization, backbone choices, and intra-node hyperparameters. Code is released.","tokens_in":26416,"tokens_out":6664,"duration_ms":70819,"significance":"The conceptual contribution is useful: viewing NIC modules as BayesNet edges and learning inter- and intra-node structure in a differentiable way is a coherent extension of prior slimmable-NIC work, and the learned intra-node structures (Table III, Fig. 10) offer a concrete design insight that resonates with later hand-crafted schemes such as ELIC. The two-stage optimization in Sec. III-E is sensible, and releasing code is a strength. However, the central scalability claims are currently supported by incomplete measurement: the adaptive control module's compute is not included in MAC totals, the bit cost of transmitting G is asserted rather than measured, and the reported comparisons lack repeated-run variance. These issues are fixable and should be addressed before publication.","major_comments":[{"comment":"The complexity loss in Eq. (33) sums C(G_inter) terms and omits the Adaptive Control Module, and App. B-B3 ('We utilize ptflops to tally the MAC figures') does not state whether the VGG-based g_c, the merge module, and the MLP are included. Since the ABC+DA curves in Figs. 4 and 6 are plotted against kMACs/pixel, an unmeasured per-pixel MAC cost for g_c can shift the low-complexity points to the right and narrow the claimed control span. The note in Sec. VI-B1 that the control module causes 'marginally reduced performance' is qualitative; please instrument the full encode path and report total MACs, and ideally wall-clock time, for both the w/o-DA and +DA variants at every complexity level.","section":"Sec. V-C, Eq. (33), App. B-B3"},{"comment":"The formulation in Eq. (8) includes log p_G(\\hat G) in the rate loss, but the implemented loss in Eq. (31) does not contain this term, and Sec. V-C dismisses the generative structure cost as negligible without measurement. The decoder must receive G for the generative dependencies p(\\hat y|\\hat z, G) and p(\\hat x|\\hat y, G), so the bitstream necessarily contains the inter-node and intra-node structure indices. At BPP around 0.3, even a few hundred bits per image is non-negligible for BD-Rate comparisons. Please report the compressed size of the transmitted G per image and per level, or provide a measured entropy model for it, and explicitly show whether this cost is negligible.","section":"Sec. V-C and Eq. (8)"},{"comment":"The MAC axis in Figs. 4 and 6 reflects inter-node backbone cost only: Eq. (33) defines C(G_intra) as constant and omits it from L_C, and Sec. IV-B/App. B show that varying S_intra changes the number of parallel stages and latency while the MACs of the dynamic masked convolution remain equal to those of a standard 2D convolution (Eq. 27). The claim of 'full computational scalability across all NIC components' therefore requires a clearer separation of metrics: the headline MAC comparisons do not exercise the intra-node autoregressive scalability, which appears only in the speed-based comparison (Fig. 8, Table III). Please state explicitly which components are traded off in each experiment and report both MAC and latency for all components.","section":"Sec. IV-B, Eq. (33), Figs. 4/6"},{"comment":"The empirical comparisons are reported as single runs without error bars or multiple seeds, even where the differences are very small (Table III, where BPP differences of 0.003-0.005 separate methods). To support the claim that the learned intra-node structures 'consistently demonstrate slightly better performance,' repeated-seed or repeated-evaluation variance should be reported. In addition, the baseline description is internally inconsistent: Sec. VI-B1 first states that the autoregressive modules are 'same as [17]' across all frameworks and then says SlimCAE uses a 'slimmable context model' according to the original paper; please specify exactly which context model each baseline uses and how λ-scheduling was disabled.","section":"Sec. VI-B1, VI-D, Table III"}],"minor_comments":[{"comment":"The sentence beginning 'Moreover, the decompression speed is also reported to evaluate the efficiency of It's important to note...' is incomplete and should be rewritten.","section":"Sec. VI-D"},{"comment":"The phrase 'more than 30% higher compression performance than BPG' is ambiguous; it should be phrased as 'more than 30% BD-rate savings over BPG' or similar according to the metric used.","section":"Sec. VI-B2"},{"comment":"The note that reported decompression speeds are effectively quadrupled because images are loaded as 32-bit float tensors should be stated in the main text, at least in the caption of Fig. 8, to avoid misleading readers.","section":"App. B-B3"},{"comment":"The table layout with 'Stages - 2 4 10' and the row of method names is hard to parse; explicit column headings for each method group would improve readability.","section":"Table III"}],"recommendation":"major_revision","confidential_remarks":"This manuscript extends the authors' ECCV 2024 BaSIC paper, and the adaptive-control components appear to be the main journal-level novelty; the revision should state more explicitly which contributions are new relative to [15]. I recommend major revision rather than rejection because the core method is coherent and the missing measurements (controller MACs, G side-information bits, repeated runs) are within the scope of a revision and could be supplied with the released code."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper is a real extension, not a repackaging. The new piece is the adaptive control module that conditions the BayesNet structure on device, data, and task, plus the two-stage training that makes the larger search space stable. The dynamic masked convolution for the intra-node graph is the most original technical idea, and it is tested properly: at equal parallel stages, the learned structures match or beat hand-crafted autoregressive models, and at 10 stages they approach the BPP of the sequential MaskConv baseline. That holds up, and the ablation across backbones adds credibility.\n\nThe soft spots are real but not fatal. The stress-test note lands: the VGG-based data controller g_c is not in the complexity loss in Equation 33, and the ptflops tally in Appendix B-B3 does not say which modules are included. Since ABC-L7 is far cheaper than ABC-L0, a full-resolution VGG-style controller could matter a lot at the low-complexity end. The paper even admits the controller adds overhead but never measures it. This is missing measurement, not a demonstrated failure, but it directly affects the central claim of broad complexity control span.\n\nThe second issue is also valid: Equation 8 includes log p_G(hat G) in the rate, but Section V-C drops it on the assertion that the structure bits are negligible. They probably are small, but the paper needs to quantify them, especially at low bitrates. The reader's weakest-assumption flag is correct.\n\nSmaller concerns: no error bars or multiple seeds, and the comparisons to SlimCAE and ELFIC rely on reimplementations with modified training schedules. That makes claims like \"marginally better\" less solid. The fixed-MAC comparison in the high-complexity regime shows ABC underperforming Cheng2020 and Checkerboard, which the paper acknowledges honestly.\n\nWho gets value from this: anyone working on computational scalability in learned image compression, efficient autoregressive models, or deployment of NIC on heterogeneous devices. It is a solid framework paper with a credible learning-based contribution, despite the missing accounting. A serious referee should engage with it, and the revision should ask for end-to-end MACs including the controller, measured structure bit overhead, and at least a few seeds or error bars on the key comparisons.","headline":"A serious, well-built extension of the ECCV BaSIC line that deserves a referee, but only if the authors add the missing MAC and bit-cost measurements for the adaptive controller and transmitted structures.","tokens_in":26935,"tokens_out":2845,"would_cite":true,"duration_ms":33459,"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":"By learning the graph structure of a neural image codec viewed as a Bayesian network, ABC makes every component computationally scalable—from 444 kMACs/pixel down to a minimal mode—while keeping rate–distortion performance competitive.","keywords":["neural image compression","computational scalability","Bayesian network structure learning","slimmable networks","autoregressive context model","rate-distortion-complexity optimization","multi-task coding","dynamic masked convolution"],"falsifier":"Run the arithmetic coder on the decoder-side structure samples $\\hat{G}$—the inter-node categorical edge selections and the intra-node topological indices—at each of the eight ABC complexity levels and compare the resulting bits with the total BPP; if that overhead reaches even a small percentage of the bitstream at ABC-L7 or at low bit-rates, the claim that structure cost is negligible and does not affect the rate–distortion trade-off is falsified.","tokens_in":25855,"feed_emoji":"🖼️","tokens_out":13974,"duration_ms":126195,"temperature":0.7,"pith_summary":"ABC claims that the computational cost of a neural image codec can be made granularly and adaptively scalable by treating the codec as a Bayesian network and learning the network's graph structure, rather than hand-designing complexity levels. It introduces two learned structures: a heterogeneous bipartite graph that selects channel widths in the neural backbone, and a homogeneous multipartite graph that assigns latent positions to parallel stages in the autoregressive context model. An adaptive control module picks the graph according to a device budget, input difficulty, or downstream vision task. On Kodak and CLIC, the codec spans eight complexity levels from 444 kMACs/pixel downward and reaches BD-Rate at or below scalable baselines across most MAC levels, while the learned autoregressive graphs use fewer bits than hand-crafted parallel context models at equal stage counts. If these claims hold, one trained model can replace a family of separately trained codecs and adjust its computation at deployment time.","feed_headline":"One neural codec now spans eight compute levels","feed_subtitle":"Learned BayesNet structure lets one model serve high-end and edge devices at competitive bitrates.","key_machinery":"The carrying object is the learned BayesNet structure $G$, decomposed into the inter-node graph and the intra-node graph. The inter-node graph is parameterized by categorical distributions $\\pi$ over candidate edges per node pair; because each candidate is a slimmable network at a distinct channel width, sampling an edge samples a computational budget, and expected complexity is computed by mixing the candidate MACs with the same weights (Equation 18). The intra-node graph is parameterized by per-location categorical topological indices $T_{c,h,w}$; the number of partites, $\\dim \\pi_{c,h,w} = S_{\\mathrm{intra}}$, sets the number of parallel stages, and the dynamic masked convolution of Equation (24) zeros out any kernel position whose topological index is not strictly smaller than the current node's, preserving acyclicity. The generative structure samples needed at decode time are claimed to be simple discrete variables, excluded from the rate loss. A two-stage optimizer first trains the rate–distortion objective for any sampled graph, then selects a finite set of near-Pareto-optimal structures and fine-tunes inside that set to stabilize the large search space.","core_discovery":"The paper's central claim is that computational scalability in neural image compression can be achieved by learning the directed acyclic graph $G$ that defines conditional dependencies among the codec's variables, split into two learnable structures. The inter-node BayesNet is a heterogeneous bipartite graph: each adjacent pair of nodes is linked by several candidate neural-network edges of different sizes, a categorical distribution selects one edge, and each candidate is implemented as a slimmable network at a different channel width, so the selected graph directly sets a MAC budget. The intra-node BayesNet is a homogeneous multipartite graph over the latent tensor: every spatial/channel location carries a topological index drawn from a categorical distribution, dependencies point only from earlier to later partites, and the number of partites equals the number of parallel autoregressive stages; a dynamic masked convolution (Equation 24) enforces acyclicity while keeping the same parameter and MAC count as a standard convolution. The adaptive control module generates $G$ conditioned on a computation-budget controller $\\Phi_C$, a task controller $\\Phi_T$, and the input $x$, and the decoder-side structure samples are described as simple discrete variables with negligible storage. Experiments report BD-Rate at or below scalable baselines across most MAC levels, with learned intra-node graphs matching or beating hand-crafted checkerboard and channel-wise context models at equal parallel stages.","pith_inferences":["Editorial inference: The paper's exclusion of $\\hat{G}$ from the rate loss is an assumption, not a measurement; a direct count of the bits spent on decoder-side structure at low bit-rates and at ABC-L7 would show whether the reported BD-Rate survives in the regime where ABC claims its biggest advantage.","Editorial inference: The learned intra-node graphs consistently converge to interlaced channel-spatial patterns, which suggests a transferable design rule for parallel autoregressive context models; one could test it by initializing hand-crafted models with interlaced assignments and comparing final bit-rate.","Editorial inference: The complexity loss is defined on differentiable MAC counts, so the paper's scalability claims are about arithmetic cost rather than wall-clock time; a complete deployment comparison would need to charge the adaptive control module's own computation and memory overhead against the MAC budget, which the paper acknowledges can reduce performance at some thresholds.","Editorial inference: The architecture-agnostic claim rests on ablations over hyperprior-style backbones; a natural extension is to apply the same inter/intra-node structural prior to transformer-based or lossless-hybrid codecs, where context dependency patterns differ."],"forward_implications":["A single ABC model can be deployed at eight selectable complexity levels, and because the four slimmable backbone networks are controlled independently, the model offers $5^4 = 625$ configurable MAC levels for hitting a hardware budget precisely.","The learned intra-node graph at ten parallel stages reaches a bit-rate close to the serial MaskConv baseline while keeping the autoregressive model parallel, so learned context structure narrows the rate–distortion gap that hand-crafted parallel models leave open.","Adaptive control lets one codec serve multiple regimes: data-adaptive ABC improves BD-Rate on CLIC under low MAC budgets, and task-adaptive ABC supports classification and instance segmentation across a wider MAC span than the task-slimmable baseline.","The structure-learning layer is not tied to one backbone: ablations on three different backbone designs show the same comparative trends, so the approach can be grafted onto other hyperprior-style codecs."],"supporting_citations":[{"why":"Supplies the hyperprior generative model $p(x|y)p(y|z)p(z)$ that ABC treats as the BayesNet substrate and extends.","marker":"[1]"},{"why":"Supplies the slimmable-network channel-width mechanism used to implement the heterogeneous inter-node edges.","marker":"[9]"},{"why":"Supplies the joint autoregressive and hyperprior architecture on which ABC's implementation is built and later ablated.","marker":"[17]"},{"why":"Supplies the checkerboard parallel context model that the learned intra-node graph is compared against at equal stage counts.","marker":"[27]"},{"why":"Supplies the channel-wise parallel autoregressive model used as another intra-node baseline.","marker":"[26]"},{"why":"Supplies the unevenly grouped channel-checkerboard context model used as the 10-stage reference for the learned intra-node graph.","marker":"[3]"},{"why":"Supplies the flexible codec baseline with rate-distortion-complexity optimization and data adaptivity that ABC compares against over MAC budgets.","marker":"[12]"},{"why":"Supplies the task-adaptive slimmable multi-task codec that ABC compares against for classification and instance segmentation.","marker":"[6]"},{"why":"Supplies the continuous relaxation and differentiable architecture-search technique on which the inter-node graph optimization is modeled.","marker":"[46]"},{"why":"Supplies the Monte-Carlo objective used to optimize the discrete intra-node topological indices.","marker":"[50]"}],"fun_headline_variants":["Learned BayesNet structure tunes neural codec compute on the fly","Adaptive graph lets one codec run from edge to cloud efficiently","Bayesian structure learning makes neural compression compute-scalable","One model, any device: BayesNet structure adapts compute per input","Neural codec learns its own complexity graph for each task and device"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise, stated in Section V-C, is that the decoder-side structure samples $\\hat{G}$ are simple discrete variables whose storage is negligible, so they are excluded from the rate loss; if transmitting those choices costs real bits, the reported bitrate savings shrink and the competitive-compression claim needs revision.","fun_headline_variants_meta":{"raw":{"variants":["Learned BayesNet structure tunes neural codec compute on the fly","Adaptive graph lets one codec run from edge to cloud efficiently","Bayesian structure learning makes neural compression compute-scalable","One model, any device: BayesNet structure adapts compute per input","Neural codec learns its own complexity graph for each task and device"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000208,"raw_usage":{"total_tokens":1458,"prompt_tokens":1052,"completion_tokens":406,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":668,"completion_tokens_details":{"reasoning_tokens":315}},"tokens_in":668,"tokens_out":406,"duration_ms":4649,"temperature":1.0,"reasoning_tokens":315,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T19:41:29.498054+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the arithmetic coder on the decoder-side structure samples $\\hat{G}$—the inter-node categorical edge selections and the intra-node topological indices—at each of the eight ABC complexity levels and compare the resulting bits with the total BPP; if that overhead reaches even a small percentage of the bitstream at ABC-L7 or at low bit-rates, the claim that structure cost is negligible and does not affect the rate–distortion trade-off is falsified.","supporting_citations":[{"cited_title":"Slimmable compressive autoencoders for practical neural image compression,","cited_arxiv_id":null,"evidence_quote":"Supplies the slimmable-network channel-width mechanism used to implement the heterogeneous inter-node edges."},{"cited_title":"Improving multiple machine vision tasks in the compressed domain,","cited_arxiv_id":null,"evidence_quote":"Supplies the checkerboard parallel context model that the learned intra-node graph is compared against at equal stage counts."},{"cited_title":"Channel-wise autoregressive entropy models for learned image compression,","cited_arxiv_id":null,"evidence_quote":"Supplies the channel-wise parallel autoregressive model used as another intra-node baseline."},{"cited_title":"Elfic: A learning-based flexible image codec with rate-distortion-complexity optimization,","cited_arxiv_id":null,"evidence_quote":"Supplies the flexible codec baseline with rate-distortion-complexity optimization and data adaptivity that ABC compares against over MAC budgets."},{"cited_title":"Slimmable multi-task image compression for human and machine vision,","cited_arxiv_id":null,"evidence_quote":"Supplies the task-adaptive slimmable multi-task codec that ABC compares against for classification and instance segmentation."},{"cited_title":"Variational inference for monte carlo objectives,","cited_arxiv_id":null,"evidence_quote":"Supplies the Monte-Carlo objective used to optimize the discrete intra-node topological indices."}],"review_version":2}