{"id":"3e7c4728-7fc6-433f-bb60-a1b4f87fab09","arxiv_id":"2412.17270","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"AsymLLIC uses a two-stage training scheme to replace complex decoder modules with simpler ones, cutting decoder MACs to 51.47 GMACs while keeping RD performance close to VVC.","lead":"AsymLLIC is a learned image compression system with a deliberately light decoder, 51.47 GMACs and 19.65M parameters, whose rate-distortion performance is close to the VVC standard. It matters because practical deployment of learned codecs is often limited by decoder compute on low-end devices, and this design shifts the compute burden to the encoder side.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The central complexity claim is untestable as reported: Table I omits TCM-small, the paper's own base architecture, so the claimed superiority over 'previous LIC methods' cannot be verified.","rationale":"The paper is a credible engineering contribution: it reports RD curves, ablations of the synthesis decoder and hyperprior path, and a concrete asymmetric training strategy. The reported numbers, if accurate, do support a lightweight decoder with RD performance near VVC. I do not see an internal inconsistency that would warrant rejection. The most load-bearing gap is the missing TCM-small baseline. TCM-small is the architecture from which AsymLLIC is derived and is already used as a baseline in Figure 3 and Section III-B, so its absence from Table I is not a matter of unavailable data. Without that row, the central comparative claim 'significantly lower decoding complexity than previous LIC methods' and 'optimal balance' cannot be substantiated. The reader's weakest assumption about the two-stage asymmetric training scheme is a reasonable secondary concern, but it is less decisive: even if end-to-end training of the same asymmetric architecture produced better RD, the paper's reported empirical results would not be invalidated, only shown to be suboptimal. The TCM-small omission, by contrast, directly bears on whether the claimed complexity advantage over previous methods is real. Conditional acceptance with a request for the missing baseline remains the appropriate outcome.","tokens_in":8727,"tokens_out":14974,"duration_ms":155177,"concrete_test":"Add a TCM-small row to Table I using the same 768×512 MAC-counting methodology, the same Kodak BD-rate anchor, and the same separation of decoder vs. total parameters and MACs. Then compare: if TCM-small's decoder MACs are within roughly 20% of AsymLLIC's 51.47 GMACs, or if TCM-small achieves a BD-rate better than -18.68% at comparable or lower decoder complexity, the headline claim of 'significantly lower decoding complexity' and 'optimal balance' is weakened. If TCM-small is substantially more complex or worse in BD-rate, the claim is supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central quantitative claim is that AsymLLIC needs only 51.47 GMACs and 19.65M decoder parameters to achieve -18.68% BD-rate over BPG, with 'significantly lower decoding complexity than previous LIC methods' and 'an optimal balance' between performance and complexity. However, the comparison in Table I omits TCM-small, the exact architecture that Section III-B and Figure 3 identify as the starting point for AsymLLIC. TCM-small appears in Figure 3c/e as a baseline, and its synthesis-decoder complexity is plotted, so the authors have the data; yet no TCM-small row appears in Table I. If TCM-small has a decoder MAC count close to or below 51.47 GMACs, or a BD-rate that is better at similar or lower complexity, then the claimed advantage over previous LIC methods is overstated. As reported, the reader cannot tell whether AsymLLIC improves on its base model or merely trades away RD performance for a modest complexity reduction. This omission is load-bearing because the paper's contribution is defined as a complexity-performance tradeoff, not just a stand-alone BD-rate number.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes AsymLLIC, a learned image compression architecture with an asymmetric encoder-decoder design that moves computational load to the encoder and uses a lightweight decoder. The method is built on TCM-small and introduces a two-stage training strategy in which the synthesis decoder is first fine-tuned with a distortion-only loss while the encoder and entropy model are frozen, and then the hyperprior decoder and context model are lightweighted and trained together with the hyperprior encoder. The authors report a BD-rate of -18.68% over BPG on Kodak with 19.65M decoder parameters and 51.47 GMACs, claim near-VVC performance, and include ablations on the synthesis decoder and hyperprior path structures.","tokens_in":8948,"tokens_out":3662,"duration_ms":34058,"significance":"If the reported complexity-performance trade-off is accurate, AsymLLIC is a useful contribution to practical learned image compression: it demonstrates that decoding complexity can be substantially reduced while retaining RD performance near traditional codecs, and the two-stage training idea is simple and potentially transferable to other LIC models. The paper evaluates on three datasets and includes ablation studies that isolate design choices. However, the verification is incomplete: the direct baseline TCM-small is omitted from the main complexity comparison, the staged training procedure is not ablated against joint training, and the small performance differences in the ablations are not accompanied by variance estimates or released code/weights. These gaps currently prevent the reader from verifying the central claim that AsymLLIC achieves the 'optimal balance' between compression performance and decoding complexity.","major_comments":[{"comment":"Table I omits TCM-small, which Section III-B identifies as the base architecture that AsymLLIC starts from. Since TCM-small's synthesis-decoder complexity is already plotted in Figure 3(e), the authors demonstrably have the data to include a TCM-small row. Without this row, the central claim of 'lowest decoding complexity' and 'optimal balance' cannot be verified: if TCM-small has comparable or lower decoder MACs, or better BD-rate at similar complexity, the claimed advantage over previous LIC methods is overstated. Please add TCM-small (with the same evaluation protocol and input size) to Table I and discuss the comparison explicitly.","section":"Section III-B, Table I"},{"comment":"The two-stage asymmetric training strategy is not validated. Stage 1 freezes the encoder and entropy model and trains only the lightweight synthesis decoder with a distortion-only loss; stage 2 then trains the hyperprior decoder and context model together with ha. The paper provides no ablation comparing this staged procedure with (i) joint end-to-end training of the full asymmetric network, or (ii) fine-tuning all decoder modules simultaneously from the TCM-small checkpoint. Without such an ablation, the reader cannot tell whether the staged training is necessary or whether it converges to a worse optimum than joint training. Please add an ablation that isolates the effect of the training scheme on the final RD performance.","section":"Section III-A, Figure 2"},{"comment":"The reported performance differences are small (e.g., 0.18 dB in Table II) and no variance or statistical significance is reported. Since the paper does not release code or trained weights, the reader cannot assess whether the proposed structure's advantage over the RBU baseline is within run-to-run variability. Please report the standard deviation across at least three independent training runs, or release trained models, to support the claimed superiority of the proposed decoder structures.","section":"Tables II and III"}],"minor_comments":[{"comment":"The word 'evitably' should be 'inevitably'.","section":"Abstract"},{"comment":"The caption of Figure 2(d) contains the typo 'Aymmetric'; it should be 'Asymmetric'.","section":"Figure 2"},{"comment":"The caption has a typo: 'comparsion' should be 'comparison'.","section":"Table I caption"},{"comment":"The legend lists Balle18, Minnen18, Cheng20, Xie21, Qian22, TinyLIC, and Ours, but the text claims comparisons with TCM; clarify whether TCM is included in Figure 4 and, if so, why it is missing from the legend.","section":"Figure 4"},{"comment":"The ablation for the hyperprior path (Table III) varies the number of slices and the structure simultaneously; it would be clearer to state which factors are held fixed when comparing 'slice number = 5, Ours' with the other rows.","section":"Section IV-C"}],"recommendation":"major_revision","confidential_remarks":"The omission of TCM-small from Table I is the most serious issue: the paper's contribution is defined as a complexity-performance trade-off relative to its direct parent architecture, and without that baseline the headline claim is not verifiable. The two-stage training scheme also needs an ablation to justify its necessity. These are fixable with additional experiments, so I recommend major revision rather than rejection. The paper may benefit from a clearer statement of the evaluation protocol for the MAC/parameter counts of all methods, since the comparison in Table I mixes numbers that may come from different sources."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know. First, the paper does something real: it shows a learned image codec can offload most decoder complexity to the encoder by replacing the synthesis decoder, hyperprior decoder, and context model with lighter versions, and it reports a 51.47 GMACs decoder with -18.68% BD-rate over BPG on Kodak. That is a useful engineering data point for low-end deployment. Second, the central complexity claim is not actually checkable as written because Table I omits TCM-small, the exact base architecture the authors start from. Figure 3 compares against TCM-small, so the data exists, but the table doesn't list its BD-rate or decoder MACs. That is a reporting flaw, not necessarily a flaw in the method.\n\nWhat is new: the asymmetric training strategy — first fine-tune the lightweight synthesis decoder with a distortion-only loss, then train the hyperprior decoder and context model with rate-distortion loss — and the reversed-pyramid decoder channel design. Removing shifted windowing from Swin blocks is simple and well-motivated. The ablations in Tables II and III are meaningful: the proposed decoder gets the best PSNR at second-lowest MACs among the compared structures, and the 5-slice context model with their simplifications beats fewer-slice variants at similar complexity. These are honest engineering choices backed by controlled comparisons within the paper.\n\nSoft spots, in proportion. The missing TCM-small row is the big one. The contribution is defined as a complexity-performance tradeoff, so without TCM-small in Table I the reader cannot tell whether AsymLLIC improves on its parent or simply trades away RD for a modest complexity cut. That is load-bearing for the headline claim. Also: no code or weights, no variance across runs, and the \"comparable to VVC\" language is too loose — the RD curves in Figure 4 sit below VTM, not really on top of it. The two-stage training scheme is plausible but the paper never compares it to joint training of the full asymmetric network, leaving the central training assumption untested. These are not fatal objections; the method is sensible, and the reported numbers are internally consistent.\n\nWho this is for: anyone working on practical learned codecs, especially for edge devices. It deserves a serious referee, but the referee should demand the TCM-small baseline, code/weights, and a sharper VVC comparison. If those are added, this is a solid paper; as is, the headline complexity advantage is underdetermined.","headline":"Solid engineering paper with a real asymmetric decoder contribution, but the missing TCM-small baseline makes its headline complexity claim underdetermined as written.","tokens_in":9494,"tokens_out":2244,"would_cite":true,"duration_ms":23796,"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":"AsymLLIC shows that a learned image codec can keep its heavyweight encoder and still match VVC with a decoder that costs only 51.47 GMACs and 19.65M parameters.","keywords":["learned image compression","asymmetric encoder-decoder","lightweight decoder","two-stage training","rate-distortion optimization","Swin Transformer simplification","hyperprior context model","decoder complexity"],"falsifier":"Run the released model on a mobile CPU and measure per-image decode time against TCM under identical conditions; if AsymLLIC is not faster, or if the Kodak BD-rate cannot be reproduced near 18.68%, the central claim fails.","tokens_in":8511,"feed_emoji":"🗜️","tokens_out":8080,"duration_ms":65274,"temperature":0.7,"pith_summary":"Learned image codecs usually pair a heavy encoder with a heavy decoder, which makes decoding slow on low-end devices. This paper proposes AsymLLIC, a learned codec that deliberately shifts computational cost to the encoder and keeps the decoder light, using a two-stage training scheme that gradually swaps complex decoder modules for simpler ones. On Kodak, AsymLLIC reportedly beats BPG by 18.68% BD-rate while using only 51.47 GMACs and 19.65M parameters for decoding, and its rate-distortion curves sit close to VVC. If correct, this is the first learned codec to decouple encoder and decoder complexity to this degree, suggesting that efficient decoding on constrained devices does not have to sacrifice compression performance.","feed_headline":"Learned codec cuts decoder MACs to 51.47G, matches VVC","feed_subtitle":"AsymLLIC shifts computation to the encoder, reaching near-VVC quality with a 51.47-GMACs decoder.","key_machinery":"The central mechanism is the asymmetric two-stage training strategy. In stage one, the synthesis decoder is swapped for a lighter structure and fine-tuned with a distortion-only loss while the encoder and entropy model are frozen; in stage two, the hyperprior decoder and context model are swapped for lighter versions and retrained together with the hyperprior encoder using the full rate-distortion loss. This staged substitution is what lets the asymmetric architecture hold onto RD performance while cutting decoder MACs from 145.71 in TCM to 51.47 in AsymLLIC.","core_discovery":"The paper claims that the decoder side of a learned image codec can be drastically simplified without a proportionate loss in rate-distortion performance, provided the simplification is done with an asymmetric training schedule rather than by simply shrinking the network. In AsymLLIC, the complex synthesis decoder is first replaced by a lightweight version and fine-tuned alone with a distortion-only loss while the encoder and entropy model are frozen; then the hyperprior decoder and context model are replaced by lighter versions and retrained jointly with the hyperprior encoder under the full rate-distortion loss. The resulting model reports an 18.68% BD-rate improvement over BPG on Kodak with a decoder of 51.47 GMACs and 19.65M parameters, cutting decoder MACs to about one third of the TCM baseline while keeping RD performance close to VVC. The paper interprets this as evidence that computation can be effectively offloaded to the encoding side in learned image compression.","pith_inferences":["Because the stage-one synthesis decoder is fine-tuned with distortion only, that step may act as a form of knowledge distillation from the complex decoder; if so, the same gain could be obtained by direct distillation, which would be a cheaper alternative to the two-stage schedule.","The paper's MACs and parameter counts are theoretical; actual decoding latency on a given device also depends on memory bandwidth, kernel implementation, and quantization, so the practical speedup over TCM could differ from the 2.8x MAC ratio.","The reversed pyramid channel structure and removal of shifted window attention suggest that global receptive field matters less on the reconstruction path than on the analysis path; this asymmetry might transfer to other image restoration tasks.","The claim that the design methodology applies to 'any LIC model' is broader than what the experiments show, since only TCM-small is tested; verifying it on stronger baselines like full TCM or transformer-only codecs would test the generalization."],"forward_implications":["Decoder-side complexity in learned image codecs can be treated as a separate design variable, allowing a heavy encoder on a server and a light decoder on client devices.","The same two-stage training recipe can be applied to other symmetric learned codecs beyond TCM-small, potentially giving them similarly light decoders without retraining from scratch.","A decoder of 51.47 GMACs and 19.65M parameters brings learned image compression within reach of mobile and embedded hardware that could not run symmetric models like TCM.","The authors state the asymmetric design can be extended to learned video compression, which could make efficient video decoding on low-end devices feasible."],"supporting_citations":[{"why":"the TCM mixed transformer-CNN base architecture whose decoder this paper modifies.","marker":"[11]"},{"why":"introduced the scale hyperprior that the entropy model is built on.","marker":"[5]"},{"why":"introduced joint autoregressive and hierarchical priors, which motivates the slice-based context model.","marker":"[6]"},{"why":"introduced discretized Gaussian mixture likelihoods used in the rate-distortion loss.","marker":"[7]"},{"why":"the VVC standard whose rate-distortion performance the paper claims to match.","marker":"[4]"},{"why":"the BPG/HEVC intra anchor used for BD-rate calculations.","marker":"[3]"},{"why":"a lightweight learned codec used as a comparison baseline.","marker":"[10]"},{"why":"an enhanced invertible encoding baseline used in comparisons.","marker":"[8]"}],"fun_headline_variants":["Asymmetric training slims decoder to 51.47 GMACs","Learned codec offloads compute to encoder, keeps VVC parity","Decoder MACs cut to one-third with asymmetric training","AsymLLIC: lightweight decoder matches VVC at 51.47 GMACs","Shift complexity to encoder for simpler learned decoding"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The two-stage training procedure must converge to reconstruction quality as good as jointly training the full asymmetric network would; if the frozen encoder's latents are incompatible with the simplified decoder, the reported rate-distortion gains could degrade substantially.","fun_headline_variants_meta":{"raw":{"variants":["Asymmetric training slims decoder to 51.47 GMACs","Learned codec offloads compute to encoder, keeps VVC parity","Decoder MACs cut to one-third with asymmetric training","AsymLLIC: lightweight decoder matches VVC at 51.47 GMACs","Shift complexity to encoder for simpler learned decoding"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000163,"raw_usage":{"total_tokens":1221,"prompt_tokens":901,"completion_tokens":320,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":517,"completion_tokens_details":{"reasoning_tokens":229}},"tokens_in":517,"tokens_out":320,"duration_ms":3073,"temperature":1.0,"reasoning_tokens":229,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T05:38:05.335891+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the released model on a mobile CPU and measure per-image decode time against TCM under identical conditions; if AsymLLIC is not faster, or if the Kodak BD-rate cannot be reproduced near 18.68%, the central claim fails.","supporting_citations":[{"cited_title":"Learned image compression with mixed transformer-cnn architectures,","cited_arxiv_id":null,"evidence_quote":"the TCM mixed transformer-CNN base architecture whose decoder this paper modifies."},{"cited_title":"Learned image compression with discretized gaussian mixture like- lihoods and attention modules,","cited_arxiv_id":null,"evidence_quote":"introduced discretized Gaussian mixture likelihoods used in the rate-distortion loss."},{"cited_title":"Versatile video coding– towards the next generation of video compression,","cited_arxiv_id":null,"evidence_quote":"the VVC standard whose rate-distortion performance the paper claims to match."},{"cited_title":"Overview of the high efficiency video coding (hevc) standard,","cited_arxiv_id":null,"evidence_quote":"the BPG/HEVC intra anchor used for BD-rate calculations."},{"cited_title":"Enhanced invertible encoding for learned image compression,","cited_arxiv_id":null,"evidence_quote":"an enhanced invertible encoding baseline used in comparisons."}],"review_version":1}