{"id":"57595085-eeaa-482b-843d-a56ec171042f","arxiv_id":"2412.00437","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"DeepFGS is a learned image codec that produces a single fine-grained scalable bitstream, outperforming prior scalable codecs on Kodak in PSNR and MS-SSIM.","lead":"Researchers introduce DeepFGS, a learned image compression system whose single compressed stream can be cut at many points, with each chunk still decoding into a usable image of a different quality level. The design separates image information into essential and optional layers, then orders the optional layers by importance so quality rises smoothly as more bits arrive.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 'truncate at any position' claim is unsupported: no bitstream syntax shows how arithmetic-coded channel prefixes are located and decoded after arbitrary byte truncation; experiments use pre-selected channel counts, not true bitstream truncation.","rationale":"The reader's weakest assumption identifies the same load-bearing gap: the bitstream syntax for truncation is unspecified, and the experiments do not actually truncate at the byte level. My read does not change the conditional verdict. The proposed method is plausible and the channel-level RD gains are meaningful, but the contribution is explicitly framed as a scalable bitstream, and a bitstream has byte-level semantics. An arithmetic coder is not automatically prefix-decodable at arbitrary positions: its output is designed to be decoded from the beginning, and a byte prefix can be decoded into a prefix of the symbol sequence only with careful design. The paper does not provide that design, and the 'use channel 0-k' protocol sidesteps the question. This is a concrete, correctable deficiency rather than a demonstrated flaw: there may exist a natural syntax, such as independently arithmetic-coded channel segments with a parseable header, that makes truncation at segment boundaries work, though 'any position' would still need stronger guarantees. Because the missing artifact is a specification and an empirical test, conditional acceptance is appropriate; the paper should not be rejected outright, nor accepted without the bitstream-level evidence.","tokens_in":7376,"tokens_out":4711,"duration_ms":52313,"concrete_test":"Obtain the actual encoder/decoder (requires released code or a detailed bitstream specification) and perform a byte-truncation audit on Kodak: encode each image to a full bitstream, truncate at every byte offset, and decode with no external channel-count information. Record (i) the fraction of truncations that decode without error, (ii) whether the decoded channel count is recoverable from the prefix, and (iii) whether PSNR and MS-SSIM increase monotonically with truncation length. If any byte offset yields a decode failure or a quality regression, the 'any position' claim is false. Additionally, provide a formal bitstream syntax (segment headers, termination rules, or range-coder flush semantics) and prove that any prefix uniquely identifies a valid channel subset; without such a specification, the existing channel-selection RD curves do not demonstrate bitstream scalability.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim is that a single encoded bitstream can be truncated at any position and still decode to a useful, monotonically improving image. For this to hold, the arithmetic-coded representation must be prefix-decodable at byte granularity: from a prefix of the bitstream alone, the decoder must recover how many latents or channels are present and remain synchronized. The paper never specifies this syntax. The mutual entropy model and Eqs. (4)-(5) sum rates R(y_b) + R(y^{\\le i}_s) for channel prefixes, but a rate is not a bitstream layout. Figure 1 and the RD experiments decode 'channel 0-191', 'channel 0-223', etc., with k known in advance and evaluated at 8-channel or 1-channel intervals. That is a channel-selection protocol, not byte-level truncation. Without an explicit coder design—e.g., independently terminated channel segments with locatable boundaries, or a single arithmetic coder whose symbol ordering and flush behavior make every byte prefix correspond to a valid channel subset—an arbitrary truncation can land inside a codeword and cause invalid or mis-synchronized decoding. The one-pass encoding claim also depends on the decoder knowing k without side information; if the channel count is signaled externally, that overhead is unaccounted. Thus the central 'continuously scalable bitstream via one-pass encoding' claim is not yet substantiated by the evidence presented.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"DeepFGS proposes a learned image compression framework that generates a single flexible bitstream whose prefixes can be decoded into images of progressively increasing quality. The method separates latent features into a basic part and a scalable part using a feature separation backbone, applies an information rearrangement strategy to order scalable channels by importance, employs a mutual entropy model that conditions scalable latents on basic latents, and reuses the decoder through a feature fusion module to reduce complexity. Experiments on the Kodak dataset report rate-distortion curves showing gains over traditional scalable codecs (SVC and SHVC) and prior learned scalable models (Jia et al., Mei et al., Guo et al., Zhang et al.) in both PSNR and MS-SSIM, and near-competitive performance with non-scalable learned codecs. The paper's central claim is that this is a fine-grained, one-pass scalable coding method where the bitstream can be truncated at any position and still decode to a complete image.","tokens_in":7622,"tokens_out":5269,"duration_ms":51164,"significance":"If validated, a one-pass bitstream that supports fine-grained truncation at byte granularity would be practically valuable for adaptive streaming and progressive transmission. The training strategy of sampling random truncation points is a sensible way to induce a channel ordering, and the decoder-multiplexing idea is a reasonable complexity reduction. The paper is empirical and does not claim formal derivations; its value rests on the strength of the experimental evidence. However, the manuscript currently lacks a specification of the actual bitstream syntax, which is essential for the central truncation claim, and the experimental evaluation does not include variance estimates or released artifacts.","major_comments":[{"comment":"The central claim that the bitstream can be truncated at any position is not substantiated. The paper describes rates R(yb) and R(ys^{<=i}) in Eqs. (3)-(5) as sums of estimated entropies, but an entropy estimate is not a bitstream layout. The experiments decode pre-selected channel counts (e.g., 'Use channel 0-191' in Fig. 1 and the 8-channel or 1-channel intervals in Fig. 6) with k known in advance, which is a channel-selection protocol rather than byte-level truncation. Without specifying how arithmetic-coded channel segments are terminated, concatenated, and located after an arbitrary byte cut, a truncation can land inside a codeword and cause desynchronization. Please provide a precise bitstream syntax, clarify how the decoder determines the number of channels present, and report experiments that truncate the actual encoded bitstream at byte positions.","section":"Introduction and Experiments (Fig. 1 and Fig. 6)"},{"comment":"The claim that no side information is needed because the channel ordering is learned during training is ambiguous. If the permutation is a fixed global ordering applied to every image, the statement is plausible, but the paper does not state this explicitly and the visualization in Fig. 3 is anecdotal. If the ordering is computed per image at inference time, the ordering index must be transmitted, and that overhead is not accounted for in the rate calculation. Please clarify whether the rearrangement is a fixed trained permutation and discuss how the decoder knows both the ordering and the number of decoded channels from the truncated bitstream.","section":"Information Rearrangement Strategy"},{"comment":"The rate-distortion comparisons are presented as single curves with no error bars, multiple training runs, or statistical significance tests. This is particularly relevant for the ablation study in Fig. 7 and the comparisons to prior scalable methods, where the claimed improvements need to be robust to training seed and initialization. Please provide at least three runs per configuration with mean and standard deviation, or state clearly that the curves are from a single run and temper the corresponding claims.","section":"Experiments and Fig. 6"},{"comment":"The sampling distribution of j in Eq. (5) is not specified, and Eq. (4) counts R(yb) once in every summand, which disproportionately weights the base-layer rate in the multi-term loss. The effect of this weighting on the learned channel ordering is not analyzed. Please state the sampling distribution (e.g., uniform over the range) and discuss how the choice of w(i) and the truncation schedule affect the granularity and rate range of the final RD curve.","section":"Eqs. (4)-(5)"}],"minor_comments":[{"comment":"The legend in Fig. 6 contains garbled citation strings such as '/s48 /s46 /s54/s48'; these should be replaced with readable labels that map to the references in the caption.","section":"Figure 6"},{"comment":"The notation 'j is sampled from [C1, C2] ∩ N' is confusing; it should be 'j is sampled uniformly from {C1, C1+1, ..., C2}' or the actual distribution should be stated.","section":"Eq. (5)"},{"comment":"The architecture of the mutual entropy model is shown in Fig. 4 but not described in the text; a brief explanation of how the hyperprior and prior fusion are combined would improve readability.","section":"Figure 4 and Mutual Entropy Model"},{"comment":"The statement that 'our model arranges the channels in descending entropy order' should be tempered because the entropy ordering is an emergent property of training rather than an explicit sorting operation; suggest phrasing such as 'the learned ordering tends to have decreasing entropy'.","section":"Figure 3 discussion"},{"comment":"There are minor language issues, including 'the two locations share the same decoder' grammar, and the phrase 'nearly continuously adjustable' should be defined (e.g., 'the RD curve is sampled at intervals of one channel').","section":"General"},{"comment":"The Experiments section mentions 'more results on the high-definition data set will be given in the Appendix,' but no appendix is included in the manuscript.","section":"Appendix"}],"recommendation":"major_revision","confidential_remarks":"The paper's main contribution is an interesting engineering system, but the truncation claim is the load-bearing element and it is currently unsupported at the bitstream level. If the authors cannot demonstrate true byte-level truncation with a defined entropy-coding syntax, the title and abstract overclaim and should be revised. I would also encourage the editor to require the release of code and model weights, given the empirical nature of the paper and the impossibility of reproducing the curves from the current text."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear colleague,\n\nDeepFGS is a genuine step forward for fine-grained scalable learned image compression. The idea of training a single model so that prefixes of the latent channel ordering form a smooth quality ladder is new among the cited works, and the feature separation backbone, the optimization-driven rearrangement loss, mutual entropy model, and decoder reuse are all coherent and well ablated. On Kodak, it beats the scalable baselines clearly and lands close to non-scalable cheng2020. That part is credible.\n\nThe main soft spot is the gap between the 'truncate at any position' claim and what is actually demonstrated. The experiments decode at channel boundaries (8-channel or 1-channel intervals), not arbitrary byte truncations. There is no bitstream syntax showing how arithmetic-coded channel segments are terminated, located, and decoded when k is unknown. If each channel is an independently terminated arithmetic segment with fixed order, channel-boundary truncation works and byte-level truncation only works to the end of a segment (unless you add a parser). The paper needs to state this or weaken the claim. This is a revision issue, not a fatal one.\n\nOther issues are minor: no code/weights, single-run curves without error bars, and Eq. (5)'s sampling range '[C1, C2]' is a typo. The no-side-information argument is fine for a fixed ordering, but the decoder still needs to know k, which is part of the missing spec.\n\nThis deserves a serious referee. The method and results are worth careful checking, and the revision path is clear: clarify the coder layout, fix the typo, release code if possible. Recommend acceptance with major revision.","headline":"A genuinely fine-grained scalable learned codec with strong Kodak results, but the 'truncate at any position' claim needs a bitstream-syntax reality check.","tokens_in":8189,"tokens_out":3072,"would_cite":true,"duration_ms":29534,"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":"DeepFGS is a learned image codec that produces a single bitstream which can be truncated at nearly any channel boundary and decoded into an image whose quality grows smoothly with the number of channels received, while coming close to the…","keywords":["learned image compression","scalable coding","fine-grained scalability","bitstream truncation","feature separation","mutual entropy model","decoder multiplexing","rate-distortion optimization"],"falsifier":"Encode a set of held-out images with a trained DeepFGS model, decode every prefix of each stream, and record PSNR; if any prefix scores below the previous prefix (or below the basic-layer reconstruction), the forward-dependence ordering has failed for that image. A sharper test is to measure the per-channel entropies of the encoded scalable latents and compare their rank order with the fixed order learned during training: low rank correlation across images would show that the no-side-information transfer assumption does not hold universally.","tokens_in":7131,"feed_emoji":"🖼️","tokens_out":9204,"duration_ms":88361,"temperature":0.7,"pith_summary":"This paper tries to establish that learned image compression can be made finely scalable without the usual penalty: one encoding pass produces a bitstream whose prefixes are each useful, and decoding more of the stream (more channels of the latent representation) yields smoothly better reconstructions. The proposed framework, DeepFGS, separates image information into basic features and scalable features, re-orders the scalable channels by importance during training, and conditions the entropy model of the scalable part on the basic part. If correct, this means a network sender can serve many bandwidths from a single stored stream, and the rate–distortion loss relative to non-scalable codecs can be kept small. The paper reports that DeepFGS outperforms prior learned and traditional scalable codecs in PSNR and MS-SSIM and nearly closes the quality gap to a strong non-scalable learned codec.","feed_headline":"One bitstream, any bandwidth: learned codec scales by truncation","feed_subtitle":"Single-pass encoding yields continuous rate-quality points, with little loss against non-scalable learned codecs.","key_machinery":"The carrying mechanism is the combination of a feature separation backbone and a channel-wise information rearrangement strategy: the backbone splits the image into basic features $y_b$ and scalable features $y_s$, and the rearrangement—implemented not by physically permuting features but by a sampled training loss over prefixes—makes each channel of $y_s$ depend only on earlier channels, turning the bitstream into a quality ladder that can be cut at each channel boundary. A second mechanism is the mutual entropy model, a conditional entropy coder in which $\\hat{y}_s$ is coded using probabilities conditioned on $\\hat{y}_b$, lowering the rate by the mutual information between the two latents. A third is decoder multiplexing with a feature fusion module, allowing the basic-only and full reconstructions to share one decoder.","core_discovery":"The central claim is that fine-grained scalability can be built into a learned codec by construction rather than by training multiple codecs or iterating a recurrent network. The authors claim that the feature separation backbone yields two latent sets, the basic features $y_b$ and the scalable features $y_s$; the information rearrangement strategy, driven by a sampled prefix loss, orders the channels of $y_s$ so that each prefix $y_s^{\\le i}$ adds a smooth quality gain; and the mutual entropy model encodes $\\hat{y}_s$ conditional on $\\hat{y}_b$, using $H(\\hat{y}_s|\\hat{y}_b)=H(\\hat{y}_s)-I(\\hat{y}_s,\\hat{y}_b)$ to reduce the rate. With decoder reuse through the feature fusion module, this single framework is claimed to outperform existing scalable codecs and to approach the rate–distortion performance of non-scalable learned models.","pith_inferences":["Editorial inference: the truncated-stream property makes the codec a natural fit for network-adaptive delivery and edge caching, where a server can send a longer byte range when bandwidth allows; the paper does not implement such a protocol or measure end-to-end latency.","Editorial inference: the learned channel ordering is fixed after training and content-agnostic; a per-image or per-content-class ordering could improve the rate–distortion trade-off further, at the cost of side information that the current design deliberately avoids.","Editorial inference: the same basic-versus-scalable separation and mutual conditioning could be carried into learned video compression, using the previous decoded frame as the 'basic' condition for the current frame's scalable latents; the paper leaves this uncharted.","Editorial inference: the demonstration of 'any position' truncation is at channel boundaries; true byte-level truncation inside an arithmetic-coded segment would require an entropy-coding resynchronization mechanism that the paper does not specify."],"forward_implications":["A single DeepFGS bitstream can be cut at any channel boundary and still decode to a complete image, with quality rising smoothly from the basic layer up to the full stream.","Bandwidth adaptation requires no re-encoding and no storage of multiple versions of the same image, because one pass produces the entire scalable stream.","Coding the scalable latents conditioned on the basic latents saves rate to the extent that $I(\\hat{y}_s,\\hat{y}_b)$ is nonzero, and the reported experiments show this improves the rate–distortion curve.","The gap between scalable and non-scalable learned compression can be small: DeepFGS is reported near an attention-based non-scalable codec and close to its own single-rate variant.","Reusing the decoder with feature fusion keeps parameter and computation cost below that of multi-decoder scalable systems."],"supporting_citations":[{"why":"Layered scalable auto-encoder baseline that DeepFGS compares against and whose multiple-codec design motivates decoder multiplexing.","marker":"[2]"},{"why":"RNN-based variable-rate codec representing the coarse-grained scalable alternative with high encoding–decoding complexity.","marker":"[3]"},{"why":"Scale-hyperprior model used in the motivating experiment showing that latent channel entropies are disordered in non-scalable codecs.","marker":"[7]"},{"why":"Learning-based scalable codec with independent basic/enhancement entropy models; the mutual entropy model improves on that independent treatment.","marker":"[8]"},{"why":"Attention-based non-scalable learned codec used as the performance reference that DeepFGS approaches in rate–distortion.","marker":"[14]"},{"why":"Non-scalable learned codec included in the rate–distortion comparison to show the gap to scalable coding.","marker":"[15]"},{"why":"SVC standard providing the traditional scalable codec baseline that DeepFGS outperforms.","marker":"[17]"},{"why":"Spatial channel attention module whose design inspires the feature-level redundancy reduction and feature fusion modules.","marker":"[6]"}],"fun_headline_variants":["Single pass, many rates: DeepFGS scales learned compression","Fine-grained bitstream truncation: one model, any rate","DeepFGS: channel reordering unlocks smooth scalable coding","Scalable learned codec: one pass, many rates, no retraining","Truncate anytime: DeepFGS approaches non-scalable rate-distortion"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The scheme rests on the assumption that the channel ordering and mutual-entropy model learned during training remain valid for every image seen at inference, so the decoder can reconstruct any prefixed subset of the bitstream without side information about which channels were used or how the arithmetic-coded segments were synchronized.","fun_headline_variants_meta":{"raw":{"variants":["Single pass, many rates: DeepFGS scales learned compression","Fine-grained bitstream truncation: one model, any rate","DeepFGS: channel reordering unlocks smooth scalable coding","Scalable learned codec: one pass, many rates, no retraining","Truncate anytime: DeepFGS approaches non-scalable rate-distortion"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000915,"raw_usage":{"total_tokens":3899,"prompt_tokens":888,"completion_tokens":3011,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":504,"completion_tokens_details":{"reasoning_tokens":2919}},"tokens_in":504,"tokens_out":3011,"duration_ms":20969,"temperature":1.0,"reasoning_tokens":2919,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T05:22:29.328238+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Encode a set of held-out images with a trained DeepFGS model, decode every prefix of each stream, and record PSNR; if any prefix scores below the previous prefix (or below the basic-layer reconstruction), the forward-dependence ordering has failed for that image. A sharper test is to measure the per-channel entropies of the encoded scalable latents and compare their rank order with the fixed order learned during training: low rank correlation across images would show that the no-side-information transfer assumption does not hold universally.","supporting_citations":[{"cited_title":"Layered image compression using scalable auto-encoder,","cited_arxiv_id":null,"evidence_quote":"Layered scalable auto-encoder baseline that DeepFGS compares against and whose multiple-codec design motivates decoder multiplexing."},{"cited_title":"Joint autoregressive and hierarchical priors for learned image compression,","cited_arxiv_id":null,"evidence_quote":"Scale-hyperprior model used in the motivating experiment showing that latent channel entropies are disordered in non-scalable codecs."},{"cited_title":"Learning-based scalable image compression with latent-feature reuse and prediction,","cited_arxiv_id":null,"evidence_quote":"Learning-based scalable codec with independent basic/enhancement entropy models; the mutual entropy model improves on that independent treatment."},{"cited_title":"Learned image compression with discretized gaussian mixture likelihoods and attention modules,","cited_arxiv_id":null,"evidence_quote":"Attention-based non-scalable learned codec used as the performance reference that DeepFGS approaches in rate–distortion."},{"cited_title":"Text of iso/iec 14496-10: 2005/fdam 3 scalable video coding,","cited_arxiv_id":null,"evidence_quote":"SVC standard providing the traditional scalable codec baseline that DeepFGS outperforms."},{"cited_title":"Cbam: Convolu- tional block attention module,","cited_arxiv_id":null,"evidence_quote":"Spatial channel attention module whose design inspires the feature-level redundancy reduction and feature fusion modules."}],"review_version":1}