{"id":"78743ec6-682f-4f62-a620-f76f1da041d4","arxiv_id":"2505.01742","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Easz compresses images by erasing and squeezing patches on the edge, then reconstructs them on a server with a lightweight transformer at flexible ratios and low edge cost.","lead":"This paper proposes Easz, an image compression framework that erases parts of an image on an edge device and reconstructs the missing content on a server with a lightweight transformer. It aims to make learned-style compression practical on weak IoT hardware by shifting computation off the edge and supporting flexible compression ratios.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Reconstruction-quality claim is confounded: no-reference metrics may reward erasure-induced smoothing, and Kodak results come after Kodak fine-tuning (Sec. IV-D).","rationale":"The reader's weakest assumption correctly identifies the quality evaluation as the precarious link in the central claim. My stress-test agrees: the paper's own statement that reference-based metrics are avoided because erasure impacts them (Sec. IV-A) is a warning that the method may trade fidelity for perceptual smoothness. The fine-tuning on Kodak before reporting Kodak results (Sec. IV-D and Table II) is a concrete methodological flaw that directly undermines the headline quality numbers. The efficiency and adaptability claims are on firmer ground because they are based on direct testbed measurements (Sec. IV-B) and the design of the erase-squeeze process, which is genuinely lightweight on the edge. Therefore the verdict remains CONDITIONAL: the paper should be accepted only if the reconstruction quality claim survives an independent evaluation with a clean protocol (no test-set fine-tuning) and metrics that are not trivially gamed by erasure. I did not find an additional decisive flaw in the complexity analysis; the arithmetic inconsistency in the worked example (Sec. III-B) is a minor correctness issue that does not affect the qualitative conclusion of reduced attention complexity.","tokens_in":11066,"tokens_out":6183,"duration_ms":58993,"concrete_test":"Run a forced-choice human perceptual study (e.g., ITU-T P.910) comparing Easz-processed images against the original compressed baselines at matched BPP on Kodak and CLIC, using a reconstruction model trained without any Kodak fine-tuning. If human raters do not significantly prefer Easz, the reconstruction quality claim should be downgraded.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim of improved reconstruction quality is the least secure part of the paper. In Sec. IV-A the authors explicitly avoid PSNR/SSIM when comparing with compression baselines because 'removing content would generally impact reference-based metrics', and instead rely on Brisque, Pi, and Tres. These no-reference metrics are sensitive to smoothness and artifact removal; erasing blocks and reconstructing with a transformer can reduce compression noise and thereby inflate scores without preserving true detail. More concretely, Sec. IV-D reports that fine-tuning the reconstruction model on the Kodak dataset improves training loss, and Table II then reports performance on the Kodak dataset. Unless the fine-tuned model is withheld from the Kodak evaluation, the Kodak rows in Table II are not independent and the claimed quality advantage is partly a test-set adaptation. The efficiency and adaptability claims (edge latency, power, mask flexibility) are supported by direct measurements and do not depend on this critique, but the 'reconstruction quality' advantage is load-bearing for the paper's overall conclusion and is not yet established.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Easz, an edge-server image compression framework in which the edge device performs only a cheap erase-and-squeeze operation on image patches using a row-based conditional sampler, while a server-side lightweight transformer reconstructs the erased content. The framework is designed to wrap existing codecs such as JPEG, BPG, and learned compressors, and is evaluated on a Jetson TX2 plus RTX 2080Ti testbed. The authors claim that Easz reduces edge latency, power, and memory, provides flexible compression-level adjustment without model switching, and improves reconstruction quality relative to both super-resolution methods and standard codecs.","tokens_in":11184,"tokens_out":4206,"duration_ms":42827,"significance":"The core architectural idea is practically motivated and potentially valuable: moving essentially all neural computation to the server while keeping the edge operation as a simple masking and repacking step is a clean way to address the resource constraints of IoT devices. The measured edge latency, power, and memory advantages reported in Fig. 6 are concrete and credible, and the model-size reduction to 8.7MB is a useful contribution. The flexibility of the row-based sampler is also a strength. However, the paper's headline claim of improved reconstruction quality is not currently established because the evaluation procedure leaks test-set information and relies on no-reference metrics that are acknowledged to favor the erase-and-reconstruct pipeline. The significance of the paper therefore depends on whether the evaluation can be repaired.","major_comments":[{"comment":"The reconstruction model is fine-tuned on the Kodak dataset (Fig. 7d) and then Table II reports the headline quality comparisons on the same Kodak dataset. This is a test-set leakage: the Kodak rows in Table II do not measure generalization to unseen images. To support the quality claim, the authors must either fine-tune only on a held-out portion of Kodak and evaluate on a disjoint test portion, or fine-tune on a different dataset and report results on Kodak and CLIC without Kodak fine-tuning. Without this change, the Kodak quality comparisons in Table II are not independent evidence.","section":"Sec. IV-D, Fig. 7d, Table II"},{"comment":"The paper explicitly avoids PSNR and SSIM when comparing with compression baselines because 'removing content would generally impact reference-based metrics,' and instead uses the no-reference metrics Brisque, Pi, and Tres. No-reference metrics can reward smoothing and artifact removal even when true detail is lost, so the reported quality improvements may partly reflect the smoothing effect of erasure and reconstruction rather than faithful regeneration. The authors should report reference-based metrics (PSNR, SSIM, and ideally LPIPS) for all methods at matched bitrates, and should also provide a perceptual user study or other validation that the no-reference gains correspond to genuine quality improvement. This is load-bearing because the conclusion that Easz 'enhances reconstruction quality' rests on these metrics.","section":"Sec. IV-A, Tables I-II, Fig. 8"},{"comment":"The complexity derivation contains a dimensional error. The text states that after two-stage patchification 'there will be (hw)^2/n^2 × n^2/b^2 sub-patches,' but the total number of sub-patches in an h×w image should be hw/b^2, not (hw)^2/b^2. The subsequent numeric example for n=32, b=4 gives 1,048,576 required attention operations, whereas the correct value is 262,144. The qualitative conclusion that patchification reduces attention complexity is correct, but the equations and the claimed '4096 times' reduction must be corrected and recomputed.","section":"Sec. III-B, complexity analysis"},{"comment":"The comparison with super-resolution methods in Table I does not state the reduction ratio, scale factor, or bitrate at which each method is evaluated. Without matched experimental conditions, the reported PSNR and SSIM differences cannot be interpreted. The authors should specify the erase ratio or scale factor for Easz and for each SR baseline, and ideally report bitrate so that the comparison is rate-constrained.","section":"Sec. IV-C, Table I"}],"minor_comments":[{"comment":"The caption of Fig. 7d says 'MSE,' but the y-axis is labeled 'Loss value' and the training loss in Eq. (2) is L1 plus LPIPS, not MSE. Please align the caption, axis label, and loss definition.","section":"Sec. IV-D, Fig. 7d"},{"comment":"The phrase 'non-reference perceptual metrics' should be 'no-reference perceptual metrics' throughout the section.","section":"Sec. IV-A"},{"comment":"Reference [26] is cited as 'K. Alex and H. Geoffrey'; the standard citation is A. Krizhevsky and G. Hinton, 'Learning multiple layers of features from tiny images,' 2009. Please correct the author names.","section":"Reference [26]"},{"comment":"The inter-row constraint uses T both as the number of sampled columns per row and as the index of the last sample in the previous row. Please clarify the notation, for example by writing the previous row's samples as a set and requiring separation from all of them.","section":"Sec. III-A, Eq. (1)"}],"recommendation":"major_revision","confidential_remarks":"The architectural idea and edge-efficiency measurements are the strongest parts of the paper. The main concern is evaluation integrity: fine-tuning on Kodak and then reporting Kodak results, combined with the reliance on no-reference metrics, makes the central quality claim unsupported as written. This is fixable with a properly split evaluation and additional metrics, so I recommend major revision rather than rejection. I would also ask the authors to correct the complexity arithmetic before the paper is reconsidered."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nHere is my read of Easz. The new piece is a generalized row-based erase sampler for edge-to-server image compression, plus a lightweight two-stage patchified transformer that reconstructs erased blocks on the server. The special cases reduce to known diagonal masks and 2x super-resolution, so the novelty is incremental, but the generalization to arbitrary erase ratios with one model is real and useful for IoT.\n\nWhat the paper does well is measure the edge-side cost. The Jetson TX2 numbers are direct, repeated 24 times, and the comparison against MBT and Cheng-anchor is on point: erase-and-squeeze costs almost nothing on the edge, and the power and memory savings are large. The server-side reconstruction being heavier is acknowledged up front. That part of the paper should survive review.\n\nThe soft spot is the quality evaluation, and it is load-bearing. Section IV-D shows the reconstruction model was fine-tuned on Kodak ('Fig. 7d' reports MSE during Kodak fine-tuning). Table II then reports Kodak results for the '+Proposed' rows. Unless the fine-tuning split is disjoint from the evaluation split, those rows are test-set adaptation, not independent results. The paper does not say a separate split was used. That is a real confound and needs to be fixed before the quality claim is taken seriously.\n\nRelatedly, the paper avoids PSNR/SSIM for codec comparisons and uses Brisque, Pi, and Tres instead. As the authors note, erasure lowers reference-based metrics, but no-reference metrics can reward the smoothing and artifact removal that erasure plus transformer reconstruction provides. That means the perceptual-quality wins in Table II and Fig. 8 could partly reflect noise removal rather than real detail preservation. Human-rated or region-based reference metrics would address this.\n\nOne smaller issue: the complexity worked example in Section III-B gives 1,048,576 calculations after patchification for n=32, b=4, but plugging their own formula O(hw*n^2/b^4) yields 262,144. The text says a 4096x reduction from the original 4.29e9, which matches 1,048,576, but the formula and example don't agree. It's a fixable arithmetic inconsistency, but it needs correcting.\n\nThere is also no variance reported for the Table II numbers, and the claim of compatibility with all existing compressors is tested on only four. Minor.\n\nOverall: the efficiency contribution is credible and likely useful for edge vision systems. The reconstruction-quality advantage is not yet established. I'd send this to peer review with the expectation of major revision on the evaluation: separate the fine-tuning split, add human or reference-based quality checks, fix the complexity arithmetic. The systems-side contribution deserves referee time.","headline":"The edge-efficiency story is credible and worth engaging; the reconstruction-quality claim is confounded by Kodak fine-tuning and no-reference metrics.","tokens_in":11801,"tokens_out":2702,"would_cite":true,"duration_ms":26560,"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":"Easz erases image patches on the edge and rebuilds them on the server","keywords":["image compression","edge computing","transformer","erase-and-squeeze","patch erasing","perceptual quality","IoT devices","conditional sampling"],"falsifier":"Take a held-out test set the model has never seen, fine-tune on a separate training set, then compute PSNR and SSIM between the original and Easz-reconstructed images at the same bitrates as JPEG/BPG; if reference-based metrics show Easz performs worse than the baselines while the no-reference metrics show it performing better, the claimed quality improvement would be shown to be an artifact of the chosen metrics. Alternatively, decoding with the model trained without the Kodak fine-tuning step and comparing on Kodak would expose whether test-set tuning drives the reported gains.","tokens_in":10781,"feed_emoji":"🖼️","tokens_out":5412,"duration_ms":47836,"temperature":0.7,"pith_summary":"Easz is an image coding framework that moves the computationally heavy part of neural compression off the edge device and onto the server. On the edge, it applies a row-based conditional sampler to erase a chosen fraction of image sub-patches, squeezes the survivors together, and compresses the result with any existing codec such as JPEG or BPG. The server then runs a lightweight transformer-based encoder-decoder that reconstructs the erased content, restoring the full image. The paper argues that this design avoids the multi-second encoding and model-switching costs of learned compressors on devices like the Jetson TX2, while improving perceptual quality at matched bitrates. If correct, Easz offers a practical way to get learned-quality image compression onto resource-constrained IoT devices.","feed_headline":"Easz erases image patches on the edge and rebuilds them on the server","feed_subtitle":"A row-based sampler squeezes the image cheaply, while an 8.7MB transformer on the server restores it, beating JPEG and learned codecs.","key_machinery":"The load-bearing components are (1) the row-based conditional sampler that generates the binary erase mask, choosing for each row column positions at least δ away from earlier picks in the same row and at least Δ away from the previous row's picks, which yields flexible, well-distributed erasure with a single reconstruction model; (2) the two-stage patchify process, which splits the image into n×n patches and then into b×b sub-patches and runs attention within each patch, reducing the theoretical attention complexity from O((hw)^2) to O(hw·$n^{2}$/$b^{4}$), making pixel-level reconstruction feasible; and (3) the lightweight 8.7MB transformer encoder-decoder, trained with L1 plus LPIPS loss, that takes embeddings of un-erased sub-patches plus zero vectors at erased positions and outputs the restored sub-patches.","core_discovery":"Easz's central claim is that erasing a deliberately spaced subset of image sub-patches at the sender, squeezing the survivors, and reconstructing the erased content on the receiver with a lightweight transformer is enough to match or beat both traditional codecs (JPEG, BPG) and learned compression models (MBT, Cheng-anchor) in perceptual quality at the same bitrate, while cutting the sender's computation to nearly zero. The paper argues that the row-based conditional sampler — which keeps erased blocks non-adjacent both along a row and across neighboring rows — preserves enough local structure that the receiver's model can regenerate the missing detail, and that because erase ratio is just a parameter of the sampler, one trained model serves all compression levels without model switching. In evaluation, Easz improves the perceptual scores of every baseline it is prepended to, and the end-to-end latency on a Jetson TX2 to server testbed is about 89% lower than the neural baselines, with the erase-and-squeeze step taking 0.7% of total latency.","pith_inferences":["The row-based sampler's structure suggests a direct connection to coding theory: it builds a constrained sampling pattern that preserves local correlations; one could test whether optimizing the δ and Δ spacings with a rate-distortion objective rather than fixed thresholds yields further quality or bitrate gains.","The paper's quality comparisons rely on no-reference perceptual metrics; a natural extension is to pair Easz with reference-based or task-based evaluation, such as downstream object detection on reconstructed images, to see whether the perceptual gains translate to machine-vision performance.","Since the reconstruction is trained on CIFAR-10 and fine-tuned on Kodak, the framework's performance on out-of-distribution imagery, such as medical, satellite, or 4K video frames, is an open question; evaluating across diverse domains would test how general the generative prior is.","The two-stage patchify complexity reduction is independent of the erasure scheme and could be reused as a generic way to make pixel-level transformers tractable, not only in compression but in any dense prediction task."],"forward_implications":["Edge devices with no GPU can perform compression-level switching instantly, since changing the erase ratio is just changing a parameter of the sampler, not loading a new model.","Existing deployed compressors (JPEG, BPG, and learned codecs) get a perceptual quality boost at roughly constant bitrate when Easz is prepended, so the framework can be adopted without replacing the codec stack.","The server-side reconstruction model at 8.7MB is small enough to run on modest server GPUs, and the paper's latency breakdown suggests most of its end-to-end latency is reconstruction, which would shrink on faster datacenter hardware.","Because the erase mask is a tiny binary map (128 bytes at 32×32), the sender can adapt bitrate per image or per region with essentially no extra signaling cost."],"supporting_citations":[{"why":"D. Minnen, J. Ballé, and G. Toderici's joint autoregressive hierarchical learned codec (MBT) is one of the two neural baselines Easz is compared against.","marker":"[13]"},{"why":"Z. Cheng et al.'s discretized Gaussian mixture likelihood codec (Cheng-anchor) is the other neural baseline in the quality and latency comparisons.","marker":"[14]"},{"why":"A. Vaswani et al.'s 'Attention is all you need' supplies the transformer attention mechanism used in the reconstruction encoder-decoder.","marker":"[24]"},{"why":"R. Zhang et al.'s LPIPS is the perceptual loss, combined with L1, used in the training objective for the reconstruction model.","marker":"[25]"},{"why":"A. Krizhevsky's CIFAR-10 dataset is used to pretrain the reconstruction model so it acquires generative capabilities.","marker":"[26]"},{"why":"The Kodak image suite is used both to fine-tune the reconstruction model and as a primary test set for compression comparisons.","marker":"[27]"},{"why":"The CLIC dataset provides a second, independent evaluation set for the compression performance comparison.","marker":"[28]"},{"why":"A. Mittal et al.'s BRISQUE is the main no-reference perceptual metric used to compare Easz against compression baselines.","marker":"[29]"},{"why":"Y. Blau et al.'s Pi metric from the PIRM challenge is another no-reference perceptual metric used in the quality comparison.","marker":"[30]"},{"why":"S. A. Golestaneh et al.'s Tres provides the third no-reference perceptual metric used in the evaluation.","marker":"[31]"}],"fun_headline_variants":["Easz: erase on the edge, rebuild on the server","Patch-erase compression shifts compute to server","Transformer image codec cuts edge compute to 0.7%","Easz matches JPEG quality while edge does almost no work","Easz: edge erases, server rebuilds, no model switching"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The claimed quality advantage rests on the load-bearing assumption that the chosen no-reference metrics (BRISQUE, Pi, Tres) are a trustworthy measure of reconstruction quality, because the paper deliberately avoids PSNR/SSIM for its compression comparisons — erasing content lowers those reference metrics by construction — and because the reconstruction model is fine-tuned on the Kodak test set before Kodak results are reported.","fun_headline_variants_meta":{"raw":{"variants":["Easz: erase on the edge, rebuild on the server","Patch-erase compression shifts compute to server","Transformer image codec cuts edge compute to 0.7%","Easz matches JPEG quality while edge does almost no work","Easz: edge erases, server rebuilds, no model switching"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000333,"raw_usage":{"total_tokens":1855,"prompt_tokens":956,"completion_tokens":899,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":572,"completion_tokens_details":{"reasoning_tokens":815}},"tokens_in":572,"tokens_out":899,"duration_ms":9019,"temperature":1.0,"reasoning_tokens":815,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T04:12:02.279942+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a held-out test set the model has never seen, fine-tune on a separate training set, then compute PSNR and SSIM between the original and Easz-reconstructed images at the same bitrates as JPEG/BPG; if reference-based metrics show Easz performs worse than the baselines while the no-reference metrics show it performing better, the claimed quality improvement would be shown to be an artifact of the chosen metrics. Alternatively, decoding with the model trained without the Kodak fine-tuning step and comparing on Kodak would expose whether test-set tuning drives the reported gains.","supporting_citations":[{"cited_title":"Joint autoregressive and hier- archical priors for learned image compression,","cited_arxiv_id":null,"evidence_quote":"D. Minnen, J. Ballé, and G. Toderici's joint autoregressive hierarchical learned codec (MBT) is one of the two neural baselines Easz is compared against."},{"cited_title":"Learning multiple layers of features from tiny images,","cited_arxiv_id":null,"evidence_quote":"A. Krizhevsky's CIFAR-10 dataset is used to pretrain the reconstruction model so it acquires generative capabilities."},{"cited_title":"Kodak Lossless True Color Image Suite","cited_arxiv_id":null,"evidence_quote":"The Kodak image suite is used both to fine-tune the reconstruction model and as a primary test set for compression comparisons."},{"cited_title":"Challenge on Learned Image Compression","cited_arxiv_id":null,"evidence_quote":"The CLIC dataset provides a second, independent evaluation set for the compression performance comparison."},{"cited_title":"The 2018 pirm challenge on perceptual image super-resolution,","cited_arxiv_id":null,"evidence_quote":"Y. Blau et al.'s Pi metric from the PIRM challenge is another no-reference perceptual metric used in the quality comparison."},{"cited_title":"No-reference image quality assessment via transformers, relative ranking and self- consistency,","cited_arxiv_id":null,"evidence_quote":"S. A. Golestaneh et al.'s Tres provides the third no-reference perceptual metric used in the evaluation."}],"review_version":1}