{"id":"4d4be35a-310a-4531-b988-f21b63f067f2","arxiv_id":"2508.18968","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"An extension of SCF image coding to 4:2:0 chroma formats, using luma-guided chroma prediction and luma-chroma range side information, beats HEVC-SCC by 5.66% average bitrate on screen content.","lead":"This paper extends the soft context formation lossless screen content coder to YCbCr 4:2:0 by coding brightness first and then using brightness information to predict and model color. In tests on 173 screen content images it needs about 5.7% less bitrate than the HEVC screen content standard.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"CRC hyperparameters are tuned on SIQAD and then SIQAD is included in the reported averages, so the precise 5.66% advantage and the SIQAD row of Table IV are partly test-set-selected; the qualitative claim likely survives, but the magnitude and generalization of the gain are not yet established.","rationale":"The reader's weakest assumption correctly identifies the most load-bearing issue: CRC parameters are tuned on SIQAD, and SIQAD is then included in the final evaluation and aggregate. Because the measured contribution of the two new tools (CRC and LMAP) is only 0.81%, the reported 5.66% average advantage over HEVC could be partly inflated by test-set selection. However, the per-dataset results for the four datasets not used for tuning (SCID, SC-Text, SC-Mixed, HEVC CTC) show consistent advantages over HM, so the central qualitative claim is not refuted. The concern is about the precision and generalizability of the numerical claim, not about a fundamental flaw in the coding approach. A leave-one-dataset-out re-evaluation would settle whether the advantage and the CRC benefit persist outside the tuning distribution. I therefore do not change the reader's CONDITIONAL verdict.","tokens_in":8538,"tokens_out":10172,"duration_ms":95757,"concrete_test":"Re-run the Table IV evaluation with leave-one-dataset-out validation: for each of the five datasets, fix CRC parameters to the values that minimize average bitrate on the other four datasets, then report the held-out dataset's bitrate and the pooled average, together with a CRC on/off ablation on each held-out set. If the pooled SCF 420 advantage over HM remains above 2% and CRC is beneficial on a majority of held-out datasets, the concern is resolved; if CRC flips negative or the advantage drops below 2%, the claim should be qualified to the tuning distribution.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim rests on Table IV. The CRC parameters (b=4, p=64, sCRC=64) were chosen in Table III by minimizing average bitrate on SIQAD, and SIQAD is then one of the five datasets in Table IV and is part of the pooled 'All' row. The ablation in Table V shows the entire CRC gain is only 0.56% (1.591 vs 1.582 bpp) and the combined CRC+LMAP gain is only 0.81% (1.595 vs 1.582 bpp). These gains are small enough that parameter selection on the evaluation set could materially alter the reported average. Table II already shows LMAP slightly worsens prediction MAE on Kodak, so on content with weaker luma-chroma correlation the new tools could become a net loss. The four non-SIQAD rows still show SCF 420 ahead of HM by 3.43%–10.31%, so the qualitative 'outperforms HEVC on every dataset' claim may survive; what is not independently established is the exact magnitude of the advantage and whether CRC/LMAP generalize beyond the tuning distribution. No code or data are released, so the reported bpp numbers cannot be independently reproduced.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper extends the soft context formation (SCF) lossless screen-content coder from RGB 4:4:4 to YCbCr 4:2:0. The luma plane is coded first with SCF, and the Cb/Cr planes are coded jointly as color pairs. Two new tools are proposed: a luma-guided predictor (LMAP) for chroma samples, and a luma-dependent chroma range coding (CRC) scheme that transmits side information about which luma-chroma combinations occur in blocks, allowing the coder to discard impossible symbols from its histograms. Evaluation on five screen-content datasets (173 images) reports that the proposed SCF 420 outperforms both VTM 17.2 and HM-16.21+SCM-8.8 on every dataset, with HEVC-SCC needing 5.66% more bitrate on average than SCF 420.","tokens_in":8907,"tokens_out":4150,"duration_ms":40005,"significance":"If the reported results hold, this is a useful and nontrivial extension: SCF was previously limited to 4:4:4 formats, and 4:2:0 is the dominant chroma format in video and conferencing applications. The algorithmic description is clear, the ablation in Table V isolates the contributions of CRC (0.56%) and LMAP (0.25%), and the comparisons use standard reference codecs. However, the overall advantage over HEVC is modest, the CRC hyperparameters appear to be selected on a dataset that is also used in the final evaluation, and no per-image variance or statistical significance is reported. The qualitative claim that SCF 420 beats HEVC on the evaluated datasets is plausible, but the exact magnitude and generalization of the gain are not yet established.","major_comments":[{"comment":"There is a direct inconsistency in the parameter-selection description. The text says: \"we evaluate the dataset SIQAD with multiple parameter combinations as visualized in Table III\" and chooses the best average bitrate on SIQAD. The caption of Table III, however, says the parameters were obtained by training on the SCID dataset. If the parameters were indeed tuned on SIQAD, then the SIQAD row in Table IV and the pooled 'All' row are partly test-set-selected, inflating the reported 1.582 bpp and the 5.66% HEVC comparison. If the parameters were tuned on SCID, the same circularity applies to the SCID row. This must be corrected. Because Table V shows the combined CRC+LMAP gain is only 0.81% (1.595 vs 1.582 bpp), the choice of hyperparameters could plausibly change the outcome on a given dataset. The authors should separate tuning and test data, report results with the tuning dataset exclu","section":"Section IV, Tables III and IV"},{"comment":"The central quantitative claim is that HEVC requires 5.66% more bitrate on average, with dataset-specific differences as small as 3.43% (SC-Mixed) and as large as 10.31% (SC-Text). However, only dataset-averaged bitrates are reported; there are no per-image results, standard deviations, confidence intervals, or paired significance tests. With 173 images and small average differences, the claim should be supported by a per-image distribution or at least by per-dataset variance. Without this, the reader cannot tell whether the reported advantage is consistent or driven by a few images. This is load-bearing for the magnitude of the contribution and should be added in the revision.","section":"Table IV"},{"comment":"Table II shows that LMAP improves chroma MAE on SCID (1.317 to 1.175) but slightly worsens it on Kodak (0.708 to 0.711). The method is intended for screen content, so this alone is not fatal, but the evaluation datasets include 'mixed' and 'textual' webpage stimuli from [19], which may contain natural-image-like regions. The paper does not report how often the LMAP condition in Eq. (5) is active on each dataset, nor whether any individual images are harmed by LMAP or CRC. Given the small average gains in Table V, the authors should report the distribution of gains/losses per image, and specifically how many images are worse with the proposed tools than without them.","section":"Section III-B, Table II"}],"minor_comments":[{"comment":"Typo: 'A veraged' should be 'Averaged'.","section":"Abstract"},{"comment":"The caption says 'obtained by training on SCID dataset' while the text says the parameters are chosen on SIQAD. Even if this is a typo, it must be fixed because it affects the interpretation of the evaluation.","section":"Table III caption vs Section IV"},{"comment":"The floor operation in Eq. (4) is not typeset clearly. Please define the rounding/flooring explicitly. Similarly, the precision scaling parameter sCRC would benefit from a one-sentence explanation of how it quantizes luma values.","section":"Section III-C, Eq. (4)"},{"comment":"The version designation for HEVC is inconsistent: Table IV and the text use HM-16.21+SCM-8.8, while Table VI lists 'HM 21.18 SCF-8.8'. Please use the same version string throughout.","section":"Table VI"},{"comment":"The conversion from RGB 4:4:4 to YCbCr 4:2:0 uses ffmpeg, but the exact filter (e.g., yuv420p with default chroma siting) is not specified. Since lossless coding results depend on the exact subsampling and chroma siting, this should be documented.","section":"Section IV, preprocessing"}],"recommendation":"major_revision","confidential_remarks":"The paper is an experimental codec paper without code or data release. Given the small gains claimed for the new tools and the parameter-selection issue, I would ask the editor to require per-image results or at least per-dataset variance, and to clarify the training/evaluation split. The core idea is sound and the paper is likely publishable after these revisions."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a competent engineering contribution. The authors take SCF, which natively does 4:4:4 RGB, and make it work for 4:2:0 YCbCr by coding luma first and then CbCr as pairs. The NMI analysis justifying the plane splitting is sensible. The new pieces are the luma-guided MAP prediction (LMAP) and the luma-based chroma range coding (CRC) side information. Both are clearly explained and the ablation isolates their contributions: CRC alone saves 0.56%, LMAP plus CRC saves 0.81%. The main gain is just adapting SCF to 4:2:0 at all, which is the point.\n\nThe paper is honest about its methods and includes enough detail to reimplement. The comparisons use reference codecs under common test conditions, which is the right baseline. I believe the qualitative result: SCF 420 beats HM and VTM on these screen content datasets, and the magnitude is plausible.\n\nThe soft spot is the CRC parameter selection. They tune b, p, sCRC on SIQAD (Table III) and then include SIQAD in the final average (Table IV). That is a form of test-set selection. It likely inflates the reported 5.66% average advantage over HM, and the SIQAD row in particular is partly optimized. The stress-test note is correct about that. However, the other four datasets also show SCF 420 ahead, so the qualitative 'wins on every dataset' claim probably survives. What is not established is the exact magnitude and how much CRC/LMAP generalize beyond screen content similar to SIQAD. Table II already hints LMAP slightly increases MAE on Kodak natural images, so on mixed content the new tools could be a net loss.\n\nAnother minor issue: no code or data released, and no per-image variance. Given the small ablation gains, a skeptic can't verify the margin of error. That said, the paper is not slippery; it states exactly what it did. The self-citations are fine; prior SCF work is RGB 4:4:4 only, so the 4:2:0 extension is new.\n\nWho is this for? People working on lossless screen content coding, especially if they want a 4:2:0-capable pixel-wise coder. It's a niche but real audience. I'd send it to peer review, but with a request to re-evaluate without SIQAD in the parameter selection, or at least report per-dataset results excluding the tuning set. Also ask for variance or confidence intervals. I would not desk-reject it.","headline":"Solid 4:2:0 extension of SCF with a likely real but slightly over-stated bitrate advantage; the tuning-on-SIQAD issue is a genuine soft spot.","tokens_in":9407,"tokens_out":3133,"would_cite":false,"duration_ms":24915,"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":"Soft context formation is extended to lossless 4:2:0 screen content by coding luma first and using luma-guided chroma prediction, beating HEVC-SCC by 5.66% bitrate on average.","keywords":["lossless image compression","screen content coding","soft context formation","4:2:0 chroma subsampling","luma-guided prediction","chroma range coding","arithmetic coding","HEVC screen content coding"],"falsifier":"Compress a held-out set of screen-content frames containing photographic backgrounds and many unique colors with SCF 420 both with and without CRC; if the no-CRC version ties or beats the CRC version, the range side information is not generally beneficial. Alternatively, re-run the five datasets with the CRC parameters excluded from the training set and check whether the 5.66% margin over HEVC-SCC shrinks materially.","tokens_in":1857,"feed_emoji":"🖥️","tokens_out":5581,"duration_ms":89941,"temperature":0.7,"pith_summary":"The paper extends the soft context formation (SCF) coder, which previously handled only RGB 4:4:4 color triplets, to lossless YCbCr 4:2:0 screen content. It does this by coding the luma plane first with SCF, then coding the Cb/Cr pair jointly, guided by measured normalized mutual information between channels. Two additions carry the gain: luma-guided median-adaptive prediction for chroma, and luma-dependent Chroma Range Coding, which sends compact side information about which luma-chroma combinations actually occur so impossible symbols are excluded from arithmetic-coding histograms. Averaged over 173 screen-content images from five datasets, the proposed method uses 5.66% less bitrate than HEVC with its screen content coding extension, and it also outperforms VVC in lossless mode. If correct, this establishes SCF-based coding as the most compression-efficient lossless 4:2:0 screen-content method among the compared codecs.","feed_headline":"Luma-guided coder beats HEVC-SCC by 5.66%","feed_subtitle":"Soft context formation gains 4:2:0 support with luma-guided chroma prediction and compact range side information.","key_machinery":"The load-bearing mechanism is a two-pass pipeline: the luma plane is coded first, and a downsampled, quantized luma map then drives chroma prediction and side-information generation. LMAP uses equality of downsampled luma values to choose the top or left chroma neighbor as predictor, and doubles the count of the predicted color in the histograms. CRC partitions the Cb and Cr ranges, divides the image into blocks, and transmits binary range images marking each occurring quantized-luma and chroma combination; removing entries absent from these ranges elevates the estimated probability of every real symbol.","core_discovery":"SCF coding is extended to YCbCr 4:2:0 by coding the Y plane with the existing SCF stages, then coding the Cb and Cr pair together as color doubles. Two enhancements exploit luma-chroma correlation: LMAP, which substitutes the top or left chroma neighbor when the downsampled luma value equals that neighbor's luma value, and chroma range coding (CRC), which transmits binary images marking which quantized luma-chroma combinations occur per block. CRC lets the histogram-based probability models exclude impossible chroma entries, raising the odds of real colors. Reported results: 1.582 bpp on average over five screen-content datasets, versus 1.672 bpp for HEVC-SCC and 1.728 bpp for VTM 17.2.","pith_inferences":["A natural extension not explored in the paper is per-block or per-image adaptation of the CRC parameters; textual content showed the largest gain (10.31% over HEVC), so content-adaptive settings could widen the margin where luma-chroma structure is strong and shrink it where it is weak.","The same histogram-restriction idea could transfer to near-lossless or lossy screen-content coding, or to palette and intra-block-copy modes in video codecs, where side information about possible luma-chroma combinations could sharpen entropy coding.","Because Cb-Cr mutual information stays relatively high for screen content even after the YCbCr transform, coding the two chroma planes jointly rather than separately is the structurally important choice; other pixel-wise lossless codecs that code planes independently could gain from the same pairing.","A held-out test crossing from pure screen content toward photographic backgrounds would likely reduce the measured advantage, since the paper already notes LMAP leaves natural images nearly unaffected and CRC relies on few occurring luma-chroma combinations."],"forward_implications":["SCF-based lossless coding now covers YCbCr 4:2:0, the common video chroma format, without per-plane workarounds that lose the benefit of joint color-triplet coding.","On the five datasets tested, SCF 420 outperforms both HEVC-SCC and VVC reference implementations in lossless screen-content coding.","The luma-guided chroma prediction mainly helps screen content: it improves chroma match rates from 81.3% to 82.7% on SCID while leaving natural-image prediction almost unchanged.","The CRC side-information overhead is small on screen content: disabling it costs only 0.56% bitrate, while the total gain over HEVC-SCC is 5.66%.","Encoding time for SCF 420 is slower than HEVC-SCC but about ten times faster than VTM 17.2; decoding is symmetric rather than asymmetric."],"supporting_citations":[{"why":"Supplies the soft context formation coder itself: pattern matching, color palette, residual coding, and context-based probability estimation.","marker":"[8]"},{"why":"Supplies the enhanced color palette modeling and the key idea that removing impossible entries from histograms raises probabilities, which CRC builds on.","marker":"[9]"},{"why":"Supplies the luma-to-chroma cross-component prediction approach that motivates the proposed luma-guided chroma prediction.","marker":"[12]"},{"why":"Supplies the SCID screen-content dataset used for the channel mutual-information analysis and for final evaluation.","marker":"[17]"},{"why":"Supplies the SIQAD screen-content dataset used for selecting the CRC parameters b, p, and sCRC, and for final evaluation.","marker":"[18]"},{"why":"Supplies the textual and mixed webpage stimuli used as two of the five evaluation datasets.","marker":"[19]"},{"why":"Supplies the HEVC screen-content coding common test conditions and dataset, and the HEVC-SCC reference baseline.","marker":"[20]"},{"why":"Supplies the VTM 17.2 reference codec used as the VVC lossless baseline.","marker":"[23]"},{"why":"Supplies the common test conditions for lossless coding used to configure the VTM baseline fairly.","marker":"[24]"}],"fun_headline_variants":["Lossless 4:2:0 coding beats HEVC-SCC by 5.66%","Luma-guided SCF extension cuts 5.66% vs HEVC-SCC","Soft context coding now handles 4:2:0 with luma guidance","New luma-guided coder saves 5.66% over HEVC-SCC on 4:2:0","4:2:0 lossless coding: luma-guided prediction cuts 5.66% vs HEVC"],"cache_read_input_tokens":11136,"weakest_assumption_plain":"The reported advantage depends on the CRC side-information settings having been tuned on one screen-content dataset, which is also included in the final averaged numbers; images with less luma-chroma structure could make that side information cost more than it saves.","fun_headline_variants_meta":{"raw":{"variants":["Lossless 4:2:0 coding beats HEVC-SCC by 5.66%","Luma-guided SCF extension cuts 5.66% vs HEVC-SCC","Soft context coding now handles 4:2:0 with luma guidance","New luma-guided coder saves 5.66% over HEVC-SCC on 4:2:0","4:2:0 lossless coding: luma-guided prediction cuts 5.66% vs HEVC"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000879,"raw_usage":{"total_tokens":3658,"prompt_tokens":787,"completion_tokens":2871,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":531,"completion_tokens_details":{"reasoning_tokens":2747}},"tokens_in":531,"tokens_out":2871,"duration_ms":16604,"temperature":1.0,"reasoning_tokens":2747,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T16:02:38.530138+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compress a held-out set of screen-content frames containing photographic backgrounds and many unique colors with SCF 420 both with and without CRC; if the no-CRC version ties or beats the CRC version, the range side information is not generally beneficial. Alternatively, re-run the five datasets with the CRC parameters excluded from the training set and check whether the 5.66% margin over HEVC-SCC shrinks materially.","supporting_citations":[{"cited_title":"Screen content compression based on enhanced soft context formation,","cited_arxiv_id":null,"evidence_quote":"Supplies the soft context formation coder itself: pattern matching, color palette, residual coding, and context-based probability estimation."},{"cited_title":"Enhanced color palette modeling for lossless screen content compression,","cited_arxiv_id":null,"evidence_quote":"Supplies the enhanced color palette modeling and the key idea that removing impossible entries from histograms raises probabilities, which CRC builds on."},{"cited_title":"Adaptive cross component linear model for chroma intra-prediction in VVC,","cited_arxiv_id":null,"evidence_quote":"Supplies the luma-to-chroma cross-component prediction approach that motivates the proposed luma-guided chroma prediction."},{"cited_title":"SCID: A database for screen content images quality assessment,","cited_arxiv_id":null,"evidence_quote":"Supplies the SCID screen-content dataset used for the channel mutual-information analysis and for final evaluation."},{"cited_title":"Perceptual quality assessment of screen content images,","cited_arxiv_id":null,"evidence_quote":"Supplies the SIQAD screen-content dataset used for selecting the CRC parameters b, p, and sCRC, and for final evaluation."},{"cited_title":"Webpage saliency,","cited_arxiv_id":null,"evidence_quote":"Supplies the textual and mixed webpage stimuli used as two of the five evaluation datasets."},{"cited_title":"Common test conditions for screen content coding,","cited_arxiv_id":null,"evidence_quote":"Supplies the HEVC screen-content coding common test conditions and dataset, and the HEVC-SCC reference baseline."},{"cited_title":"VVC test model reference software (VTM)","cited_arxiv_id":null,"evidence_quote":"Supplies the VTM 17.2 reference codec used as the VVC lossless baseline."},{"cited_title":"JVET common test conditions and software reference configurations for lossless, near lossless, and mixed lossy/lossless coding,","cited_arxiv_id":null,"evidence_quote":"Supplies the common test conditions for lossless coding used to configure the VTM baseline fairly."}],"review_version":1}