{"id":"01c826e7-86d8-46da-8229-7b0a5ec4edf1","arxiv_id":"2412.13508","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"A tri-branch invertible block that processes low-frequency image information in YCbCr color space, plus an all-zero high-frequency mapping during upscaling, improves learned image rescaling.","lead":"This paper introduces a new building block for image rescaling networks that routes color and brightness information through separate branches, and it reports higher reconstruction quality on standard benchmarks when plugged into existing models. The method is designed as a drop-in replacement for the invertible blocks used in IRN and SAIN, so a broad set of existing rescaling systems could adopt it.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Claimed 0.9–1.3 dB gain over IRN conflates tri-branch block with all-zero training strategy; need a re-trained IRN baseline to substantiate the plug-and-play claim.","rationale":"The reader's weakest assumption focuses on whether the all-zero mapping discards information needed for dense textures, pointing to the ×4 Urban100 deficit. That is a valid concern, but the more immediately load-bearing issue is the confounded comparison underlying the headline ×2 gain. The paper's own ablation (Table 3) reveals that the all-zero test-time trick alone gives IRN a 0.58–0.86 dB boost on the cited benchmarks, so the claimed 0.9–1.3 dB improvement over vanilla IRN cannot be attributed to the tri-branch block without a properly re-trained baseline. This affects the central 'plug-and-play' claim: if the gain is mostly from the training objective, then replacing only the block in an existing model would not yield the advertised improvement. The proposed test directly settles this by retraining IRN under the same all-zero regime. The verdict remains CONDITIONAL because the method plausibly works, but the conditions should include this isolation experiment and a tempered SOTA claim in view of the ×4 Urban100 result. The reader's verdict already requested clarification of code and SOTA, so the overall verdict is unchanged; however, the specific requirement for a re-trained baseline is an additional condition not articulated by the reader, hence 'partial' agreement.","tokens_in":14154,"tokens_out":10600,"duration_ms":93512,"concrete_test":"Retrain IRN from scratch using the exact all-zero training objective described for T-IRN (z = 0 in reconstruction, no Ldistr, same λ1/λ2, same 800k iterations and learning-rate schedule, same DIV2K 128×128 crops) and compare against T-IRN on Set14, Urban100, and DIV2K at ×2. If the PSNR gap between T-IRN and this retrained all-zero IRN is about 0.3–0.5 dB, the headline 0.9–1.3 dB improvement is mostly attributable to the all-zero strategy rather than the tri-branch block, and the plug-and-play claim must be revised. If the gap remains above 0.9 dB, the block's contribution is confirmed.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central empirical claim in Section 4.2 is that T-IRN achieves 0.9–1.3 dB PSNR improvement over IRN at ×2, which is used to support the abstract's assertion that T-InvBlocks 'can be seamlessly integrated into existing rescaling models, improving performance.' However, T-IRN differs from IRN in two ways: (i) the tri-branch architecture and (ii) the all-zero mapping training strategy, which removes the distribution-matching loss Ldistr (Section 3.4). The paper's ablation in Table 3 shows that simply applying an all-zero latent to the original IRN at test time (without retraining) improves PSNR by 0.58–0.86 dB on Set14/Urban100/DIV2K. Thus, a large fraction of the claimed 0.9–1.3 dB gain is due to the all-zero strategy, not the tri-branch block. The correct baseline for isolating the block's contribution is IRN retrained with the same all-zero objective and identical training conditions. Without this baseline, the 'plug-and-play' claim is overstated: the block's own contribution appears to be roughly 0.3–0.5 dB (comparing T-IRN to IRN with all-zero in Table 3), and even that comparison is confounded because IRN(all-zero) was not retrained. The abstract's unqualified 'state of the art' claim is also undermined by the ×4 results: T-IRN is worse than IRN on Urban100 (31.19 vs 31.41 dB, Table 1), directly contradicting 'advances the state of the art' for that benchmark.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper presents a plug-and-play tri-branch invertible block (T-InvBlock) for learned image rescaling. The block replaces the dual-branch additive/affine coupling in IRN by splitting the Haar low-frequency component into luminance (Y) and chrominance (Cb/Cr) channels and coupling them with the high-frequency branch through three invertible update equations (Eqs. 2-3). The authors also replace the latent high-frequency variable with an all-zero tensor during upscaling, which removes the distribution-matching loss. They insert T-InvBlock into IRN and SAIN, producing T-IRN and T-SAIN, and report PSNR/SSIM gains on Set5, Set14, BSD100, Urban100, and DIV2K for both x2 and x4 rescaling, as well as on DIV2K with JPEG compression at quality factors 30 to 90. The paper claims a 0.9 to 1.3 dB improvement over IRN at x2 and 0.4 to 0.6 dB over SAIN under lossy compression.","tokens_in":14500,"tokens_out":6134,"duration_ms":52642,"significance":"The manuscript is a competent empirical architecture paper in a well-studied area. The invertibility of Eqs. (2)-(3) is correctly constructed as a sequential coupling with an unambiguous inverse, and the parameter count is nearly unchanged. The JPEG-compression results in Table 2 are consistently positive across all tested quality factors and both scales, which is the strongest evidence for practical value. The ablation design in Table 3 is a useful first step, and the paper promises code via GitHub. The main weakness is attribution: the method varies the architecture and the training objective jointly, and the paper does not retrain the original IRN or SAIN with the all-zero objective, so the headline gain is not cleanly attributable to T-InvBlock alone. If the missing baseline is added, the paper would make a clean and useful contribution.","major_comments":[{"comment":"The claimed 0.9 to 1.3 dB gain over IRN at x2 conflates two interventions. Table 3 shows that IRN with an all-zero latent at test time (IRN (all-zero)) already improves PSNR over the standard IRN by 0.58 dB on Set14, 0.63 dB on Urban100, and 0.86 dB on DIV2K, without retraining. Comparing T-IRN with this row leaves only 0.33, 0.50, and 0.28 dB of residual gain. Because IRN (all-zero) is not retrained with the distribution-loss-free objective used for T-IRN, even this residual is not an unbiased estimate of the block's contribution. Please retrain IRN with the same all-zero objective, training schedule, and loss weights, and report it as the primary baseline. The same issue affects T-SAIN: Table 4 does not state whether SAIN (all-zero) is retrained, even though T-SAIN changes both the block and the GMM.","section":"Section 4.2, Table 3"},{"comment":"The all-zero strategy is motivated by an assertion about SAIN's official GMM weights, namely that the 'GMM actually learns a near-zero mean unimodal distribution.' No plot, variance value, or other quantitative evidence is provided for this load-bearing claim. If the claim is used to justify discarding the learned latent model, it should be substantiated. Similarly, the paper's appeal to IRN's ablation studies should cite the specific table or section in Xiao et al. (2023) rather than a general reference. Otherwise, the motivation should be stated as a hypothesis, not as an established fact.","section":"Section 3.3"},{"comment":"The unqualified 'state of the art' claim in the abstract is not supported by the paper's own x4 results. In Table 1, T-IRN underperforms IRN on Urban100 at x4 (PSNR 31.19 vs 31.41, SSIM 0.9132 vs 0.9157) and ties IRN on BSD100 PSNR (31.64). The text acknowledges this by saying 'some datasets do not fully surpass,' but the abstract and conclusion do not carry the necessary qualification. Please either qualify the claim to x2 and to most x4 datasets, or report an aggregate statistic over the full benchmark set that justifies the state-of-the-art phrasing.","section":"Abstract and Section 4.2, Table 1"}],"minor_comments":[{"comment":"The reported RGB-PSNR improvement of 0.6 to 1.1 dB over IRN is not shown in any table or figure; please provide these numbers explicitly, either in the main text or in supplementary material.","section":"Section 4.2"},{"comment":"The notation x_c for chrominance is easy to confuse with the concatenation of the Cb and Cr channels; consider using a more explicit label such as x_chroma or x_cbcr throughout.","section":"Section 3.2, Figure 2"},{"comment":"Please state explicitly what the 'Param' column counts, since T-IRN uses 1.57M vs IRN's 1.67M at x2 but 4.67M vs 4.35M at x4; calling this 'similar' is acceptable but should be accompanied by the exact numbers and the model footprint definition.","section":"Section 4.1, Tables 1 and 2"},{"comment":"The loss weights lambda1 = 1 and lambda2 = 0.25 are hand-chosen. Please state whether identical weights and training schedules were used for the IRN and SAIN baselines, or whether those baselines were tuned separately.","section":"Section 3.4, Eq. (4)"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is within the journal's scope. I have no concerns about citation practice or novelty disclosure. The main risk is overclaiming in the abstract relative to the x4 results, and the missing retrained all-zero baseline is a standard but necessary fix. I would be willing to review a revised version."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe paper's real novelty is the tri-branch invertible block: after Haar, the low-frequency branch is split into Y and CbCr, each processed with its own coupling functions. That's a clean, plausible architectural change, and it's not in the cited prior work. The all-zero latent strategy is borrowed from an IRN ablation, but the paper applies it systematically to both IRN and SAIN, and it works. The ablations (Table 3) do separate the contributions: YCbCr alone hurts slightly, all-zero alone helps ~0.6-0.9 dB, and the full T-InvBlock adds another ~0.3-0.5 dB on top. That's honest, and the math of Eqs. 2-3 is sound — the block is genuinely invertible.\n\nThe soft spots are real but not fatal. The main claim of \"0.9 to 1.3 dB over IRN\" conflates the architecture with the training strategy. The right baseline for isolating the block is IRN retrained with the all-zero objective; Table 3's \"IRN (all-zero)\" is just test-time inference with zeros, not a retrained model. So the block's actual contribution is likely closer to the 0.3-0.5 dB gap between T-IRN and that informal baseline, and even that is confounded. The abstract's \"state of the art\" is also too strong: at x4, T-IRN is worse than IRN on Urban100 (31.19 vs 31.41). The SAIN claim about the GMM learning a near-zero unimodal distribution is asserted without evidence — either show the fitted weights or drop the claim.\n\nNone of this sinks the paper. The method is simple, reproducible in principle (code link is provided), and the gains under JPEG compression for T-SAIN are consistent and meaningful (0.4-0.6 dB at x2, 0.2-0.5 dB at x4). A serious referee should ask for the retrained IRN(all-zero) baseline, trimmed claims, and evidence for the GMM statement. With those, it's a solid incremental contribution to image rescaling.\n\nFor you: worth a skim if you work on invertible rescaling, but the takeaway is \"architectural tweak plus training trick,\" not a paradigm shift. I'd send it to review.","headline":"A modest but real architectural improvement; the headline gain is mostly the all-zero trick, so the claims need trimming.","tokens_in":15080,"tokens_out":2030,"would_cite":false,"duration_ms":17153,"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":"The paper claims that separating the low-frequency branch into luminance and chrominance, and feeding an all-zero tensor for high frequencies at upscaling, improves invertible image rescaling by up to 1.3 dB over IRN and up to 0.6 dB over…","keywords":["image rescaling","invertible neural network","tri-branch invertible block","all-zero mapping","YCbCr color space","lossy compression","JPEG"],"falsifier":"Train or evaluate T-IRN at ×4 scale and compare per-image PSNR against IRN on Urban100, where the paper's own table shows T-IRN trailing on average (31.19 vs 31.41 dB); if the deficit concentrates in images with dense repetitive textures, then the all-zero mapping is losing high-frequency information that a sampled latent preserves. A direct control experiment—sampling $z$ from the trained GMM or Gaussian during upscaling instead of using zeros and measuring the PSNR difference—would separate the benefit of the tri-branch block from the benefit of the zero strategy.","tokens_in":13947,"feed_emoji":"🖼️","tokens_out":6397,"duration_ms":51413,"temperature":0.7,"pith_summary":"The paper tries to establish that the standard way of running invertible image rescaling—modeling high-frequency image details as random noise drawn from a Gaussian or a mixture—is unnecessary. It argues that the low-frequency branch, when split into luminance (Y) and chrominance (CbCr) and processed through a three-branch invertible coupling block, can carry all the information needed to reconstruct the high-resolution image. During upscaling the high-frequency latent is set to an all-zero tensor instead of being sampled, and the claim is that this both simplifies training and improves reconstruction. If true, any existing rescaling model built on invertible blocks can be upgraded by swapping in the new block and dropping its distribution-matching loss, gaining roughly 0.9–1.3 dB at ×2 scale over IRN and 0.2–0.6 dB over SAIN under JPEG compression without adding parameters.","feed_headline":"Zero-noise invertible block lifts image rescaling by 1.3 dB","feed_subtitle":"Splitting low-frequency RGB into Y and CbCr lets a drop-in block beat IRN and SAIN without more parameters.","key_machinery":"The T-InvBlock is the central object: a tri-branch invertible coupling block that operates on three variables, the luminance component $x_y$, the chrominance component $x_c$, and the high-frequency wavelet detail $x_h$, with coupling functions $\\phi, \\theta, \\nu, \\pi, \\rho, \\epsilon, \\eta, \\mu$ parameterized by densely connected convolutional blocks. Its forward map (Equation 2) updates $x_c$ then $x_y$ then $x_h$ using all previously updated branches, and its inverse (Equation 3) unwinds those updates, so the block is exactly invertible. The all-zero mapping strategy sets the latent $z$ to $0$ during upscaling, removing the distribution-matching loss and the GMM sampler; the paper argues this forces the LR image to be the sole carrier of information across the rescaling pipeline.","core_discovery":"The central claim is that the high-frequency latent variable in an invertible rescaling network does not need a learned or fixed probability model: setting $z = 0$ at upscaling time, together with a tri-branch block that treats luminance, chrominance, and high-frequency detail as separate interacting streams, produces better reconstruction than sampling from a Gaussian (IRN) or a learned Gaussian mixture (SAIN). The paper reports that the resulting T-IRN outperforms IRN by about 0.9–1.3 dB PSNR at ×2 scale across Set5, Set14, BSD100, Urban100, and DIV2K, and that T-SAIN beats SAIN by 0.2–0.6 dB under JPEG compression at quality factors 30–90. At ×4 scale the gains are smaller and not universal, with T-IRN trailing IRN on Urban100 PSNR, which the paper acknowledges as a partial exception. The discovery is framed as a plug-and-play architectural result: replacing the existing invertible block in IRN or SAIN with T-InvBlock and switching to the all-zero strategy yields the improvement without changing the overall network design.","pith_inferences":["The zero-latent result points toward a reinterpretation of invertible rescaling as deterministic learned compression: if the LR image can carry all needed detail, then the latent space is a code rather than a generative prior.","Because the all-zero strategy makes upscaling deterministic and removes the GMM, inference should be faster and more reproducible; a timing comparison would test this directly.","The ×4 gap on Urban100 suggests a testable boundary: images with dense, repetitive textures may be exactly where the LR image cannot absorb all high-frequency information, so comparing per-image PSNR against texture-density metrics would reveal the failure mode.","The YCbCr benefit is likely to transfer to other compression codecs that operate in chroma-separated color spaces, such as HEVC or AVIF, but that transfer is not tested in the paper."],"forward_implications":["Any invertible rescaling model that currently samples high-frequency latents can replace its invertible blocks with T-InvBlocks and disable the sampler, obtaining reconstruction gains without increasing parameter count.","Training becomes simpler: the distribution-matching loss used by IRN and the GMM-based latent model used by SAIN can be dropped entirely.","In lossy-compression scenarios the YCbCr decomposition aligns the network's internal color space with JPEG's, so robustness to compression artifacts improves most at low quality factors.","At ×4 scale, the benefit shrinks and is inconsistent across datasets, so the method's advantage is strongest for ×2 rescaling."],"supporting_citations":[{"why":"Supplies the IRN baseline and the vanilla invertible block that T-InvBlock replaces, and the comparison target for the main ×2/×4 results.","marker":"Xiao et al. 2020"},{"why":"Supplies the SAIN baseline with its GMM-based latent model and compression simulator, which T-SAIN modifies.","marker":"Yang et al. 2023"},{"why":"Provides the IRN ablation showing that an all-zero tensor can match or beat Gaussian sampling, the paper's stated motivation for the all-zero strategy.","marker":"Xiao et al. 2023"},{"why":"Provides the affine coupling layer used as the template for both the vanilla invertible block and the T-InvBlock updates.","marker":"Dinh, Sohl-Dickstein, and Bengio 2016"},{"why":"Provides the additive coupling layer that the low-frequency branch of IRN is based on and that the tri-branch design extends.","marker":"Dinh, Krueger, and Bengio 2014"},{"why":"Supplies the densely connected convolutional blocks that parameterize the coupling functions in the invertible blocks.","marker":"Wang et al. 2018"},{"why":"Supplies the DIV2K training and validation data used for all reported experiments.","marker":"Agustsson and Timofte 2017"},{"why":"Defines the JPEG compression standard whose YCbCr color space motivates the branch split and whose codec is used in the compression experiments.","marker":"Wallace 1991"}],"fun_headline_variants":["Tri-branch invertible block improves rescaling by 1.3 dB","Zero-noise upscaling with YCbCr split beats Gaussian models","Drop-in tri-branch block gains 1.3 dB without extra parameters","All-zero high-frequency wins in image rescaling"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that setting the high-frequency latent to zeros during upscaling discards no information needed for reconstruction, meaning the invertible block can push all essential high-frequency detail into the low-resolution image itself.","fun_headline_variants_meta":{"raw":{"variants":["Tri-branch invertible block improves rescaling by 1.3 dB","Zero-noise upscaling with YCbCr split beats Gaussian models","Drop-in tri-branch block gains 1.3 dB without extra parameters","All-zero high-frequency wins in image rescaling"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000779,"raw_usage":{"total_tokens":3482,"prompt_tokens":1025,"completion_tokens":2457,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":641,"completion_tokens_details":{"reasoning_tokens":2381}},"tokens_in":641,"tokens_out":2457,"duration_ms":16690,"temperature":1.0,"reasoning_tokens":2381,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T13:03:26.461651+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train or evaluate T-IRN at ×4 scale and compare per-image PSNR against IRN on Urban100, where the paper's own table shows T-IRN trailing on average (31.19 vs 31.41 dB); if the deficit concentrates in images with dense repetitive textures, then the all-zero mapping is losing high-frequency information that a sampled latent preserves. A direct control experiment—sampling $z$ from the trained GMM or Gaussian during upscaling instead of using zeros and measuring the PSNR difference—would separate the benefit of the tri-branch block from the benefit of the zero strategy.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the IRN baseline and the vanilla invertible block that T-InvBlock replaces, and the comparison target for the main ×2/×4 results."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the SAIN baseline with its GMM-based latent model and compression simulator, which T-SAIN modifies."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the IRN ablation showing that an all-zero tensor can match or beat Gaussian sampling, the paper's stated motivation for the all-zero strategy."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the JPEG compression standard whose YCbCr color space motivates the branch split and whose codec is used in the compression experiments."}],"review_version":1}