{"id":"1d6c1b03-54f8-41c9-b68e-521eda19422b","arxiv_id":"2505.19120","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A frequency-separated dual-branch transformer with a learnable fusion module reports state-of-the-art or near-state-of-the-art demoiréing on FHDMi and UHDM with only about 6 million parameters.","lead":"Freqformer, a transformer for removing moiré patterns from photos of screens, splits the image into high-frequency texture and low-frequency color parts and restores each with its own network branch. On two public high-resolution benchmarks it reports the best or second-best scores with roughly 6 million parameters, a favorable accuracy-to-size trade-off.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The claimed 0.1x scale-robustness of the low-frequency branch is load-bearing but unsupported: without decomposition level L or a real protocol, the resize training strategy—and the UHDM/FHDMi gains built on it—could collapse.","rationale":"The reader's weakest_assumption is the right target. The central SOTA claim depends less on the specific transformer blocks than on the frequency-split training scheme: the low branch is deliberately trained and, by bullet 4, inferred at roughly 0.1x scale. If that scale-robustness premise is wrong, the model either sees an input that lacks information needed for color correction (if inference is resized) or faces a train/test mismatch (if inference is full-resolution). Both scenarios would invalidate the dual-branch efficiency story and could change the UHDM numbers, where the reported gains over ESDNet-L are small. The concrete test isolates the premise by directly measuring I_l reconstruction error with the official L and real data. Table 3's resize-vs-crop comparison is a useful sanity check, but it does not quantify information loss or reveal the decomposition level, so it cannot substitute for this test. I considered the implausible Haar baseline in Sec. 4.2 (PSNR 13.32 vs input 17.97) as an alternative concern; it is a real red flag suggesting a broken comparison protocol, but it mainly undermines the motivation for the new decomposition rather than the reported SOTA numbers themselves. The resize premise is more load-bearing for the central claim. Therefore the reader's CONDITIONAL verdict stands without adjustment.","tokens_in":12729,"tokens_out":10985,"duration_ms":91619,"concrete_test":"Reproduce the Sec. 3.2 bullet 2 experiment with the official code: take the FHDMi and UHDM test sets, compute I_l with the official decomposition level L, bicubic-downsample to 0.1x (for UHDM, 384x216), upsample back to full resolution, and report median and minimum PSNR over all images. Repeat for L=1,2,3,4 if L is not documented. If the median reconstruction PSNR is materially below ~45 dB, the scale-robustness premise fails and the resize-based training/inference protocol—and the Table 1 marginal gains on UHDM—require re-evaluation. Also verify whether the test-time low branch receives full-resolution or 512px input; the two descriptions in the paper must be reconciled.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The architecture's main efficiency and performance claim rests on Sec. 3.2 bullet 2: the low-frequency component I_l 'can be downsampled to a much lower resolution (e.g., 0.1x) and still be accurately reconstructed,' with reconstruction PSNR '~50 dB'. This justifies training and, per bullet 4, inferring the low branch at a small scale. The supporting experiment has no protocol: the decomposition level L (never stated in the paper), interpolation method, dataset, subset size, and variance are all omitted. For UHDM, 0.1x means 3840x2160 to roughly 384x216; if the actual L-level I_l contains energy above that grid's Nyquist rate, resizing irreversibly destroys color and moiré information and the low branch cannot recover it, regardless of network capacity. The reported UHDM margins are small (SSIM 0.8018 vs 0.7985, LPIPS 0.2434 vs 0.2454), so a failed premise could plausibly change the ranking. Section 3 also contradicts bullet 4: it says both branches operate at full resolution during testing, while bullet 4 says the low branch uses the same small scale at training and inference. The test-time protocol is therefore ambiguous. Table 3 gives indirect support (resize outperforms crop for the low branch) but does not quantify information loss or isolate the scale-robustness premise.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Freqformer, a dual-branch Transformer for image demoiréing that first decomposes the input into high-frequency and low-frequency components via a recursive, dilated low-pass convolution, then processes each branch with a tailored training strategy (crop for high frequency, resize for low frequency), and finally fuses the branches with a learnable Frequency Composition Transform (FCT). The authors report state-of-the-art results on the FHDMi and UHDM benchmarks with a compact 6.065M-parameter model, and support the design with ablations on the frequency decomposition, the FCT, the resize strategy, and the proposed SA-CA module.","tokens_in":13025,"tokens_out":3604,"duration_ms":22976,"significance":"If the results hold, Freqformer is a meaningful contribution to single-image demoiréing: it combines a frequency-decomposition design with an efficient dual-branch transformer and demonstrates competitive or superior performance against both dedicated demoiréing methods and recent general restoration models at a lower parameter count. The paper is strengthened by a public code release, comparisons on two public high-resolution benchmarks, and a coherent set of ablations that isolate the contributions of the main components. However, the central efficiency claim relies on a scale-robustness premise for the low-frequency branch that is asserted with minimal experimental detail, and the reported UHDM margins over the second-best method are small, so the SOTA claim needs stronger evidence and a clarified inference protocol.","major_comments":[{"comment":"The claim that the low-frequency component can be downsampled to 0.1x and reconstructed at ~50 dB PSNR is load-bearing for the resize-based training strategy and the low-scale inference, but the supporting experiment is not described: no decomposition level L, dataset, interpolation method, subset size, or variance is reported. Please provide the full protocol and results, including the value of L used, or alternatively reframe this premise as an empirical observation supported by the branch-level results in Table 3 rather than as a general property of the decomposition.","section":"Sec. 3.2, bullet 2"},{"comment":"The test-time protocol is contradictory: Sec. 3 states that during testing both branches operate on the full-resolution image, while Sec. 3.2 bullet 4 states that the low-frequency branch uses the same small-scale input during both training and inference. Please specify the exact inference pipeline, particularly the input resolution of the low branch, how its output is upsampled, and how the upsampled features are aligned with the high-branch features in the FCT before the final reconstruction. This is also needed to interpret the reported 2.49 TFLOPS for 4K inference and the UHDM results.","section":"Sec. 3, first paragraph vs. Sec. 3.2, bullet 4"},{"comment":"No error bars, repeated runs, or statistical significance tests are reported. On UHDM, Freqformer's advantage over ESDNet-L is small (SSIM 0.8018 vs. 0.7985; LPIPS 0.2434 vs. 0.2454), so single-run numbers do not establish that the ranking is stable. Please report at least three runs with mean and standard deviation, or a paired comparison, for the UHDM experiments.","section":"Table 1"},{"comment":"The number of decomposition levels L is never specified in the paper, although it controls the cutoff frequency of the low-frequency component and therefore directly affects the validity of the 0.1x resizing claim and the reproducibility of the method. Please state the value of L in Sec. 3.1 and list it in the implementation details.","section":"Sec. 3.1, Eq. (2)"}],"minor_comments":[{"comment":"There are several typos in this section, including 'awesome benifits' and 'reals' (for 'reveals'), which should be corrected.","section":"Sec. 3.2"},{"comment":"The phrase 'Ultr-High-Definition' is missing a hyphen and should read 'Ultra-High-Definition'.","section":"Sec. 2.1"},{"comment":"The expression 'P ostF usion' appears with artificial spacing; this formatting should be fixed to 'PostFusion' throughout.","section":"Sec. 3.1"},{"comment":"The row labeled 'Freqformer-L (e2e)' is confusing because no 'Freqformer-L' variant is defined; please clarify whether this is an end-to-end version of Freqformer without frequency decomposition or a different variant.","section":"Table 4"},{"comment":"Some notation in Figure 3 (e.g., the projected features with prime symbols and the subscripts) is difficult to read at the printed size; consider enlarging or simplifying the notation in the figure.","section":"Fig. 3"}],"recommendation":"major_revision","confidential_remarks":"The paper is a solid empirical contribution, but the missing value of L and the contradictory inference description are genuine reproducibility and correctness concerns. The authors can address them without changing the method: add the decomposition level, give the full protocol behind the 50 dB claim, and clarify the test-time resolution of the low branch. The small UHDM margins also make the lack of repeated runs worth addressing. I recommend major revision rather than rejection because the core idea is coherent and the ablations already provide indirect support for the resize strategy."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper is worth reading: it introduces a recursive convolution-based frequency split for demoiréing, trains the low-frequency branch on downsampled images, and reports SOTA on FHDMi and UHDM with a 6M-parameter model and public code. The central idea is fresh relative to the wavelet-based prior work (WDNet, AWUDN, MBWDN), and the ablations make a coherent case that resize-based training helps the low branch (Table 3) and that the learnable FCT adds a little. The numbers are plausible, and the code release means the results can be checked.\n\nThe soft spots are real but mostly fixable. The load-bearing claim that the low-frequency component is scale-robust—downsampled to 0.1x with ~50 dB reconstruction—is asserted without protocol: the decomposition level L is never given, and the interpolation method, dataset, subset size, and variance are absent. The stress-test worry is right: if L is small, the low-frequency image still contains energy above the 0.1x Nyquist rate, and resizing destroys it. Table 3 gives indirect support, but it doesn't isolate the premise. Also, Sec. 3 says both branches operate at full resolution at test, while Sec. 3.2 bullet 4 says the low branch uses the same small scale at training and inference; that contradiction needs resolving. The Haar baseline (PSNR 13.32, SSIM 0.2113) is implausibly bad—worse than the input—which suggests the newly trained Haar model may be broken or the training setup is unfair; a comparison against an existing wavelet method like WDNet would be more informative. No error bars anywhere.\n\nThese issues don't sink the central claim, but they need to be answered. A referee should ask for the decomposition level and a full protocol for the 0.1x experiment, clarify the test-time pipeline, and rerun the Haar ablation with a proper baseline. The paper deserves peer review, not desk rejection, because the method is novel, the results are measurable, and the code is public. If I worked on demoiréing I'd cite it after the revision. I'd maybe bring it to reading group to discuss the decomposition.","headline":"A novel, compact demoiréing transformer with credible SOTA results, undermined by an undocumented scale-robustness claim and a training/inference contradiction.","tokens_in":13565,"tokens_out":4903,"would_cite":false,"duration_ms":44714,"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":"A dual-branch Transformer that splits moiré into high-frequency texture and low-frequency color can beat larger demoiréing networks on FHDMi and UHDM.","keywords":["image demoiréing","frequency decomposition","dual-branch transformer","low-frequency color distortion","high-frequency texture","channel attention","image restoration","FHDMi"],"falsifier":"Take a diverse set of FHDMi and UHDM moiré images, apply the paper's L-level decomposition, downsample the low-frequency component to 0.1x, resize it back, and measure reconstruction PSNR against the original low-frequency component. If typical values fall well below roughly 50 dB, or if images with strong low-frequency scene texture break the scale-robustness, the resize-based training strategy degrades and the dual-branch design loses its justification.","tokens_in":12547,"feed_emoji":"🖼️","tokens_out":4387,"duration_ms":38256,"temperature":0.7,"pith_summary":"Freqformer argues that moiré artifacts split cleanly by frequency: high-frequency components carry localized texture corruption, low-frequency components carry smooth color distortions. The paper builds a dual-branch Transformer that trains on these two components separately, using random crops for the high-frequency branch and heavily downsampled images for the low-frequency branch, then fuses the branches with a learnable Frequency Composition Transform. On the FHDMi and UHDM benchmarks, Freqformer reports the best PSNR, SSIM, and LPIPS on FHDMi and best SSIM/LPIPS on UHDM, with 6.065M parameters, smaller than most compared models. The paper's point is that explicit frequency separation, not a monolithic mapper, is what makes efficient high-resolution demoiréing feasible.","feed_headline":"Freqformer splits moiré into texture and color to beat bigger models","feed_subtitle":"A 6M-parameter dual-branch transformer reaches top scores on two high-resolution demoiréing benchmarks.","key_machinery":"The load-bearing object is the recursive frequency decomposition of Eq. (2): starting with $I_l^0 = I_m$, each level applies a dilated 3x3 convolution with the binomial kernel $k$ to get $I_l^i$, and defines the high-frequency residual as $I_h^i = I_h^{i-1} + I_l^{i-1} - I_l^i$. After $L$ levels, $I_l$ is the smooth color component and $I_h$ the texture component, both at full resolution. This decomposition does the work of making the two branches trainable with different strategies, and it feeds the learnable FCT, which combines the branch features before the final PixelShuffle reconstruction.","core_discovery":"The paper claims that the right decomposition of a moiré image is into a low-frequency image that keeps color information and is robust to downsampling, and a high-frequency residual that keeps texture and is spatially local. This decomposition, built from L-level dilated convolutions with a fixed binomial kernel, separates moiré into components with different training needs: crop-based training for the high-frequency branch and resize-based training for the low-frequency branch. A learnable FCT then fuses the two branches in feature space instead of adding pixel outputs, which avoids accumulating reconstruction errors. With this design and a lightweight Spatial-Aware Channel Attention module, the paper reports results on FHDMi and UHDM that beat prior demoiréing and general restoration models, including much larger ones, at 6.065M parameters.","pith_inferences":["The scale-robust low-frequency assumption, if it generalizes, implies that any global color correction task (white balance, tint removal, cross-ISP color shifts) could be trained on downsampled images and then applied at full resolution with the same network.","The decomposition may transfer to other screen-capture artifacts such as flicker banding or subpixel aliasing, since those also have a smooth color component and a local texture component.","A testable extension would be to measure how far the resize factor can be pushed (0.05x, 0.02x) before the roughly 50 dB reconstruction PSNR of the low-frequency branch degrades, giving a direct compute-versus-quality trade-off for 8K inputs.","The paper compares against the 2-level Haar transform and reports a large drop (test PSNR 13.32), but it does not sweep the kernel or level count; varying those could reveal how much of the gain comes from this specific decomposition design."],"forward_implications":["If the decomposition claim is right, demoiréing can be solved by two smaller specialized networks rather than one large holistic network.","Resize-based low-frequency training means the color-correction branch can run at low resolution during training without hurting final quality, cutting memory and compute on 4K inputs.","The learnable FCT removes the need for a fixed inverse transform, so frequency fusion can be tuned to the data instead of being tied to Haar or other predefined wavelets.","The branch-level results in the paper (low branch 29.90 PSNR, high branch 29.20) suggest each branch is independently strong, so the gains are not just from the fusion module."],"supporting_citations":[{"why":"Supplies the design inspiration for the recursive frequency decomposition that separates low- and high-frequency components.","marker":"[37]"},{"why":"Provides the FHDMi dataset used for training and evaluation, along with the FHDe2Net baseline.","marker":"[2]"},{"why":"Provides the UHDM dataset and the ESDNet/ESDNet-L baselines for ultra-high-definition demoiréing.","marker":"[7]"},{"why":"Motivates the channel-aware design by showing the green channel is less affected by moiré patterns.","marker":"[9]"},{"why":"Represents the wavelet-based dual-branch baseline that the paper argues insufficiently exploits frequency semantics.","marker":"[11]"},{"why":"Represents the multibranch wavelet baseline that the paper argues is computationally heavy and limited by Haar separation.","marker":"[12]"},{"why":"Provides the deep supervision loss formulation used in the first training stage.","marker":"[4]"}],"fun_headline_variants":["Freqformer decomposes moiré to beat larger models","Split moiré by frequency: 6M model wins benchmarks","Frequency split demoiréing: compact transformer tops charts","Texture and color separated: small model, big results"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method assumes the low-frequency component can be downsampled to about one-tenth its size and resized back with roughly 50 dB PSNR, so training on resized low-frequency images is equivalent to training on full-resolution inputs; this is supported only by a one-sentence experiment without protocol details.","fun_headline_variants_meta":{"raw":{"variants":["Freqformer decomposes moiré to beat larger models","Split moiré by frequency: 6M model wins benchmarks","Frequency split demoiréing: compact transformer tops charts","Texture and color separated: small model, big results"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000194,"raw_usage":{"total_tokens":1354,"prompt_tokens":947,"completion_tokens":407,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":563,"completion_tokens_details":{"reasoning_tokens":338}},"tokens_in":563,"tokens_out":407,"duration_ms":4756,"temperature":1.0,"reasoning_tokens":338,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T14:20:09.998617+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a diverse set of FHDMi and UHDM moiré images, apply the paper's L-level decomposition, downsample the low-frequency component to 0.1x, resize it back, and measure reconstruction PSNR against the original low-frequency component. If typical values fall well below roughly 50 dB, or if images with strong low-frequency scene texture break the scale-robustness, the resize-based training strategy degrades and the dual-branch design loses its justification.","supporting_citations":[{"cited_title":"Chan, and Chen Change Loy","cited_arxiv_id":null,"evidence_quote":"Supplies the design inspiration for the recursive frequency decomposition that separates low- and high-frequency components."},{"cited_title":"Fhde 2 net: Full high definition demoireing network","cited_arxiv_id":null,"evidence_quote":"Provides the FHDMi dataset used for training and evaluation, along with the FHDe2Net baseline."},{"cited_title":"Towards efficient and scale-robust ultra-high-definition image demoiréing","cited_arxiv_id":null,"evidence_quote":"Provides the UHDM dataset and the ESDNet/ESDNet-L baselines for ultra-high-definition demoiréing."},{"cited_title":"P-bic: Ultra-high-definition image moiré patterns removal via patch bilateral compensation","cited_arxiv_id":null,"evidence_quote":"Motivates the channel-aware design by showing the green channel is less affected by moiré patterns."},{"cited_title":"Wavelet-based dual-branch network for image demoiréing","cited_arxiv_id":null,"evidence_quote":"Represents the wavelet-based dual-branch baseline that the paper argues insufficiently exploits frequency semantics."},{"cited_title":"Multibranch wavelet-based network for image demoiréing","cited_arxiv_id":null,"evidence_quote":"Represents the multibranch wavelet baseline that the paper argues is computationally heavy and limited by Haar separation."},{"cited_title":"Image demoireing with learnable bandpass filters","cited_arxiv_id":null,"evidence_quote":"Provides the deep supervision loss formulation used in the first training stage."}],"review_version":1}