{"id":"870f2422-f660-4478-947c-e0eef673ceb5","arxiv_id":"2412.14630","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":6,"one_line_summary":"CycleRDM reports competitive perceptual quality across nine image restoration and enhancement tasks using a three-stage diffusion process with wavelet-domain calibration, though its superiority claim is weakened by uncontrolled baselines.","lead":"CycleRDM is a diffusion-based model that aims to handle nine image restoration and enhancement tasks in one framework, using multiple denoising stages plus wavelet and text guidance. The paper claims strong results with only 500 training images per task, but the comparisons are not controlled and the code is not yet available.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Head-to-head comparisons with published baselines are not controlled: training-set sizes, test sets, and joint-vs-per-task training differ, so the data-efficiency superiority claim is unsubstantiated.","rationale":"The reader's weakest_assumption identifies the same core problem: published baseline scores are not a valid control for a 500-image training protocol, especially when test sets differ. My stress-test analysis confirms this is the load-bearing weakness. The paper never retrains baselines, uses a nonstandard deblurring test set (BSD instead of GoPro), and omits whether training is joint or per-task. I could not find a secondary concern more central than this evaluation-protocol gap. The architecture description is coherent, and the ablations show internal consistency for the loss terms and multi-stage design, so the problem is not the method's internal logic but the external validity of the comparisons. I concur with the reader's REJECT verdict, with MODERATE confidence: the reported numbers do not rule out that CycleRDM is good; they just fail to substantiate the specific claim of data-efficient unified superiority. The headline claim is exactly what the abstract promises, so the gap affects the paper's central contribution.","tokens_in":17685,"tokens_out":1810,"duration_ms":69712,"concrete_test":"Retrain DA-CLIP (or a standard task-specific baseline such as Restormer for denoising/deblurring and IDT for deraining) on exactly the same 500-image training subset used for CycleRDM on each task, using the same patch size, epochs, and optimizer settings, and evaluate on the same test sets (including BSD for deblurring). If the retrained baselines match or exceed CycleRDM's reported metrics, the claimed data-efficiency superiority fails. As a second check, train CycleRDM jointly on all nine tasks with a single shared model and report per-task numbers; if joint training degrades results, the paper's 'unified model' claim is not supported.","verdict_should_be":"REJECT","load_bearing_attack":"The paper's central claim is that CycleRDM is 'significantly superior' across nine restoration/enhancement tasks using only ~500 training images per task. This claim rests on comparing CycleRDM's numbers against published baseline scores, but the comparison is not controlled in three specific ways. First, the deblurring evaluation is performed on the BSD dataset rather than the GoPro test set used by most published deblurring baselines (Section 4.1), so the 'state-of-the-art' PSNR of 29.056 in Table 2(c) is not comparable to published GoPro numbers. Second, baseline scores such as MAXIM, Restormer, and IDT are taken from their original papers under full-data training protocols, while CycleRDM reports 500-image training, but none of these baselines is retrained on the same 500-image subset, so any observed advantage could reflect training-set composition, test-set choice, or metric computation rather than the proposed architecture. Third, the paper never states whether all nine tasks are trained jointly in one model or separately per task; the phrase 'unified' in the abstract requires a single model, but Table 1 and the experimental sections read like per-task training runs. If separate models were trained per task, the central data-efficiency and unification claims collapse to a collection of task-specific results. Table 7 shows CycleRDM beating DA-CLIP and IR-SDE on average PSNR by 1.4 and 4.3 dB, but those baselines' numbers are also not reproduced under the same protocol. On its own terms, the method could still be a reasonable unified framework; the weakness is that the headline superiority claim is not established by the evidence presented.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes CycleRDM, a three-stage diffusion inference framework for unified image restoration and enhancement across nine degradation types. The method combines a multi-stage degraded-to-normal mapping, wavelet-domain low-frequency calibration, a feature gain module for high-frequency components, CLIP-based multimodal text guidance, and Fourier-domain losses. The authors claim that CycleRDM is significantly superior to task-specific and unified baselines on reconstruction and perceptual metrics while using only up to 500 training images per task. The paper reports quantitative comparisons on deraining, dehazing, denoising, deblurring, raindrop removal, inpainting, low-light enhancement, underwater enhancement, and backlit enhancement, along with ablations of the proposed components.","tokens_in":18072,"tokens_out":4096,"duration_ms":37599,"significance":"If the central claim were fully supported, a single architecture that outperforms task-specific state-of-the-art methods while training on only ~500 images per task would be a practically valuable advance for all-in-one image restoration. The paper also contributes a plausible architectural combination of multi-stage diffusion, wavelet-domain refinement, and multimodal guidance, and it includes ablation studies demonstrating that each component contributes to the reported performance. However, the empirical evidence as presented does not yet substantiate the data-efficiency and superiority claims, because the comparisons are not controlled and a key training-protocol detail is missing.","major_comments":[{"comment":"The deblurring experiment is evaluated on the BSD dataset, whereas most compared baselines (MTRNN, DID-Anet, MSDI-Net, DeepRFT, etc.) report results on GoPro in their original papers. The paper does not state that the baseline numbers were recomputed on BSD. Consequently, the apparent advantage of CycleRDM (e.g., 29.056 vs. DeepRFT's 28.995) mixes different test sets and is not a valid head-to-head comparison. The authors should either re-evaluate all methods on the same test set or report results on the standard GoPro benchmark.","section":"§4.1 and Table 2(c)"},{"comment":"The paper compares CycleRDM, trained on up to 500 random images per task, against published baseline numbers obtained with full training datasets. No baseline is retrained on the same 500-image subsets or under the same training protocol. As a result, the central claim of being 'significantly superior' with only a small number of training samples is not supported; the observed differences could be due to training-set composition, test-set choice, or metric computation rather than the proposed architecture. A controlled comparison, at least for the unified baselines IR-SDE and DA-CLIP, is required to substantiate the data-efficiency claim.","section":"§4.1 and Tables 2-6"},{"comment":"The manuscript never states whether all nine tasks are trained jointly in one shared model or whether separate models are trained per task. The abstract's 'unified' framework and the claim of 'requiring only a small number of training samples' depend on this detail. If separate models are trained per task, the unification claim reduces to a collection of task-specific results, and the total amount of training data across tasks is not 'small' in the unified sense. The authors must clarify the training protocol and, if separate models are used, adjust the claims accordingly.","section":"§4.1, Table 1, and §4.4"},{"comment":"Table 1 lists the Backlight task with a training phase of 0 images, and Section 4.4 reports backlit enhancement results on BackLit300 without describing any training procedure for that task. A diffusion model cannot be trained with zero training images; this is either a typo, an omission of a zero-shot or adaptation mechanism, or a fundamental gap in the experimental description. Please explain how the backlight model was obtained and how the results in Table 5(b) were produced.","section":"Table 1 and §4.4"},{"comment":"The content loss Lcontent in Eq. (10) includes MSE and SSIM terms, and the reported evaluation metrics are PSNR and SSIM. This means part of the reported distortion metrics are directly optimized during training. While this is common in image restoration, the paper's claim of 'significantly superior' reconstruction quality is weakened when comparisons are made without matched training protocols. The authors should explicitly acknowledge this optimization-evaluation overlap and discuss whether the baselines also optimize SSIM/MSE during training.","section":"Eq. (10) and Tables 2-6"}],"minor_comments":[{"comment":"There are several typographical and grammatical errors, such as 'enhancement. enhancement.' and 'And CyclerRDM can reach an effective balance...' These should be corrected before publication.","section":"§4.4 and Fig. 5"},{"comment":"The caption of Table 2 does not state which test set is used for each task, and the section text inconsistently names datasets (e.g., 'RESIDE-6k' and 'RESIDE-6K'). Please standardize the dataset terminology and clearly specify the evaluation protocol for each row.","section":"§4.3 and Table 2"},{"comment":"The sentence describing the forward diffusion time steps says 'we only set the time step of the forward diffusion process to 200' but later says 'we can perform fewer forward diffusion time steps in the third stage.' Please clarify the exact number of training and inference steps for each stage.","section":"§3.2"},{"comment":"The limitation section states that the model has not been trained to recover multiple degradations in the same scene. This limitation should be mentioned in the abstract or introduction, as it affects the claim of generalization to realistic mixed degradations.","section":"§5"}],"recommendation":"major_revision","confidential_remarks":"The paper appears to be an early preprint with a promising architecture but seriously incomplete empirical validation. The key claims of unified training and data efficiency are not verifiable from the current text. The authors should be asked to either provide controlled comparisons with retrained baselines or substantially temper the claims. The backlight training issue is particularly concerning and needs explicit clarification."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know about this paper. First, the CycleRDM architecture is genuinely new as a combination: two-stage diffusion mapping from degraded to rough normal to fine normal, then a third stage doing wavelet low-frequency calibration, plus a feature gain module on wavelet high-frequency, CLIP text guidance, and Fourier losses. Each piece is borrowed from known methods, but the specific cascade is not in the cited literature, and the ablations suggest the multi-stage design earns its keep at least on distortion metrics. Second, the central claim that this works with ~500 images per task and is 'significantly superior' is not supported by the evidence as presented.\n\nWhat the paper does well: the experiments span nine tasks, which is a lot of work. The ablations are informative, and the limitation section honestly acknowledges that mixed degradations in one scene are not handled.\n\nThe soft spot is the evaluation. Baselines are published numbers trained on full datasets or different test sets. Deblurring is tested on BSD while most baselines report GoPro; backlit enhancement uses zero training images and yet is compared with methods that trained on backlit data; and the paper never says whether one joint model was trained for all nine tasks or separate per-task models. 'Unified' requires a single model, and the experimental section reads like per-task runs. None of the baselines were retrained on the same 500-image subsets or evaluated on the same test sets, so any advantage could come from protocol differences rather than the architecture. There are also small internal inconsistencies (the IR-SDE low-light PSNR differs between Table 4 and Table 7) and duplicated phrases that an editor would want cleaned up.\n\nOne more thing: the content loss includes MSE and SSIM, which are also the reported metrics. That is standard supervised practice and not a fatal flaw, but it means the reported gains are optimized, not predicted.\n\nThis paper is for researchers working on unified restoration architectures, especially those interested in multi-stage diffusion and wavelet-domain refinement. It could be a useful baseline to build on once the evaluation is redone with same-protocol comparisons, code release, and a clear statement of joint vs per-task training.\n\nI would send it to peer review, not desk reject, because the architecture is novel enough and the experiments broad enough to merit referee time. But I would expect major revision: controlled comparisons, standard benchmarks, and a realistic claim. My own verdict would be reject as submitted.","headline":"Plausible new architecture for unified restoration, but the headline data-efficiency claim is built on uncontrolled published-baseline comparisons; needs major revision, not desk rejection.","tokens_in":18589,"tokens_out":2472,"would_cite":false,"duration_ms":16160,"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":"CycleRDM claims that a single three-stage diffusion framework, trained on at most 500 images per degradation, outperforms or matches task-specific and unified baselines across nine restoration and enhancement tasks.","keywords":["image restoration","image enhancement","diffusion model","wavelet transform","multi-stage inference","all-in-one restoration","frequency-domain calibration","small-sample training"],"falsifier":"Run a controlled comparison in which every baseline and CycleRDM are trained on exactly the same 500-image subset and evaluated on the same standard benchmark for each task, including GoPro for deblurring, then check whether CycleRDM still holds its reported average PSNR, SSIM, LPIPS, and FID leads.","tokens_in":17476,"feed_emoji":"🖼️","tokens_out":10188,"duration_ms":60534,"temperature":0.7,"pith_summary":"The paper proposes CycleRDM, a diffusion-based framework intended to unify six linear image restoration tasks (dehazing, deraining, deblurring, denoising, raindrop removal, inpainting) with three blind enhancement tasks (low-light, underwater, backlight). Its central claim is that a three-stage inference process—first mapping the degraded image to a rough normal image, then to a finer normal image, then calibrating in the wavelet low-frequency domain—lets one architecture reach top-tier or near-top-tier quality on all nine tasks using only about 500 training images per task. The authors argue this matters because existing unified models either need large training sets or remain biased toward one task type, and diffusion models used as direct generators risk unstable content. If the claim holds, a practical all-in-one restoration system could be adapted to new degradations with very little paired data.","feed_headline":"One diffusion model restores nine degradations with 500 training shots","feed_subtitle":"Three-stage wavelet-calibrated design claims near-best quality on nine tasks trained on 500 images each","key_machinery":"The load-bearing mechanism is a three-stage cycle-reconstruction diffusion process. Stage 1 maps the degraded image to a rough normal domain; Stage 2 maps that rough estimate to the normal domain using the degradation prior learned in Stage 1; Stage 3 performs the final calibration in the wavelet low-frequency domain after a discrete wavelet transform splits the image into low-frequency structure and high-frequency detail. The low-frequency component is refined with a short 10-step diffusion pass, while a feature gain module made of residual dense blocks cleans redundant features from the high-frequency component before the inverse wavelet transform reassembles the output. Training is stabilized by a multimodal text guidance loss in CLIP space, a content loss combining MSE and SSIM, and a Fourier loss on amplitude and phase, all added to the diffusion noise-prediction loss.","core_discovery":"CycleRDM learns a coarse-to-fine mapping from the degraded domain to the normal domain. In Stage 1 the low-quality image conditions a diffusion process that produces a rough normal estimate; in Stage 2 that estimate conditions a second diffusion pass that learns the residual gap to the normal domain. Stage 3 applies a discrete wavelet transform to the Stage 2 output, performs a short 10-step diffusion calibration on the low-frequency component under the learned degradation prior, and passes the high-frequency component through a feature gain module built from residual dense blocks to suppress redundant detail; an inverse wavelet transform then assembles the high-quality image. Training combines a diffusion noise-prediction loss, a CLIP-based multimodal text guidance loss, a content loss in CLIP feature space plus SSIM, and a Fourier loss on amplitude and phase of the spectrum. The paper reports that with no more than 500 randomly selected training images per task, and zero for backlight enhancement, CycleRDM obtains the best or second-best scores on most benchmarks and the best average PSNR, SSIM, LPIPS, and FID against the two unified baselines it evaluates.","pith_inferences":["If the small-data claim transfers to a controlled setting, an obvious extension is to measure how performance scales with training subset size (100, 200, 500 images) to find the minimum viable data budget for each degradation.","The wavelet-split design suggests a general recipe for other generative restoration models: spend diffusion steps on low-frequency structure and use deterministic feature cleanup on high frequencies; this could be tested by swapping the feature gain module for other denoisers.","Because the paper evaluates deblurring on BSD rather than GoPro and does not state whether one joint model or separate per-task models were trained, a direct controlled comparison on the standard benchmark with identical training subsets would clarify the scope of the unification claim.","The CLIP text-guidance component likely makes results sensitive to prompt wording; a robustness test across positive and negative prompt sets would show whether the reported perceptual gains are stable."],"forward_implications":["If the central claim holds, a single architecture can be adapted to a new degradation by training on roughly 500 paired images, which would drastically lower the data cost of all-in-one restoration systems.","The multi-stage design reduces the performance gap between linear restoration tasks and blind enhancement tasks, as shown by the paper's average PSNR, SSIM, LPIPS, and FID across eight tasks improving on both unified baselines.","Performing final calibration in the wavelet low-frequency domain with only 10 denoising steps suggests that high-frequency detail can be handled deterministically, lowering inference cost while preserving perceived quality.","The combination of distortion metrics (PSNR, SSIM) and perceptual metrics (LPIPS, FID, MUSIQ, VIF) improving together indicates the framework can be both faithful and visually pleasing, not just one or the other."],"supporting_citations":[{"why":"Supplies the mean-reverting SDE unified restoration baseline that CycleRDM is compared against on all tasks.","marker":"[21]"},{"why":"Supplies the vision-language controlled multi-task restoration baseline, the main unified comparison method.","marker":"[22]"},{"why":"Provides the wavelet-based diffusion design for low-light enhancement that motivates the wavelet low-frequency calibration.","marker":"[11]"},{"why":"Provides the wavelet-domain diffusion image restoration approach that the final-stage wavelet calibration extends.","marker":"[10]"},{"why":"Provides the generative diffusion prior baseline for unified restoration and enhancement.","marker":"[3]"},{"why":"Supplies the RESIDE-6k dehazing dataset and the FFA-Net baseline for the dehazing experiments.","marker":"[31]"},{"why":"Supplies the Rain100H paired deraining dataset used for the 500-image training and 100-image test.","marker":"[44]"},{"why":"Supplies the BSD dataset used for deblurring evaluation in place of the GoPro benchmark.","marker":"[54]"},{"why":"Supplies the LOLv1 paired low-light training set used to test the small-data enhancement claim.","marker":"[38]"},{"why":"Supplies the LSUI underwater image enhancement dataset used for training and testing.","marker":"[29]"}],"fun_headline_variants":["CycleRDM: wavelet-calibrated diffusion unifies restoration and enhancement","One diffusion model, 500 shots: wavelet-calibrated restoration for 9 tasks","Wavelet-calibrated CycleRDM: few-shot unified restoration for 9 degradations","CycleRDM: 500 images, 9 tasks, wavelet-cleaned diffusion","Few-shot wavelet diffusion model unifies 9 restoration and enhancement tasks"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The central claim assumes that published baseline scores from full-data training can be fairly compared with CycleRDM numbers obtained from 500 random images per task, even though the paper evaluates deblurring on BSD rather than GoPro, uses zero training images for backlight enhancement, and does not state whether one joint model or separate per-task models were trained.","fun_headline_variants_meta":{"raw":{"variants":["CycleRDM: wavelet-calibrated diffusion unifies restoration and enhancement","One diffusion model, 500 shots: wavelet-calibrated restoration for 9 tasks","Wavelet-calibrated CycleRDM: few-shot unified restoration for 9 degradations","CycleRDM: 500 images, 9 tasks, wavelet-cleaned diffusion","Few-shot wavelet diffusion model unifies 9 restoration and enhancement tasks"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000885,"raw_usage":{"total_tokens":3840,"prompt_tokens":979,"completion_tokens":2861,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":595,"completion_tokens_details":{"reasoning_tokens":2757}},"tokens_in":595,"tokens_out":2861,"duration_ms":14986,"temperature":1.0,"reasoning_tokens":2757,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T12:03:27.288912+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run a controlled comparison in which every baseline and CycleRDM are trained on exactly the same 500-image subset and evaluated on the same standard benchmark for each task, including GoPro for deblurring, then check whether CycleRDM still holds its reported average PSNR, SSIM, LPIPS, and FID leads.","supporting_citations":[{"cited_title":"Image restoration with mean-reverting stochastic differential equations","cited_arxiv_id":null,"evidence_quote":"Supplies the mean-reverting SDE unified restoration baseline that CycleRDM is compared against on all tasks."},{"cited_title":"Gustafsson, Zheng Zhao, Jens Sj¨ olund, and Thomas B","cited_arxiv_id":null,"evidence_quote":"Supplies the vision-language controlled multi-task restoration baseline, the main unified comparison method."},{"cited_title":"Low-light image enhancement with wavelet-based diffusion models","cited_arxiv_id":null,"evidence_quote":"Provides the wavelet-based diffusion design for low-light enhancement that motivates the wavelet low-frequency calibration."},{"cited_title":"Wavedm: Wavelet-based diffusion models for image restoration","cited_arxiv_id":null,"evidence_quote":"Provides the wavelet-domain diffusion image restoration approach that the final-stage wavelet calibration extends."},{"cited_title":"Generative diffusion prior for unified image restoration and enhancement","cited_arxiv_id":null,"evidence_quote":"Provides the generative diffusion prior baseline for unified restoration and enhancement."},{"cited_title":"Ffa-net: Feature fusion attention 29 network for single image dehazing","cited_arxiv_id":null,"evidence_quote":"Supplies the RESIDE-6k dehazing dataset and the FFA-Net baseline for the dehazing experiments."},{"cited_title":"Deep joint rain detection and removal from a single image","cited_arxiv_id":null,"evidence_quote":"Supplies the Rain100H paired deraining dataset used for the 500-image training and 100-image test."},{"cited_title":"Real-world video deblurring: A benchmark dataset and an efficient recurrent neural network","cited_arxiv_id":null,"evidence_quote":"Supplies the BSD dataset used for deblurring evaluation in place of the GoPro benchmark."},{"cited_title":"U-shape transformer for underwater image enhancement","cited_arxiv_id":null,"evidence_quote":"Supplies the LSUI underwater image enhancement dataset used for training and testing."}],"review_version":1}