{"id":"07d0d7f4-4526-4b24-be9f-10203694c445","arxiv_id":"2411.17687","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"A controlled diffusion model, GenDeg, generates a large dataset of diverse paired degradations that improves out-of-distribution performance of all-in-one image restoration models when used as training data.","lead":"The paper trains a diffusion model to synthesize realistic degraded images from clean photos, then builds a large augmented dataset called GenDS that improves out-of-distribution generalization of all-in-one image restoration models. This is notable because real paired degradation data is scarce, and the results suggest a practical path to more robust restoration.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"No matched-scale control: Table 2 and Fig. 6 show GenDS (existing + ~550k generated) beats existing data alone, but never compare against adding an equal number of classically synthesized degradations, so the OOD gains cannot yet be attributed to GenDeg's diffusion-based degradation manifold.","rationale":"The reader's weakest assumption is that GenDeg-generated degradations overlap the real-world OOD manifold, with the t-SNE of Fig. 3 cited as the only qualitative evidence. I agree that this is a soft spot, but I think the sharper load-bearing issue is attribution: even if the generated patterns do overlap the OOD manifold, the paper has not shown that the diffusion generator is responsible for the improvements, because no matched-scale baseline exists. The strongest empirical claim, as stated, is only that training on GenDS improves OOD performance relative to existing data alone; that claim is supported by Table 2 across five architectures. However, the paper's contribution and novelty rest on GenDeg specifically, and without a control that adds an equal volume of non-diffusion synthetic degradations, the diffusion-specific mechanism is unproven. This does not overturn the paper's practical value, and the existing evidence is substantial enough to keep a conditional verdict, but it should be made an explicit condition of acceptance. I therefore keep the reader's CONDITIONAL verdict unchanged while sharpening the required check.","tokens_in":14592,"tokens_out":10985,"duration_ms":116119,"concrete_test":"Build a control dataset of roughly 550k pairs by applying classical degradation models to the same ~120k clean images used for GenDeg: atmospheric scattering for haze with random transmission maps, parametric rain/snow streak rendering, motion-blur convolution kernels, gamma/retinex-based low-light simulation, and raindrop overlays. Sample μ/σ from the same histograms, apply the same filtering thresholds, and generate five degradation types per clean image. Train NAFNet and PromptIR on (existing + control) and (existing + GenDeg) under identical schedules, seeds, and compute budgets, then report OOD LPIPS/FID and PSNR/SSIM on the Table 2 test sets. If the classical control matches GenDS within noise, the reported gains are a scale/diversity effect rather than a diffusion-specific effect; if GenDS clearly outperforms the control, the domain-overlap assumption is supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that GenDS materially improves out-of-distribution restoration performance. The evidence compares models trained on existing datasets alone against models trained on GenDS, which is a strict superset with roughly 3-4x more samples and, crucially, multiple degradation types per clean scene. Section 4.1 (Table 2) and Section 4.2 (Fig. 6, Table 4) never include a control where the same ~120k clean images are degraded by classical parametric/physically based generators at matched sample counts, matched μ/σ histograms, and the same five-degradations-per-image scheme. Without such a control, the reported OOD improvements could be driven primarily by dataset size, by the multi-degradation pairing, or by generic diversity, rather than by the specific diffusion-based degradation patterns that GenDeg learns. The paper's own observation that motion blur gains are negligible because existing real data is already sufficient (Sec. 4.1) is consistent with a volume-saturation explanation. Table 4 only compares existing-only, GenDeg-only, and GenDS, all at different sizes, so it does not isolate the generator. Thus the load-bearing assumption is not merely that generated patterns overlap the OOD manifold; it is that the diffusion-generated content, rather than the sheer scale of added data, causes the gains. If a classical control matches GenDS, the paper's novelty and its degradation-manifold story collapse, even though the dataset may still be a useful resource.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes GenDeg, a latent diffusion model that synthesizes degraded images from clean images by conditioning on a text prompt and on the mean (µ) and standard deviation (σ) of the degradation map. The authors generate about 550k degraded samples across six degradation types (haze, rain, snow, motion blur, low-light, raindrops) and combine them with existing restoration datasets to form GenDS, a dataset of roughly 750k paired images. They train five restoration models (NAFNet, PromptIR, a Swin-based model, DA-CLIP, and Diff-Plugin) on GenDS and compare against models trained on existing datasets alone, reporting improved LPIPS/FID on several out-of-distribution (OOD) benchmarks. They also analyze scaling behavior, the effect of µ/σ conditioning, and the domain gap between existing, generated, and OOD degradations.","tokens_in":14930,"tokens_out":8914,"duration_ms":76006,"significance":"If the empirical claim is robust, GenDS is a large-scale practical resource and the paper offers a useful recipe for improving OOD generalization in all-in-one restoration. The evaluation is broad (five architectures and nine OOD test sets), the OOD benchmarks are external and disjoint from GenDeg's training data, and the µ/σ statistics are fitted to existing training data rather than to OOD results, so the evaluation is not circular. However, the study does not isolate the contribution of the diffusion-based degradation manifold from dataset scale and multi-degradation pairing, and it lacks error bars or significance tests. The significance of the reported gains is therefore contingent on additional control experiments.","major_comments":[{"comment":"The reported OOD gains are not attributable to GenDeg's diffusion-based degradation synthesis because every comparison pairs 'existing data alone' against 'existing data plus roughly 550k GenDeg samples.' No control adds an equal number of classically or physically synthesized degradations from the same ~120k clean images, with the same five-degradations-per-image scheme and matched µ/σ distributions. Under this design, the improvements could be driven by dataset size, by generic diversity, or by multi-degradation pairing rather than by the learned diffusion manifold. The saturation after ~100k samples in Fig. 6 and the near-zero motion-blur differences in Table 2 (e.g., NAFNet GoPro FID 28.2 versus 28.7) are consistent with a volume-saturation explanation. Please add a matched-scale classical synthesizer control and provide per-degradation quantitative measures of distributional overlap (e.g., Wasserstein distance between generated and OOD feature distributions), since Fig. 3 is qualitative and shown only for haze.","section":"Sec. 4.1, Table 2; Sec. 4.2, Fig. 6, Table 4"},{"comment":"GenDS gives each clean image five generated degradations that were not present in its original training set, so a model trained on GenDS sees the same scene under multiple corruption types while the existing-data baseline does not. This structural difference alone could improve an AIOR model's ability to disentangle scene content from degradation, independent of whether the degradations come from GenDeg or from any other generator. Please ablate this factor by training on a dataset with the same 120k clean images each paired with five classically synthesized degradations, or by otherwise matching the multi-degradation-per-image structure.","section":"Sec. 3.2"},{"comment":"The paper repeatedly describes improvements as 'significant' but reports no error bars, confidence intervals, or significance tests. Several individual OOD cells show small or opposite changes (e.g., NAFNet GoPro FID 28.2 versus 28.7; Swin GoPro FID 31.7 versus 36.2), and the aggregate gains in Table 2 appear to come from single runs. Please report variance over at least three seeds for the core comparisons, or provide paired significance tests, so that the central claim can be properly assessed.","section":"Table 2; Sec. 4.1"},{"comment":"The statement that GenDS-trained models often produce 'richer colors than the ground truth,' which is used to justify discarding PSNR/SSIM in favor of LPIPS/FID, is concerning because it indicates a systematic deviation from the reference image. If the synthetic data biases models toward more saturated or more colorful outputs, LPIPS and FID gains could partly reflect a stylistic preference rather than improved restoration accuracy. Please quantify the color shift (e.g., mean saturation difference or ∆E) and report PSNR/SSIM for the same models in the main paper, at least for the headline comparisons.","section":"Sec. 4.1, qualitative comparisons"}],"minor_comments":[{"comment":"The filtering thresholds and the µ/σ binning ranges are said to be empirically determined and given in the supplementary, but they are not stated in the main text. Please report them in the main paper or state explicitly that they were fixed before looking at any OOD test results.","section":"Sec. 3.2"},{"comment":"The notation ᾱ_t is not explicitly defined (presumably the cumulative noise schedule), and the weighting term in Eq. (4) should be either derived or cited; the behavior at the boundary t = 0 should also be clarified.","section":"Sec. 3.1, Eqs. (3)-(4)"},{"comment":"Each cell gives 'LPIPS/FID' as 'value/value', but the captions do not label which number is LPIPS and which is FID; please make this explicit and format the tables so that the two metrics are visually distinguishable.","section":"Tables 2 and 3"},{"comment":"The t-SNE visualization is shown only for hazy samples; the caption and text should state this clearly, and the t-SNE hyperparameters (e.g., perplexity and seed) should be reported for reproducibility.","section":"Fig. 3"},{"comment":"The row labels 'Existing data', 'GenDeg data', and 'GenDS data' should be defined precisely, including whether GenDeg data is post-filtering and whether it uses the same clean images as the existing data.","section":"Sec. 4.2, Table 4"},{"comment":"The sentence 'Airnet and [12] used contrastive loss' is grammatically malformed; it should read 'AirNet [12] used a contrastive loss' or similar.","section":"Sec. 2.1"}],"recommendation":"major_revision","confidential_remarks":"The stress-test concern about the missing matched-scale control is valid and is the main obstacle to acceptance. The paper is otherwise a solid empirical study with a potentially useful dataset; if the authors add a classical synthesizer control at matched sample count and report error bars, I would be willing to support acceptance."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is the first credible attempt to use a diffusion model to synthesize paired degradations for all-in-one restoration, and the dataset they are building (roughly 750k pairs) plausibly helps. But the headline comparison is confounded by data volume, so the paper is a solid resource contribution with a weaker attribution claim.\n\nWhat's actually new: fine-tuning InstructPix2Pix to condition on clean image, text prompt, and the mean/std statistics of the degradation map, then filtering the outputs. That is a practical, reproducible recipe. The GenDS dataset, with five degradations per clean scene, is genuinely new and, if released, will be useful for the AIOR subfield. The experiments are broad: five models, six degradations, both generative and non-generative architectures, and the OOD gains are consistent across most cells (LPIPS mostly down, FID mostly down). They also honestly report that motion blur gains are negligible because existing real data is already sufficient, and they report within-distribution performance staying roughly flat—that is the kind of check many data-augmentation papers skip.\n\nSoft spots, in proportion. First, the main comparison is not matched in scale. GenDS is existing data plus roughly 550k generated samples, so the OOD gains could come from sheer volume, from the multi-degradation-per-image pairing, or from the diffusion-specific manifold. The scaling analysis (Fig. 6) shows diminishing returns but never compares against classical parametric degradations at equal sample count and matched mean/std histograms. That missing control means the paper's more interesting claim—that diffusion-based degradation patterns are what close the domain gap—is not actually established. Second, there are no error bars or significance tests; several cells are near-zero or slightly worse (e.g., NAFNet GoPro FID 28.2 to 28.7). Third, the filtering thresholds in Sec. 3.2 are empirically determined per degradation, and the binning is fit to existing datasets; both are reasonable engineering choices but mean the pipeline has several tunable hand-set knobs. Fourth, the t-SNE evidence for domain-gap bridging is qualitative and uses the same DA-CLIP features that define the Wasserstein distances in Table 1, so it is not an independent confirmation.\n\nWho it is for: anyone building training sets for all-in-one restoration. It deserves a serious referee: the resource is likely useful regardless of the attribution question, and an attentive reviewer can push for the matched-scale control. My recommendation: send it to review, with the expectation that the authors add a classical-degradation control before publication.","headline":"Useful dataset, real novelty in conditioning, but the central attribution claim is confounded by data volume.","tokens_in":15460,"tokens_out":1782,"would_cite":true,"duration_ms":15724,"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":"Training all-in-one image restoration models on a 750k-sample dataset that mixes existing data with 550k diffusion-synthesized degraded images substantially improves their performance on out-of-distribution scenes, the paper reports.","keywords":["all-in-one image restoration","diffusion models","synthetic data generation","degradation synthesis","out-of-distribution generalization","image deweathering","low-light enhancement","dataset scaling"],"falsifier":"Evaluate a GenDS-trained model and an existing-data-only model on a held-out degradation produced by a different physical mechanism than any training dataset—for example, haze with a scattering model whose parameters fall outside the $\\mu/\\sigma$ histogram ranges, or rain with a new streak geometry—and compare LPIPS/FID. If the GenDS-trained model does not beat the baseline there, the claim that the synthetic data bridges the gap to genuinely novel degradations is falsified. A direct precursor check is to compute the Wasserstein distance between DA-CLIP features of GenDS samples and such a held-out set; the paper's own metric would show whether the bridge exists.","tokens_in":14408,"feed_emoji":"🖼️","tokens_out":8023,"duration_ms":66389,"temperature":0.7,"pith_summary":"The paper claims that a large training set of synthetic degraded images produced by a diffusion model can fix a core weakness of all-in-one image restoration models: they generalize poorly to degradation patterns and scenes not seen during training. It proposes GenDeg, a degradation and intensity-aware conditional diffusion model that adds controlled haze, rain, snow, motion blur, low-light, and raindrops to clean images, and GenDS, a 750k-sample dataset built from GenDeg plus existing restoration data. The reported result is that five different restoration models trained on GenDS improve out-of-distribution metrics compared with training on existing datasets alone, while within-distribution performance is maintained or improved. If true, this provides a scalable way to obtain paired training data for corruptions where real paired capture is impractical.","feed_headline":"Synthetic degradations boost restoration on unseen scenes","feed_subtitle":"GenDeg trains on 550k synthetic degraded images; models then handle real haze, rain, snow, blur, and low light.","key_machinery":"The load-bearing mechanism is the $\\mu/\\sigma$ conditioning of GenDeg. The generator is trained with the latent-diffusion objective of Eq. (1) on a union of existing restoration datasets, conditioned on the clean image, a scene caption plus degradation phrase, and a one-hot encoding of $\\mu$ and $\\sigma$ binned into 128 levels (plus a null-prompt bin). At inference, $\\mu_{\\mathrm{gen}}$ is sampled from the histogram of a randomly chosen dataset for the target degradation and $\\sigma_{\\mathrm{gen}}$ is sampled conditional on that $\\mu$ bin, producing degradation intensities that match realistic ranges. A Structure Correction Module aligns generated images with their clean counterparts for smooth degradations like haze, blur, and raindrops, while rain, snow, and low-light use a VAE-passed clean image instead. This mechanism is what turns a generic image editor into a controllable paired-data generator that can cover a wider degradation manifold than the original datasets.","core_discovery":"The paper's central claim is that the domain gap experienced by all-in-one image restoration models is largely a data problem, and that a diffusion model can manufacture the missing data. GenDeg is a latent diffusion model that takes a clean image, a text description of the scene and degradation, and a two-number control signal—the mean intensity $\\mu$ and spatial standard deviation $\\sigma$ of the degradation map $c_{\\mathrm{map}} = |x_{\\mathrm{in}} - c_{\\mathrm{img}}|$—and produces a degraded version of that image. Sampling $\\mu$ and $\\sigma$ from histograms of existing datasets and generating roughly 550k additional degraded images from about 120k clean scenes yields GenDS, a paired dataset of over 750k images. The empirical claim is that five restoration architectures trained on GenDS show lower LPIPS and FID on out-of-distribution test sets than the same architectures trained on existing data alone, with the largest gains on haze, low-light, and raindrops, the categories with the smallest and least diverse real datasets.","pith_inferences":["The paper does not test whether GenDeg can generate degradation types beyond the six it was trained on; a natural next experiment is to condition the same pipeline on new prompts and check whether the histogram-sampling machinery transfers to unseen corruptions.","One could turn the $\\mu/\\sigma$ controls into an automatic curriculum, generating training samples at the degradation intensities where a current model's out-of-distribution error is highest, rather than sampling from historical histograms.","A reader might infer that the generation budget should be allocated inversely to the amount of real diverse data available per degradation, since the observed gains are largest for haze, low-light, and raindrops—the categories with the smallest real datasets."],"forward_implications":["Models trained on GenDS should transfer better to real-world haze, rain, snow, motion blur, low light, and raindrop images outside the original training distribution, which is the practical setting for autonomous driving and surveillance.","Because GenDS contains multiple degraded versions of the same clean scene, it enables controlled studies of degradation-invariant representations and makes it possible to isolate scene content from corruption type in what a model learns.","The scaling analysis indicates that the first roughly 100k synthetic samples deliver most of the out-of-distribution gain, so dataset builders can expect diminishing returns beyond that point.","The near-flat motion-blur numbers suggest that synthetic data helps most where real diverse paired data is scarce; for a degradation already well covered by real data, extra synthetic samples add little."],"supporting_citations":[{"why":"Supplies the text-and-image-conditioned latent diffusion formulation that GenDeg adapts for degradation generation.","marker":"[5]"},{"why":"Provides the pretrained latent diffusion model and VAE backbone that GenDeg is built on.","marker":"[40]"},{"why":"Provides the degradation-aware features used to define and measure the out-of-distribution gap and to evaluate models.","marker":"[32]"},{"why":"Contributes the Structure Correction Module idea used to align generated degraded images with clean images.","marker":"[18]"},{"why":"One of the restoration baselines retrained on GenDS to measure the dataset's effect.","marker":"[10]"},{"why":"One of the restoration baselines retrained on GenDS to measure the dataset's effect.","marker":"[35]"},{"why":"A diffusion-based restoration baseline retrained on GenDS to show the dataset benefits generative restoration models.","marker":"[29]"},{"why":"A real hazy test set used as out-of-distribution evaluation data for dehazing.","marker":"[1]"},{"why":"A real hazy video dataset used as out-of-distribution evaluation data for dehazing.","marker":"[60]"},{"why":"A real motion blur dataset used in training and in out-of-distribution evaluation.","marker":"[34]"}],"fun_headline_variants":["Diffusion-generated degradations train restoration for unseen scenes","Synthetic defect data helps all-in-one restoration generalize better","GenDeg's 550k synthetic degraded images close the domain gap","AI-fabricated rain, snow, and blur sharpen real-world restoration","Diffusion-made degradation diversity powers robust image fixing"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The claim rests on the assumption that degradation patterns drawn from the $\\mu/\\sigma$ histograms of existing datasets lie in the same part of degradation space as real-world out-of-distribution test images, so that adding them to training closes the domain gap rather than merely interpolating among already-known synthetic patterns.","fun_headline_variants_meta":{"raw":{"variants":["Diffusion-generated degradations train restoration for unseen scenes","Synthetic defect data helps all-in-one restoration generalize better","GenDeg's 550k synthetic degraded images close the domain gap","AI-fabricated rain, snow, and blur sharpen real-world restoration","Diffusion-made degradation diversity powers robust image fixing"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000444,"raw_usage":{"total_tokens":2268,"prompt_tokens":987,"completion_tokens":1281,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":603,"completion_tokens_details":{"reasoning_tokens":1198}},"tokens_in":603,"tokens_out":1281,"duration_ms":9670,"temperature":1.0,"reasoning_tokens":1198,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T11:49:20.872097+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Evaluate a GenDS-trained model and an existing-data-only model on a held-out degradation produced by a different physical mechanism than any training dataset—for example, haze with a scattering model whose parameters fall outside the $\\mu/\\sigma$ histogram ranges, or rain with a new streak geometry—and compare LPIPS/FID. If the GenDS-trained model does not beat the baseline there, the claim that the synthetic data bridges the gap to genuinely novel degradations is falsified. A direct precursor check is to compute the Wasserstein distance between DA-CLIP features of GenDS samples and such a held-out set; the paper's own metric would show whether the bridge exists.","supporting_citations":[{"cited_title":"In- structpix2pix: Learning to follow image editing instructions","cited_arxiv_id":null,"evidence_quote":"Supplies the text-and-image-conditioned latent diffusion formulation that GenDeg adapts for degradation generation."},{"cited_title":"Simple baselines for image restoration","cited_arxiv_id":null,"evidence_quote":"One of the restoration baselines retrained on GenDS to measure the dataset's effect."},{"cited_title":"Promptir: Prompting for all-in- one image restoration","cited_arxiv_id":null,"evidence_quote":"One of the restoration baselines retrained on GenDS to measure the dataset's effect."},{"cited_title":"Diff-plugin: Revitalizing details for diffusion-based low-level tasks","cited_arxiv_id":null,"evidence_quote":"A diffusion-based restoration baseline retrained on GenDS to show the dataset benefits generative restoration models."},{"cited_title":"Learning to restore hazy video: A new real-world dataset and a new method","cited_arxiv_id":null,"evidence_quote":"A real hazy video dataset used as out-of-distribution evaluation data for dehazing."}],"review_version":1}