{"id":"75edc3e4-d9d9-4095-ad26-20c1008cba1b","arxiv_id":"2501.15510","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Pre-training image restoration encoders to classify the degradation type of the input, then fine-tuning for restoration, yields consistent PSNR gains across CNN and transformer architectures.","lead":"Most image restoration models ignore an obvious clue: what kind of damage (haze, rain, noise, blur) is in the image. This paper pre-trains them to classify that damage, then fine-tunes them for restoration, gaining up to about 2.5 dB on all-in-one tasks and 6.5 dB on mixed damage.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The main tables lack an equal-compute baseline: DCPT adds 100k pre-training iterations on the exact same fine-tuning data, so the reported 2–6 dB gains are not yet isolated from extra training exposure.","rationale":"The paper's central claim is that degradation classification is a cheap and effective pre-training signal, and the strongest evidence is the consistent PSNR improvement across architectures and benchmarks. The most load-bearing assumption is that this improvement is caused by the classification pre-training rather than by the extra 100k iterations of exposure to the same training distribution. The paper partially addresses this with ablations: Table 11 shows generation-only pre-training hurts and classification-only helps, and Table 12 shows a joint DC-Train variant is weaker than the two-stage DCPT. However, none of these controls equalizes total training iterations against a plain restoration baseline. Because the pre-training data is identical to the fine-tuning data, extra exposure is a real confound, and the absence of a from-scratch 850k baseline leaves the mechanism unproven. This is an addressable experimental gap, not a flaw in the method's internal logic, so the conditional verdict is appropriate. I credit the authors for releasing code, reporting ablations, and providing cross-architecture consistency; those features make the result plausible but do not remove the need for an equal-compute control.","tokens_in":23310,"tokens_out":3081,"duration_ms":30233,"concrete_test":"Train NAFNet and PromptIR from scratch with the DCPT fine-tuning recipe (128x128 patches, batch size 32, AdamW, cosine decay from 3e-4 to 1e-6) for 850k iterations on the same 5D/10D datasets, and also retrain the published baselines under the same fine-tuning schedule. If the 850k from-scratch model reaches DCPT-level PSNR (e.g., about 29.7–30.3 dB on 5D), the reported gains are an iteration-budget artifact; if it stays near the current 750k baseline (27.6–28.1 dB), the degradation-classification pre-training is vindicated as the active ingredient.","verdict_should_be":"UNCHANGED","load_bearing_attack":"DCPT's protocol is 100k pre-training iterations followed by 750k fine-tuning iterations (Appendix B.2), while the baseline numbers in Tables 2, 3, and 7 appear to be trained for 750k iterations from scratch. Appendix B.1 states explicitly that the all-in-one dataset used for DCPT is identical to the fine-tuning dataset, with only a slight sampler modification. Thus the DCPT pipeline sees every training image for 100k additional iterations, plus a classification loss and a pixel-reconstruction loss during those iterations. The ablations in Tables 11 and 12 show that the classification stage matters and that generation-only pre-training is harmful, and the DC-Train comparison controls for the presence of the classification loss, but none of these experiments includes a plain restoration baseline trained for 850k iterations under the same schedule and patch sampling. A gain concentrated in tasks such as dehazing and low-light, which have abundant training pairs, could plausibly come largely from the extra 100k exposures. Consequently, the central claim that degradation classification is the active ingredient is not yet secured: the missing control is a from-scratch baseline trained for the same total number of iterations on the same data.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Degradation Classification Pre-Training (DCPT), a two-stage pre-training method for universal image restoration. In the classification stage, an encoder (a restoration backbone without its restoration head) is trained so that a lightweight decoder can predict the degradation type from multi-level encoder features; in the generation stage, the encoder is trained with an L1 pixel loss to reconstruct clean images. After 100k pre-training iterations, the encoder is fine-tuned for 750k iterations on downstream restoration tasks. The paper reports consistent PSNR/SSIM gains over base backbones on 5D all-in-one restoration (Table 2), 10D all-in-one restoration (Table 3, +2.55 dB), mixed-degradation CDD restoration (Table 7, up to +5.58 dB), single-task deblurring, and cross-task transfer with a retained, frozen classifier (Table 8). It also includes ablations on decoder architecture, multi-level feature extraction, and loss stages (Tables 9-12), together with a motivating kNN experiment suggesting that restoration models can classify degradation types.","tokens_in":23624,"tokens_out":12636,"duration_ms":111010,"significance":"If the empirical gains replicate under a properly matched training budget, this is a valuable and unusually simple result: a cheap, dataset-intrinsic label—degradation type—can serve as an effective pre-training signal for universal restoration, and keeping the classifier after pre-training enables a form of cross-task transfer. The paper's strengths include its breadth across CNN and Transformer backbones, the explicit ablations of the classification and generation stages (Tables 10-12), the monotonic relationship reported in Table 13, and the public release of code and models. There is no circularity in the evaluation: the pretext objective is degradation classification while the downstream metric is restoration quality on external test sets. The main open risk is that DCPT's protocol gives the pre-trained models 100k additional iterations of exposure to the same data, and the paper does not yet provide an equal-compute control.","major_comments":[{"comment":"DCPT is not compared against an equal-compute baseline. Appendix B.1 states that the pre-training dataset for the all-in-one setting is identical to the fine-tuning dataset, with only slight modifications in the sampler; Appendix B.2 states that DCPT runs for 100k iterations and the subsequent fine-tuning runs for 750k iterations. The baseline rows in Tables 2, 3, 7, and 15 are therefore trained for 750k iterations from scratch while DCPT models receive 100k additional iterations of exposure to the same training images plus an auxiliary classification and pixel-reconstruction loss. The DC-Train ablation in Table 12 also does not equalize total iterations. Please add from-scratch baselines trained for 850k iterations under the same data, patch sampling, batch size, and learning-rate schedule, and also report a matched-budget DC-Train variant. If the gains survive this control, the attribution to degradation classification is supported; without it, the empirical headline of the paper is not secured.","section":"Appendix B.1, Appendix B.2; Tables 2, 3, 7, 15"},{"comment":"The provenance of the baseline numbers is unclear. The paper says that the authors adopt the same training policy for different backbones in the all-in-one setting, but it does not state whether the non-DCPT rows (SwinIR, NAFNet, Restormer, PromptIR, and the competing methods in Tables 2, 3, and 7) were re-trained by the authors under this exact policy or are quoted from their original papers. If they are quoted, differences in training duration, patch size, optimizer, or data sampler can account for part of the reported gap. Please specify which rows were produced under the shared protocol and, for any quoted numbers, re-train the corresponding backbone under the same protocol so that the comparison is controlled.","section":"Appendix B.2; Tables 2, 3, 7"},{"comment":"The method description and the pseudocode disagree on the optimization procedure. Section 3.2 says that the classification and generation stages are alternated within one pre-training iteration to avoid the encoder receiving two distinct gradient flows simultaneously. In the code in Appendix D, however, l_pix and l_cls are accumulated into the same l_total and one backward pass is executed through both encoder forward computations, so the encoder receives gradients from both losses in the same optimizer step. Please clarify which protocol was used in the experiments and align the text with the released code; this affects reproducibility and the interpretation of the stage ablation in Table 11.","section":"Section 3.2 vs Appendix D"},{"comment":"The transfer-learning experiment lacks a complete protocol. Appendix B.2 specifies the source-task training (100k iterations, batch size 8) but does not give the target-task training budget, learning-rate schedule, patch size, or the initialization protocol for the Supervised row. The large DC-guided gains in Table 8, for example the improvement from 25.44 dB to 30.36 dB for denoising-to-deblurring transfer, can only be interpreted if both rows use the same number of target-task iterations and the same fine-tuning schedule. Please provide these details and include matched-budget baselines.","section":"Section 4.5 and Table 8"}],"minor_comments":[{"comment":"There are several typos: 'unkown' in Section 3.1 should be 'unknown'; 'pipline' in Section 3.2 should be 'pipeline'; and 'Insturct-NAFNet' in the Figure 6 legend should be 'Instruct-NAFNet'.","section":"Section 3.1, Section 3.2, Figure 6"},{"comment":"The asterisk on 'PromptIR*' is never defined; please add a footnote explaining what it means.","section":"Table 4"},{"comment":"Table 13 reports a single trajectory without error bars or a statement about seeds; given that the main tables also report single runs, please state the number of seeds or provide variance estimates for the headline experiments.","section":"Table 13"},{"comment":"The sentence 'The results of the non-DC-guided experiments presents' should use the singular verb 'present'.","section":"Section 4.5"},{"comment":"The monotonic relationship in Table 13 is correlational: increasing pre-training iterations simultaneously changes classification accuracy, feature quality, and the amount of pixel-reconstruction exposure, so the causal wording 'performance improved as the initial degradation classification accuracy increased' should be tempered or supported by an intervention that decouples classification accuracy from iteration count.","section":"Appendix A.3"}],"recommendation":"major_revision","confidential_remarks":"The equal-compute control is the single gate for my recommendation. I would be willing to move to accept if the 850k from-scratch control preserves most of the reported gains. I also note that the manuscript header says 'Published as a conference paper at ICLR 2025'; if the journal submission is not an explicitly extended version, the editor should verify the relationship to the ICLR publication and the journal's policy on prior conference publication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's my take on the DCPT paper.\n\nThe new thing is simple and likely useful: instead of masked-image reconstruction, pre-train a restoration encoder by classifying the degradation type of the input with a lightweight decoder, and keep that decoder around for transfer. The motivation experiment is a nice touch—randomly initialized models already classify degradation above chance, and restoration training sharpens that separation. Applying DCPT to SwinIR, NAFNet, Restormer, and PromptIR gives consistent gains on the 5D/10D all-in-one tasks and on the CDD mixed-degradation benchmark. The ablations are honest: generation-only pre-training hurts, classification-only helps, and the DC-Train comparison suggests the pre-training/fine-tuning split itself matters. Code release is a plus.\n\nThe soft spot is the one the stress test flagged. DCPT pre-trains for 100k iterations on the same data it later fine-tunes on for 750k, while the reported baselines are 750k from scratch. So the headline 2–6 dB gains are not isolated from extra training exposure. The ablations soften this—generation-only pre-training with the same extra 100k iterations makes things worse, which argues against a pure 'more compute' explanation. But none of these is a plain from-scratch run for 850k iterations under the identical schedule and patch sampling. I'd want that single experiment before quoting the gains as attributable to the classification objective.\n\nThe other issues are minor: no error bars in the main PSNR/SSIM tables (Table 1 has them, the rest don't), a few typos, and the transfer-learning section could use more explanation. None of these change the core idea.\n\nWho this is for: anyone working on universal image restoration or pre-training recipes for low-level vision. It won't reshape theory, but as a practical default it's worth a serious look. I'd send it to review and ask for the equal-compute baseline plus a variance estimate. The idea deserves to be in the literature; the magnitude claims just need one more control.","headline":"A practical pre-training recipe for restoration with consistent gains, but the headline numbers need an equal-compute control before I'd trust the magnitude.","tokens_in":24060,"tokens_out":3076,"would_cite":true,"duration_ms":28123,"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":"This paper argues that classifying the type of degradation in an image is a nearly free and highly effective pre-training objective for universal image restoration, yielding consistent gains of over 2 dB across architectures and as much…","keywords":["image restoration","degradation classification","pre-training","all-in-one restoration","mixed degradation","transfer learning","weak supervision"],"falsifier":"Run an equal-compute control in which the same backbone trains from scratch on the same paired data for 850,000 iterations with no classification loss; if its PSNR matches or surpasses DCPT's fine-tuned numbers, the degradation-classification signal is not the active ingredient.","tokens_in":23135,"feed_emoji":"🖼️","tokens_out":10661,"duration_ms":82673,"temperature":0.7,"pith_summary":"This paper argues that classifying the type of degradation in an input image is a powerful and nearly free pre-training signal for universal image restoration. The proposed DCPT method first trains a restoration network's encoder so that a lightweight decoder can name the degradation (haze, rain, noise, blur, low light, and more) from multi-level features, while a pixel-reconstruction loss preserves the ability to generate clean images. Across five-task and ten-task all-in-one (single model, many degradations) benchmarks, DCPT raises CNN and Transformer backbones by at least 2.08 dB on average, reaching 2.55 dB on the ten-task setting, and produces an average gain of 6.53 dB across the 11 degradation categories of the CDD mixed-degradation benchmark. The wider point is that restoration models already contain degradation-discrimination information, and making that information explicit before fine-tuning improves restoration and transfer across unseen and mixed damage.","feed_headline":"Naming the damage first lifts image restoration by 6.53 dB","feed_subtitle":"Keeping the degradation classifier lifts all-in-one and mixed-damage restoration on CNN and Transformer backbones.","key_machinery":"The central mechanism is the degradation classifier that DCPT retains after pre-training instead of discarding it as masked-image-modelling methods do. Named DCPT, the framework has two alternating stages per iteration: (a) a classification stage in which a lightweight decoder (ResNet18 with LayerNorm) predicts the degradation from scaled multi-level encoder features under Focal Loss, and (b) a generation stage in which a convolution maps the deepest feature to a clean image under an L1 pixel loss. The retained classifier is then used in DC-guided training, where a frozen decoder adds a cross-entropy term during fine-tuning so the model continues to distinguish clean from degraded inputs; this term is what carries cross-degradation transfer.","core_discovery":"The authors establish that degradation classification is a latent capability of image restoration networks and can be turned into a pre-training objective. Randomly initialized SwinIR, NAFNet, Restormer, and PromptIR classify five degradations at 52 to 71 percent accuracy, and the same models trained on a three-task all-in-one task exceed 94 percent while also recognizing unseen degradation types. DCPT exploits this by pre-training an encoder with a lightweight ResNet18-style classifier on multi-level features using Focal Loss, alternating with a simple pixel-generation stage; the encoder then initializes a restoration model, and the classifier is kept for DC-guided training. The paper's central claim is that this degradation-classification prior, rather than external prompt embeddings or generative priors, is what drives the observed restoration gains.","pith_inferences":["The paper leaves implicit that any paired restoration dataset already supplies the classification label, so DCPT could be applied to synthetic super-resolution or real-world mixed degradation by treating the corruption recipe or a coarse degradation tag as the class.","The correlation the paper reports between pre-training classification accuracy and final PSNR suggests classification accuracy on a held-out degradation set could be used as an early-stopping signal or as a cheap proxy for downstream restoration quality.","A direct testable extension is to keep the frozen classifier at inference time and use its predicted degradation distribution to route the restoration model to task-specific heads or prompts, which the DC-guided training results suggest would work.","If DCPT's gains come from classification prior rather than extra compute, then combining degradation classification with masked-image modeling or contrastive objectives in the same pre-training could compound the benefits; the paper does not test this."],"forward_implications":["CNN and Transformer backbones both improve: average gains of 2.08 dB or more on five-task all-in-one restoration (Table 2).","On ten-task all-in-one restoration, DCPT-NAFNet improves the NAFNet baseline by 2.55 dB average PSNR, with gains on all ten degradations (Table 3, Tables 16-17).","On the CDD mixed-degradation benchmark, DCPT-NAFNet has the highest PSNR among the compared universal restoration models, with an average gain of 6.53 dB over the NAFNet baseline across the 11 degradation categories (Table 18).","DC-guided training shrinks the cross-task transfer penalty: a denoising-trained Restormer loses only 2.56 dB when transferred to deblurring with DC guidance, versus 7.48 dB without it (Table 8).","Pre-training needs only the degradation type already implicit in paired restoration datasets, and the paper reports 20 epochs of pre-training plus 50 epochs of fine-tuning, compared with 1200 epochs for the IDR baseline."],"supporting_citations":[{"why":"PromptIR, the all-in-one restoration backbone and baseline used for the 5D experiments and for the classification-motivation study.","marker":"(Potlapalli et al., 2023)"},{"why":"Restormer, the transformer backbone used for 5D, single-task, and cross-degradation transfer experiments.","marker":"(Zamir et al., 2022)"},{"why":"NAFNet, the CNN backbone used for the 10D and CDD mixed-degradation experiments.","marker":"(Chen et al., 2022)"},{"why":"AirNet, the all-in-one baseline whose dataset assembly and repeated-sampler ratios DCPT inherits.","marker":"(Li et al., 2022)"},{"why":"IDR, the multi-degradation baseline compared on 5D tasks and in the transfer discussion.","marker":"(Zhang et al., 2023)"},{"why":"DACLIP, the CLIP-prompted NAFNet baseline outperformed on 10D all-in-one restoration.","marker":"(Luo et al., 2023b)"},{"why":"OneRestore, source of the CDD benchmark and a key mixed-degradation baseline.","marker":"(Guo et al., 2025)"},{"why":"Focal Loss, the classification loss chosen for long-tail degradation datasets.","marker":"(Lin et al., 2017)"}],"fun_headline_variants":["Degradation classification pre-training lifts image restoration by 6.53 dB","Classify damage first, then restore: DCPT pre-training for universal restoration","Pre-training on degradation type boosts restoration gains up to 6.53 dB","DCPT: Pre-training via degradation classification for all-in-one restoration","Keep the classifier, boost restoration: DCPT pre-training"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The gains come from learning to classify degradation types, not merely from the extra 100,000 pre-training iterations on the same datasets that fine-tuning also uses.","fun_headline_variants_meta":{"raw":{"variants":["Degradation classification pre-training lifts image restoration by 6.53 dB","Classify damage first, then restore: DCPT pre-training for universal restoration","Pre-training on degradation type boosts restoration gains up to 6.53 dB","DCPT: Pre-training via degradation classification for all-in-one restoration","Keep the classifier, boost restoration: DCPT pre-training"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000294,"raw_usage":{"total_tokens":1723,"prompt_tokens":969,"completion_tokens":754,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":585,"completion_tokens_details":{"reasoning_tokens":657}},"tokens_in":585,"tokens_out":754,"duration_ms":7074,"temperature":1.0,"reasoning_tokens":657,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T14:13:16.138275+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run an equal-compute control in which the same backbone trains from scratch on the same paired data for 850,000 iterations with no classification loss; if its PSNR matches or surpasses DCPT's fine-tuned numbers, the degradation-classification signal is not the active ingredient.","supporting_citations":[{"cited_title":"Promptir: Prompting for all-in-one image restoration","cited_arxiv_id":null,"evidence_quote":"PromptIR, the all-in-one restoration backbone and baseline used for the 5D experiments and for the classification-motivation study."}],"review_version":1}