{"id":"c260f749-004b-4ca6-9e0e-d49e33e242c3","arxiv_id":"2506.20302","paper_version":2,"verdict":"REJECT","confidence":"MODERATE","novelty_score":2.0,"correctness_risk":"high","formal_verification":"none","parameter_count":1,"one_line_summary":"TDiR wraps PromptIR in a DDPM and reports mixed benchmark results, but the method is underspecified and the data contradict some claims.","lead":"A team combines a diffusion framework with a transformer-based denoiser (PromptIR) and reports results on underwater enhancement, denoising, and deraining. Their own tables show lower SSIM than baselines at high noise and on deraining, and the method's training details are incomplete.","discovery_kind":"incremental","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq. (5) omits the degraded input from the training loss, so the model as described is an unconditional denoiser; no reverse-process conditioning is specified, leaving the central restoration claim unsupported.","rationale":"The central claim requires the generative model to be conditional on the specific degraded input; otherwise the task is unconditional generation, not restoration. The paper's Eq. (5) trains a denoiser to predict Gaussian noise from a forward-noised clean image, and no sampling-time term involving the degraded input is given. This is an internal-consistency problem rather than a disagreement with consensus: an unconditional DDPM trained on clean images cannot select the clean image matching a particular corrupted input. The reader's weakest-assumption analysis identifies this same load-bearing gap. I considered whether the internal metric contradictions (e.g., TDiR SSIM 0.653 vs. PromptIR 0.799 at σ=50, and 0.947 vs. 0.979 on Rain100L) are more directly falsifying, but the conditioning omission is more fundamental: if the model is not conditional, all positive reported numbers are suspect regardless of how the tables are read. No code, data, or hyperparameters are released, so there is no independent support to offset this. The proposed test—re-implementing the specified protocol and checking whether outputs correlate with inputs—would settle whether the concern lands. Because the concern supports the reader's REJECT verdict rather than changing it, the verdict remains UNCHANGED.","tokens_in":11125,"tokens_out":6322,"duration_ms":72669,"concrete_test":"Re-implement the exact Section 3 protocol: generate ey by forward noising clean images only, train fθ(ey, γ) with Eq. (5), and sample with no reference to the observed degraded image x. Evaluate on UIEB and Rain100L. If the outputs are uncorrelated with the inputs and PSNR is near the level of a randomly sampled clean image, the conditioning gap is confirmed and Table 1's 22.90 dB cannot be reproduced. If the authors intended to condition, the test should instead trace where x enters fθ in Figure 1 and rewrite the loss as ||fθ(x, ey, γ) − ε||1 with the reverse sampler explicitly conditioning on x.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 3 defines the forward process on a clean image y0=y and then trains fθ with ||fθ(ey, γ) − ε||1, where ey is the forward-noised version of the clean image. The observed degraded input x (underwater image, noisy image, or rainy image) never appears in the loss function, and no term involving x is specified in the reverse process. Without x, the network has no information about which clean image it should produce; at inference it can at best sample from the marginal clean-image distribution. The sentence 'we build on this network by adding a path to the decoder part to concatenate the input and the noise timestep' hints at conditioning, but the formal specification does not state how the degraded input enters fθ. If the path exists, Eq. (5) and the surrounding description are incomplete; if it does not exist, the reported UIEB and Rain100L results cannot be produced by the described system. This is the load-bearing condition for the abstract's claim that the diffusion model 'surpasses current methods in performance.' The reader's weakest-assumption analysis identifies exactly this gap, and the paper's own tables also undercut the 'outperforms all' phrasing (for example, TDiR SSIM is 0.653 versus PromptIR's 0.799 at σ=50, and 0.947 versus 0.979 on Rain100L), but the missing conditioning is the more fundamental blocker: a model that never sees the degraded input cannot perform restoration at all.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes TDiR, an image-restoration method that uses a transformer-based U-Net, specifically the PromptIR architecture, as the denoiser inside a DDPM-style diffusion framework. The method is evaluated on three tasks — underwater image enhancement (UIEB and Test-60), Gaussian denoising (BSD68), and deraining (Rain100L) — and compared with 18 prior methods using PSNR, SSIM, UIQM, and UCIQE. The central claim, stated in the abstract and conclusion, is that the diffusion model combined with transformers outperforms current methods on these benchmarks.","tokens_in":11501,"tokens_out":2847,"duration_ms":31465,"significance":"If the central claim were supported, the contribution would be incremental: the main novelty is replacing the usual U-Net denoiser with PromptIR and adding a decoder path that concatenates the input and the noise timestep. The paper does not provide code, data, or a procedure detailed enough to reproduce the experiments, and it reports no ablations or parameter-free derivations. The manuscript's own tables contain numbers that contradict the 'outperforms all' claim in multiple cells. The work has no machine-checked proofs or reproducible code to weigh in its favor. Its potential value is a possible empirical data point for transformer-based diffusion denoisers, but as written the evidence does not establish even that modest claim.","major_comments":[{"comment":"The training loss is written as ||fθ(ey, γ) − ε||1, where ey is the forward-noised version of the clean image y. The observed degraded input x — the underwater, noisy, or rainy image — never appears in the loss function, and no term involving x is specified in the reverse process. Without x, the network has no information about which clean image it should produce; at inference it can at best sample from the marginal clean-image distribution. The sentence 'we build on this network by adding a path to the decoder part to concatenate the input and the noise timestep' is the only hint of conditioning, but the formal specification does not state how the degraded input enters fθ. This missing conditioning is load-bearing for the paper's central claim: a model that never sees the degraded input cannot perform restoration at all.","section":"Section 3, Eq. (5)"},{"comment":"The manuscript repeatedly claims that TDiR 'outperforms all benchmark methods,' but the reported numbers contradict this. In Table 2, at σ=50, TDiR reports SSIM 0.653 while PromptIR reports 0.799 and AirNet reports 0.797. In Table 3, TDiR reports SSIM 0.947 on Rain100L, while PromptIR* reports 0.979 and AirNet reports 0.967. Section 5.2 acknowledges that the model is outperformed by a single-task denoiser, yet the same paragraph asserts that the model outperforms all benchmark methods. The quantitative evidence in the paper itself undercuts the central claim, so the positive conclusions in the abstract and Section 6 are not supported by the presented data.","section":"Tables 2 and 3"},{"comment":"The diffusion process is severely underspecified. The paper does not state the number of diffusion timesteps T, the noise scheduler β_t, the sampling procedure (DDPM, DDIM, or other), the number of sampling steps at test time, or how the timestep γ is embedded into the PromptIR network. Section 4.1 mentions a pre-trained encoder and training only part of the decoder, but it does not specify which part, what learning rate or schedule was used, what patch size or data augmentation was applied, or how the prompt blocks interact with the diffusion conditioning. Without these details, the experiments cannot be reproduced, which is a fundamental requirement for the empirical claims made in the paper.","section":"Section 4.1 and Section 3"}],"minor_comments":[{"comment":"The caption contains a typo: 'shops' should be 'shows.'","section":"Fig. 1 caption"},{"comment":"The baseline name 'PrompIR' is misspelled; it should be 'PromptIR.'","section":"Table 1"},{"comment":"Equation (4) is written as q(yt|y0) = N(yt, sqrt(α_t)y0, (1 − α_t)I), but the standard DDPM form uses N(yt; sqrt(α_t)y0, (1 − α_t)I), and the missing semicolon makes the notation ambiguous. Also, the subscript 1 in Eq. (5) is unusual and should be clarified.","section":"Section 3, Eq. (4)"},{"comment":"The evaluation details are incomplete: it is not stated whether PSNR and SSIM are computed on the luminance channel or on RGB, nor how the no-reference metrics are applied to the different tasks.","section":"Section 4.2"},{"comment":"The Data Availability and Code Availability statements both say 'NA,' which makes the experimental results non-reproducible; at minimum, the authors should state which trained models or evaluation scripts, if any, will be released.","section":"Declarations"},{"comment":"The reference list has inconsistent formatting; for example, [21] and [30] lack venue and year information or use different styles. Some citations are also incomplete, such as [41] which has no publication venue.","section":"References"}],"recommendation":"reject","confidential_remarks":"The manuscript appears to be an early draft. The missing conditioning in Eq. (5) is not a minor omission: it means the described system cannot perform the claimed restoration, and the internal contradictions between the text and Tables 2 and 3 further undermine the central claim. The authors would need to rewrite the methodology with a fully specified conditioning mechanism, retrain and re-evaluate the model, correct the quantitative claims, and provide enough experimental detail to make the work reproducible before this could be considered for peer review. The paper is also very close to PromptIR in architecture, and the incremental contribution ('adding a path to the decoder') is not enough to justify publication without stronger evidence and a clearer positioning relative to existing diffusion-restoration literature."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper is a thin combination of PromptIR and a standard DDPM, and as written it doesn't actually describe a restoration model. The training loss in Eq. (5) takes ey, the forward-noised clean image, and predicts the noise; the degraded input x never appears. The reverse process is not specified either. The one sentence about adding a path to concatenate 'the input and the noise timestep' hints at conditioning, but the formal description and Figure 1 don't show it. So either Eq. (5) is missing a term, or the model as described is an unconditional denoiser that could not remove rain or underwater color cast.\n\nWhat the paper does well is mostly organizational: it runs experiments on three tasks, five benchmarks, 18 baselines, and gives qualitative figures. The idea of using PromptIR as the denoiser inside diffusion is reasonable, and the reported PSNR numbers on UIEB and BSD68 are in the right ballpark. But the paper's own tables undercut its claims. On BSD68 at sigma=50, TDiR SSIM is 0.653 versus PromptIR's 0.799; on Rain100L, TDiR SSIM is 0.947 versus PromptIR's 0.979, yet the text says the model outperforms all benchmark methods. That is a direct contradiction.\n\nThe methodological gaps are not minor. There is no diffusion timestep count, noise schedule, or sampling procedure. Training details are limited to 'batch size 32, Adam, L1 loss.' No code or data is released. The only novel component, the added skip path, is described in a single sentence. The paper itself notes blocking artifacts and performance drops, but doesn't quantify them.\n\nThis is not reproducible or verifiable from the manuscript. The central claim is unsupported by the described system, and the reported numbers don't match the conclusions. For readers, this could be a useful negative example of how to mismatch a training loss with a goal, but not as a method to build on. It should be desk rejected. If the authors clarify the conditioning, release code, and fix the table inconsistencies, a revised submission might deserve a look, but this version does not.","headline":"As written, TDiR's training loss never sees the degraded input, so the central restoration claim is unsupported; the tables also contradict the 'outperforms all' text.","tokens_in":11978,"tokens_out":2453,"would_cite":false,"duration_ms":25841,"reading_group":"no","serious_thinker":"no","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A transformer-based diffusion model claims to beat 18 state-of-the-art methods on three restoration tasks — underwater enhancement, Gaussian denoising, and deraining — with top PSNR/SSIM on UIEB, the best UIQM on Test-60, and leading…","keywords":["image restoration","diffusion models","transformers","underwater image enhancement","image denoising","image deraining","generative models","all-in-one image restoration"],"falsifier":"Train and run the pipeline exactly as Eq. (5) is written — noised clean images in, predicted Gaussian noise out — while logging every tensor that enters the U-Net during the reverse process; if the degraded photograph never appears at training or inference, the output cannot depend on the specific input, and the UIEB and Rain100L gains must be inherited from the pretrained encoder or the clean-image prior rather than from conditioning. A quicker check a reader can do without retraining: run the reverse chain from the same starting noise on two very different degraded images, and see whether the outputs are nearly identical, which would show the input is not conditioning the restoration.","tokens_in":10973,"feed_emoji":"🖼️","tokens_out":18032,"duration_ms":167343,"temperature":0.7,"pith_summary":"TDiR asks whether one generative architecture can restore three very different kinds of degradation — underwater color cast, Gaussian noise, and rain streaks — more effectively than the specialist methods built for each. The paper's answer is yes: a denoising diffusion model whose denoiser is a transformer U-Net reports the best scores among 18 compared methods on five standard benchmarks, including 22.90 dB PSNR and 0.8724 SSIM on the UIEB underwater set, a UIQM of 2.73 on Test-60, and leading results on BSD68 denoising and Rain100L deraining. If these numbers hold, the practical payoff is one generalist model that replaces task-specific restorers in downstream applications such as detection and classification that demand clean input imagery. The mechanism is the standard diffusion forward-reverse loop, with a transformer U-Net trained to predict the Gaussian noise added to clean images, which leaves the question of how the actual degraded photograph enters that loop as the crux of the claim.","feed_headline":"Diffusion plus transformer tops 18 image-restoration rivals","feed_subtitle":"One generalist model claims top scores on underwater, noisy, and rainy benchmarks — one checkpoint for three restoration jobs.","key_machinery":"The load-bearing object is the denoising diffusion loop with a transformer U-Net as its denoiser. In the forward pass, Gaussian noise is added to a clean image under the scheduler $\\beta_t$, giving the closed-form noised sample $y_t = \\sqrt{\\bar\\alpha_t}\\,y_0 + \\sqrt{1-\\bar\\alpha_t}\\,\\epsilon$; the denoiser $f_\\theta$ — a four-level transformer U-Net with prompt modules, reused from the PromptIR architecture, plus a decoder path that concatenates the input with the noising timestep — is trained to predict the added noise $\\epsilon$ under the $\\ell^1$ loss of Eq. (5). The encoder stays frozen at its pretrained state, so only part of the decoder is actually trained. The reverse process iteratively subtracts the predicted noise to return a restored image, and this identity — predicting the noise is the same as learning to score the data distribution, and a transformer attention backbone can approximate that score across several degradation families — is what carries the argument.","core_discovery":"The paper's central claim is that a diffusion model with a transformer U-Net denoiser produces state-of-the-art restored images on three tasks at once, without a separate network per degradation. The forward process adds Gaussian noise to a clean image over $T$ steps, $q(y_t|y_{t-1}) = \\mathcal{N}(y_t; \\sqrt{1-\\beta_t}\\,y_{t-1}, \\beta_t I)$, closing to $q(y_t|y_0) = \\mathcal{N}(y_t; \\sqrt{\\bar\\alpha_t}\\,y_0, (1-\\bar\\alpha_t)I)$, so any noised sample can be drawn directly. A transformer U-Net — the PromptIR restoration architecture plus a decoder path that concatenates the input with the noise timestep — is trained to predict the added noise $\\epsilon$ under the $\\ell^1$ loss $\\|f_\\theta(\\tilde y, \\gamma) - \\epsilon\\|_1$, and the reverse process converts Gaussian noise back into a restored image. The reported numbers are: 22.90 dB PSNR and 0.8724 SSIM on UIEB, outranking all compared underwater methods; UIQM 2.73 on Test-60 against a next-best near 1.65; the best PSNR among all compared methods on BSD68 at every tested noise level; and 37.43 dB PSNR on Rain100L. The paper interprets these results as showing that diffusion's ability to sample from multimodal degradation distributions, steered by transformer attention, transfers in a single generalist model across noise, rain, and underwater color shift.","pith_inferences":["An open question the paper leaves unresolved is what tensor is actually concatenated at the decoder: if it is only the running noisy sample rather than the original degraded photograph, the reported gains come from the pretrained backbone and the clean-image prior, making an ablation that feeds the degraded image at every reverse step the decisive experiment.","A testable extension: feed two very different inputs, such as a green-cast underwater scene and a rain-streaked street, through the same reverse chain and measure output similarity — a conditioning-free sampler should return near-identical outputs.","The paper's own tables imply a trade it does not discuss: the $\\sigma=50$ row shows PSNR 28.62 with SSIM 0.653, well below the baseline's 0.799, so the high-noise win sacrifices structural fidelity; evaluating the outputs on detection or matching tasks would show whether that trade is acceptable in practice.","If the conditioning path is made explicit, the same architecture points toward a single model spanning more degradation families such as haze, blur, and low light, because the transformer U-Net already handles several prompt-driven restoration tasks and the diffusion loop is task-agnostic."],"forward_implications":["One checkpoint replaces three: the same transformer-diffusion model posts top scores on underwater enhancement, Gaussian denoising, and deraining benchmarks where specialists previously dominated.","The biggest margin is perceptual: UIQM 2.73 on Test-60 against a next-best near 1.65 means the visible gain is largest where no ground-truth reference constrains the output.","Adaptation is cheap in design: with the encoder frozen and only part of the decoder trained, a new degradation type could be added without retraining the whole network.","On Rain100L the paper reports 37.43 dB PSNR, above the all-in-one baseline it starts from and above a version trained on deraining alone, which the paper reads as evidence that the diffusion loop adds capacity beyond task mixing."],"supporting_citations":[{"why":"Supplies the DDPM forward noising schedule, the closed-form sampling of $y_t$, and the noise-prediction objective that Eq. (5) instantiates.","marker":"[30]"},{"why":"Foundational formulation of diffusion as a nonequilibrium thermodynamics process, grounding the backward denoising stage the paper adopts.","marker":"[31]"},{"why":"PromptIR, the transformer U-Net used as the diffusion denoiser and the all-in-one baseline whose numbers the paper builds on and reports beating in Tables 1–3.","marker":"[33]"},{"why":"UIEB, the underwater benchmark that provides the training references and the PSNR/SSIM and Test-60 evaluation reported in Table 1.","marker":"[37]"},{"why":"BSD68, the standard Gaussian-denoising test set on which the Table 2 comparisons are reported.","marker":"[45]"},{"why":"Cited as the Rain100L deraining dataset used for training and for the Table 3 evaluation.","marker":"[46]"},{"why":"Prior application of a denoising diffusion probabilistic model to underwater enhancement, the line TDiR extends by using a transformer denoiser.","marker":"[32]"}],"fun_headline_variants":["TDiR: one diffusion model restores underwater, noisy, rainy images","Transformer diffusion beats 18 methods on three image-restoration tasks","Single checkpoint: state-of-the-art denoising, deraining, underwater fix","TDiR diffusion tops benchmarks for noise, rain, and color cast","Generalist transformer diffusion outguns task-specific restoration nets"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The entire restoration claim rests on the degraded photograph actually being fed to the denoiser: the training loss in Eq. (5) only predicts the Gaussian noise added to a clean image and never mentions the underwater or rainy input, so if the observed image never enters the network, the model has no information about the specific corruption it must remove.","fun_headline_variants_meta":{"raw":{"variants":["TDiR: one diffusion model restores underwater, noisy, rainy images","Transformer diffusion beats 18 methods on three image-restoration tasks","Single checkpoint: state-of-the-art denoising, deraining, underwater fix","TDiR diffusion tops benchmarks for noise, rain, and color cast","Generalist transformer diffusion outguns task-specific restoration nets"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000155,"raw_usage":{"total_tokens":1250,"prompt_tokens":1016,"completion_tokens":234,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":632,"completion_tokens_details":{"reasoning_tokens":155}},"tokens_in":632,"tokens_out":234,"duration_ms":2874,"temperature":1.0,"reasoning_tokens":155,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T22:51:13.215124+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train and run the pipeline exactly as Eq. (5) is written — noised clean images in, predicted Gaussian noise out — while logging every tensor that enters the U-Net during the reverse process; if the degraded photograph never appears at training or inference, the output cannot depend on the specific input, and the UIEB and Rain100L gains must be inherited from the pretrained encoder or the clean-image prior rather than from conditioning. A quicker check a reader can do without retraining: run the reverse chain from the same starting noise on two very different degraded images, and see whether the outputs are nearly identical, which would show the input is not conditioning the restoration.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the DDPM forward noising schedule, the closed-form sampling of $y_t$, and the noise-prediction objective that Eq. (5) instantiates."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Foundational formulation of diffusion as a nonequilibrium thermodynamics process, grounding the backward denoising stage the paper adopts."},{"cited_title":"In: Oh, A., Neumann, T., Globerson, A., Saenko, K., Hardt, M., Levine, S","cited_arxiv_id":null,"evidence_quote":"PromptIR, the transformer U-Net used as the diffusion denoiser and the all-in-one baseline whose numbers the paper builds on and reports beating in Tables 1–3."},{"cited_title":"IEEE Transactions on Image Processing 29, 4376–4389 (2020) https://doi.org/10.1109/TIP.2019.2955241","cited_arxiv_id":null,"evidence_quote":"UIEB, the underwater benchmark that provides the training references and the PSNR/SSIM and Test-60 evaluation reported in Table 1."},{"cited_title":"IJCV (2009) 16","cited_arxiv_id":null,"evidence_quote":"BSD68, the standard Gaussian-denoising test set on which the Table 2 comparisons are reported."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Cited as the Rain100L deraining dataset used for training and for the Table 3 evaluation."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Prior application of a denoising diffusion probabilistic model to underwater enhancement, the line TDiR extends by using a transformer denoiser."}],"review_version":1}