{"id":"a416a098-c61a-4b32-a84c-f69a7f82afa1","arxiv_id":"2505.08526","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A two-stage diffusion pipeline, DCSR, removes solver and noise biases from low-resolution data using an imbalanced SDEdit step, then upscales the corrected fields with cascaded conditional diffusion models.","lead":"Researchers combine diffusion-model correction with cascaded super-resolution to clean and upscale low-fidelity scientific data, such as noisy climate fields or coarse PDE solutions. The pipeline needs only high-fidelity reference data, not paired low-fidelity examples, and it aims to handle several bias types with one model.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"IPD hyperparameters (t1, t2, Te, metric) are selected on the test LFLR data via Algorithm 2 and Figure 3, so the headline fidelity gains may reflect tuning rather than the method.","rationale":"The reader's weakest assumption pins the load-bearing issue to the existence and selection of t1 and t2: the IPD mechanism relies on an alignment between perturbed distributions, but there is no constructive condition and Algorithm 2 searches on the test data. I agree that this is the most load-bearing concern. My own reading of Theorem 1 confirms that its bound grows with ||e||^2, σ^2(t1), and σ^2(t2), so it does not by itself show that correction improves on the input; Proposition 1 is clean but restricted to Gaussian bias. The missing comparison against [49] and [60] further weakens the claim of versatility, though that is a completeness issue rather than an internal inconsistency. The proposed experiments are otherwise substantive and the IPD idea is plausible; with validation-set hyperparameter selection and the omitted baselines added, the central claim could be adequately supported. Therefore the conditional verdict stands, and I do not recommend moving to accept or reject on the basis of this stress-test pass.","tokens_in":23217,"tokens_out":4052,"duration_ms":44664,"concrete_test":"Re-run the 1D advection ablation with a strict validation split: construct an LFLR validation set disjoint from the 100 test samples, select t1, t2 (and, separately, Te and the metric M) using Algorithm 2 solely on that validation set, freeze those hyperparameters, and then evaluate on the held-out test set. Report TVD, RMSE, and MELRw with the frozen times. If performance on held-out data is within noise of BPD or of the uncorrected input, the central correction claim is not supported. As a sanity check, also compare the Algorithm 2 optimum against random or zero (t1, t2) within [0, Te] on the same held-out data to verify that the search is not exploiting test-set noise.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that IPD corrects diverse biases depends on the existence of t1<t2 with p(u_l(t1)) close to p(tilde_u_h(t2)); no constructive condition is provided, and the theoretical results do not supply one. In practice, Algorithm 2 selects t1 and t2 by minimizing a metric between the perturbed test LFLR distribution and the perturbed HFLR reference, and Section 4.1 fixes Te=0.2 and the MELRw metric from test-based boxplots in Figure 3. This makes the correction-step validation circular: the same test samples and closely related metrics are used both to tune and to evaluate. A grid search can always find a minimum even when genuine distributional alignment does not exist, and the chosen times may be overfit to the noise realizations in the small test set. Moreover, Theorem 1 does not establish bias removal: the bound is essentially ||e||^2 plus nonnegative terms, so the corrected point is not guaranteed to be closer to the HFLR target than the input is. Proposition 1 handles only Gaussian bias. Thus the strongest remaining evidence for the central claim is empirical, and that evidence is weakened by test-set hyperparameter selection and by omission of the closest baselines [49] and [60].","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes a two-stage diffusion-based pipeline, DCSR, for correcting and super-resolving low-fidelity, low-resolution (LFLR) data. In the correction stage, an \"imbalanced perturbing and denoising\" (IPD) variant of SDEdit perturbs the LFLR input for a time t1 and then denoises with a score model trained on high-fidelity low-resolution (HFLR) data for an independent time t2. In the super-resolution stage, a cascaded SR3 model upsamples the corrected field. The method is designed to require no physics knowledge and only HFHR/HFLR reference data for training, not paired LFLR data. Algorithm 2 selects t1 and t2 by minimizing a distribution-alignment metric between perturbed LFLR and HFLR sets; Section 4.1 additionally fixes the metric and terminal search time Te empirically. The theoretical results, Proposition 1 and Theorem 1, provide conditional L2 error bounds for the correction step. Numerical experiments on 1D advection, 2D linear elasticity, 2D Navier-Stokes, and ERA5 climate data compare DCSR with interpolation and SR3-only baselines.","tokens_in":23436,"tokens_out":5569,"duration_ms":59905,"significance":"The idea of a physics-free, unpaired correction-then-super-resolution pipeline is timely and potentially useful, and the IPD variant of SDEdit is a plausible and clearly presented contribution. The paper also provides explicit theoretical statements and an ablation of IPD versus balanced perturbing and denoising (BPD), which are strengths. However, the central empirical claim is weakened by two issues: the main hyperparameters t1, t2, Te, and the metric M are selected on the test LFLR datasets, and the reported errors are computed on the same samples; and the theoretical bounds do not, as stated, guarantee that the corrected output is closer to the target than the biased input. If the authors can address the test-set tuning issue with a proper validation protocol and compare against the closest unpaired-debiasing baselines [49] and [60], the contribution would be valuable. As it stands, the evidence for the headline \"versatile fidelity improvement\" claim is not yet conclusive.","major_comments":[{"comment":"The hyperparameters t1, t2, Te, and the metric M are selected using the test LFLR and HFLR datasets, and the reported errors are computed on the same samples. In Section 4.1, Figure 3 is a boxplot of TVD on the test LF solutions, and the text states that \"Consequently, IPD with MELRw metric and Te=0.2 is adopted in Algorithm 3 for the rest numerical examples.\" This is data leakage: Algorithm 2 searches for the pair (t1,t2) that minimizes a distributional metric on the test data, and the final evaluation uses those same test samples. The method may still work, but the current protocol cannot distinguish genuine correction capability from tuning to the test set. Please introduce a validation split or an explicit tuning protocol that does not touch the test samples, and report performance on truly held-out LFLR data.","section":"Algorithm 2 and Section 4.1"},{"comment":"Theorem 1 bounds the correction error by e^{2 Ls t2} [ ||e||^2 + sigma^2(t2) delta + (sigma^2(t1)+sigma^2(t2)) C_lambda ], where e is the bias. This is an upper bound that grows with ||e||^2; it does not establish that the corrected output is closer to the target HFLR data than the input, nor does it show that the bias is removed. The text in Section 3.1 says the theorem characterizes a trade-off, but the abstract and contributions claim that robustness and effectiveness are \"validated theoretically.\" Please either weaken the theoretical claim to a conditional stability bound or add a result that compares the correction error with ||e||^2 and shows a contraction regime.","section":"Theorem 1 (Section 3.1)"},{"comment":"The load-bearing premise of IPD is that there exist t1 < t2 such that p(ul(t1)) aligns closely with p(uh(t2)) while t2 remains small enough to preserve large-scale structure. The manuscript asserts this alignment qualitatively and then searches for such times in Algorithm 2, but provides no constructive condition or quantitative characterization of when such times exist. Since Algorithm 2 will always find a minimizer on a finite grid even if the distributions do not genuinely align, the reader cannot tell whether the reported gains reflect a real alignment or overfitting to the test noise realizations. Please add either a theoretical criterion for alignability, or a sensitivity/ablation study (e.g., comparing the selected times with random or fixed times, and plotting the alignment metric as a function of t1 and t2) that demonstrates the mechanism is operative beyond the grid-search minimum.","section":"Section 3.1, paragraph on IPD and Algorithm 2"},{"comment":"The two closest works, [49] and [60], perform unpaired debiasing at low resolution followed by probabilistic super-resolution, which is exactly the same two-stage architecture as DCSR. The paper differentiates itself by claiming versatility and bias removal, but no numerical comparison against these methods is provided in Section 4; the only baselines are LFLR+Interp and LFLR+SR, neither of which performs a debiasing step. A direct comparison on at least one of the PDE or climate benchmarks is needed to support the claim that DCSR is preferable to or more general than these existing approaches. If such a comparison is infeasible, the paper should state the obstacle explicitly and restrict its claims accordingly.","section":"Sections 1.3 and 4"}],"minor_comments":[{"comment":"The third paired dataset is written as {˜uh_{32,i}, ˜uh_{64,i}} in Algorithm 4 but as {˜uh_i, ˜uh_{64,i}} in Section 3.3. Please make the notation consistent, since ˜uh normally denotes the 32x32 HFLR data.","section":"Algorithm 4 and Section 3.3"},{"comment":"In Eq. (27) of the proof, the term sigma^2(t) delta should presumably be sigma^2(t2) delta, since the reconstruction bound is stated with t2 in Proposition 1 and Lemma 2. This is a typographical inconsistency that should be fixed.","section":"Proof of Theorem 1"},{"comment":"There are several typographical errors, including \"isotrophic\" instead of \"isotropic\", \"repling\" instead of \"relying\", \"Algotirhm\" instead of \"Algorithm\", and \"userspecific\" instead of \"user-specific\". A careful proofreading pass is recommended.","section":"Throughout"},{"comment":"The text says \"The study is conducted on all six types of LF data,\" but Figure 3 and the surrounding discussion show results only for the LW solver with pink noise, with other cases deferred to the appendix. Please either include the additional cases in the main text or state clearly that Figure 3 is representative and that the appendix contains the full ablation.","section":"Section 4.1"}],"recommendation":"major_revision","confidential_remarks":"The relation to [49] and [60] is delicate: the paper presents DCSR as a unified alternative, but without a numerical comparison the novelty over those works is difficult to assess. I would ask the editor to encourage the authors to add such a comparison or to explicitly scope the claim. The test-set hyperparameter selection is the most serious methodological issue; if it is addressed with a clean validation protocol, the paper could become publishable."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper has a genuinely useful idea: decouple the perturbation and denoising times in SDEdit, so you can inject enough noise to screen out bias without destroying large-scale structure. That's a real and modest extension of existing purification methods, and the two-stage debias-plus-super-resolution pipeline is sensibly assembled. The experiments cover physically diverse cases—advection, elasticity, Navier-Stokes, and ERA5 climate fields—which is a strength. The theory is not window dressing: Proposition 1 is a clean statement for Gaussian bias, and Theorem 1 is at least an honest upper bound rather than a fitted claim. The paper is clearly written and the authors engage with the closest prior work by name.\n\nThat said, the central empirical claim is weakened by a methodology problem. Algorithm 2 selects t1, t2, and the metric M by running a grid search over the test LFLR and HFLR data, and Figure 3 is used to justify Te=0.2 and MELRw. The reported TVD and RMSE numbers are then computed on the same samples. That's data leakage. You can always find a minimum on a small test grid even when no true distributional alignment exists, and the chosen times may be overfit to noise realizations. The paper would be much stronger if this selection moved to a validation split.\n\nSecond, the paper omits the closest baselines. [49] and [60] are the same two-stage framework, and [49] is, to my knowledge, the paper that introduced \"debias coarsely, sample conditionally.\" The authors describe differences but never benchmark against either. That matters because the main novelty over those works is the IPD correction step, and you need to see whether it actually helps relative to the state of the art.\n\nThird, Theorem 1 does not establish that correction improves the input. The bound contains ||e||^2 plus nonnegative terms, so the corrected point could be farther from the target than the LFLR input is. Proposition 1 only covers Gaussian bias and relies on the score being well-trained. The theoretical case for general bias removal is therefore weaker than the introduction suggests. The authors should either soften the claim or supply a bound that shows improvement.\n\nThese are fixable issues. The IPD mechanism is the kind of practical trick that could inform future work, and the paper deserves a serious referee. I would not accept it as is, but I would engage with a revision that moves hyperparameter selection to a validation set, adds comparisons to [49] and [60], and reframes the theory as conditional rather than general.","headline":"Genuinely useful IPD idea, but test-set hyperparameter tuning undermines the headline numbers; worth a serious referee after a validation-set redo.","tokens_in":24034,"tokens_out":1535,"would_cite":true,"duration_ms":17030,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["60H10","65M99","68T07"],"pacs":[],"model":"deepseek-v4-flash","headline":"A two-step diffusion pipeline can lift low-fidelity, low-resolution data to high-fidelity, high-resolution output without the governing equations or paired training examples.","keywords":["diffusion models","super-resolution","bias correction","score-based generative models","SDEdit","imbalanced perturbing and denoising","scientific computing","climate data"],"falsifier":"Take a fixed clean high-fidelity dataset and add a strongly non-Gaussian bias, for instance sparse localized spikes with heavy-tailed amplitudes. If no pair $t_1<t_2$ within the search interval makes the alignment metric small while $t_2$ stays below the structure-preserving threshold, then IPD corrections will fail to approach the high-fidelity data; comparing IPD against balanced perturbing-denoising on that bias family and checking whether the Theorem 1 bound is violated would settle whether the method removes diverse biases as claimed.","tokens_in":22917,"feed_emoji":"📈","tokens_out":11916,"duration_ms":105848,"temperature":0.7,"pith_summary":"This paper claims that low-fidelity, low-resolution data can be corrected and upscaled by a two-step diffusion pipeline that never uses the governing physics or paired low- and high-fidelity examples. Correction runs a forward diffusion for a short time $t_1$ to let added noise dominate unknown biases, then denoises with the backward probability-flow ODE for an independent time $t_2$, using a score model trained only on clean high-fidelity low-resolution samples; keeping $t_1 < t_2$ is what allows large-scale structure to survive while bias is removed. Super-resolution then chains three conditional diffusion models (cascaded SR3) to reinsert fine-scale detail. The paper supports the correction step with an $L_2$ error bound and an exact variance-addition argument for Gaussian biases, and tests the full pipeline on three PDE benchmarks and a climate reanalysis dataset.","feed_headline":"Two diffusion passes: strip bias, then restore fine detail","feed_subtitle":"No paired data or physics equations needed: one score model removes unknown biases, a cascaded diffusion chain restores resolution.","key_machinery":"The central mechanism is imbalanced perturbing and denoising (IPD): perturb the low-fidelity input forward for time $t_1$, then run the backward probability-flow ODE for the larger time $t_2$, with the score model trained on clean high-fidelity samples. The load-bearing identity is distribution alignment, $p(u_l(t_1)) \\approx p(\\tilde{u}_h(t_2))$, and the inequality $t_1 < t_2$ is what lets a small perturbation erase bias while a moderate denoising interval keeps large-scale structure intact. In practice, the two times are selected on test data by minimizing a distributional metric, and the experiments adopt a weighted mean energy log ratio. The super-resolution half is cascaded SR3: three conditional diffusion models, each trained on pairs created by a known downsampling operator at magnification $\\times 2$, refine the corrected $32\\times32$ field back to $256\\times256$.","core_discovery":"The central claim is that a score-based diffusion model trained on the target high-fidelity distribution can remove diverse, unknown biases from low-fidelity data if forward perturbation and backward denoising run for two independent times $t_1$ and $t_2$ instead of one shared time. For $0 < t_1 < t_2 < 1$, perturbing $u_l = \\tilde{u}_h + e$ for time $t_1$ gives $u_l(t_1)$, and the paper argues the distribution $p(u_l(t_1))$ aligns with $p(\\tilde{u}_h(t_2))$; running the probability-flow ODE from $t_2$ back to $0$ then maps the perturbed sample toward $\\tilde{u}_h$. Theorem 1 states that, with probability at least $1-\\lambda$, $\\|\\hat{u}_l(t_1,t_2)-\\tilde{u}_h\\|_2^2 \\le e^{2L_s t_2}[\\|e\\|_2^2 + \\sigma^2(t_2)\\delta + (\\sigma^2(t_1)+\\sigma^2(t_2))C_\\lambda]$, where $\\delta$ is the score training loss and $C_\\lambda$ is a dimension-dependent concentration constant. Proposition 1 shows that for Gaussian bias $e\\sim N(0,\\gamma^2 I)$ with $\\sigma^2(t_2)=\\sigma^2(t_1)+\\gamma^2$, the expected correction error is at most $e^{2L_s t_2}\\sigma^2(t_2)\\delta$, vanishing as the training loss goes to zero. This is the mechanism by which IPD can beat balanced perturbing-denoising, whose single time cannot simultaneously screen out the bias and preserve structure.","pith_inferences":["Going beyond the paper: the alignment condition $p(u_l(t_1)) \\approx p(\\tilde{u}_h(t_2))$ presumes the bias can be mimicked by injected Gaussian noise at some stage of the forward SDE; biases with strong spatial correlation or heavy tails may require a learned bias model or repeated correction passes rather than a single IPD pair.","Going beyond the paper: the grid search for $(t_1,t_2)$ is a calibration step on the data itself; a natural extension is to learn a mapping from dataset statistics to the optimal times, removing the test-time search and reducing inference cost.","Going beyond the paper: IPD's logic is distribution-translation rather than physics-specific, so the same correction step could serve as a general-purpose purifier for any target domain where clean samples exist, such as transferring bias between numerical solvers or observational products.","Going beyond the paper: the error bound contains the factor $e^{2L_s t_2}$, so the practical robustness seen on 2D fields is not automatically guaranteed for high-dimensional or strongly nonlinear 3D problems, where the Lipschitz constant of the learned score can be large."],"forward_implications":["One correction model, trained only on clean high-fidelity low-resolution snapshots, handles numerical errors from different PDE solvers and white, pink, and brown noise without retraining or knowing the bias type.","Because bias is removed before upscaling, the super-resolution stage sees data already consistent with the target distribution, so systematic error is not amplified by the finer-grid model.","In the Gaussian-bias case, the correction error bound tends to zero as the score training loss tends to zero, meaning the bias is removed completely in the ideal score limit.","Reported experiments show IPD consistently gives lower total variation distance and narrower variability than balanced perturbing and denoising across metrics and search horizons.","The cascaded SR3 chain makes the method applicable to arbitrary resolution gaps by adding more $\\times2$ stages, as demonstrated by the $32\\times32$ to $256\\times256$ example."],"supporting_citations":[{"why":"Supplies the base SDEdit procedure of perturbing with noise and denoising with a target-domain diffusion model, which IPD modifies.","marker":"[40]"},{"why":"Supplies the SR3 conditional diffusion super-resolution model and the cascaded refinement strategy used in the second step.","marker":"[44]"},{"why":"Provides the score-based SDE framework, reverse-time SDE, probability-flow ODE, and score-matching objective used throughout.","marker":"[47]"},{"why":"Represents the guided-diffusion purification approach that the paper identifies as computationally expensive, motivating IPD.","marker":"[50]"},{"why":"Represents the adversarial diffusion bridge approach that requires prior knowledge of the bias, which IPD avoids.","marker":"[27]"},{"why":"A two-stage unpaired downscaling method using OT-based debiasing, used as a comparison point for the correction step.","marker":"[49]"},{"why":"A two-stage unpaired super-resolution method using enhanced DDIB, used as a comparison point for task-specific correction.","marker":"[60]"},{"why":"Supplies the chi-square concentration inequality that turns the noise norm into the high-probability constant in Theorem 1.","marker":"[23]"},{"why":"Supplies the ERA5 reanalysis climate dataset on which the real-data fidelity experiments are run.","marker":"[17]"}],"fun_headline_variants":["Two diffusion times: strip bias, then upscale","Two-stage diffusion: correct bias, then boost resolution","Two times t1,t2: fix bias, then refine detail","Bias correction and super-resolution in one diffusion model","Imbalanced perturbation: two ticks for bias-free upscaling"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The correction step rests on the premise that some perturbation time $t_1$ and some larger denoising time $t_2$ make the noise-perturbed low-fidelity distribution align closely with the noise-perturbed high-fidelity distribution while $t_2$ is still small enough that backward denoising preserves large-scale structure, and the paper gives no constructive condition guaranteeing such a pair exists for a given bias.","fun_headline_variants_meta":{"raw":{"variants":["Two diffusion times: strip bias, then upscale","Two-stage diffusion: correct bias, then boost resolution","Two times t1,t2: fix bias, then refine detail","Bias correction and super-resolution in one diffusion model","Imbalanced perturbation: two ticks for bias-free upscaling"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000729,"raw_usage":{"total_tokens":3309,"prompt_tokens":1032,"completion_tokens":2277,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":648,"completion_tokens_details":{"reasoning_tokens":2196}},"tokens_in":648,"tokens_out":2277,"duration_ms":14129,"temperature":1.0,"reasoning_tokens":2196,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T21:52:40.129338+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a fixed clean high-fidelity dataset and add a strongly non-Gaussian bias, for instance sparse localized spikes with heavy-tailed amplitudes. If no pair $t_1<t_2$ within the search interval makes the alignment metric small while $t_2$ stays below the structure-preserving threshold, then IPD corrections will fail to approach the high-fidelity data; comparing IPD against balanced perturbing-denoising on that bias family and checking whether the Theorem 1 bound is violated would settle whether the method removes diverse biases as claimed.","supporting_citations":[{"cited_title":"Im- age super-resolution via iterative refinement","cited_arxiv_id":null,"evidence_quote":"Supplies the SR3 conditional diffusion super-resolution model and the cascaded refinement strategy used in the second step."},{"cited_title":"Debias coarsely, sample conditionally: Statistical downscaling through optimal transport and probabilis- tic diffusion models","cited_arxiv_id":null,"evidence_quote":"A two-stage unpaired downscaling method using OT-based debiasing, used as a comparison point for the correction step."},{"cited_title":"Adaptive estimation of a quadratic functional by model selection","cited_arxiv_id":null,"evidence_quote":"Supplies the chi-square concentration inequality that turns the noise norm into the high-probability constant in Theorem 1."},{"cited_title":"The era5 global reanalysis","cited_arxiv_id":null,"evidence_quote":"Supplies the ERA5 reanalysis climate dataset on which the real-data fidelity experiments are run."}],"review_version":1}