{"id":"b1e33b32-6cc1-434f-9fb2-aac901127031","arxiv_id":"2412.12550","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"A training scheme for diffusion restoration models that feeds the model its own generated intermediate states during training, reducing the train/test input gap and improving fidelity across five restoration tasks.","lead":"This paper trains diffusion image restoration models on the model's own backwards-process outputs instead of simple forward-corrupted images, so training and testing inputs match. The authors report large gains over prior restoration models on five tasks, especially in fidelity metrics.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Data-consistent training's error-reduction proof is definitional: Eq. (6) zeroes the input cumulative error by construction, but the claim that this captures all accumulated error is imported from [27] and not re-derived for ResShift.","rationale":"The reader's weakest_assumption identifies exactly the same load-bearing concern: the paper's Eq. (6) defines away the input cumulative error by construction, and the claim that this is the only error conduit is inherited from [27] without verification for ResShift. I agree because the proof of the method's core benefit, 'training loss directly optimizes cumulative error,' depends on the recursion in Eqs. (2)-(3) being a faithful model of error propagation in the specific diffusion backbone. The paper provides no independent derivation of that recursion, and it is not obvious that the recursion holds when the backward process is nonlinear and the error amplification mu_t >= 1. If the recursion fails, the method still has value as a distribution-matching training strategy, but the stronger claim of eliminating accumulated error is unsupported. The reader also raised other issues (denoising protocol on SIDD/Urban100, no error bars, no code), but those are secondary: they threaten the empirical convincibility of one task, not the conceptual correctness of the central argument. The recommended verdict remains CONDITIONAL, because the concern is real but addressable: a careful re-derivation and experiments with per-step error measurements could validate or refute the recursion. No change to the reader's verdict is needed.","tokens_in":17109,"tokens_out":9275,"duration_ms":78562,"concrete_test":"Analytical re-derivation: For a single trained ResShift model, expand the unrolled backward computation and express the error of the final output as a function of per-step errors. Check whether the condition for the cumulative error to equal the modular error (Eq. 7) reduces to the definitional identity D(x_back_t, x_train_t)=0 or requires additional assumptions (e.g., Lipschitz continuity of f_theta, monotonicity of the shift schedule, or independence of noises). If additional assumptions are needed, identify which one fails for ResShift by measuring the Lipschitz constant of f_theta along the backward trajectory. Empirical cross-check: Train two SISR models on RealSR, one with data-consistent training and one with standard training, then on the test set compute at each step the distance between x_back_t and the forward-process input x_forw_t derived from the ground truth.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central theoretical claim is that using backward-process samples as training inputs makes the input cumulative error zero (Eq. 6) and therefore the training loss directly optimizes cumulative error (Eq. 7). This rests on Eqs. (2)-(3), which define the input cumulative error at step t as D(x_back_t, x_train_t). By setting x_train_t = x_back_t, Eq. (6) makes this distance zero by definition. The concern is not that the definition is false, but that the paper treats this definitional identity as if it proves that no other accumulated error remains. A backward-process sample x_back_t is produced by the network's own repeated evaluations; errors from every earlier step are folded into that input. If the network at step t is imperfect, those earlier errors can interact with the step-t mapping in ways not captured by the single distance D(x_back_t, x_train_t). For instance, if f_theta at step t is not a contraction, a small input perturbation can be amplified by mu_t, so the cumulative error should depend on the Jacobian of f_theta, not just on the distance between inputs. Eq. (7) asserts mu_t * 0 = 0 by construction, but Eq. (2) is a claim about how errors actually propagate; it must hold for the specific backbone (ResShift) and schedule used. The paper never shows that ResShift's backward process satisfies the recursion; it merely imports it from [27], which analyzes a different (non-shift) diffusion formulation. If the recursion does not hold exactly, then optimizing the loss in Eq. (8) does not directly optimize the true cumulative error, and the reported SOTA results could be due to a weaker but useful property (training and test input distributions matching) rather than the claimed elimination of accumulated error. This is the most load-bearing concern because both the method's motivation and the claim of direct optimization rely on it.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a training methodology for denoising diffusion models (DDMs) in image restoration, called data-consistent training. The key idea is to replace the forward-process training input x_forw_t with the backward-process sample x_back_t at each diffusion step, so that the training distribution matches the testing distribution. The authors argue that this makes the input cumulative error zero by construction (Eq. 6), so that the training loss directly optimizes the cumulative error (Eq. 7). They instantiate the method on the ResShift backbone and report state-of-the-art or competitive PSNR/SSIM/LPIPS/FID/CLIPIQA numbers on five restoration tasks: super-resolution, denoising, deraining, dehazing, and dual-camera super-resolution. An efficient variant that amplifies a one-step error by a factor of T-t is also introduced. The paper includes an ablation with a latent diffusion model backbone and memory/time comparisons.","tokens_in":17340,"tokens_out":4537,"duration_ms":39038,"significance":"If the central claim is correct, the contribution is attractive: a simple, backbone-agnostic training recipe that reduces the train/test input discrepancy in iterative restoration models, with consistent gains across five tasks and both ResShift and LDM backbones. The paper is also honest in reporting an efficient version with reduced memory and computation, and it includes a useful computational-cost table. However, the theoretical justification is currently definitional: Eq. (6) sets the input cumulative error to zero by choosing the training input equal to the backward input, and the recursion in Eq. (2) is imported from prior work without verification for the ResShift backbone. The empirical evaluation lacks error bars or significance tests, and the denoising protocol has a train/test degradation mismatch. These issues are load-bearing for the paper's claimed guarantee, though the empirical recipe may still be salvageable.","major_comments":[{"comment":"The derivation that data-consistent training removes cumulative error is a definitional identity rather than a substantive result. Eq. (6) states that the input cumulative error is D(x_back_t, x_train_t); since x_train_t is set to x_back_t, this quantity is necessarily zero. Eq. (7) then equates cumulative error with the modular error. This does not prove that no accumulated error remains in the backward trajectory; it only shows that the paper's chosen error recursion returns zero when its two arguments are identical. The manuscript needs to establish that the recursion in Eq. (2) is a faithful description of how errors propagate through the specific backbone and schedule, and that a zero value of the input-distance term accounts for all error contributed by previous iterations. Without that, the central theoretical claim is vacuous.","section":"Sec. 4.1, Eqs. (5)-(7)"},{"comment":"The cumulative-error recursion is stated as 'proved in [27]' and is imported without verification for ResShift. ResShift's backward process is not the same diffusion formulation analyzed in [27], and the paper does not show that the recursion holds for it. A backward input x_back_t is produced by repeated evaluations of the same network f_theta, so errors from earlier iterations can interact nonlinearly with the step-t mapping; the distance D(x_back_t, x_t) between two inputs does not, in general, determine how those earlier errors propagate. The authors should either re-derive the recursion under ResShift's specific update, state the necessary Lipschitz/contraction conditions, or provide an independent numerical test that Eq. (2) actually accounts for propagated error.","section":"Sec. 3.1, Eqs. (2)-(3)"},{"comment":"The denoising setup is inconsistent: training uses SIDD, while testing uses Urban100 with synthetic Gaussian noise of standard deviation 50. SIDD is a real-smartphone-noise dataset, so the train-time degradation distribution does not match the test-time Gaussian corruption; this mismatch can disadvantage task-specific denoisers and makes the comparison hard to interpret. Additionally, no error bars, repeated runs, or significance tests are reported. Given the large reported PSNR gains (e.g., Table 1, 32.25 dB vs. 30.81 dB for MAXIM; Table 2, 34.22 dB vs. 33.91 dB for DA-CLIP), the authors should report variance across at least three training runs or perform paired tests on the test images to confirm that the improvements are not within run-to-run noise.","section":"Sec. 5.1 and Tables 1-5"},{"comment":"The efficient variant depends on two hand-set quantities: it assumes mu_t = 1 in Eq. (2) and sets the amplification factor lambda = T - t. No derivation is given for either choice, and the statement that this ensures 'the strength of error is the same with that in consistent diffusion' is not substantiated. The ablation in Table 7 only compares no amplification, random amplification, and the proposed lambda = T - t; it does not sweep lambda, so the sensitivity of the efficient variant to this assumed propagation coefficient is unknown.","section":"Sec. 4.2, Eqs. (9)-(11)"},{"comment":"The ablation results are internally inconsistent with the main-table narrative in a way that undercuts the claim that the proposed ResShift variant is the best. In Table 6, 'LDM+DCT' achieves 27.60 dB on RealSR, which is higher than 'ResShift+DCT (Ours)' at 25.96 dB from Table 5. Moreover, LDM improves from 25.19 to 27.60 dB (+2.41 dB) under DCT, while ResShift improves only from 24.59 to 25.96 dB (+1.37 dB). This does not by itself invalidate the method, but it complicates the claim that the proposed training is uniformly beneficial and demands an explanation, especially since the main tables highlight only the ResShift-based results.","section":"Table 6 and Sec. 5.2"}],"minor_comments":[{"comment":"Reference [33] for the SIDD dataset is incorrect: it cites a paper about detecting sensitive data exfiltration, not the Single Image Denoising Dataset of Abdelhamed et al. The correct dataset citation should be provided.","section":"References"},{"comment":"There are several typos: 'miles-stone', 'specfically', and 'Compasion Methods' should read 'milestone', 'specifically', and 'Comparison Methods'; the backbone name is also inconsistently spelled as both 'Resshift' and 'ResShift'.","section":"Sec. 2.1 and Sec. 5.1"},{"comment":"Figures 1 and 3 show KL divergence on the y-axis, while Sec. 5.1 states that mean squared error is used as the loss metric beta_t. Please clarify whether the modular and cumulative error plots are computed with KL divergence, MSE, or another metric, since the paper's error definitions in Eq. (1)-(3) use a generic 'D'.","section":"Figs. 1 and 3, Sec. 5.1"},{"comment":"The indexing in Eq. (3) is confusing: the left-hand side is xi_cumu_{t+1}, but the right-hand side is D(x_back_t, x_t), where both arguments are at the same time index t. Please realign the index convention so that the input error that propagates from iteration t+1 to iteration t is unambiguously defined.","section":"Sec. 3.1, Eq. (3)"},{"comment":"The caption of Table 1 says 'We compare 8 general and 4 task-specific restoration models,' but the table lists 12 baseline methods plus Ours; please recount or reclassify the comparison methods.","section":"Table 1"}],"recommendation":"major_revision","confidential_remarks":"The stress-test concern is on target: the central proof in Eqs. (5)-(7) is definitional, and the error-propagation recursion is imported from [27] without verification for ResShift. That said, the empirical recipe is plausible and the experiments are extensive, so the paper is not beyond repair. I would require (1) an honest reformulation of what the method guarantees, either replacing the circular proof with a direct train/test consistency argument or deriving the recursion for ResShift; (2) fixing the denoising protocol and adding error bars; and (3) explaining the Table 6 inconsistency before publication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Candid read: the contribution is real but the headline theory is not. The paper proposes training a diffusion restoration model on its own backward-chain samples instead of one-step forward corruptions. That is a clean, natural idea—scheduled sampling for diffusion restoration—and it delivers strong gains across five tasks. The efficient variant, which amplifies the one-step error by (T-t), is a sensible approximation, and the ablations against the LDM backbone show the trick isn't locked to ResShift. I'd credit the empirical work: extensive tables, a memory/time comparison, and honest acknowledgment that [27] also notices the modular-cumulative gap.\n\nWhere it gets soft: the formal error-propagation argument in Sec. 4.1 is close to circular. Eq. (6) defines the \"input cumulative error\" as the distance between the backward sample and the training input; by setting them equal, the distance is zero by construction. That does not prove the backward trajectory carries no residual error from earlier steps; it just renames that error out of existence. Eq. (2)'s recursion is imported from Li and van der Schaar, and the paper never shows it holds for ResShift's residual-shift schedule. If mu_t is actually Jacobian-dependent or schedule-dependent, Eq. (7) doesn't follow. The useful property here is probably training/test input distribution matching, not zero accumulated error.\n\nSmaller issues: no error bars or significance tests; the denoising protocol trains on SIDD and tests on Urban100 with synthetic Gaussian noise, which is an odd mismatch that may disadvantage the baselines; no code release. All fixable.\n\nBottom line: this is worth a serious referee. The idea is simple, general, and supported by enough experiments to be taken seriously, but the authors should be pressed to either prove the recursion for their backbone or reframe the contribution as empirical training-test alignment. I'd bring it to a reading group, and I'd cite it if I were working on diffusion restoration.\n\nRecommendation: engage—send it to peer review, but flag the theoretical claim as the main revision point.","headline":"A simple, empirically effective training trick that is over-theorized: the proof of zero cumulative error is a definitional tautology, but the method itself deserves attention.","tokens_in":18057,"tokens_out":2711,"would_cite":true,"duration_ms":24723,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims that training a diffusion restoration model on its own backward-process inputs, instead of one-step forward-corrupted inputs, makes the cumulative test-time error the direct training objective.","keywords":["image restoration","denoising diffusion models","data-consistent training","cumulative error","modular error","error propagation","super-resolution","image denoising"],"falsifier":"At test time on a data-consistently trained model, record $D(x_t^{\\mathrm{back}}, x_t^{\\mathrm{train}})$ and the per-step modular error simultaneously; if total output error keeps growing even where the recorded distance is zero and modular errors are small, the recursion in Eq. (7) has left out a real source of error accumulation.","tokens_in":16814,"feed_emoji":"🖼️","tokens_out":13376,"duration_ms":102263,"temperature":0.7,"pith_summary":"Denoising diffusion restoration models are trained on inputs produced by a one-step forward corruption, but at test time they are fed inputs produced by their own multi-step backward process. The paper argues that this mismatch lets small per-step errors accumulate, producing the shape and color distortions seen in diffusion-based restoration. It proposes data-consistent training: generate the training input for step $t$ by running the backward process, exactly as testing does, so the input cumulative error term becomes zero by construction and the standard per-step loss directly optimizes the error that accumulates at test time. Using ResShift as the backbone, the paper reports higher accuracy than previous restoration models on five tasks: super-resolution, denoising, deraining, dehazing, and dual-camera super-resolution. It also provides an efficient two-step approximation that trades a small accuracy loss for training cost close to the baseline.","feed_headline":"Train diffusion models on their own outputs to cut image errors","feed_subtitle":"Using backward-process inputs as training data turns test-time error into the loss; the gain appears on five tasks.","key_machinery":"The load-bearing object is the backward-process sample $x^{\\mathrm{back}}_t$, the input that the model's own sequential inverse process would produce at step $t$ during testing. Training on this sample instead of the one-step forward-corrupted sample $x^{\\mathrm{forw}}_t$ makes the input cumulative error in Eq. (6) identically zero, so the recursion in Eq. (2) collapses to $\\hat{\\xi}^{\\mathrm{cumu}}_t = \\hat{\\xi}^{\\mathrm{mod}}_t$ (Eq. (7)); the ordinary per-step loss then minimizes the cumulative error. The efficient version keeps the same idea but replaces the full backward rollout with one forward step plus one backward step and amplifies the resulting one-step error by $\\lambda = T - t$ to approximate the accumulated error.","core_discovery":"The paper argues that the train/test data inconsistency, not the network architecture, is the main source of the accumulated error that distorts diffusion-based restoration. In traditional training the loss minimizes the modular error $\\xi^{\\mathrm{mod}}_t = D(f_\\theta(x^{\\mathrm{forw}}_t),\\mathrm{GT}_t)$, but testing runs the whole chain so the real error is the cumulative error $\\xi^{\\mathrm{cumu}}_t = \\xi^{\\mathrm{mod}}_t + \\mu_t \\xi^{\\mathrm{cumu}}_{t+1}$, with $\\mu_t \\ge 1$. By using the backward-process input $x^{\\mathrm{back}}_t$ as $x^{\\mathrm{train}}_t$, the paper obtains $D(x^{\\mathrm{back}}_t,x^{\\mathrm{train}}_t)=0$, making the cumulative error equal to the modular error, so the ordinary loss becomes a direct optimizer of the test-time error. The paper reports experiments on five restoration tasks with ResShift, and an ablation with latent diffusion, to demonstrate that claim.","pith_inferences":["A testable extension the authors leave implicit: the zeroing argument is tied to their chosen definition of cumulative error, so applying data-consistent training under an error recursion that includes a separate stochastic-transition term would show whether the gains come specifically from closing the input gap.","Because the backward inputs used for training are produced by the current weights, early training epochs see low-quality trajectories; a warm-start schedule that begins with forward-process training before switching to data-consistent training is a natural refinement the paper does not test.","The efficient version's choice $\\lambda = T - t$ assumes the one-step error grows linearly with the remaining number of steps; a per-task adaptive schedule for $\\lambda$ could recover part of the gap between the efficient and full versions."],"forward_implications":["If the central claim is correct, data-consistent training should transfer to any diffusion restoration backbone, and the paper's ablation shows it helps both ResShift and LDM.","Because the change is to training inputs rather than to network architecture or sampling, test-time inference cost stays exactly the same as the original DDM.","The efficient variant gives most of the benefit at near-baseline training cost: 19.8 GB and 28 h versus 16.5 GB and 27 h for ResShift, compared with 41.7 GB and 90 h for the full version.","The method is positioned as a general fix for the shape and color distortions that come from error accumulation, so it should apply to other iterative restoration pipelines without per-task architectural changes."],"supporting_citations":[{"why":"Supplies the modular/cumulative error decomposition and the proof that the input cumulative error is propagated with a factor of at least one, the foundation for the paper's zeroing argument.","marker":"[27]"},{"why":"Provides the ResShift backbone used in the main experiments and in the comparison tables; it also defines the shifting schedule used to generate backward-process samples.","marker":"[73]"},{"why":"Establishes the standard DDPM training and sampling formulation whose train/test input mismatch the paper targets.","marker":"[16]"},{"why":"Defines the deterministic iterative sampling process that produces the test-time backward inputs the paper matches during training.","marker":"[45]"},{"why":"Serves as the second backbone in the ablation study, where adding data-consistent training improves latent diffusion results as well.","marker":"[42]"},{"why":"Represents a diffusion-based restoration baseline in the comparisons that the paper must beat to support its accuracy claims.","marker":"[82]"},{"why":"Another diffusion-based restoration baseline; its traditional forward-process training is the paradigm the paper argues causes cumulative error.","marker":"[34]"}],"fun_headline_variants":["Cut diffusion restoration errors by training on backward-process inputs","Make diffusion models learn from their own accumulated errors","Align diffusion training and testing to fix shape and color distortions","Data-consistent training: use backward inputs to reduce restoration drift","Train diffusion on its own outputs to eliminate restoration distortions"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole scheme rests on the definition that the only way past errors enter a step is the distance between the backward-generated input and the training input; if mistakes can carry over through any other channel, setting that distance to zero does not make cumulative error equal modular error.","fun_headline_variants_meta":{"raw":{"variants":["Cut diffusion restoration errors by training on backward-process inputs","Make diffusion models learn from their own accumulated errors","Align diffusion training and testing to fix shape and color distortions","Data-consistent training: use backward inputs to reduce restoration drift","Train diffusion on its own outputs to eliminate restoration distortions"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000256,"raw_usage":{"total_tokens":1554,"prompt_tokens":900,"completion_tokens":654,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":516,"completion_tokens_details":{"reasoning_tokens":576}},"tokens_in":516,"tokens_out":654,"duration_ms":6395,"temperature":1.0,"reasoning_tokens":576,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T13:58:05.081848+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"At test time on a data-consistently trained model, record $D(x_t^{\\mathrm{back}}, x_t^{\\mathrm{train}})$ and the per-step modular error simultaneously; if total output error keeps growing even where the recorded distance is zero and modular errors are small, the recursion in Eq. (7) has left out a real source of error accumulation.","supporting_citations":[{"cited_title":"On error prop- agation of diffusion models","cited_arxiv_id":null,"evidence_quote":"Supplies the modular/cumulative error decomposition and the proof that the input cumulative error is propagated with a factor of at least one, the foundation for the paper's zeroing argument."},{"cited_title":"Resshift: Efficient diffusion model for image super- resolution by residual shifting","cited_arxiv_id":null,"evidence_quote":"Provides the ResShift backbone used in the main experiments and in the comparison tables; it also defines the shifting schedule used to generate backward-process samples."},{"cited_title":"High-resolution image synthesis with latent diffusion models","cited_arxiv_id":null,"evidence_quote":"Serves as the second backbone in the ablation study, where adding data-consistent training improves latent diffusion results as well."},{"cited_title":"Selective hourglass mapping for universal image restoration based on diffusion model","cited_arxiv_id":null,"evidence_quote":"Represents a diffusion-based restoration baseline in the comparisons that the paper must beat to support its accuracy claims."},{"cited_title":"Diff-plugin: Revitalizing details for diffusion-based low-level tasks","cited_arxiv_id":null,"evidence_quote":"Another diffusion-based restoration baseline; its traditional forward-process training is the paradigm the paper argues causes cumulative error."}],"review_version":1}