{"id":"bf6ec6d9-a80b-413b-a432-b4bd46c2509b","arxiv_id":"2412.08149","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"AsyncDSB replaces the single shared noise schedule in diffusion Schrödinger bridge inpainting with a per-pixel schedule steered by predicted image gradients, improving FID by about 3% to 14% over the I2SB baseline.","lead":"AsyncDSB proposes an image inpainting method that gives each pixel its own noise schedule in a diffusion Schrödinger bridge, restoring high-detail pixels before low-detail ones. A smart generalist might read it because it challenges the widespread assumption that all pixels should denoise in sync, and it reports modest FID gains over prior inpainting models.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Per-pixel schedule breaks score calibration: Eqs. (12)-(13) feed a score sθ trained on global β_t with per-pixel variances, so the update is not a valid Schrödinger bridge posterior unless Σ_{t,i,j}=Σ_t is shown; without that, the FID gains may be an artifact of miscalibrated denoising.","rationale":"The reader's weakest_assumption isolates exactly the gap I consider load-bearing: per-pixel time changes are used without proving that the trained score remains calibrated or that the bridge coupling is preserved. My concern is therefore in agreement. I did not identify a fatal flaw: the idea is plausible, the ablation in Table 3 attempts to control for gradient conditioning, and the qualitative figures are suggestive. However, the theoretical gap is not merely cosmetic. If the per-pixel schedule changes the noise level that sθ sees, the method is not sampling from the learned Schr\\\"odinger bridge; it is an ad hoc denoiser with extra gradient guidance. That would weaken the paper's central claim that asynchronous schedules are the cause of the improvement, even if the FID numbers reproduce. The concrete test I propose directly measures whether the score is miscalibrated under the per-pixel schedule, which would settle whether the concern lands. Because the issue is addressable with a validation-pair experiment and, if confirmed, with retraining or score conditioning, the appropriate verdict remains CONDITIONAL rather than a rejection.","tokens_in":15654,"tokens_out":9641,"duration_ms":107618,"concrete_test":"On held-out (x0,x1) validation pairs with the center mask, compute the per-pixel variances σ²_{t,i,j} and σ̄²_{t,i,j} induced by the proposed β_{t,i,j} and check whether Σ_{t,i,j}=σ²_{t,i,j}σ̄²_{t,i,j}/(σ²_{t,i,j}+σ̄²_{t,i,j}) equals the global Σ_t at the same t. If it does not, feed the per-pixel-scheduled x_t to sθ and compare its output against the ground-truth score (x_t-x0)/σ̄²_{t,i,j}; a large error indicates sθ is evaluated out-of-distribution and Eq. (13) is not the trained bridge posterior.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central mechanism rests on Eqs. (11)-(13): each pixel receives its own β_{t,i,j}, and posterior sampling uses sθ(x_{t,i,j}, t) trained on the global schedule. For this to sample the learned Schrödinger bridge, the per-pixel marginal q(x_{t,i,j}|x0,x1) must have the same noise level Σ_{t,i,j}=σ²_{t,i,j}σ̄²_{t,i,j}/(σ²_{t,i,j}+σ̄²_{t,i,j}) as the global Σ_t used in training at every global step t. The paper never defines β_{t,i,j} explicitly as a function of t and τ_{i,j}, and never proves this equality or adjusts the score network input for per-pixel noise level. Consequently, at a fixed global t, different pixels carry different effective noise levels, so the UNet receives a mixture it never saw during training. The update in Eq. (13) is then not the posterior of any single diffusion bridge, and the claimed 3%-14% FID improvements could come from uncalibrated denoising or the extra gradient conditioning rather than from a valid asynchronous schedule. The empirical support is also thin: no error bars, seeds, or statistical tests, and τ_min,τ_max are tuned per mask on the evaluated benchmark.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper analyzes the noise schedule of I2SB and argues that a single global schedule beta_t is mismatched to the asynchronous way image pixels are restored. It proposes AsyncDSB, which (i) completes the gradient map of the corrupted region with an adversarially trained module, (ii) converts gradient magnitudes into per-pixel shift parameters tau_{i,j} via Eq. (11), and (iii) applies per-pixel posterior sampling with pixel-dependent variances sigma_{t,i,j}. Experiments on CelebA-HQ and Places2 report FID improvements over I2SB, with relative gains of roughly 3% to 14% depending on mask type.","tokens_in":15961,"tokens_out":6717,"duration_ms":69392,"significance":"If correct, the idea that per-pixel asynchronous schedules can better align the theoretical and practical restoration processes in diffusion Schrodinger bridges is a useful and non-obvious contribution to image inpainting. The paper identifies a real phenomenon: the empirical restoration speed of I2SB lags the theoretical schedule, and this lag differs across frequency bands. The proposed fix is simple, practical, and testable, and the authors include a useful control (I2SB plus gradient condition) to separate the effect of the asynchronous schedule from the effect of the extra gradient input. The main strength is the clean diagnosis of restoration asynchrony and the straightforward gradient-prior schedule. However, the paper does not yet establish that the per-pixel shifted schedule is a valid Schrodinger bridge, and the reported gains are not accompanied by measures of statistical uncertainty.","major_comments":[{"comment":"The per-pixel noise schedule beta_{t,i,j} is never explicitly defined as a function of t and tau_{i,j}. The text says that tau_{i,j} transforms the traditional noise schedule into a pixel-wise form, but no equation gives beta_{t,i,j} in terms of beta_t and tau_{i,j}. This is needed both for reproducibility and for checking the validity of the bridge. More importantly, Eq. (13) evaluates the score as s_theta(x_{t,i,j}, t)/sigma_t using the global sigma_t, while the variances in Eq. (12) are per-pixel sigma_{t,i,j}. For the posterior update to be a valid discretization of a Schrodinger bridge, the score must correspond to the per-pixel marginal q(x_{t,i,j} | x0, x1) at the per-pixel noise level Sigma_{t,i,j}; using a score trained on the global schedule on a mixture of pixels at different effective noise levels is not justified. The authors need to either prove that the per-pixel time shifts preserve the score calibration, retrain or condition the score on per-pixel noise levels, or restrict the theoretical claims to the empirical heuristic they actually implement.","section":"Step 2, Eqs. (11)-(13)"},{"comment":"The derivation of restoration speed drops the stochastic Term3 with the statement that its expectation is zero. For individual trajectories, the noise term contributes to the per-pixel restoration order, and Eq. (11) assigns schedules to individual pixels based on gradient magnitude alone. An expectation-based argument therefore does not establish that high-gradient pixels restore earlier along a typical sample path. If the intended claim is about the expected restoration speed, the paper should say so explicitly and relate the measured SSIM derivative in Figure 2 to the expectation rather than to individual pixel behavior.","section":"Theoretical Analysis, Eq. (7)"},{"comment":"The headline FID improvements are reported without error bars, seeds, or statistical tests; the reproducibility checklist explicitly answers 'no' to the use of statistical tests. Many absolute FID differences are small (e.g., 2.2 to 1.9 on CelebA-HQ center, 8.6 to 8.3 on Places2 center) and could lie within run-to-run variation. Additionally, tau_min and tau_max are tuned per mask type and dataset (Figure 6, Table 2), with no validation split described. The authors should provide multiple seeds with means and variances, and a significance test for the I2SB comparison, especially because the 3%-14% relative improvement is the central empirical claim.","section":"Experiments, Table 1 and Reproducibility Checklist"},{"comment":"The I2SB+gradient baseline in Table 3 is not described in sufficient detail: it is unclear how the gradient map is injected into the UNet, whether the same gradient completion network is used, and whether the baseline is trained with the same objective as AsyncDSB. Without these details, Table 3 cannot cleanly attribute the improvement to the asynchronous schedule rather than to the way the gradient conditioning is implemented.","section":"Ablation Study, Table 3"},{"comment":"The claim that AsyncDSB alleviates the schedule-restoration mismatch is, in part, enforced by design: Eq. (11) assigns earlier schedules to pixels with larger gradients, so observing in Figure 7 that high-frequency pixels restore earlier than low-frequency ones is not an independent confirmation of the mechanism. The FID improvements in Table 1 are the relevant empirical evidence; the text should not present Figure 7 as a verification of the scheduling hypothesis without acknowledging that the schedule directly encodes that hypothesis.","section":"Figure 7 and Section 'Can AsyncDSB solve the schedule-restoration mismatch issue?'"}],"minor_comments":[{"comment":"The notation in Eq. (2) is ambiguous because both variances are rendered as sigma^2_t; based on Eq. (6), the first should be sigma-bar^2_t and the second sigma^2_t. Please use distinct symbols consistently.","section":"Preliminaries, Eq. (2)"},{"comment":"Several cross-references are missing or blank, such as 'Section' and 'our insight of Section' in the Introduction and Methodology sections. Please fill in the appropriate section numbers.","section":"Theoretical Analysis and Abstract"},{"comment":"The abstract and text state a '3% - 14% improvement' over I2SB, but some entries in Table 1 report gains in the ‰ range (e.g., IS 6.71 to 6.74) while others are in the % range. The reporting should use a consistent relative measure and avoid mixing percentages with permille.","section":"Table 1"},{"comment":"The figure would be easier to interpret with explicit axis labels for tau_min and tau_max, a colorbar with units, and a legend or note explaining that blank cells are invalid because tau_min must not exceed tau_max.","section":"Figure 6"},{"comment":"The checklist states that all source code is included in an appendix and will be made public, but no code appendix appears in the manuscript. Either include the code or remove these claims.","section":"Reproducibility Checklist"},{"comment":"The integral in Eq. (13) is typeset incorrectly; it should be written as an integral from t - 1/T to t, not as 'Z t/(t-1/T)'. Please fix the notation.","section":"Eq. (13)"}],"recommendation":"major_revision","confidential_remarks":"The paper has an interesting core idea and a plausible mechanism, but the per-pixel bridge validity is not established and the empirical evidence lacks statistical rigor. If the authors can define beta_{t,i,j} explicitly, address the score-calibration issue, and provide variance-aware comparisons, the contribution would be much stronger. As it stands, I would not recommend acceptance without these revisions."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Let me give you my read of AsyncDSB. The core idea is new and worth knowing: instead of one global noise schedule for all pixels, each pixel gets its own schedule shifted by a tau_{i,j} derived from predicted gradient magnitude. The motivation is grounded in an observation that high-frequency pixels restore earlier than low-frequency ones during I2SB sampling (their Figure 3). And they do run a controlled ablation: adding the gradient map as a UNet condition to I2SB (I2SB+gradient) does not reproduce the gains, so the improvement is specifically from the asynchronous schedule, not just the extra conditioning.\n\nWhere it gets soft is the theory. The stress-test note is right: they train the score network on a global beta_t schedule, then at inference feed the network xt at global time t while using per-pixel variances from a different local schedule. The posterior sampling in Eq. (13) is only the I2SB posterior if the pixel's noise level equals the training-time noise level at that t. They never prove that, and nothing in the paper corrects for the mismatch. That means the per-pixel update is a heuristic, not a tested bridge. It may work anyway—empirically it does—but the paper frames it more formally than it is.\n\nThe empirical support is also thinner than the headline suggests. FID gains over I2SB are 3-14% depending on mask and dataset, several around 3-4%, with no error bars or significance tests (the checklist admits that). And tau_min/tau_max are tuned per mask class on the evaluation set; Figure 6 shows sensitivity. Code is not released.\n\nStill, I don't see a load-bearing flaw. The mechanism is transparent, the ablation is fair, and the gains are consistent. The paper overclaims theoretical backing, but the empirical finding is plausible and the asynchronous-schedule idea could transfer to other image-to-image tasks.\n\nWho it's for: anyone working on diffusion bridges or schedule design for restoration. I'd give it a serious referee—the idea is novel and the experiments are enough to warrant scrutiny. But it needs a revision that either proves the per-pixel posterior validity or reframes the method as a heuristic and adds error bars.","headline":"Novel per-pixel asynchronous schedule idea that works empirically, but the bridge validity is not proven and the evaluation needs error bars.","tokens_in":16509,"tokens_out":2849,"would_cite":true,"duration_ms":29810,"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":"Existing diffusion Schrödinger bridge inpainting applies one noise schedule to all pixels, but AsyncDSB assigns each pixel its own shifted schedule based on gradient magnitude, improving FID by about 3% to 14% over I2SB.","keywords":["image inpainting","diffusion Schrödinger bridge","noise schedule","pixel-asynchronous schedule","gradient prior","CelebA-HQ","Places2"],"falsifier":"Train the same AsyncDSB pipeline but replace the gradient-derived $\\tau$ map with a random or shuffled per-pixel shift that has the same distribution and no correlation with image structure. If FID remains at the same level, the improvement comes from adding per-pixel variation rather than from restoring high-frequency pixels first. Alternatively, directly measure per-pixel restoration times in AsyncDSB trajectories; the mechanism is refuted if high-frequency pixels are not restored earlier than low-frequency pixels on average.","tokens_in":15445,"feed_emoji":"🖼️","tokens_out":7692,"duration_ms":61831,"temperature":0.7,"pith_summary":"The paper argues that diffusion Schrödinger bridge inpainting, as instantiated by I2SB, is suboptimal because it applies the same noise schedule $\\beta_t$ to every pixel while real restoration is asynchronous: high-frequency pixels (edges, contours) come together before low-frequency pixels (flat color, texture fill). It proposes AsyncDSB, which first predicts the gradient of the masked region, converts gradient magnitude into a per-pixel time shift $\\tau_{i,j}$, and then runs the bridge with a pixel-specific schedule $\\beta_{t,i,j}$. On CelebA-HQ and Places2, this raises FID by roughly 3% to 14% over I2SB, with the biggest gains on masks that destroy structure. If correct, the result shows that aligning the noise schedule to the per-pixel order of restoration is a direct and cheap lever for inpainting quality.","feed_headline":"Per-pixel schedules improve diffusion inpainting FID by 3-14%","feed_subtitle":"High-frequency pixels restore first, closing the gap between theory and practice in Schrödinger bridge inpainting.","key_machinery":"The central object is the pixel-asynchronous noise schedule $\\beta_{t,i,j}$, built by shifting a global schedule per pixel through a predicted-gradient time shift $\\tau_{i,j}$ (Eq. 11). The shift map comes from a gradient completion network $f_{\\theta_g}$ trained adversarially with feature matching, following the edge-prediction approach of EdgeConnect. Each pixel then has its own accumulated variances $\\bar{\\sigma}_{t,i,j}^2$ and $\\sigma_{t,i,j}^2$ (Eq. 12), and inference uses the pixel-wise posterior sampling recursion of Eq. (13). The mechanism works by allocating early steps to high-gradient pixels and later steps to low-gradient pixels, explicitly modeling the temporal ordering observed in real restoration.","core_discovery":"The paper's central discovery is a schedule-restoration mismatch in I2SB. Using the analytic Gaussian posterior of the Schrödinger bridge, it derives that the restoration speed of a pixel is $\\partial x_t/\\partial t = -\\beta_t (x_0 \\odot x_m)/(\\bar{\\sigma}_t^2 + \\sigma_t^2)$, so $\\beta_t$ sets the global pace while the pixel value sets the amplitude. Measuring the derivative of SSIM across denoising time on CelebA-HQ shows the practical restoration curve lags the theoretical one, and the lag is concentrated in mid- and low-frequency pixels, meaning pixels are not restored synchronously. The paper's fix is to give each pixel its own time shift $\\tau_{i,j}$, computed by normalizing a Gaussian-filtered predicted gradient map to $[\\tau_{\\min}, \\tau_{\\max}]$, and to run posterior sampling with a per-pixel schedule. This re-aligns practical and theoretical restoration speeds and improves FID.","pith_inferences":["The same per-pixel schedule idea could transfer to other bridge-based image translation tasks, such as super-resolution or deblurring, where restoration order is also spatially non-uniform.","A cheaper proxy for the gradient map — e.g., distance to the mask boundary or local variance — might capture most of the gain, which would separate the benefit of asynchrony from the specific gradient predictor.","The paper does not prove that per-pixel shifted schedules preserve the Schrödinger bridge coupling or that the score network stays calibrated for pixels at different effective times; if that proof fails, the empirical gains would still stand but the theoretical explanation would need revision."],"forward_implications":["AsyncDSB improves FID by roughly 3% to 14% over I2SB on CelebA-HQ and Places2 across center, half, wide, and narrow masks, with larger gains on masks that remove structure.","The asynchronous schedule visibly narrows the gap between theoretical and practical restoration-speed curves, particularly for mid- and low-frequency pixels.","The framework is not limited to gradient guidance; the authors state that any $\\beta_{t,i,j}$ strategy matching the pixel-asynchronous generation process can be plugged in.","For higher mask ratios, the optimal schedule shifts earlier (larger $\\tau_{\\min}$ and $\\tau_{\\max}$), indicating that mask-aware schedule selection is a promising extension the paper leaves for future work."],"supporting_citations":[{"why":"This citation supplies the I2SB diffusion Schrödinger bridge formulation and the baseline that the paper extends and compares against.","marker":"(Liu et al. 2023)"},{"why":"This citation provides the adversarial edge and gradient completion training with feature matching loss used for the gradient prior network $f_{\\theta_g}$.","marker":"(Nazeri et al. 2019)"},{"why":"This citation supplies the DDPM posterior sampling recursion that I2SB and AsyncDSB reuse for inference (Eq. 5 and Eq. 13).","marker":"(Ho, Jain, and Abbeel 2020)"},{"why":"This citation provides the RePaint baseline and the mask and evaluation protocol (center, half, wide, and narrow masks at 256x256) used in the experiments.","marker":"(Lugmayr et al. 2022)"}],"fun_headline_variants":["AsyncDSB: per-pixel schedules cut inpainting FID by 3-14%","Fix schedule mismatch: async pixel noise boosts inpainting FID","High-frequency pixels first: new inpainting schedule improves FID","Pixel-asynchronous Schrödinger bridge reduces inpainting error"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"All of the reported gains rest on the assumption that a score network trained with the global, synchronous schedule still gives correct per-pixel denoising predictions when each pixel is advanced at its own shifted time; the paper does not prove this, it only demonstrates the empirical improvement.","fun_headline_variants_meta":{"raw":{"variants":["AsyncDSB: per-pixel schedules cut inpainting FID by 3-14%","Fix schedule mismatch: async pixel noise boosts inpainting FID","High-frequency pixels first: new inpainting schedule improves FID","Pixel-asynchronous Schrödinger bridge reduces inpainting error"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000232,"raw_usage":{"total_tokens":1551,"prompt_tokens":1069,"completion_tokens":482,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":685,"completion_tokens_details":{"reasoning_tokens":402}},"tokens_in":685,"tokens_out":482,"duration_ms":4773,"temperature":1.0,"reasoning_tokens":402,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T18:08:50.241804+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train the same AsyncDSB pipeline but replace the gradient-derived $\\tau$ map with a random or shuffled per-pixel shift that has the same distribution and no correlation with image structure. If FID remains at the same level, the improvement comes from adding per-pixel variation rather than from restoring high-frequency pixels first. Alternatively, directly measure per-pixel restoration times in AsyncDSB trajectories; the mechanism is refuted if high-frequency pixels are not restored earlier than low-frequency pixels on average.","supporting_citations":[],"review_version":1}