{"id":"ce9b91d5-c71b-4bc0-bc18-95f14fb587b1","arxiv_id":"2506.15346","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A conditional image-to-image Schrödinger bridge reconstructs subsurface velocity models from seismic signals and a smoothed starting model, needing only a few network evaluations.","lead":"The paper trains a generative model that turns a blurry starting map of underground sound speed, together with recorded seismic waves, into a sharp velocity image. It matters because it gives seismic imaging a fast, mathematically grounded way to use the starting model that geophysicists already build, something AI-based inversion methods previously handled by guesswork.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The central claim is conditional on an unrealistic distortion prior: Section 5.1.3 admits Eq. 25 has no real-world basis, and Section 6.4 shows sensitivity to it, so the theoretical endpoint guarantee need not transfer to real smooth models.","rationale":"The reader's weakest_assumption pinpoints the distortion operator as the load-bearing element, and I agree. The entire cI2SB construction is a conditional Schrödinger bridge whose prior is defined by Eq. 25; the inference guarantee applies only when the bridge starts from that prior. Real-world smooth models are not draws from this family, as the authors concede, and Section 6.4 provides direct evidence of sensitivity to the operator. This is more fundamental than the numerical typos in the tables (which, while concerning, mainly affect the precision of the comparative claims) and more central than the lack of a seismic-conditioning ablation (which would clarify the role of dobs but does not invalidate the bridge construction). Because the paper is honest about the limitation and the method is clearly demonstrated within its synthetic setup, a CONDITIONAL verdict remains appropriate: the work is acceptable as a proof-of-concept, but its central claim of providing an FWI scheme with a guaranteed posterior endpoint must be understood as conditional on the chosen distortion prior. No change from the reader's verdict is warranted.","tokens_in":18060,"tokens_out":7211,"duration_ms":74814,"concrete_test":"Train or use the released cI2SB checkpoints on OpenFWI test data, but generate csmooth with a physically realistic pipeline instead of Eq. 25: run 5–10 iterations of standard full waveform inversion (or reverse-time migration) from a smooth starting model on the same synthetic seismic data, take the resulting low-wavenumber velocity estimate as csmooth, and evaluate MAE/SSIM against the ground-truth velocity models. Compare these numbers to the matched-distortion results in Tables 2–3 and to the out-of-distribution results in Table 6. If the physically realistic csmooth yields degradation comparable to or larger than the Table 6 shifts (e.g., SSIM decreases by more than 0.02–0.05 on several datasets), the concern is confirmed and the central claim must be scoped explicitly to the synthetic distortion family; if performance holds, the limitation is less severe.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The most load-bearing assumption is that the test-time starting model csmooth is drawn from the parametric distortion family pdist(c1|c0)=S_k(γz+(1−γ)c0) of Eq. 25. Algorithm 3 initializes c_N=csmooth and runs the reverse bridge; the Schrödinger bridge guarantee (Sections 2.4–3) only holds when the initial point is sampled from the prior pprior induced by that same distortion operator. Real initial velocity models from migration or tomography have structured, spatially correlated errors (low-wavenumber bias, fault smearing) that are not i.i.d. Gaussian noise followed by a fixed Gaussian blur. The authors explicitly state the operator 'does not have any basis in real-world data' (Section 5.1.3), and their own Section 6.4 shows that shifting the kernel size range to 16–24 or 0–8 degrades metrics (e.g., many SSIM values drop by 0.02–0.05 relative to the matched-distortion results in Tables 2–3). Consequently, the headline claim that the framework 'provides an acoustic FWI scheme' with a guaranteed posterior endpoint is not supported for realistic inversion inputs; it is a demonstration on a synthetic, self-defined prior. This is a correctness risk for the real-world applicability of the method, not merely a question of empirical generalization, because the theoretical endpoint guarantee itself depends on the distortion operator. The absence of any physics-based data-consistency term in the loss makes the method unable to correct for an out-of-distribution prior.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes cI2SB, a conditional Image-to-Image Schrödinger Bridge for acoustic full waveform inversion. It adapts the I2SB training objective to take seismic data and a smoothed velocity model as conditional inputs, and uses a classifier-free guidance style weighting to trade off conditional and unconditional denoising. The inference starts from a smoothed velocity model and terminates at a sample from the learned posterior over velocity models conditioned on the seismic data. Experiments on OpenFWI subsets compare the method against InversionNet and a reimplemented conditional SGM, reporting improved SSIM and sample fidelity with a few NFEs. The paper also studies the effect of NFE count and guidance scale.","tokens_in":18224,"tokens_out":7953,"duration_ms":64477,"significance":"If the empirical claims hold, the paper provides a more principled way to incorporate initial velocity models into diffusion-based FWI than current heuristics, and the few-NFE property is practically valuable. The derivation is a straightforward application of the cDSB framework to the I2SB setting, and the authors are transparent about the synthetic nature of the distortion prior. Strengths include a reproducible code repository, a clean statement of the theoretical backbone, and a systematic NFE study. However, the central guarantee is conditional on the training distortion operator, which the authors admit has no real-world basis; this limits the significance unless addressed.","major_comments":[{"comment":"The distortion operator D^γ_k of Eq. (25) is explicitly described as 'not hav[ing] any basis in real-world data', and Table 6 shows that using different kernel sizes degrades performance. The theoretical guarantee that the inference process ends at a posterior sample (stated in the Abstract and Section 3) holds only when the initial smoothed model is drawn from the same parametric distortion family used in training. For real FWI initial models from migration or tomography, whose errors are structured and spatially correlated, there is no reason the bridge endpoint will be the posterior. The paper should either incorporate a physics-based data-consistency term into the loss or sampling, or explicitly reframe the main claim as conditional on the distortion model.","section":"Section 5.1.3 and Section 6.4"},{"comment":"The claim that cI2SB 'outperforms' the conditional SGM 'by statistically significant margin' is not supported by the reported results. No error bars or repeated runs are provided for the deterministic baselines (InversionNet and cSGM); only the stochastic cI2SB results are averaged over 50 seeds, but their variance is not reported. Moreover, in Table 3, for FlatVelA the cSGM reports lower MAE and MSE than cI2SB OT-ODE sampling, so the statement 'in all studied cases' is factually inaccurate. Please provide confidence intervals or a proper statistical comparison, and soften the claim accordingly.","section":"Section 5.2 and Tables 2-3"},{"comment":"The NFE study contains numerical inconsistencies that undermine the quantitative claims. For example, in Table 4 the FlatVelB row at NFE=1 lists MAE=1.63e-2 and MSE=1.71e-2, which is impossible for normalized data; in Table 5 the FlatFaultB row at NFE=50 lists MAE=3.46e-3, an order of magnitude lower than the neighboring rows, suggesting a missing digit. Please correct these values and re-verify all numbers.","section":"Tables 4 and 5"},{"comment":"The claim that cI2SB achieves superior perceptual quality with fewer NFEs is supported only by visual inspection of Fig. 12. To make this claim quantitative, please report a perceptual metric such as FID or LPIPS, or at least report the variance of the standard metrics over the validation set.","section":"Section 6.2"}],"minor_comments":[{"comment":"The variance terms σ^2_t and σ^2_t use identical notation, which is ambiguous. Please follow the I2SB convention of σ_t^2 and σ̄_t^2.","section":"Equation (17)"},{"comment":"The description of the InversionNet reimplementation as passing a constant timestep value to the network is unclear; InversionNet is not a diffusion model. Please clarify the exact architecture and training procedure used.","section":"Section 5.1.4"},{"comment":"The text reports wcond=100, while the caption of Figure 8 refers to wcond=10. Please reconcile this inconsistency.","section":"Section 5.1.1 vs. Supplementary Figure 8"},{"comment":"There are numerous typos, e.g., 'existense' in Section 1 and inconsistent rendering of 'Schrödinger' as 'Schr¨ odinger'.","section":"Throughout"},{"comment":"For the stochastic cI2SB results averaged over 50 seeds, please report standard deviations or interquartile ranges in addition to the means.","section":"Section 5.1.5"}],"recommendation":"major_revision","confidential_remarks":"The paper is a solid application of existing Schrödinger bridge machinery, but the statistically unsupported claims and the admitted mismatch between the synthetic distortion prior and real FWI inputs are serious. The numerical typos in Tables 4 and 5 need to be checked against the original results. I recommend major revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"What you should know: this is an application paper, not a theoretical breakthrough. The cI2SB construction is exactly what it says: the conditional Schrödinger bridge of Shi et al. plus the I2SB training shortcut of Liu et al., adapted to acoustic FWI with a smoothed velocity model as the bridge prior. The authors are upfront about that lineage, and the contribution is in the adaptation, the reweighted classifier-free guidance, and the seismic application.\n\nThe paper earns credit on the empirical side. Experiments cover ten OpenFWI variants, two model sizes, an NFE sweep, an out-of-distribution distortion study, and released code. The low-NFE results are the strongest part: cI2SB reaches competitive regression metrics at NFE=1–5, which is practically valuable. The perception-distortion discussion in Section 6.2 is also honest and useful; they show that more NFEs do not improve MAE/MSE/SSIM and explain why.\n\nThe soft spots are real but not fatal. The central 'theoretically guaranteed posterior' claim only holds if the test-time smooth model is actually sampled from Eq. (25): Gaussian blur with small i.i.d. noise. The authors explicitly say this construction has no real-world basis, and Section 6.4 shows that shifting the kernel range degrades results. That means the guarantee is a matched-distribution statement, not a statement about real initial velocity models from migration or tomography. They need to either test on realistic smooth models or reframe the contribution as bridge inversion under a synthetic prior.\n\nSecond, the 'statistically significant margin' claim appears without error bars or significance tests. Tables 2–5 also contain entries that look like typos, and those need correction before publication. Third, the 'sample fidelity superior to supervised learning' claim rests on visual inspection; there is no perceptual metric. Fourth, there is no clean ablation that isolates the marginal contribution of seismic data, since the smoothed model is itself highly informative. The guidance scale is acknowledged to be heuristic, and that is acceptable if framed as a practical control.\n\nWho is this for? Researchers working on data-driven FWI or on conditional image-to-image generation with Schrödinger bridges. It is a careful, reproducible application of known machinery, not a new theory.\n\nMy recommendation: send it to peer review. It deserves referee time. I would ask for error bars, corrected tables, a realistic smooth-model experiment, and an ablation of the conditioning signal before acceptance.","headline":"A credible application paper that borrows its theory from cDSB/I2SB and delivers a useful FWI demo, but the headline posterior guarantee only holds under a distortion prior the authors admit is unrealistic.","tokens_in":18930,"tokens_out":3984,"would_cite":true,"duration_ms":39429,"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":"A conditional Schrödinger bridge turns a smoothed velocity model into a posterior sample over ground truth using only a few network evaluations.","keywords":["full waveform inversion","Schrödinger bridge","conditional generative modeling","seismic velocity estimation","image-to-image translation","diffusion models","classifier-free guidance","OpenFWI"],"falsifier":"Run the released cI2SB model starting from a smooth velocity model produced by a different mechanism than pdist — for example, migration or traveltime tomography, or a Gaussian blur with kernel size 16–24 or 0–8 — and compare MAE/MSE/SSIM with the in-distribution results in Tables 2–3; a drop to the levels of Table 6's out-of-distribution rows would falsify the claim that the inference is a general posterior sampler rather than a memorization of the training distortion.","tokens_in":17651,"feed_emoji":"🌊","tokens_out":6534,"duration_ms":59416,"temperature":0.7,"pith_summary":"The paper tries to establish that the conditional Image-to-Image Schrödinger Bridge (cI2SB) gives acoustic full waveform inversion a principled way to use a smoothed starting velocity model: inference begins at that smoothed model and ends, in finite time, at a sample from the posterior over true velocity models given the observed seismograms. The authors derive cI2SB from the amortized conditional Schrödinger bridge problem, adapt the I2SB training pipeline to conditional simulation, and add a classifier-free-guidance-style weighting that trades sample diversity against fidelity. On the OpenFWI benchmark, the bridge outperforms a reimplemented conditional diffusion model and matches or beats a supervised baseline in perceptual quality while needing only a handful of neural function evaluations. The payoff, if the claim holds, is that diffusion-based inversion no longer requires a long noising chain to a Gaussian prior or a heuristic recipe for injecting the initial model.","feed_headline":"Schrödinger bridge inverts seismic data in a few network calls","feed_subtitle":"A conditional bridge from smoothed to true velocity models beats diffusion baselines on OpenFWI at 5–50 steps.","key_machinery":"The central object is the conditional Image-to-Image Schrödinger Bridge (cI2SB): a Schrödinger bridge interpolating between the distribution of ground-truth velocity models and the distribution of smoothed velocity models, with the observed seismic data serving as a conditioning variable that does not evolve along the bridge (dy = 0). The load-bearing identity is the analytic posterior q(xt|x0,x1) of the I2SB formulation, which lets the model sample intermediate velocity states directly during training and gives the DDPM-style recursion used at inference. The training procedure (Alg. 2) learns conditional and unconditional drifts in a single network, alternating between prediction with and without the data channel, and the sampling procedure (Alg. 3) blends the two predictions with a guidance scale η. This machinery converts the previously heuristic 'how to include the initial velocity model' into a boundary condition of a stochastic optimal transport problem.","core_discovery":"The central claim is that the cI2SB framework, derived from the amortized conditional Schrödinger bridge problem, provides an acoustic FWI scheme whose inference starts from a smoothed velocity model and terminates at a sample from the posterior over ground-truth velocity models conditioned on observed seismic data, and that on OpenFWI this scheme outperforms a reimplemented conditional diffusion model while reaching sample fidelity superior to a supervised approach with only a few neural function evaluations. The bridge is constructed by fixing the forward reference measure to the acoustic velocity field while the seismic data channel stays constant, and by taking the endpoint prior to be the parametric distortion pdist(c1|c0) applied to ground truth. Because the bridge posterior q(ct|c0,c1) is analytic, training reduces to denoising score matching with paired (c0, c1, dobs) triples, and inference runs the DDPM recursion backward from csmooth. The authors report that the guidance scale η controls the diversity of reconstructed velocity fields, and that objective reweighting (wcond) is necessary to keep the network attending to the seismic data.","pith_inferences":["We infer that the same conditional-bridge construction should transfer to other seismic inverse problems (elastic FWI, migration-image conditioning) wherever a paired smooth-to-true training set exists, since the method only needs pjoin and pdist.","The paper's Table 6 suggests a practical calibration recipe: before field deployment, measure the actual distortion between available smooth models and true velocity using wells, and retrain the bridge with that empirical pdist; otherwise the endpoint guarantee does not apply.","The authors' observation that the network ignores the data channel unless wcond is large hints at a general pitfall: with informative priors, joint conditional/unconditional training is biased toward the unconditional solution, and explicit loss reweighting may be needed in any bridge-based posterior sampler."],"forward_implications":["Diffusion-based acoustic inversion can start from an arbitrary smoothed velocity model and still be guaranteed, in the continuous-time limit, to produce a posterior sample over true velocity models given the observed data.","The cost of high perceptual fidelity drops from hundreds of network evaluations (typical for Gaussian-prior diffusion) to roughly 5–50, making the method practical for large 2D surveys.","The guidance scale η gives practitioners a single hyperparameter to trade sample diversity against fidelity, with η = 1 collapsing the output to a near-deterministic reconstruction.","Because the distortion operator is explicit, the same framework can be retargeted to other 'smooth start' priors — migration images, tomography models — by retraining with a matching pdist.","At matched NFEs, the bridge model consistently beats the conditional diffusion baseline on all tested OpenFWI dataset families by a statistically significant margin."],"supporting_citations":[{"why":"Supplies the amortized conditional Schrödinger bridge formulation (Eqs. 13–14) that cI2SB is built on.","marker":"[24]"},{"why":"Provides the I2SB framework and the analytic bridge posterior q(xt|x0,x1) used for training and inference.","marker":"[25]"},{"why":"OpenFWI supplies the velocity-model/seismogram dataset used in all experiments.","marker":"[26]"},{"why":"Classifier-free guidance is the inspiration for the conditional/unconditional training and weighted sampling in Algs. 2–3.","marker":"[19]"},{"why":"The conditional diffusion model (cSGM) that the paper reimplements and outperforms in Tables 2–3.","marker":"[22]"},{"why":"InversionNet, the supervised baseline whose reimplementation the bridge matches in regression metrics and beats in perceptual quality.","marker":"[10]"},{"why":"Diffusion Schrödinger bridge discretization and score-matching fitting that underlie the conditional half-bridge derivation.","marker":"[35]"}],"fun_headline_variants":["Schrödinger bridge cuts seismic inversion to few network calls","Conditional bridge outperforms diffusion in acoustic FWI","Few-step seismic imaging via Schrödinger bridge","Schrödinger bridge beats diffusion for seismic velocity inversion"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that a real-world smoothed velocity model is a sample from the Gaussian-blur-plus-noise distortion family used in training; the paper states this construction has no basis in real-world data, and its Table 6 shows performance degrades when the smoothing kernel leaves the training range.","fun_headline_variants_meta":{"raw":{"variants":["Schrödinger bridge cuts seismic inversion to few network calls","Conditional bridge outperforms diffusion in acoustic FWI","Few-step seismic imaging via Schrödinger bridge","Schrödinger bridge beats diffusion for seismic velocity inversion"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000801,"raw_usage":{"total_tokens":3591,"prompt_tokens":1082,"completion_tokens":2509,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":698,"completion_tokens_details":{"reasoning_tokens":2459}},"tokens_in":698,"tokens_out":2509,"duration_ms":20029,"temperature":1.0,"reasoning_tokens":2459,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T19:37:40.453761+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the released cI2SB model starting from a smooth velocity model produced by a different mechanism than pdist — for example, migration or traveltime tomography, or a Gaussian blur with kernel size 16–24 or 0–8 — and compare MAE/MSE/SSIM with the in-distribution results in Tables 2–3; a drop to the levels of Table 6's out-of-distribution rows would falsify the claim that the inference is a general posterior sampler rather than a memorization of the training distortion.","supporting_citations":[{"cited_title":"Conditional simulation using diffusion schr¨ odinger bridges,","cited_arxiv_id":null,"evidence_quote":"Supplies the amortized conditional Schrödinger bridge formulation (Eqs. 13–14) that cI2SB is built on."},{"cited_title":"Controllable seismic velocity synthesis using generative dif- fusion models,","cited_arxiv_id":null,"evidence_quote":"The conditional diffusion model (cSGM) that the paper reimplements and outperforms in Tables 2–3."},{"cited_title":"Inversionnet: An efficient and accurate data-driven full waveform inversion,","cited_arxiv_id":null,"evidence_quote":"InversionNet, the supervised baseline whose reimplementation the bridge matches in regression metrics and beats in perceptual quality."},{"cited_title":"Diffusion schr¨ odinger bridge with applications to score-based generative modeling,","cited_arxiv_id":null,"evidence_quote":"Diffusion Schrödinger bridge discretization and score-matching fitting that underlie the conditional half-bridge derivation."}],"review_version":2}