{"id":"40748f00-002d-47b2-8f9a-f653ecc569d7","arxiv_id":"2501.00941","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"UB-Diff generates paired velocity maps and seismic waveforms from unbalanced data using a shared latent space plus a two-step training scheme, and reports better FID and downstream inversion scores than prior methods.","lead":"This paper introduces UB-Diff, a diffusion model that generates paired seismic waveforms and velocity maps from training data where one modality is plentiful and the other is scarce. It matters because data scarcity and modality imbalance are common in scientific fields, and synthetic paired data could help train subsurface imaging models.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The pairwise reliability claim rests on an indirect InversionNet proxy and two visual examples; generated (velocity, seismic) pairs are never checked against wave physics in aggregate, so the co-latent alignment could be failing silently.","rationale":"The strongest claim is that UB-Diff produces reliable, useful paired data. The method's structural keystone is the co-latent z: after self-supervised training on abundant velocity maps (Eq. 2), the seismic decoder is expected to map z to the corresponding waveform after only 1k/5k fine-tuning steps. The reader identified this as the weakest assumption, and I agree it is load-bearing. My stress-test sharpens it: the paper's pairwise evidence is not actually about pairwise physical consistency. Table 3 measures whether a neural network trained on generated pairs can then invert original seismic data. That is a reasonable practical proxy, but it cannot certify that each generated waveform is the wave-physics response of the paired velocity. The only direct comparison against a physical forward model is Figure 6, which is two samples with no aggregate statistics. FID improvements (Table 2) are about marginal distributions and cannot detect per-pair mismatch. Therefore, the central claim is conditional on a direct consistency check. I considered other issues (post hoc freeze flag, FID on seismic waveforms, no error bars, no code); these weaken confidence but are secondary. The proposed forward-modeling check would settle whether the co-latent alignment actually holds for generated data. If it passes, the central claim is substantially supported; if it fails, the method is only generating individually plausible samples, not reliable pairs. Since this is exactly the kind of experiment the paper should have included, the verdict should remain conditional pending that check.","tokens_in":17077,"tokens_out":8737,"duration_ms":83938,"concrete_test":"Sample 1,000 unconditional pairs from UB-Diff and from MT-Diff on each OpenFWI subset. For each generated velocity map, run an acoustic finite-difference forward solver with the OpenFWI survey geometry to compute a reference seismogram. Compute the normalized L2 error and structural correlation between the generated seismic and the reference, reporting means and error bars. Compare UB-Diff and MT-Diff against original test pairs (upper bound) and shuffled pairs (lower bound). If UB-Diff's mismatch is not clearly below MT-Diff and near the original-pair level, the co-latent alignment fails for generated data and the pairwise reliability claim is unsupported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"UB-Diff's paired output is produced by decoding a single co-latent z=E(ma) (Eq. 1), where the encoder and velocity decoder are trained almost exclusively on velocity maps (Eq. 2) and the seismic decoder Ds is fine-tuned on only n=1,000 or 5,000 pairs (Eq. 3). There is no explicit alignment loss and no seismic encoder, so the claim that this velocity-only latent supports faithful seismic decoding rests entirely on the near-linear latent relationship from Feng et al. (2022b). The only quantitative pairwise metric is Table 3: train InversionNet on 10,000 generated pairs and test on the original dataset. That is an indirect proxy; a high-capacity network can partially absorb mismatched pairs or exploit dataset statistics, so good MAE/MSE/SSIM do not prove that each generated seismic is the physical response of its paired velocity. The only direct physical check is Figure 6, which shows two examples and no aggregate error bars. If Ds overfits the fine-tuning set, generated pairs can look plausible individually while mismatching each other for unseen structures, and the FID improvements in Table 2 would not reveal it. Thus the central claim of reliable and useful paired generation is not yet established.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes UB-Diff, a latent diffusion model for jointly generating paired velocity maps and seismic waveforms from unbalanced training data. The architecture is a one-in-two-out encoder-decoder that maps a velocity map to a co-latent code, which is decoded by a CNN to a velocity map and by a transformer to a seismic waveform. Training proceeds in two steps: a self-supervised reconstruction on all available velocity maps, followed by a fine-tuning step on a small number of paired seismic samples; the diffusion model is then trained in the co-latent space. On five OpenFWI datasets, the authors report substantially lower FID than DDPM, Stable Diffusion, and MT-Diffusion, and improved downstream FWI metrics when InversionNet is trained on the generated pairs.","tokens_in":17354,"tokens_out":8874,"duration_ms":80019,"significance":"If the reported results hold, UB-Diff is a meaningful step toward generative data augmentation for full-waveform inversion and for multi-modal scientific data in general, where paired data are scarce and one modality is abundant. The two-step training scheme is simple and the empirical FID gains are large and consistent across datasets, which is encouraging. The paper also provides a useful comparison of diffusion baselines in this setting. However, the current evaluation does not directly demonstrate that the generated pairs are physically consistent, and the statistical basis for the claims is limited to single runs; these gaps currently make the paper a promising candidate rather than a settled result.","major_comments":[{"comment":"The central claim of 'reliable' paired generation is not established. The seismic decoder Ds is fine-tuned on only 1,000 or 5,000 pairs (Eq. 3), while the encoder and velocity decoder are trained almost exclusively on velocity maps (Eq. 2). The paper justifies the shared latent via the near-linear relationship of Feng et al. (2022b), but it does not verify that this relationship holds for the learned E or after fine-tuning. Table 3 is an indirect proxy: a high-capacity InversionNet can partially absorb mismatched pairs, and the paper does not compare with InversionNet trained on real pairs. Figure 6 contains only two examples. Please add an aggregate physics-based check, e.g., compute the physical forward response of generated velocity maps and compare it with the generated seismic waveforms over many samples, reporting MAE/MSE and correlation for UB-Diff and baselines; also report the InversionNet performance upper bound using real paired data.","section":"UB-Diff Framework, Eq. (3)"},{"comment":"The FID scores for seismic waveforms are reported without specifying the feature extractor. Standard FID uses an ImageNet-trained Inception network, which is not meaningful for seismic waveforms, and the paper does not explain how the five-channel seismic data are fed into the FID computation. Please specify the feature extractor or use a modality-appropriate metric (e.g., a trained seismic autoencoder feature space, or Kernel Inception Distance). Until this is clarified, the FID improvements in Tables 2 and 5 are difficult to interpret.","section":"Experimental Setup, Metrics"},{"comment":"All quantitative results are single runs without error bars, seeds, or significance tests, so the wording 'significantly outperforms' in the abstract and conclusion is unsupported. Moreover, the freeze flag F is selected after seeing the test results ('We both freeze or not for the second step of training and report a better result', Training Setting), which is a form of test-set selection. Please report multiple seeds with means and standard deviations, and choose F (and other hyperparameters) on a validation split rather than on the test set.","section":"Experimental Results, Tables 2–5"}],"minor_comments":[{"comment":"There are several typos, including 'Optimizaiton' in the section heading and 'Experiential Setup' in the appendix; please proofread the text.","section":"Throughout"},{"comment":"Equation (4) is hard to parse because the loss terms are nested without clear parentheses and the symbol θ*_ma is undefined (it should presumably be θ*_self); please rewrite the equation and define all symbols.","section":"Section 3.2, Eq. (4)"},{"comment":"Figures 2 and 4 are low-resolution in the current version, making the arrows, labels, and the distinction between training and inference paths hard to read; please provide higher-quality figures.","section":"Figures 2 and 4"},{"comment":"The main text does not report which freeze flag F was chosen for each dataset and setting; please include the chosen F values in the tables or in a small appendix table for reproducibility.","section":"Appendix, Experimental Results Discussion"},{"comment":"Feng et al. 2022a and 2022b are cited with the same title 'An intriguing property of geophysics inversion'; please disambiguate the two references and ensure each citation points to the correct version.","section":"References"},{"comment":"The paper does not mention whether code or trained models will be released; please add a code/data availability statement so that the large claimed improvements can be independently reproduced.","section":"Reproducibility"}],"recommendation":"major_revision","confidential_remarks":"The main gap is evaluative rather than conceptual; I would encourage the authors to run the physics-based pairwise validation because it directly tests the advertised contribution. The novelty relative to MT-Diffusion is moderate, but the unbalanced-data setting gives the work a distinct practical niche. The paper fits the journal's scope, and a revision with the suggested experiments would substantially strengthen it."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The real novelty here is the setup: generating paired velocity maps and seismic waveforms when only one modality is abundant. UB-Diff trains a one-in-two-out autoencoder on all the velocity data, then fine-tunes on a small paired set, and runs diffusion in the shared latent. That is a distinct combination over MT-Diffusion, which needs paired data, and single-modal generators. The FID gains in Table 2 are large and consistent, and the downstream InversionNet numbers in Table 3 mostly favor UB-Diff too. Credit where due: the authors state the freeze-flag trade-off plainly in the appendix, and the method is simple enough to reproduce in principle.\n\nNow the soft spots, in proportion. The biggest one is that the pairwise claim rests on an indirect proxy. Training InversionNet on generated pairs and testing on real data can look good even if individual pairs are physically mismatched, because a high-capacity network can absorb noise or exploit dataset statistics. The direct physical check is only two examples in Figure 6. That is thin. I would not call it a fatal flaw, but the paper's title and abstract claim reliable pairwise generation, and that is not yet demonstrated at the aggregate level. The stress-test note lands on this correctly.\n\nSecond, the freeze flag is chosen per dataset after seeing results. That is post hoc selection, and without error bars or multiple runs it is hard to know how much of the reported improvement is tuning. The FID for seismic waveforms as 1D temporal signals also needs a stated feature extractor; the paper does not say which one, and that matters for comparability.\n\nThird, the method leans on Feng et al. 2022's near-linear latent relationship, and the paper does not validate that alignment independently. That is a legitimate concern, but it is not circular: they use the prior result as motivation, not as proof.\n\nMinor but real: no code or data release, some single-run tables, and the appendix discussion honestly describes the freezes but does not quantify variance.\n\nOverall, the central method is sensible, the empirical improvements look real if not yet robustly quantified, and the core claim is plausible but overreaches slightly relative to the evidence. This paper deserves a serious referee. A good revision would add error bars, specify the FID extractor, and check pairwise physical consistency on a larger sample, ideally with a wave-equation misfit aggregated over many generated pairs. I would cite it if I worked on seismic data augmentation, and I would bring it to a reading group to discuss the evaluation gap between generation quality and physical consistency.","headline":"Credible new combination for unbalanced paired geoscience data generation, but the pairwise physical-consistency claim needs stronger evidence than the current indirect proxy and two visual examples.","tokens_in":17879,"tokens_out":633,"would_cite":true,"duration_ms":7962,"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":"UB-Diff, a diffusion model that learns a shared co-latent code from abundant velocity maps and a few paired seismic waveforms, can generate reliable paired velocity–seismic data and outperforms prior multi-modal and single-modal baselines…","keywords":["diffusion model","multi-modal generation","paired data generation","unbalanced data","seismic full-waveform inversion","co-latent representation","OpenFWI"],"falsifier":"Take a generated velocity map from UB-Diff, run it through a finite-difference seismic wave simulator to obtain a reference waveform for that velocity map, and compare pointwise (L1/L2) with the waveform UB-Diff generates from the same co-latent code. If the mismatch is comparable to the mismatch between two independent random samples, then the co-latent is not enforcing physical pairing, and the pairwise evaluation via InversionNet is hiding that failure. Alternatively, shuffle the paired fine-tuning labels randomly and retrain: if the pairwise InversionNet metrics do not degrade sharply, the method is not learning a true velocity–seismic correspondence.","tokens_in":16885,"feed_emoji":"🌍","tokens_out":7525,"duration_ms":60544,"temperature":0.7,"pith_summary":"The paper tries to solve a practical data problem in scientific computing: paired multi-modal data, such as subsurface velocity maps and their seismic waveforms, are expensive to collect in matched form, yet one modality (velocity maps) can be simulated in abundance. It claims that a diffusion model can be trained mostly on the abundant modality by first learning a shared co-latent representation with a one-in-two-out encoder-decoder, then fine-tuning on a few thousand real pairs, and finally running diffusion in that co-latent space. On five OpenFWI datasets, with 24,000–60,000 velocity maps but only 1,000 or 5,000 paired waveforms, the method reports sharply lower FID scores than MT-Diffusion and the single-modality DDPM/Stable Diffusion baselines, and the generated pairs train an FWI network (InversionNet) that performs better on the original test data. The reason to care is that if true, a wide class of inverse problems with scarce paired measurements could be fed with synthetic training pairs generated from the abundant simulator-only side.","feed_headline":"One latent code turns velocity maps into paired seismic data","feed_subtitle":"UB-Diff learns both modalities from mostly unpaired velocity data and beats prior multi-modal diffusion on OpenFWI.","key_machinery":"The load-bearing component is the one-in-two-out co-latent encoder-decoder, together with the two-step training scheme and the co-latent diffusion process. A single encoder E compresses an input velocity map into a latent vector z in $R^{{c×1×1}}$; two fully-connected layers map z into two modality-specific latent vectors z'_v and z'_s, which are decoded by a CNN decoder Dv (spatial velocity map) and a transformer-based decoder Ds (temporal seismic waveform). This architecture is what allows paired outputs to be generated from a single sampled latent. The two-step training first fits the encoder and the majority-decoder on the abundant velocity maps with an L1+L2 reconstruction loss, then fine-tunes on the paired minority data (sometimes freezing parts of the network); this lets the co-latent be shaped mostly by the abundant modality. The diffusion model runs in the co-latent space using a u-parameterization adapted from Salimans and Ho, so the generative model learns the distribution of velocity-map latents and inherits the ability to decode both modalities.","core_discovery":"The central discovery is that the generative burden can be shifted from the scarce paired data to the abundant unpaired data. Concretely, the paper proposes UB-Diff, which trains an encoder E to map the abundant velocity maps into a compact co-latent code z, and two decoders Dv and Ds that map z back to velocity maps and seismic waveforms respectively. The encoder-decoder is first trained self-supervised on velocity maps alone, then fine-tuned on the limited paired examples with a freeze/unfreeze scheme; afterwards a diffusion model is trained in the co-latent space using all velocity maps, so that sampling a denoised latent and decoding it through both decoders yields a velocity map and a matching seismic waveform. The paper's empirical claim is that on FlatVel-A, CurveVel-A, FlatFault-A, CurveFault-A, and Style-A, this scheme yields better FID for both modalities, and better MAE/MSE/SSIM when the generated pairs are used to train InversionNet, than MT-Diffusion, DDPM, and a reproduced latent diffusion baseline.","pith_inferences":["A direct physics test would strengthen the claim: feed a generated velocity map into a seismic forward solver (finite-difference wave simulation) and compare the simulated waveform with the waveform UB-Diff decodes from the same latent; the paper only validates pairwise quality indirectly through InversionNet reconstruction metrics, not by pointwise physical forward consistency.","The near-linear latent alignment premise may break for velocity structures far outside the paired fine-tuning distribution; a targeted test would fine-tune on simple flat-layer pairs and then generate complex faulted or Style-A-like structures to see whether the seismic output remains physically coupled to the velocity output.","The same two-step co-latent recipe could be applied to other unbalanced scientific pairs, e.g., abundant synthetic CO2 plume simulations with rare field monitoring data, or abundant anatomical images with scarce modality-matched labels, as long as a shared latent space can be established.","The freeze/unfreeze choice is a tunable knob that balances majority-group reconstruction against minority-group fidelity; the paper reports dataset-dependent decisions, suggesting a small validation-driven selection rule could be a useful extension."],"forward_implications":["Downstream FWI training can be bootstrapped from synthesized pairs: training InversionNet on 10,000 UB-Diff-generated pairs improved test MAE/MSE/SSIM over pairs from MT-Diffusion on nearly all datasets and settings.","The method exploits all of the majority-modality data even when only 1,000 or 5,000 paired samples exist; on FlatVel-A with 1,000 pairs it reports velocity FID of 16.45 and seismic FID of 97.03, far below the MT-Diffusion baseline.","Single-modality generation also improves: UB-Diff's velocity-map FID is lower than DDPM and Stable Diffusion on all five datasets in the 1k pair setting, and its seismic waveform FID beats SD in 7 of 10 tested configurations.","The framework should transfer to other domains where two modalities can be aligned in a common latent space, since nothing in the architecture depends on seismology specifics beyond the choice of decoders."],"supporting_citations":[{"why":"Supplies the observed near-linear relationship between seismic and velocity latent spaces, the load-bearing premise for using a shared co-latent.","marker":"Feng et al. 2022a"},{"why":"Restates the same latent-alignment result in the framework section to justify the one-in-two-out co-latent design.","marker":"Feng et al. 2022b"},{"why":"Provides the base DDPM forward and reverse diffusion formulation that UB-Diff adapts to the co-latent space.","marker":"Ho, Jain, and Abbeel 2020"},{"why":"Source of the encoder-decoder architecture inspiration and the InversionNet architecture used as the downstream FWI evaluation model.","marker":"Wu and Lin 2019"},{"why":"Supplies the u-parameterization of the diffusion reverse process and the associated training loss used by UB-Diff.","marker":"Salimans and Ho 2022"},{"why":"MT-Diffusion, the multi-modal generation baseline that UB-Diff compares against and claims to outperform on unbalanced data.","marker":"Chen et al. 2024"},{"why":"Provides the OpenFWI benchmark datasets used for all experiments.","marker":"Deng et al. 2022"},{"why":"Stable Diffusion / latent diffusion baseline reproduced for single-modality generation comparisons.","marker":"Rombach et al. 2022"},{"why":"Defines the FID metric used for macro-level evaluation of generated velocity maps and seismic waveforms.","marker":"Heusel et al. 2017"}],"fun_headline_variants":["UB-Diff: Unpaired data fuels paired seismic generation","One latent, two outputs: Unbalanced data powers paired seismic","Diffusion model turns abundant velocity maps into paired seismic data","From unpaired to paired: UB-Diff beats prior art on seismic data","Co-latent diffusion: Make the most of scarce paired geoscience data"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method assumes that a latent code learned almost entirely from velocity maps aligns well with the seismic waveform latent space, so that a short fine-tuning on 1,000 or 5,000 pairs is enough to make the decoder produce physically consistent seismic counterparts for unseen co-latent samples.","fun_headline_variants_meta":{"raw":{"variants":["UB-Diff: Unpaired data fuels paired seismic generation","One latent, two outputs: Unbalanced data powers paired seismic","Diffusion model turns abundant velocity maps into paired seismic data","From unpaired to paired: UB-Diff beats prior art on seismic data","Co-latent diffusion: Make the most of scarce paired geoscience data"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000615,"raw_usage":{"total_tokens":2924,"prompt_tokens":1082,"completion_tokens":1842,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":698,"completion_tokens_details":{"reasoning_tokens":1752}},"tokens_in":698,"tokens_out":1842,"duration_ms":10799,"temperature":1.0,"reasoning_tokens":1752,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T22:38:31.975505+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a generated velocity map from UB-Diff, run it through a finite-difference seismic wave simulator to obtain a reference waveform for that velocity map, and compare pointwise (L1/L2) with the waveform UB-Diff generates from the same co-latent code. If the mismatch is comparable to the mismatch between two independent random samples, then the co-latent is not enforcing physical pairing, and the pairwise evaluation via InversionNet is hiding that failure. Alternatively, shuffle the paired fine-tuning labels randomly and retrain: if the pairwise InversionNet metrics do not degrade sharply, the method is not learning a true velocity–seismic correspondence.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the OpenFWI benchmark datasets used for all experiments."}],"review_version":1}