{"id":"8c8b3da1-1a3f-44ee-847f-90bc6216d8c7","arxiv_id":"2505.09089","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A time-consistency discriminator guides a pretrained image diffusion model at inference time to generate realistic, stable spatiotemporal sequences without finetuning the diffusion model.","lead":"The paper trains a small discriminator to recognize physically ordered image sequences, then uses its gradient signal to steer a pretrained image diffusion model into generating time-consistent dynamics. It shows comparable quality to a video diffusion model trained from scratch on turbulence and global precipitation, with better calibration and long-run stability.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Centennial-scale stability claim rests on global-mean plots alone; local and spectral dynamics over 100-year rollouts are not validated.","rationale":"The reader's weakest assumption concerns whether one-step local discriminator guidance can keep a rollout stable for hundreds of steps. My concern is closely related but more specific: even the empirical evidence offered for that stability is thinner than the conclusion because only the global mean is quantified. Alternative weaknesses, such as per-dataset tuning of the guidance strength lambda, lack of released code, and possible unfairness of the VDM baseline, are real but secondary. The lambda values are fixed per dataset and the headline comparisons would still be informative; code release affects reproducibility but not the internal argument. The load-bearing issue is the evidential basis for 'stable centennial-scale climate simulations.' The method is local by construction, so slow spatial or spectral drift is not structurally prevented. The proposed test uses the already-computed 100-year rollouts and does not require retraining. This does not overturn the reader's conditional verdict; it sharpens the conditions under which the paper should be accepted.","tokens_in":18979,"tokens_out":8957,"duration_ms":103745,"concrete_test":"Using the existing 10 guided 100-year rollouts, compute the spatial power spectrum, the first three EOFs, and the local autocorrelation function in non-overlapping 10-year windows (2020-2030 through 2140-2150). Test whether the last three windows differ from the first window by more than the inter-run variability, for example via a permutation test on spectral slopes or EOF amplitude projections. If later decades deviate significantly, the stable-global-mean result is insufficient to support the centennial-scale claim.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's headline advantage over the VDM is the stable centennial-scale precipitation rollout (Section 5, Figure 7). The only stability diagnostic shown is the annual rolling global mean over 10 runs and one 170-year run. A stable global mean is a weak necessary condition. The discriminator is trained with m=1 on local one-step transitions (Eqs. 4-6), so the guidance exerts no explicit control on slow spatial patterns, regional biases, or multi-decadal variability. The VDM drift is visible in Figure 7, but the guided runs could in principle exhibit compensating spatial drifts, gradual loss of spectral variance, or degradation of local autocorrelation while the global mean stays flat. No spatial power spectra, EOF stability over time, or local ACF in later decades is reported. Thus 'stable centennial-scale' is currently an assertion about one aggregate statistic, not about the dynamics the method is claimed to preserve. The Discussion lists limitations on short-term skill and univariate settings, but does not flag this missing validation of the central stability claim.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes an inference-time guidance method that uses a separately trained time-consistency discriminator to steer pretrained image diffusion models toward temporally consistent autoregressive generation. The discriminator classifies whether a noised image is the next frame given the current and previous clean frames, and its logit gradient is added to the unconditional score in the reverse SDE. The method is evaluated on 2D Navier-Stokes turbulence and global ERA5 daily precipitation, with comparisons to an unconditional image diffusion model and a video diffusion model trained from scratch. Metrics include Wasserstein distances between Hovmöller-diagram rows, autocorrelation functions, CRPS, spread-skill ratios, EOFs, waiting-time distributions, bias maps, and long-run stability assessed by annual rolling global means. The paper claims comparable temporal consistency to the video diffusion model, improved calibration and lower bias, and stable centennial-scale climate simulations.","tokens_in":19165,"tokens_out":5529,"duration_ms":56425,"significance":"If the claims hold, the contribution is practically significant: it offers a way to reuse pretrained image diffusion models for spatiotemporal generation without architecture changes or finetuning, with modest inference overhead (reported as 3%-8%). The empirical evaluation is unusually broad, drawing on standard metrics from fluid dynamics and climate science, and the manuscript reports detailed hyperparameters, training configurations, and sampling pseudocode. The main uncertainty is whether the headline stability claim is supported by the evidence, given that it rests on a single aggregate statistic, and whether the reported quantitative advantages are robust to hyperparameter choices and statistical noise. These issues are addressable and do not undermine the core idea, but they are material to the paper's central claims.","major_comments":[{"comment":"The claim of 'stable centennial-scale climate simulations' is supported only by annual rolling global-mean plots. Because the discriminator is trained on local one-step transitions with m=1 (Eqs. 4-6), nothing in the method explicitly controls slow spatial patterns, regional biases, or multi-decadal variability; a flat global mean is compatible with compensating regional drifts or a gradual loss of spectral variance. The manuscript should add diagnostics for the 100-year and 170-year rollouts, such as spatial power spectra, EOF stability over time, local ACF in early versus late decades, and regional bias maps, ideally with trends of these quantities over the run. Without such validation, the stability advantage over the video DM is not established beyond a single aggregate statistic.","section":"Section 5, Figure 7"},{"comment":"The guidance strength λ and conditioning length m are tuned per dataset (λ=14 for vorticity, λ=68 for precipitation, m=1) without any sensitivity analysis. Since λ controls the relative weight of the discriminator gradient in Eq. (5), the reported improvements in calibration, bias, and long-run stability may depend critically on this choice. The authors should provide a sensitivity study over λ, and ideally m, for at least one headline metric per dataset, such as global-mean drift, ACF error, or CRPS, to demonstrate robustness and rule out hyperparameter selection effects.","section":"Sections 4-5, Table 1"},{"comment":"The quantitative comparisons of CRPS and spread-skill ratio are reported without confidence intervals or significance tests. Differences between the guided DM and video DM at individual lead times are small, and statements such as 'improved calibration' and 'slightly outperforming' are not supported by uncertainty quantification. The authors should add bootstrap confidence intervals or significance tests over the 100 forecasts for CRPS, SSR, and the error curves in Figures 14-15, so that readers can assess whether the reported differences are meaningful.","section":"Section 5, Figures 4c, 6c, 11, 12"},{"comment":"The discriminator is trained using ground-truth clean conditioning frames (Eq. 6), but during autoregressive inference it is applied to previously generated frames. This train/inference distribution shift is not discussed and is a potential source of error accumulation in long rollouts, especially for the centennial-scale claim. The authors should discuss this issue explicitly and, ideally, measure how the discriminator's classification accuracy degrades when conditioned on generated frames rather than ground-truth frames.","section":"Section 3 and Algorithm 1"}],"minor_comments":[{"comment":"Equation (13) is a tautology and does not justify Equation (14); the latter follows directly from Bayes' theorem. Consider replacing Eq. (13) with the correct factorization p(x^{n+1}|x^{(n-m):n}) = p(x^{n+1}) p(x^{(n-m):n}|x^{n+1}) / p(x^{(n-m):n}), or removing it entirely.","section":"Appendix B.1, Eq. (13)"},{"comment":"The phrase 'reserve diffusion process' should be 'reverse diffusion process' in the caption of Figure 2.","section":"Figure 2 caption and main text"},{"comment":"The time axis of Figure 7 extends beyond the ERA5 test period (2011-2020), but the ground-truth line appears constant; the authors should clarify how the ground-truth reference is represented for later years and state how the 170-year guided run is initialized.","section":"Figure 7"},{"comment":"The sampling parameters Stmin and Stmax are not defined in the table or the main text; they should be defined consistently with the stochastic sampler notation in Appendix B.3.","section":"Table 1"},{"comment":"The statement that 'all generative DMs remain sharp' is qualitative; sharpness is not defined or measured. Either define a sharpness metric or soften the claim.","section":"Section 5"}],"recommendation":"major_revision","confidential_remarks":"The core idea is promising and the evaluation is broad, but the headline stability claim needs substantially more evidence before publication. The lack of uncertainty quantification in the quantitative comparisons is pervasive and should be treated as a required revision rather than a polish item. I would also encourage the authors to release code or model checkpoints to support reproducibility."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear Colleague,\n\nYou should know this paper is worth your time. It shows that a pretrained image diffusion model, guided at inference by a small temporal discriminator, can generate realistic autoregressive dynamics without any video finetuning. The discriminator costs 3-8% extra compute. That is the actual new thing, and it is cleanly positioned against prior discriminator-guidance and temporal-GAN work.\n\nWhat the paper does well is evaluation. On 2D Navier-Stokes turbulence and global ERA5 precipitation, the metrics go beyond samples: ACF, CRPS, spread-skill ratio, EOFs, waiting-time distributions, and long rollouts. The guided model matches or beats a from-scratch video diffusion model on most of these, with better spread-skill calibration and lower bias. I see no circularity problem: the discriminator is trained on real temporal pairs and tested on independent periods.\n\nWhere I push back is the \"stable centennial-scale climate simulation\" headline. Figure 7 shows annual rolling global means over 10x100-year and one 170-year runs. A flat global mean is a weak necessary condition; it can hide regional drift, spectral variance loss, or local temporal degradation. The stress-test note is correct: there is no spatial power spectrum, no EOF stability in later decades, and no local ACF for the final century. Since the discriminator uses m=1 local one-step pairs, long-range behavior is only indirectly controlled. This is not fatal, but the claim needs either more analysis or softer wording.\n\nSmaller issues: lambda is tuned per dataset (14 vs 68) with no sensitivity analysis; quantitative results lack confidence intervals; code is not released. The derivation in Appendix B.1 contains a tautological intermediate expression, although the guidance formula itself is the standard GAN result and works.\n\nThe paper is a genuine contribution, and it deserves a serious referee. I would recommend conditional acceptance: add the missing spatiotemporal stability diagnostics, report lambda sensitivity, and ideally release code.\n\nBest.","headline":"A genuinely useful inference-time discriminator guidance for image diffusion models, with broad and serious evaluation; the centennial-stability claim overreaches its global-mean evidence.","tokens_in":19691,"tokens_out":3105,"would_cite":true,"duration_ms":31204,"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":"A time-consistency discriminator lets pretrained image diffusion models generate stable, realistic dynamics over centuries.","keywords":["time-consistency discriminator","image diffusion models","video generation","discriminator guidance","autoregressive rollout","climate simulation","precipitation","Navier-Stokes turbulence"],"falsifier":"Disable the discriminator guidance partway through a guided precipitation rollout and continue with unconditional sampling; if the autocorrelation, Hovmöller statistics, and global mean stay stable for decades, then the guidance is not what enforces long-run stability, whereas a rapid drift would support the paper's causal claim.","tokens_in":18798,"feed_emoji":"🌍","tokens_out":8967,"duration_ms":77764,"temperature":0.7,"pith_summary":"The paper claims that a small discriminator, trained only to tell whether an image follows its predecessor in a dynamical sequence, can act as an inference-time steering signal for a pretrained image diffusion model. Added to the denoising score, this guidance makes the model generate time-consistent sequences of turbulent flows and global daily precipitation, reproducing wave propagation, autocorrelation, extreme-event statistics, and low bias. The method requires no architectural changes or finetuning of the diffusion model and adds only a few percent to sampling cost. Against a video diffusion model trained from scratch, the guided image model matches temporal consistency, has better calibrated ensemble forecasts and lower biases, and sustains stable centennial-scale climate rollouts where the video model drifts.","feed_headline":"Pretrained image diffusion now yields stable century-long climate runs","feed_subtitle":"A light discriminator guides the sampler, matching video diffusion quality without any retraining.","key_machinery":"The central object is the time-consistency discriminator $D_\\theta(x_t^{n+1}; x_0^{n-1:n}, t)$, a binary classifier conditioned on the two most recent denoised frames and on the diffusion noise time. Its guidance term is the gradient with respect to $x_t^{n+1}$ of $\\log(D_\\theta/(1-D_\\theta))$, which equals $\\nabla \\log[p(x_t^{n+1}\\mid \\text{past}) / p(x_t^{n+1})]$; adding it to the unconditional score turns the reverse SDE into a conditional sampler. The discriminator is trained with cross-entropy on real next frames versus importance-sampled corrupted frames and random crops, then applied in both solver steps of the stochastic EDM sampler.","core_discovery":"On its own terms, the paper establishes that temporal consistency can be imposed on an unconditionally trained image diffusion model by adding the score-like guidance term $d_\\theta(x_t^{n+1}; x_0^{n-1:n}, t) = \\nabla_{x_t^{n+1}} \\log(D_\\theta/(1-D_\\theta))$ to the reverse SDE, where $D_\\theta$ is a discriminator trained to separate the conditional density $p(x^{n+1}\\mid x^n, x^{n-1})$ from the marginal $p(x^{n+1})$ at noise level $t$. This term is the gradient of the log-density ratio, so it steers the denoising trajectory toward frames that belong after the already-generated frames. The paper argues, and demonstrates on 2D Navier-Stokes turbulence and ERA5 daily precipitation, that this suffices to turn a pretrained image diffusion model into a dynamical emulator with realistic autocorrelation, Hovmöller structure, extreme-event waiting times, and forecast calibration, and that the resulting autoregressive rollout remains stable over more than a century, whereas a video diffusion baseline exhibits drifting global means.","pith_inferences":["If the discriminator gradient remains informative near the end of denoising, the same recipe should transfer to latent image diffusion models and to video processing tasks such as downscaling or inpainting, because it only needs a clean past frame.","The $m=1$ conditioning makes the method naturally suited to first-order Markov dynamics; systems with longer memory or slower modes may need an $m>1$ discriminator plus a long-range statistic term, which would require retesting the stability claim.","A direct extension the paper only mentions in passing: apply the same locally-trained discriminator to sampling from a video diffusion model, to see whether its centennial-scale drift is corrected by the same guidance.","The balance of results suggests that for climate emulation, long-run stability and calibration may matter more than short-lead forecast skill, so the guided image model may be the more useful configuration for century-scale studies."],"forward_implications":["Any pretrained image diffusion model with access to clean conditioning frames can be converted into a dynamical emulator without retraining; the discriminator trains separately on target data.","Long autoregressive rollouts (more than 100 years at daily steps for precipitation) remain stable under guidance, while the video diffusion baseline develops mean drift, suggesting that guidance prevents error accumulation.","Ensemble forecasts from the guided model are better calibrated (spread-skill ratio) and have lower spatial bias than the video diffusion model, though slightly worse CRPS at the shortest lead times.","The guidance adds roughly 3-8% to generation time, so it can be attached to existing pretrained models and cheap discriminators.","The recovered Hovmöller wave structures and extreme-event waiting time distributions show that the method reproduces dynamical statistics, not just pixel-level sharpness."],"supporting_citations":[{"why":"Supplies the discriminator-guidance formalism (guidance as gradient of the log density ratio) that the paper adapts to temporal consistency.","marker":"[72]"},{"why":"Provides the EDM sampler, preconditioning, and noise schedule used for both training and inference in all experiments.","marker":"[77]"},{"why":"Supplies the score-based reverse-SDE framework and the DDPM++ UNet architecture used for the diffusion models.","marker":"[17]"},{"why":"Gives the optimal discriminator ratio formula from which the guidance term is derived.","marker":"[78]"},{"why":"Introduces a separate temporal discriminator in video GANs, the direct inspiration for the time-consistency discriminator.","marker":"[41]"},{"why":"Demonstrates temporal discriminators for precipitation nowcasting and supports the use of spread-skill and CRPS evaluation.","marker":"[42]"},{"why":"Defines the video diffusion model baseline that the paper trains from scratch and compares against.","marker":"[1]"},{"why":"Supplies the ERA5 global precipitation reanalysis data used as the real-world test case.","marker":"[79]"},{"why":"Provides the numerical solver used to generate the 2D Navier-Stokes vorticity training data.","marker":"[85]"},{"why":"Defines the Hovmöller diagram used to evaluate and visualize the generated spatiotemporal wave dynamics.","marker":"[80]"}],"fun_headline_variants":["Discriminator guides image diffusion to stable climate dynamics","Time-consistency discriminator enables video from image diffusion","Century-stable climate runs from pretrained image diffusion","Image diffusion gets temporal consistency via discriminator","Pretrained diffusion plus discriminator yields realistic video"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that a discriminator trained on only the local one-step transition (the last two denoised frames) produces gradients that keep an autoregressive rollout accurate for hundreds of steps; the century-scale stability is demonstrated empirically in ten 100-year runs and one 170-year run, but it is not theoretically guaranteed, and the video diffusion baseline fails the same test.","fun_headline_variants_meta":{"raw":{"variants":["Discriminator guides image diffusion to stable climate dynamics","Time-consistency discriminator enables video from image diffusion","Century-stable climate runs from pretrained image diffusion","Image diffusion gets temporal consistency via discriminator","Pretrained diffusion plus discriminator yields realistic video"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000189,"raw_usage":{"total_tokens":1331,"prompt_tokens":934,"completion_tokens":397,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":550,"completion_tokens_details":{"reasoning_tokens":322}},"tokens_in":550,"tokens_out":397,"duration_ms":3963,"temperature":1.0,"reasoning_tokens":322,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T21:40:29.556849+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Disable the discriminator guidance partway through a guided precipitation rollout and continue with unconditional sampling; if the autocorrelation, Hovmöller statistics, and global mean stay stable for decades, then the guidance is not what enforces long-run stability, whereas a rapid drift would support the paper's causal claim.","supporting_citations":[{"cited_title":"Generative Adversarial Nets","cited_arxiv_id":null,"evidence_quote":"Gives the optimal discriminator ratio formula from which the guidance term is derived."},{"cited_title":"Skillful Precipitation Nowcasting using Deep Generative Models of Radar","cited_arxiv_id":"2104.00954","evidence_quote":"Demonstrates temporal discriminators for precipitation nowcasting and supports the use of spread-skill and CRPS evaluation."}],"review_version":1}