{"id":"666269bc-bf99-4547-834e-360c18e0303b","arxiv_id":"2506.02981","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"AstroDiff improves astronomical image restoration under atmospheric turbulence by fusing a diffusion-based generative prior with a restoration branch via SGLD, achieving lower LPIPS and BRISQUE than prior learning-based methods.","lead":"AstroDiff is a new AI method that removes blur and distortion caused by Earth's atmosphere from telescope photos of planets. It combines a generative model that knows what planets look like with a restoration model, and reports sharper results under severe turbulence than existing deep-learning approaches.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The SOTA claim is tested only on synthetic turbulence from the Chan et al. simulator; real-telescope evaluation has no baseline comparison, so transfer to real astrophotography is unverified.","rationale":"The paper has plausible components: large synthetic datasets and a two-branch diffusion design, and the simulator is a reasonable first test. But the central claim is about real astrophotography, and the only evidence that crosses the sim-to-real gap is BRISQUE on real images, a no-reference metric that rewards smoothness and is not compared against baselines. Without real paired data or a validated simulator, LPIPS wins on synthetic severe turbulence cannot carry the conclusion. The reader already identified this as the weakest assumption, and the conditional verdict is appropriate; the table discrepancy strengthens the need for code release but does not change the verdict.","tokens_in":6844,"tokens_out":6750,"duration_ms":86488,"concrete_test":"On the same Celestron NexStar 127 SLT / ZWO ASI462MC hardware, record short-exposure video of a bright planet or the Moon, register the frames, and form a lucky-imaging ground-truth stack. Run AstroDiff, DATUM, TMT, and ESTRNN on the degraded frames and compare PSNR/LPIPS; if AstroDiff does not lead, the 'best turbulence mitigation model' claim should be restricted to the simulated benchmark.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Table 3, the only quantitative SOTA comparison, is computed on AstroEVA images degraded by the turbulence simulator of Chan et al. [4]. The real telescope captures in AstroEVA are used only for a no-reference BRISQUE before/after score (Table 2) with no baseline run, so the claim that AstroDiff 'outperforms existing state-of-the-art learning-based methods' has not been demonstrated on real atmospheric seeing, which is the target domain named in the abstract and conclusion. If the simulator does not reproduce the spatial/temporal statistics of the Celestron/ZWO setup, the LPIPS advantages under simulated C_n^2 values do not transfer. This concern is compounded by an internal inconsistency: the same final model appears as 'AstroDiff' in Table 3 (medium/high PSNR 30.22/30.86) and as 'Two Step Model' in Table 4 (medium/high PSNR 30.02/29.86), so even the published synthetic numbers are not self-consistent absent released code.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes AstroDiff, a two-branch generative framework for mitigating atmospheric turbulence in astrophotography. One branch is an unconditional diffusion prior trained on clean astronomical images; the other is a conditional diffusion restoration branch trained on turbulence-degraded/clean pairs. The two branches are combined at inference time through a stochastic-gradient-Langevin-dynamics (SGLD) fusion that blends generative and restoration gradients. The authors introduce three datasets: PlanetSYN and TechSYN for training and AstroEVA, a benchmark containing real telescope captures plus turbulence-simulated ground-truth pairs. Quantitative evaluation reports PSNR, LPIPS, and BRISQUE against DATUM, TMT, and ESTRNN, and the paper claims state-of-the-art perceptual quality and structural fidelity, especially under severe turbulence.","tokens_in":7037,"tokens_out":3792,"duration_ms":40414,"significance":"If the claims are supported, the paper makes a useful contribution: it is an early application of generative diffusion priors to astrophotography-specific turbulence mitigation, and the proposed PlanetSYN/TechSYN training datasets and the AstroEVA evaluation set could be valuable assets to the community. The SGLD-based fusion of a domain-specific prior with a restoration branch is a sensible idea, and the reported LPIPS and BRISQUE improvements over the compared baselines are encouraging. However, the significance is substantially tempered by the limited and internally inconsistent quantitative evidence, the absence of error bars or significance tests, and the fact that the only head-to-head SOTA comparison is conducted on synthetic turbulence generated by the same simulator used to build the benchmark. The real-telescope evaluation is purely no-reference and includes no baseline methods.","major_comments":[{"comment":"The same final model appears as 'AstroDiff (ours)' in Table 3 with medium/high PSNR 30.22/30.86 and as 'Two Step Model' in Table 4 with medium/high PSNR 30.02/29.86. This internal inconsistency means the reported quantitative results cannot both be correct as presented. Please resolve the discrepancy and report a single, consistent set of numbers, ideally with standard deviations or confidence intervals over the AstroEVA subsets.","section":"§4.3, §4.5, Tables 3 and 4"},{"comment":"All state-of-the-art comparisons in Table 3 are on AstroEVA images degraded by the turbulence simulator of Chan et al. [4]. The real-telescope evaluation in Table 2 reports only AstroDiff's BRISQUE before/after with no baseline methods. Therefore the abstract's unqualified claim that AstroDiff 'outperforms existing state-of-the-art learning-based methods' is not demonstrated on real atmospheric seeing, which is the target domain. Either run the baselines on the real captures with a no-reference metric or carefully restrict the claim to simulated turbulence.","section":"§4.2, §4.3, Table 2"},{"comment":"The SGLD update in Eq. (2) is written as y_{t+1} = (eta_t/2)[-∇log ρ_γ(y_t) - ∇log ρ_γ(y_t | x)] + sqrt(eta_t) eta_t, which omits the current iterate y_t on the right-hand side; it is therefore not the standard SGLD update. In addition, the fusion weights, the step-size schedule eta_t, the number of refinement iterations, and the likelihood exponent gamma are never specified. These hyperparameters are essential for reproducibility and for the claim that the fusion procedure is principled. Please correct the equation and report all fusion hyperparameters.","section":"§3.3, Eq. (2)"},{"comment":"The text states that BRISQUE was used to categorize turbulence into weak, medium, and high levels, but Table 1 lists only subject/C_n^2 assignments and no BRISQUE thresholds, scores, or ranges. This makes the grouping used in Table 3 non-reproducible. Please clarify how the categories were obtained and provide the actual BRISQUE intervals for each category.","section":"§3.2, Table 1"}],"minor_comments":[{"comment":"The conclusion's phrase 'making it the best turbulence mitigation model for astrophotography' is too strong given the limited benchmark coverage and the internal inconsistencies in the quantitative tables; please temper the claim to match the demonstrated scope.","section":"Abstract and Conclusion"},{"comment":"The notation p^γ(y|x) on the right-hand side of Eq. (1) is confusing: in a posterior p(y|x) ∝ p^γ(x|y) p(y), the likelihood should be p(x|y), not p(y|x). Please clarify the notation.","section":"§3.1, Eq. (1)"},{"comment":"There are numerous typographical inconsistencies: 'AstroEV A' appears with an inconsistent space, Table 1 uses 'AstroEv', the references contain 'V AEs' and 'F oundations', and Eq. (2) uses both ρ and p for the same distribution. A careful proofreading pass is needed.","section":"Throughout"},{"comment":"The labels 'AstroDiff (Only Restoration Branch)' and 'AstroDiff (Final Architecture)' in Figure 4 are confusing because the restoration branch alone is not the full AstroDiff method. Please relabel the figures to distinguish the one-step restoration branch from the two-step fused model.","section":"Fig. 2 and Fig. 4"},{"comment":"The implementation details omit the total number of parameters, training hardware, and inference time, which are useful for assessing practicality and reproducibility.","section":"§4.1"},{"comment":"The code and results URL (https://web-six-kappa-66.vercel.app/) does not appear to be a standard repository link; please provide a stable GitHub or institutional repository address.","section":"Data availability"}],"recommendation":"major_revision","confidential_remarks":"The paper is short for the breadth of claims it makes, and the self-referential evaluation setup (the authors' own simulator and two of three baselines from the same group) will make reviewers cautious. The internal numerical inconsistency between Tables 3 and 4 is the most serious issue and must be resolved before the work can be considered. If the authors can supply consistent numbers with error bars, specify the fusion hyperparameters, and either add real-data baseline comparisons or restrict the claims accordingly, the paper could become a solid contribution."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Plainly: the paper's real contribution is three new datasets for astrophotography turbulence and the empirical observation that a dual-branch diffusion prior plus restoration model, fused with SGLD, gives better LPIPS and BRISQUE on simulated astronomical turbulence than deterministic baselines. That is worth a look. But the headline claim that it 'outperforms existing state-of-the-art' is not established. The only head-to-head quantitative comparison (Table 3) runs on synthetic images generated by the Chan et al. simulator, which is also the source of two of the three baselines. The real telescope captures are scored only with no-reference BRISQUE, with no baseline run, so transfer to real seeing is unverified. The internal numbers are also inconsistent: Table 3 reports AstroDiff at 30.22/30.86 PSNR for medium/high turbulence, while the same 'Two Step Model' in Table 4 shows 30.02/29.86. The abstract and conclusion should be qualified until that is resolved.\n\nWhat is good: the datasets are concrete and could be reused; the method is simple to understand; the LPIPS improvements are consistent and the paper honestly acknowledges the PSNR trade-off. The architecture itself is a direct adaptation of Bayesian Diffusion Models, so the novelty is not in the math but in the domain-specific training and the empirical demonstration. That is a legitimate, if modest, contribution.\n\nSoft spots beyond the table discrepancy: no error bars anywhere, so we cannot tell if the single-digit LPIPS differences are meaningful; key inference hyperparameters (fusion weights, SGLD step schedule, number of refinement steps, and the likelihood exponent gamma) are not reported; the evaluation set is narrow—one Celestron telescope, mostly Jupiter images—so generalizability is unclear. The authors also claim code is available at a Vercel page; that is not a stable archival link. The self-citation of the simulator is not a flaw by itself, but the evaluation is heavily dependent on that simulator's fidelity.\n\nWho should read this: people working on turbulence mitigation or diffusion-based image restoration will find the datasets and the negative result (perceptual gains without pixel gains) informative. It deserves a serious referee, but it needs major revision: release code and data, add error bars, specify the inference details, fix the tables, and soften the claims to match the evidence. As it stands, the evidence supports 'competitive LPIPS on simulated turbulence,' not 'best model for astrophotography.'","headline":"The datasets and LPIPS gains are real, but the SOTA claim overreaches: head-to-head numbers are synthetic-only, real-data has no baselines, and the tables disagree.","tokens_in":7595,"tokens_out":2492,"would_cite":false,"duration_ms":27806,"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":"AstroDiff, a dual-branch generative model, claims to beat existing learning-based methods on severe astronomical turbulence by fusing a diffusion prior with a conditional restoration branch via stochastic gradient Langevin dynamics.","keywords":["atmospheric turbulence mitigation","astronomical image restoration","diffusion models","Bayesian diffusion","stochastic gradient Langevin dynamics","perceptual image quality","lucky imaging","generative priors"],"falsifier":"A field test would settle it: record a long high-frame-rate burst of a bright planet, build a lucky-imaging reference from the sharpest frames, then run AstroDiff on single frames from the same burst; if the restored single frames do not at least match the lucky-imaging reference in no-reference quality and resolved detail, the claim that AstroDiff is the best turbulence mitigation model for astrophotography is falsified.","tokens_in":6618,"feed_emoji":"🔭","tokens_out":8371,"duration_ms":79378,"temperature":0.7,"pith_summary":"The paper proposes AstroDiff, a single-image generative method for removing atmospheric turbulence from astronomical photographs. It combines a diffusion prior trained only on clean planet images with a diffusion restoration branch trained on synthetic turbulence pairs, then fuses the two through stochastic gradient Langevin dynamics. The authors claim this dual-branch design preserves global structure under severe turbulence where deterministic methods over-smooth or create artifacts, and that it achieves better perceptual quality (LPIPS) and no-reference quality (BRISQUE) than existing learning-based baselines, with a mixed PSNR that is lower in low and medium turbulence. The motivation is to replace the heavy data acquisition and manual processing of lucky imaging with an automated single-frame restoration pipeline.","feed_headline":"AstroDiff beats prior models on severe-turbulence astrophotography","feed_subtitle":"Two fused diffusion branches keep planetary structure where lucky imaging needs hundreds of frames.","key_machinery":"The load-bearing mechanism is the SGLD-based fusion of two separately pre-trained diffusion branches within a Bayesian diffusion formulation. The prior branch is a U-Net (a convolutional encoder–decoder) denoising diffusion model trained on 22,512 clean astronomical images, so it knows what planets should look like; the restoration branch is a conditional diffusion model trained on 110,745 paired clean/turbulent images, so it knows how to invert the degradation. At inference, each SGLD step combines the restoration branch's data-driven gradient and the prior branch's generative gradient with additive Langevin noise, producing a sample from the joint posterior rather than a deterministic regression.","core_discovery":"In its own terms, AstroDiff is a Bayesian diffusion restoration model that writes the posterior as p(y|x) ∝ p^γ(y|x) p(y), with a conditional restoration branch supplying the likelihood and an unconditional generative prior branch supplying the prior over astronomical scenes. Inference alternates gradient descent on both terms with injected noise via SGLD, so the output must simultaneously fit the observed turbulent frame and look like a plausible planet or moon. Evaluated on the AstroEVA benchmark of 1,907 real telescope captures plus simulated turbulence at nine $C_n^{2}$ levels, the paper reports that AstroDiff achieves the lowest LPIPS in every turbulence regime (e.g., 0.07 in high turbulence vs. 0.11–0.18 for the baselines) and improves BRISQUE from 95.64 to 22.78 on real data. PSNR is mixed—lower than the best baseline in low and medium regimes, highest in high turbulence—which the authors attribute to the usual perceptual-versus-pixel-accuracy trade-off.","pith_inferences":["Editorial inference: the prior/restoration split is not astronomy-specific, so the same SGLD fusion could be applied to other degradations with scarce clean data, such as underwater imaging or low-light microscopy, though the paper does not test this.","Editorial inference: the real-data BRISQUE improvement does not establish that restored images preserve scientifically useful detail; a user study or a task such as measuring resolved surface features on Jupiter would be needed to know whether the perceptual gain carries astronomical information.","Editorial inference: since every synthetic pair comes from one turbulence simulator, the method's practical ceiling is set by how well that simulator captures anisoplanatic seeing; a direct single-frame versus lucky-imaging comparison on a real telescope with measured turbulence would be the natural next test."],"forward_implications":["Turbulence mitigation becomes a single-image, automated operation: no hundreds of high-frame-rate captures, no lucky-frame selection, no manual registration.","In severe turbulence, AstroDiff retains global coherence and structural detail, where the paper shows deterministic baselines producing artifacts or over-smoothed reconstructions.","The perceptual and no-reference gains are strongest by LPIPS and BRISQUE; applications that need pixel-level fidelity (e.g., astrometry) face a lower PSNR in low and medium regimes.","The released datasets (PlanetSYN, TechSYN, AstroEVA) give future astronomical restoration work a common training and evaluation ground."],"supporting_citations":[{"why":"Supplies the turbulence simulator and the ATSyn Static dataset used to build TechSYN; every synthetic training pair and benchmark image depends on it.","marker":"[4]"},{"why":"Supplies the Bayesian diffusion posterior formulation that motivates fusing a likelihood term with an unconditional prior.","marker":"[20]"},{"why":"Provides stochastic gradient Langevin dynamics, the inference procedure that fuses the two gradient terms.","marker":"[21]"},{"why":"Defines BRISQUE, the no-reference metric used for real-data evaluation and for labeling turbulence levels.","marker":"[22]"},{"why":"Defines LPIPS, the perceptual metric on which AstroDiff reports its main advantage over the baselines.","marker":"[25]"},{"why":"Provides the TMT baseline that AstroDiff is compared against in the benchmark.","marker":"[3]"},{"why":"Provides the ESTRNN baseline that AstroDiff is compared against in the benchmark.","marker":"[24]"},{"why":"Supplies the denoising diffusion probabilistic model and U-Net training recipe on which both branches are built.","marker":"[16]"},{"why":"Supplies the U-Net architecture used for both the prior and restoration branches.","marker":"[23]"}],"fun_headline_variants":["AstroDiff: diffusion model clears atmospheric blur in astronomy","AstroDiff outperforms prior models on turbulent telescope images","Generative diffusion restores planets from ground-based photos","AstroDiff fuses two diffusion branches to beat turbulence"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"Every synthetic training pair and the AstroEVA turbulence benchmark are generated with the simulator of reference [4]; if that simulator does not faithfully reproduce real atmospheric seeing, the reported LPIPS and BRISQUE gains may not appear on real telescope images.","fun_headline_variants_meta":{"raw":{"variants":["AstroDiff: diffusion model clears atmospheric blur in astronomy","AstroDiff outperforms prior models on turbulent telescope images","Generative diffusion restores planets from ground-based photos","AstroDiff fuses two diffusion branches to beat turbulence"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000588,"raw_usage":{"total_tokens":2722,"prompt_tokens":865,"completion_tokens":1857,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":481,"completion_tokens_details":{"reasoning_tokens":1791}},"tokens_in":481,"tokens_out":1857,"duration_ms":15118,"temperature":1.0,"reasoning_tokens":1791,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T11:11:51.500707+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A field test would settle it: record a long high-frame-rate burst of a bright planet, build a lucky-imaging reference from the sharpest frames, then run AstroDiff on single frames from the same burst; if the restored single frames do not at least match the lucky-imaging reference in no-reference quality and resolved detail, the claim that AstroDiff is the best turbulence mitigation model for astrophotography is falsified.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the turbulence simulator and the ATSyn Static dataset used to build TechSYN; every synthetic training pair and benchmark image depends on it."},{"cited_title":"Deep unsupervised learning using nonequi- librium thermodynamics,","cited_arxiv_id":null,"evidence_quote":"Provides stochastic gradient Langevin dynamics, the inference procedure that fuses the two gradient terms."},{"cited_title":"Multiscale structure guided diffusion for im- age deblurring,","cited_arxiv_id":null,"evidence_quote":"Defines LPIPS, the perceptual metric on which AstroDiff reports its main advantage over the baselines."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the TMT baseline that AstroDiff is compared against in the benchmark."},{"cited_title":"Removing atmospheric turbu- lence via space-invariant deconvolution,","cited_arxiv_id":null,"evidence_quote":"Supplies the denoising diffusion probabilistic model and U-Net training recipe on which both branches are built."}],"review_version":1}