{"id":"dda1bd60-50fc-409e-bedb-a86d0c5be216","arxiv_id":"2412.19543","paper_version":2,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"An optimization method that uses normalizing flows for feature-space density estimation and multi-start latent search to generate diverse rare samples from pretrained GANs.","lead":"Generating rare images from a pretrained GAN usually requires retraining or massive resampling; this paper instead optimizes latent vectors to push samples into low-density regions of the feature space. A normalizing flow estimates rarity, and a multi-objective loss controls rarity, diversity, and similarity to a reference, all without touching the GAN's weights.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Ablation Table 7 shows the NF rarity term alone leaves the rarity score at baseline (18.99 vs 18.88); the paper never tests the full objective without Lrare, so the central NF-density mechanism is unverified as the driver of rare generation.","rationale":"The reader's weakest assumption was that minimizing the NF log-density reliably increases the k-NN rarity score. The paper's own ablation data directly bear on this: with only Lrare, the mean rarity score is 18.99 versus a baseline of 18.88, showing essentially no improvement. This is stronger evidence than the correlation coefficient alone, because it tests the actual optimization objective rather than a static scatter plot. The missing ablation is Lsim+Ldiv without Lrare; without it, the observed gains could be entirely due to the similarity boundary (which pushes samples toward the edge of the allowed ball) and the diversity term. The paper provides code and extensive experiments, which is creditworthy, but the central mechanism is not yet isolated. The proposed test is cheap and decisive: if the no-Lrare variant achieves the same rarity score, the NF contribution is not load-bearing and the paper's framing needs revision; if the rarity score collapses, the concern is resolved. Since this is a specific, testable condition rather than a demonstrated failure, the existing CONDITIONAL verdict remains appropriate, so no verdict change is recommended.","tokens_in":21357,"tokens_out":7287,"duration_ms":66733,"concrete_test":"Re-run the FFHQ-StyleGAN2 main experiment (1,000 references, N=10, λ1=30, λ2=0.002, σ=0.1, k'=100, same NF, same best-sample selection) with the loss L = λ1 Lsim + λ2 Ldiv only, i.e., set the coefficient of Lrare to 0, while keeping the same Φreal and d* constraints. Report mean Rarity Score, LPIPS, Precision, and OOM% over the 10,000 generated samples, with per-reference paired differences and bootstrap confidence intervals. Compare against the 'Ours' row of Table 1 (RS=23.50). If the no-Lrare RS is within about 1 point of 23.50, the NF term is not load-bearing; if it drops toward the baseline (~19), the term is essential.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim requires that minimizing the normalizing-flow log-density in Eq. (2) reliably increases the k-NN-based rarity score. The paper's own ablation (Table 7) undermines this: with only Lrare, RS = 18.99, which is essentially the baseline RS of 18.88 from Table 1. The increase to 21.11 comes from adding Lsim, and to 21.28 from further adding Ldiv; the full method in Table 1 reaches 23.50. However, the paper never runs the full objective with Lrare removed (i.e., Lsim + Ldiv only), so we cannot tell whether the NF rarity term contributes anything beyond what the similarity boundary and diversity constraint already provide. If the no-Lrare variant matches the full objective, the 'utilizes normalizing flows for density estimation' contribution is not load-bearing, and the method reduces to constrained multi-start diversity optimization. This is a concrete internal gap: the paper attributes the rarity improvement to NF density estimation but provides no ablation isolating its effect within the full objective.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes DivRareGen, a training-free method for generating diverse rare samples from a pretrained GAN. It optimizes latent codes with a multi-objective loss that combines a normalizing-flow log-likelihood term as a differentiable rarity proxy, a one-sided similarity penalty anchored at a reference image, and a pairwise diversity term, subject to constraints that keep samples inside the real-data k-NN manifold and within a similarity boundary. The method is evaluated on FFHQ, AFHQ, and MetFaces with StyleGAN2 variants, reporting improved rarity scores, diversity, and increased prevalence of rare semantic attributes, alongside higher FID/KID values.","tokens_in":21660,"tokens_out":5943,"duration_ms":53058,"significance":"If the central mechanism is verified, the method would be a practically valuable, retraining-free way to explore low-density regions of a pretrained GAN with user-controllable rarity, diversity, and reference similarity. The paper contributes a clear optimization formulation, public code, extensive appendices, and a useful correlation analysis between NF-estimated density and k-NN-based rarity. However, the ablation study does not isolate the contribution of the NF rarity term, and the evaluation pipeline has selection issues (hyperparameters chosen on the evaluation metrics, rarity scores computed after excluding undefined cases). These issues currently leave the paper's main claim unproven, so the result is promising but requires substantial verification.","major_comments":[{"comment":"The ablation in Table 7 does not isolate the normalizing-flow rarity term. With Lrare alone the rarity score is 18.99, essentially equal to the baseline of 18.88 reported in Table 1, and the improvement appears only after adding Lsim (21.11) and Ldiv (21.28). Since the paper never evaluates the full objective without Lrare, the evidence is consistent with the NF term contributing nothing beyond the similarity boundary and diversity constraint; the method could reduce to constrained multi-start diversity optimization. Please add the Lsim + Ldiv configuration (without Lrare) and report both RS and LPIPS, so the role of the NF density estimator is actually tested.","section":"§4.3, Table 7"},{"comment":"The hyperparameters λ1, λ2, σ, and k′ are selected using the same rarity score and LPIPS metrics on the evaluation setup. Appendix C, Tables 9–11, reports RS/LPIPS for different values on the FFHQ-StyleGAN2 setting that later produces the headline results in Table 1, which is tuning on the test distribution and can inflate the reported improvements. Please provide a validation-based selection procedure or a sensitivity analysis that does not use the evaluation metrics for model choice, and report confidence intervals for RS, precision, recall, and FID/KID, not only for LPIPS.","section":"§4.1 and Appendix C"},{"comment":"The text in Section 4.1 states that 'The FID score decreases,' but Table 1 shows FID increasing from 4.17 to 7.38 (higher is worse), and Table 4 shows KID increasing for all three datasets. This misstates the fidelity trade-off. Please correct the wording and explicitly discuss the fidelity cost, especially because the introduction and conclusion emphasize avoiding low-quality samples and the method's precision score is high, which creates an apparent tension with the large FID/KID degradation.","section":"§4.1, Table 1"},{"comment":"Rarity-score computations in the paper exclude samples with undefined rarity scores: Appendix C states 'mean rarity score except for the undefined rarity cases,' and the method produces 11–14% out-of-manifold samples (Tables 9 and 10). If the headline RS in Table 1 is also computed on the in-manifold subset only, the observed improvement may partly reflect this filtering rather than the optimization itself. Please report the out-of-manifold percentage for the main experiments and state unambiguously how undefined rarity scores are treated in every reported RS value.","section":"§4.1 and Appendix C"}],"minor_comments":[{"comment":"The Lsim term is zero whenever d(x, x*) ≤ d*, so it only penalizes violations of the similarity boundary and does not actively attract the solution toward the reference. Please clarify whether the constraint is enforced by projection or by penalty, and why both the constraint and the Lsim term are needed.","section":"§3.2, Eqs. (1)–(2)"},{"comment":"The typeset objective in Figure 2 is missing operators between terms (the expression reads as a concatenation of −λ2 sum, +λ1 max, and log p(xi)). Please fix the equation formatting.","section":"Figure 2"},{"comment":"The full-objective RS in Table 7 is 21.28 for 100 initial latent vectors, while the corresponding setting in Table 1 reports RS 23.50 for 1,000 initial latent vectors. Please explain this difference (sample size, parameter variation, or random seed) so readers can reconcile the two tables.","section":"§4.3, Table 7"},{"comment":"The caption of Table 3 says 'Sorted in descending order of FFHQ(%),' but the rows are listed in ascending order of FFHQ(%) (0.12, 0.39, 0.88, ...). Please correct the caption or the ordering.","section":"Table 3"},{"comment":"The correlation analysis in Figure 8 excludes samples with undefined rarity scores; please state this explicitly in the main text rather than only in the figure context, since it affects the interpretation of the reported 0.815 correlation for fake samples.","section":"§4.4"}],"recommendation":"major_revision","confidential_remarks":"The paper is well organized and the supplementary material is thorough, but the central claim—that the normalizing-flow density term is what drives rare sample generation—is not supported by the provided ablation. The missing Lsim+Ldiv configuration is the key experiment that must be added. I also share concerns about hyperparameter selection on the evaluation metrics and the exclusion of undefined rarity cases from the reported scores. These are fixable within the manuscript's scope, so I recommend major revision rather than rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a practical, well-executed method for generating diverse rare images from a pretrained GAN, and the multi-objective combination of NF density on deep features with similarity and diversity constraints is genuinely new. The central mechanism is less well supported than the paper claims, because the one ablation that would isolate the NF term is missing. Still worth reviewing and likely citing.\n\nWhat's actually new: prior work either resamples (Polarity), needs text conditioning (Chang et al.), or is non-differentiable (Han et al.'s rarity score). Here the authors optimize latent vectors directly with a differentiable NF log-likelihood plus a similarity boundary and a diversity term, using multi-start to get varied outputs. The paper ships code, evaluates on three 1024x1024 datasets, and adds a careful Polarity comparison including online rejection sampling. The correlation check between NF likelihood and rarity score (0.928 for real, 0.815 for fake) is a credible sanity check.\n\nSoft spots: the stress-test is on target. In Table 7, Lrare alone gives RS 18.99 versus baseline 18.88; the gains appear only when Lsim and Ldiv are added. The paper never runs the full objective without Lrare, so we cannot tell whether the NF term contributes anything beyond the similarity boundary and diversity constraint. That is the main issue. Second, hyperparameters (lambda_1, lambda_2, k') are selected using the same evaluation metrics (RS, LPIPS) in Appendix C, which makes the headline numbers a bit optimistic. Third, there are no error bars except the LPIPS t-test, and FID degrades from 4.17 to 7.38 on FFHQ. The authors do acknowledge artifacts and limitations, which is honest, but the conclusion's claim of 'avoiding low-quality samples' is stronger than the evidence.\n\nNone of these are fatal. The method likely works, and the qualitative results are convincing. But the missing ablation should be run before publication, and repeated seeds would help.\n\nWho should read it: anyone working with pretrained GANs for data augmentation, creative generation, or rare-sample benchmarking. I would cite it.\n\nRecommendation: yes, send to peer review; ask for the Lsim+Ldiv ablation and variance reporting.","headline":"Useful, practical rare-sample generation method with real novelty, but the missing no-Lrare ablation leaves the normalizing-flow mechanism under-supported.","tokens_in":22105,"tokens_out":3527,"would_cite":true,"duration_ms":31198,"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 pretrained GAN can be steered to generate rare, diverse images by optimizing latent vectors against a learned density estimate, with no retraining or fine-tuning of the generator.","keywords":["rare sample generation","pretrained GANs","latent space optimization","normalizing flows","density estimation","multi-objective optimization","k-NN rarity score","diversity control"],"falsifier":"On the FFHQ–StyleGAN2 outputs (10,000 optimized samples), compute the Pearson correlation between flow log-likelihood and k-NN rarity score restricted to the low-likelihood tail, say the bottom 10% by $\\log p$; if the correlation collapses toward zero or reverses sign inside that tail, or if the highest-rarity samples are predominantly the kind of manifold artifacts the authors concede are 'inevitable' in the appendix, then the differentiable surrogate is not driving the metric it claims to optimize. A second decisive check: replace the flow with a smooth surrogate trained directly on k-NN rarity values; if it matches or beats the flow-driven results, the flow is incidental rather than load-bearing.","tokens_in":21199,"feed_emoji":"🎨","tokens_out":18265,"duration_ms":154329,"temperature":0.7,"pith_summary":"Generative models produce realistic images but rarely the unusual cases that sit in low-density regions of the data manifold, and methods that improve fidelity often shrink diversity. This paper claims that a pretrained GAN can be pushed to generate rare images — faces with hats, non-brown hair, extreme ages, non-frontal poses, uncommon breeds — by running gradient descent on its latent vectors with a multi-objective loss, without retraining or fine-tuning the generator. The rarity signal is a normalizing flow that estimates the log-density of real images in a deep feature space; minimizing $\\log p(x)$ drives generated samples toward low-density regions. A repulsion term between simultaneously optimized latents enforces diversity, a boundary term keeps each result similar to a chosen reference image, and a manifold constraint keeps outputs realistic. If correct, the method turns any pretrained GAN into a controllable supplier of rare, varied images, which matters for building synthetic datasets that cover underrepresented cases.","feed_headline":"Optimization turns pretrained GANs into rare-image generators","feed_subtitle":"A density-guided latent search lifts rarity and diversity while keeping image quality high, with no model retraining.","key_machinery":"The load-bearing object is the multi-objective latent objective of Eq. (2), made differentiable by replacing the k-NN rarity score with a normalizing-flow density estimate. A Glow-style flow — a stack of invertible layers that turns a simple base distribution into a complex one while giving exact per-sample log-likelihoods — is trained on VGG16-fc2 features of real images, so $\\log p(x)$ serves as a smooth stand-in for rarity; the paper reports Pearson correlations of 0.928 for real and 0.815 for fake samples between the two measures. The similarity term $(\\max\\{d(x,x^*)-d^*,0\\})^2$ with boundary $d^*$ taken from the fake k-NN manifold keeps optimized images near the reference, the diversity term $-\\sum_{j \\neq i} d(x_i,x_j)^2$ pushes multi-start solutions apart, and feasibility is enforced by requiring each optimized feature to lie inside the real k-NN manifold $\\Phi_{\\mathrm{real}}$. Multi-start initializes each run at $z^* + \\epsilon$ with $\\epsilon \\sim \\mathcal{N}(0,\\sigma^2 I)$, and Adam with a step scheduler performs the optimization.","core_discovery":"The paper's central claim is that generating rare images from a pretrained GAN reduces to a constrained, multi-objective optimization over latent vectors. Starting from a reference image $x^* = f(G(z^*))$ in a deep feature space, the method solves $$\\min_{z_i}\\; \\log p(x_i) + \\lambda_1 (\\max\\{d(x_i,x^*)-d^*,0\\})^2 - \\lambda_2 \\sum_{j \\neq i} d(x_i,x_j)^2\\quad \\text{subject to } x_i \\in \\Phi_{\\mathrm{real}},\\; d(x_i,x^*) \\leq d^*,$$ where $p$ is the normalizing-flow density of real images in that feature space, $\\Phi_{\\mathrm{real}}$ is the union of k-nearest-neighbor balls around real samples, and $d^*$ is a penalizing boundary derived from the fake k-NN manifold. Minimizing $\\log p(x_i)$ pushes samples into low-density regions, the similarity penalty reins them back toward the reference, and the pairwise repulsion spreads distinct latents apart so that multi-start optimization from $z^* + \\epsilon$ converges to visually different rare images rather than the same local optimum. Across FFHQ, AFHQ, and MetFaces at 1024×1024 resolution, this raises the k-NN rarity score and pairwise LPIPS diversity relative to baseline sampling while holding precision at 0.76–0.92, and it increases the measured prevalence of rare facial, breed, and painting-style attributes.","pith_inferences":["The 0.815 correlation between flow likelihood and k-NN rarity on fake samples is the bridge the whole method crosses; a direct check of whether that correlation survives inside the low-likelihood tail the optimizer targets would tell whether the surrogate is truly driving the evaluated metric.","Nothing in the objective is GAN-specific, so the recipe — replace a non-differentiable rarity or fidelity metric with a trained differentiable density model on deep features — should transfer to diffusion models, where low-density guidance has so far relied on class-conditional Gaussian assumptions.","Because diversity is a pure geometric repulsion in feature space, the number of distinct rare attributes reachable per reference is bounded by the generator's latent geometry; attribute diversity should saturate as $\\lambda_2$ grows while rarity eventually falls, a pattern already visible in the paper's parameter sweep.","The FID cost on FFHQ (4.17 to 7.38) is structural rather than a defect: any method that deliberately moves mass to low-density regions worsens distribution-level fidelity metrics, so applications should judge rarity-augmented data by task-level outcomes such as classifier accuracy on rare classes rather than by FID alone."],"forward_implications":["Rarity, diversity, and similarity to the reference become user-set knobs: $\\lambda_1$ controls how close results stay to the reference, $\\lambda_2$ controls how far apart the rare samples spread, and the noise scale $\\sigma$ and boundary parameter $k'$ set how far the optimizer roams.","No retraining or fine-tuning of the generator is needed; the only per-dataset cost is training the density estimator, reported as under 30 minutes on a single GPU, after which any reference latent can be turned into rare variants.","Across FFHQ, AFHQ Cat, AFHQ Dog, and MetFaces at 1024×1024, the method raises the rarity score and pairwise LPIPS diversity over baseline sampling while keeping precision between 0.76 and 0.92, compared with 0.38–0.39 for the Polarity baseline, which generates many out-of-manifold artifacts.","Measured rare-attribute prevalence increases on all datasets — hats, eyeglasses, baldness, non-brown hair, extreme age, non-white race, and non-frontal poses among faces; minor cat and dog classes; and non-Western painting styles — so synthetic data can be deliberately enriched with underrepresented cases.","A single reference yields several visually distinct rare images rather than one, because the diversity term keeps the multi-start runs from collapsing into the same local optimum."],"supporting_citations":[{"why":"Defines the k-NN rarity score that the method treats as the target concept of rarity and uses as its main evaluation metric.","marker":"Han et al. (2023)"},{"why":"The Glow normalizing-flow architecture, adapted here as the differentiable density estimator on VGG16 features.","marker":"Kingma and Dhariwal (2018)"},{"why":"Supplies the VGG16-fc2 feature extractor in whose space density, similarity, and diversity are all measured.","marker":"Simonyan and Zisserman (2015)"},{"why":"The StyleGAN architecture and the FFHQ dataset used for the main 1024x1024 experiments.","marker":"Karras, Laine, and Aila (2019)"},{"why":"Provides StyleGAN2-ADA and the MetFaces dataset used for the limited-data experiments.","marker":"Karras et al. (2020a)"},{"why":"The StyleGAN2 config-f generator used in the FFHQ experiments.","marker":"Karras et al. (2020b)"},{"why":"Polarity sampling, the main baseline whose rarity/diversity behavior and precision collapse the method is measured against.","marker":"Humayun, Balestriero, and Baraniuk (2022)"},{"why":"The quality-diversity latent optimization that motivates the similarity-boundary regularization and the multi-start formulation.","marker":"Chang et al. (2024)"},{"why":"Defines the k-NN manifold balls used both as the feasibility constraint on optimized samples and as the precision/recall evaluation metrics.","marker":"Kynk̈änniemi et al. (2019)"},{"why":"LPIPS, used to measure diversity, and the perceptual-metric evidence that Euclidean distance in deep feature space aligns with human perception.","marker":"Zhang et al. (2018)"}],"fun_headline_variants":["Rare-image generation via multi-objective latent optimization","Density-guided latent search turns GANs into rarity generators","Pretrained GANs yield rare images with no retraining","Optimizing latents with normalizing flows for rare GAN samples","Diverse rare outputs from GANs via multi-objective optimization"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method assumes the normalizing flow's density estimate, trained on ordinary real images, stays accurate in exactly the low-density regions the optimizer seeks out, so that minimizing $\\log p(x)$ reliably raises the k-NN rarity score used for evaluation — the reported correlation of 0.815 on fake samples is evidence for the surrogate but not a guarantee that the two measures agree where rarity matters most.","fun_headline_variants_meta":{"raw":{"variants":["Rare-image generation via multi-objective latent optimization","Density-guided latent search turns GANs into rarity generators","Pretrained GANs yield rare images with no retraining","Optimizing latents with normalizing flows for rare GAN samples","Diverse rare outputs from GANs via multi-objective optimization"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000305,"raw_usage":{"total_tokens":1774,"prompt_tokens":995,"completion_tokens":779,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":611,"completion_tokens_details":{"reasoning_tokens":693}},"tokens_in":611,"tokens_out":779,"duration_ms":7483,"temperature":1.0,"reasoning_tokens":693,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T00:13:00.953273+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"On the FFHQ–StyleGAN2 outputs (10,000 optimized samples), compute the Pearson correlation between flow log-likelihood and k-NN rarity score restricted to the low-likelihood tail, say the bottom 10% by $\\log p$; if the correlation collapses toward zero or reverses sign inside that tail, or if the highest-rarity samples are predominantly the kind of manifold artifacts the authors concede are 'inevitable' in the appendix, then the differentiable surrogate is not driving the metric it claims to optimize. A second decisive check: replace the flow with a smooth surrogate trained directly on k-NN rarity values; if it matches or beats the flow-driven results, the flow is incidental rather than load-bearing.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the k-NN rarity score that the method treats as the target concept of rarity and uses as its main evaluation metric."},{"cited_title":"P.; and Dhariwal, P","cited_arxiv_id":null,"evidence_quote":"The Glow normalizing-flow architecture, adapted here as the differentiable density estimator on VGG16 features."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the VGG16-fc2 feature extractor in whose space density, similarity, and diversity are all measured."},{"cited_title":"I.; Balestriero, R.; and Baraniuk, R","cited_arxiv_id":null,"evidence_quote":"Polarity sampling, the main baseline whose rarity/diversity behavior and precision collapse the method is measured against."},{"cited_title":"C.; Booth, S.; Matari \\'c , M","cited_arxiv_id":null,"evidence_quote":"The quality-diversity latent optimization that motivates the similarity-boundary regularization and the multi-start formulation."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the k-NN manifold balls used both as the feasibility constraint on optimized samples and as the precision/recall evaluation metrics."},{"cited_title":"A.; Shechtman, E.; and Wang, O","cited_arxiv_id":null,"evidence_quote":"LPIPS, used to measure diversity, and the perceptual-metric evidence that Euclidean distance in deep feature space aligns with human perception."}],"review_version":1}