{"id":"9ffe3ae2-8fea-439e-b8c4-6e4c7a325217","arxiv_id":"2501.05441","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"A minimalist GAN with a regularized relativistic loss and modern backbone matches or beats StyleGAN2 and several diffusion models on standard FID benchmarks.","lead":"An image-generation model called R3GAN strips the usual GAN training tricks down to a relativistic loss with two gradient penalties and a modern ResNet-style backbone. The authors report that this simpler recipe trains stably and beats the widely used StyleGAN2 on several standard image benchmarks while generating in one step.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The local convergence theorem does not cover the R1+R2 recipe that actually stabilizes training; the paper's own R1-only divergence shows the central stability claim rests on unproved global behavior.","rationale":"The reader's weakest assumption identifies the same load-bearing gap: the convergence theorem is local and rests on unverifiable realizability/expressiveness assumptions, while the practical stability of the R1+R2 loss is not covered by the proof. My stress test sharpens this by noting that the paper itself demonstrates divergence of RpGAN+R1—a loss configuration the theorem claims to cover—from random initialization, and that the additions which fix this divergence (R2, augmentation, schedules, EMA) are outside the analysis. This does not overturn the paper; the empirical results are substantial, and the theoretical claim is carefully scoped as local. However, it means the narrative that the loss is 'well-behaved' in the sense required for the baseline is an empirical claim, not a proven consequence. The reader's CONDITIONAL verdict already accounts for this by demanding code and further verification, so I recommend no change to the verdict.","tokens_in":54,"tokens_out":4778,"duration_ms":127639,"concrete_test":"Analyze the regularized DiracRpGAN ODE of Appendix B with both R1 and R2, i.e. v=(−ψ f′(ψθ), θ f′(ψθ)−γψ) with f(t)=−log(1+e^{-t}), and determine global stability: either exhibit a Lyapunov function for all γ>0 or find an initial point whose trajectory does not converge. If the simplest model is not globally convergent, then the R1+R2 recipe has no global guarantee even in the toy setting; if it is globally convergent, it still would not establish the finite-network case, but it would at least indicate the local theorem can be extended beyond an infinitesimal neighborhood.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that the proposed regularized relativistic loss is 'well-behaved' and removes the need for ad-hoc tricks is supported by a local convergence theorem (Appendix C) for gradient descent on the idealized objective. That theorem covers RpGAN with R1 alone or R2 alone. Yet the paper's own Figure 1 and Table 1 show that RpGAN+R1, which the theorem covers, diverges from random initialization on StackedMNIST; the authors concede in Section 2.4 that 'R1 alone is not sufficient for globally-convergent training.' The practical recipe therefore relies on adding R2, and the proof for the combined R1+R2 loss is asserted as 'trivially extended' without details. Moreover, the final model uses augmentation, EMA, cosine burn-in schedules, cutout, and BFloat16 (Appendix D), none of which appear in the convergence analysis. For finite networks, Assumption III (discriminator expressiveness and C1 equilibrium manifolds) is unverifiable, so the theorem does not establish the stability that the empirical recipe exploits. The load-bearing inference that a principled loss makes GANs easy to train is not what the proofs demonstrate; it is an empirical observation about a specific architecture and hyperparameter schedule.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper argues that GAN training need not rely on a large bag of empirical tricks. It derives a regularized relativistic GAN objective (RpGAN + R1 + R2), proves local convergence for the R1- or R2-regularized variants in the idealized setting of Mescheder et al., and then uses this loss to strip StyleGAN2 down to a minimal backbone and modernize it into R3GAN. The authors report FID results on FFHQ-256, FFHQ-64, CIFAR-10, ImageNet-32, ImageNet-64, and StackedMNIST, claiming that R3GAN surpasses StyleGAN2 and is competitive with state-of-the-art GANs and diffusion models.","tokens_in":28568,"tokens_out":5845,"duration_ms":52956,"significance":"If the empirical results are reproducible, this is a valuable contribution: it provides a simple modern GAN baseline, a careful ablation roadmap from StyleGAN2 to R3GAN, and evidence that a principled loss can replace many ad-hoc GAN tricks. The paper is commendably transparent in several respects: it reports detailed hyperparameters (Appendix D), includes training curves (Appendix G), documents negative results (Appendix E), and adapts the Mescheder et al. convergence framework to RpGAN. However, the gap between the theoretical guarantee and the actual training recipe, the unsupported ImageNet claim in the abstract, and the absence of error bars currently prevent me from endorsing the paper as is.","major_comments":[{"comment":"The abstract claims that R3GAN 'surpasses StyleGAN2 on FFHQ, ImageNet, CIFAR, and Stacked MNIST datasets,' but Tables 7 and 8 (ImageNet-32 and ImageNet-64) contain no StyleGAN2 baseline; the comparisons there are against StyleGAN-XL, BigGAN-deep, and diffusion models. The ImageNet part of the claim is therefore unsupported by the reported experiments. Please add the missing StyleGAN2 results or revise the claim to name the methods actually compared.","section":"Abstract; Sections 4.6-4.7"},{"comment":"The convergence theorem is proved for RpGAN with either R1 or R2 alone, yet the paper's own Figure 1 and Table 1 show that RpGAN+R1 diverges from random initialization on StackedMNIST. The statement in Appendix C that the combined R1+R2 case follows 'trivially' is too thin, because the final recipe relies on exactly this combination, and the practical model also uses augmentation, EMA, cosine schedules, and BFloat16 (Appendix D, Table 9) that are absent from the analysis. The load-bearing claim that the loss is 'well-behaved' and removes the need for tricks is thus not established by the proof; it is an empirical property of a specific architecture and training schedule. Please provide the combined proof or a detailed argument, and explicitly scope the theoretical claim as local and as not covering the full training recipe.","section":"Section 2.4; Appendix C, Eq. (20)"},{"comment":"All FID numbers are single runs without error bars or confidence intervals. Some of the headline improvements are modest (e.g., 2.75 vs. 3.78 on FFHQ-256, 1.95 vs. 2.39 on FFHQ-64, 1.96 vs. 2.42 on CIFAR-10), so run-to-run variation could affect the conclusions. Please provide multiple seeds and variance estimates for at least the smaller-scale experiments (StackedMNIST, CIFAR-10, FFHQ-64), and state the known variability for the larger runs.","section":"Tables 4-8; NeurIPS checklist item 7"}],"minor_comments":[{"comment":"R2 is written as R2(θ,ψ), but the gradient vector field only includes −∇ψRi and not a generator-side gradient ∇θR2; please clarify that R2 is a discriminator-only regularizer in the training algorithm.","section":"Eq. (20)"},{"comment":"There is a typo: 'evaluted' should be 'evaluated'.","section":"Table 2 caption"},{"comment":"The text says the comparison includes 'likelihood-based methods,' but Table 3 also lists GAN variants such as MEG and Adv. DSM; please clarify the category labels for these baselines.","section":"Section 4.2"},{"comment":"The checklist states that no code is available at submission time; given that the paper proposes a new baseline architecture and loss, please make the code publicly available or provide a detailed implementation plan at revision time.","section":"NeurIPS checklist item 5"}],"recommendation":"major_revision","confidential_remarks":"This is a strong empirical paper with a useful baseline, but the theoretical claims are oversold relative to what is proved and the abstract overstates the StyleGAN2 comparison on ImageNet. The main concerns are fixable: add missing baselines, provide a rigorous treatment of the R1+R2 combined loss, and add variance information. I would be willing to look at a revised version."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's my take. The paper is a genuine step for GAN practice: it shows that a stripped-down StyleGAN with a modern ResNet-ish backbone and a relativistic loss with R1+R2 penalties trains stably and gets competitive FID on standard benchmarks, often with one forward pass. The roadmap ablation on FFHQ is well done, and the appendix is unusually transparent about hyperparameters and negative results. Credit where due: combining RpGAN with R1+R2, the local convergence analysis for regularized RpGAN, and the no-tricks baseline are new relative to prior work cited.\n\nThe soft spots are real but not fatal. The convergence theorem in Appendix C covers RpGAN with R1 alone or R2 alone; the paper says the R1+R2 case follows trivially but gives no details. That matters because the paper's own StackedMNIST experiment shows RpGAN+R1 diverges; the practical stability comes from the combination, which is exactly what isn't proven. The authors are candid about this in Section 2.4, so it's not a hidden flaw, but the abstract's \"well-behaved ... loss\" and \"prove local convergence\" overstate what the math covers. The guarantee is local under realizable-case assumptions that can't be verified for finite networks. That's standard for this literature, but it means the central stability claim is empirical, not proven.\n\nEmpirically, the FID tables are single runs without error bars, and code wasn't released at submission. The abstract claims to surpass StyleGAN2 on ImageNet, but the ImageNet tables compare against StyleGAN-XL and diffusion models, not StyleGAN2. These are addressable issues. The detailed hyperparameter table and negative-results appendix help reproducibility a lot.\n\nBottom line: if the results hold up, this is an important baseline for the GAN community. The paper deserves a serious referee. I'd ask for the R1+R2 proof (or a softened claim), code release, and at least a couple of seeds on the main benchmarks. I'd also fix the abstract to say exactly which comparison was run.","headline":"A credible, well-executed modern GAN baseline with a real gap between the local convergence theory and the R1+R2 recipe that actually stabilizes training; worth serious review.","tokens_in":29092,"tokens_out":2078,"would_cite":true,"duration_ms":20227,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["68T07"],"pacs":[],"model":"deepseek-v4-flash","headline":"With the right regularized loss, GAN training needs no tricks: the minimalist R3GAN beats StyleGAN2 and matches diffusion models in one forward pass.","keywords":["GAN stability","relativistic GAN","zero-centered gradient penalty","local convergence","mode dropping","StyleGAN","FID benchmark","one-step generation"],"falsifier":"A direct test would be to train the exact trick-free R3GAN recipe on a held-out dataset such as LSUN bedrooms or ImageNet-256: if training diverges or collapses unless some removed trick is restored, the claim that the RpGAN+R1+R2 loss alone provides sufficient stability would fail; a cheaper check is to rerun the Stacked-MNIST experiment with RpGAN+R1 alone across many seeds and a tuned learning-rate schedule, since the paper's central practical justification for needing R2 is that R1-only training diverges from typical starts.","tokens_in":81,"feed_emoji":"🖼️","tokens_out":11616,"duration_ms":104681,"temperature":0.7,"pith_summary":"The paper sets out to overturn the common belief that GANs are intrinsically unstable and hard to train, and claims the real difficulty lies in two correctable choices: a brittle training objective and an outdated visual backbone. Its central proposal is a relativistic pairing loss (which judges each fake sample relative to a real sample) regularized by zero-centered gradient penalties on both real and fake data, a combination the authors prove to be locally convergent, unlike most earlier relativistic losses. That stability is then used as a license to strip away every ad-hoc trick from StyleGAN2 — mapping network, style injection, minibatch standard deviation, equalized learning rate, path-length regularization — and to replace the old backbone with modern ResNet/ConvNeXt-style blocks. The resulting minimalist baseline, R3GAN, surpasses StyleGAN2 on FFHQ, ImageNet, CIFAR-10, and Stacked-MNIST, and matches or beats several diffusion baselines while requiring only one network evaluation. If correct, the paper reframes 'GANs are difficult' as a property of outdated components rather than of adversarial training itself, and hands the field a clean starting point.","feed_headline":"A stable GAN loss drops every trick and beats StyleGAN2","feed_subtitle":"R3GAN's one-step generator outruns StyleGAN2 and diffusion baselines on FID.","key_machinery":"The central object is the regularized relativistic pairing loss (RpGAN), which compares each fake sample against a real sample rather than against a fixed decision boundary: $L(\\theta,\\psi)=\\mathbb{E}_{z\\sim p_z, x\\sim p_D}[f(D_\\psi(G_\\theta(z))-D_\\psi(x))] + \\frac{\\gamma}{2}\\mathbb{E}_{x\\sim p_D}[\\|\\nabla_x D_\\psi\\|^2] + \\frac{\\gamma}{2}\\mathbb{E}_{x\\sim p_\\theta}[\\|\\nabla_x D_\\psi\\|^2]$, where $f(t)=-\\log(1+e^{-t})$, plus the decision to apply gradient penalties on both the real and fake distributions. The convergence argument follows the established technique of linearizing the gradient vector field $v(\\theta,\\psi)=(-\\nabla_\\theta L,\\nabla_\\psi L)$ at the equilibrium $(\\theta^*, \\psi^*)$ where $p_{\\theta^*}=p_D$ and $D_{\\psi^*}$ is constant on the data support: if the Jacobian's eigenvalues all have negative real part, gradient descent converges locally at a linear rate. The paper shows that for unregularized RpGAN the Jacobian has purely imaginary eigenvalues (a concrete DiracRpGAN example runs in circles), while the penalties shift the spectrum left, yielding the local-convergence theorem. On the architecture side, the mechanism that carries the empirical result is the pairing of this stable loss with a modern symmetric ResNet-style backbone built from bilinear resampling, grouped convolutions, inverted bottlenecks, and fix-up initialization: the absence of normalization layers is compensated by careful initialization, and the absence of tricks is compensated by the well-behaved loss.","core_discovery":"The central discovery is that the combination of a relativistic pairing loss and both zero-centered gradient penalties — R1 on real data and R2 on fake data — turns GAN training into a locally convergent dynamical system, and that this one change is what makes a trick-free, modern architecture viable. The paper proves that unregularized RpGAN does not generally converge: in a Dirac-style toy setting its gradient dynamics run on circles around the equilibrium, and the Jacobian there has purely imaginary eigenvalues. Adding either R1 or R2 moves those eigenvalues into the left half-plane, giving linear local convergence under the standard realizable-case assumptions (the generator can exactly match the data, and the discriminator can detect any deviation from equilibrium). With this loss alone, the authors strip StyleGAN2 down to a bare backbone — no mapping network, style injection, minibatch stddev, equalized learning rate, or path-length regularization — then modernize it with ResNet/ConvNeXt-style blocks, grouped convolutions, and an inverted bottleneck. The resulting R3GAN reports FID 2.75 on FFHQ-256 (StyleGAN2: 3.78), 1.95 on FFHQ-64 (EDM: 2.39), 1.96 on CIFAR-10, 1.27 on ImageNet-32, and 2.09 on ImageNet-64 (EDM: 2.23), all with a single function evaluation, and it recovers all 1000 modes of Stacked-MNIST with the lowest KL divergence among the GANs tested.","pith_inferences":["A natural extension we draw, not the paper's: the same recipe of a stable adversarial loss plus a modern backbone could be carried into text-to-image generation, where one-step generators would compete with diffusion without needing consistency distillation.","Because removing the style network also removes the standard interface for editing and inversion, we expect R3GAN-type generators to need new mechanisms for controllability if they become the common baseline.","We would predict, based on the DiracRpGAN analysis, that other relativistic losses whose unregularized Jacobians have imaginary-axis eigenvalues will also exhibit non-convergence unless equipped with analogous zero-centered penalties; this is a testable classification the paper does not perform.","A cheaper empirical check of our own: monitoring $\\mathbb{E}_{x\\sim p_\\theta}\\|\\nabla_x D\\|^2$ during R1-only training should show the fake-side gradient norm exploding before the generator loss diverges, directly testing the paper's proposed failure mechanism."],"forward_implications":["GAN research can stop layering empirical tricks: a clean, modern backbone trained with a stable loss outperforms the best-known trick-laden GAN on the tested benchmarks.","One-step adversarial generators can match or beat multi-step diffusion models on FID without distillation, so the computational advantage of GANs is achievable in practice.","Loss design becomes a checkable engineering step: candidate relativistic losses can be screened for local convergence by the same eigenvalue test used here.","The paper's mode-recovery numbers on Stacked-MNIST indicate that penalizing the discriminator on both real and fake data (R1 and R2 together) is the practical route to global stability and full mode coverage, even though the local theorem holds for either penalty alone.","Because the reported FID gains are obtained without a pretrained ImageNet discriminator, the results are not attributable to feature leakage the way some state-of-the-art GANs' numbers are."],"supporting_citations":[{"why":"Defines the relativistic pairing GAN objective (RpGAN) that the paper regularizes; the starting point of the loss.","marker":"[22]"},{"why":"Supplies the local-convergence analysis technique, the R1/R2 zero-centered gradient penalties, and the DiracGAN toy model that the paper adapts.","marker":"[52]"},{"why":"Proves that RpGAN's landscape has no bad local minima corresponding to mode dropping, which is why the paper builds on it.","marker":"[81]"},{"why":"Connects R1/R2 to instance-noise smoothing of both real and fake distributions, used to justify why applying both penalties helps stability.","marker":"[64]"},{"why":"The StyleGAN2 architecture that serves as the baseline, the source of the tricks to be removed, and the main comparison point for FID.","marker":"[31]"},{"why":"ConvNeXt design study that motivates the backbone modernization choices (grouped convolution, inverted bottleneck, fewer activations, separate resampling).","marker":"[48]"},{"why":"EDM, the diffusion baseline that R3GAN beats on FFHQ-64 and ImageNet-64 while using one function evaluation; defines the comparison setting.","marker":"[33]"},{"why":"StyleGAN2-ADA, source of the non-leaky augmentation and EMA schedule used in the final training recipe.","marker":"[30]"}],"fun_headline_variants":["R3GAN: trick-free GAN beats StyleGAN2 and diffusion","Modern GAN baseline drops tricks, proves convergence, beats StyleGAN2","One loss rule: R3GAN simplifies GAN, beats StyleGAN2","Simple GAN with proven local convergence outruns StyleGAN2","R3GAN: minimalist GAN outperforms StyleGAN2 and diffusion"],"cache_read_input_tokens":31232,"weakest_assumption_plain":"The convergence proof only applies near an equilibrium where the generator already matches the data distribution almost exactly and the discriminator is expressive enough to detect any small divergence from it — a condition that cannot be verified for the finite networks actually trained, and the paper's own Stacked-MNIST runs show that the theoretically-covered R1-only variant still diverges from random initialization.","fun_headline_variants_meta":{"raw":{"variants":["R3GAN: trick-free GAN beats StyleGAN2 and diffusion","Modern GAN baseline drops tricks, proves convergence, beats StyleGAN2","One loss rule: R3GAN simplifies GAN, beats StyleGAN2","Simple GAN with proven local convergence outruns StyleGAN2","R3GAN: minimalist GAN outperforms StyleGAN2 and diffusion"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00044,"raw_usage":{"total_tokens":2286,"prompt_tokens":1050,"completion_tokens":1236,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":666,"completion_tokens_details":{"reasoning_tokens":1154}},"tokens_in":666,"tokens_out":1236,"duration_ms":10702,"temperature":1.0,"reasoning_tokens":1154,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T21:12:53.054325+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A direct test would be to train the exact trick-free R3GAN recipe on a held-out dataset such as LSUN bedrooms or ImageNet-256: if training diverges or collapses unless some removed trick is restored, the claim that the RpGAN+R1+R2 loss alone provides sufficient stability would fail; a cheaper check is to rerun the Stacked-MNIST experiment with RpGAN+R1 alone across many seeds and a tuned learning-rate schedule, since the paper's central practical justification for needing R2 is that R1-only training diverges from typical starts.","supporting_citations":[{"cited_title":"Towards a better global loss landscape of gans","cited_arxiv_id":null,"evidence_quote":"Proves that RpGAN's landscape has no bad local minima corresponding to mode dropping, which is why the paper builds on it."},{"cited_title":"Stabilizing training of generative adversarial networks through regularization","cited_arxiv_id":null,"evidence_quote":"Connects R1/R2 to instance-noise smoothing of both real and fake distributions, used to justify why applying both penalties helps stability."}],"review_version":1}