{"id":"e0a6b426-f4eb-45a0-af90-1bfbc5e8cdf2","arxiv_id":"2501.11236","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":3,"one_line_summary":"Li-CFG adds an ε-centered gradient penalty to the CFG GAN method and claims this enlarges the discriminator gradient norm, shrinking the latent neighborhood size and thereby increasing image diversity.","lead":"This paper introduces a new GAN training penalty that shifts the usual gradient penalty center from 0 or 1 to a tunable value ε, aiming to improve training stability and output diversity. The authors report lower FID scores on several image datasets compared with the CFG baseline, suggesting the trick can help GANs produce more varied images.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 3.7's radius ordering is unsupported: Proposition 3.5 substitutes an upper bound for the actual latent N-size, Eq. C19 adds a scalar penalty to a gradient vector, and Lemma 3.6's norm inequality is false in general.","rationale":"The central claim is the ordering r_{R1} > r_{R0} > r_{R\\epsilon}. For that claim to hold, Proposition 3.5 must give the actual latent N-size and Lemma 3.6 must compare those actual quantities. Neither condition is met: the proof substitutes an upper bound for the actual infimized quantity, Eq. C19 uses an undefined vector-plus-scalar update, and Lemma 3.6's norm inequality fails for admissible scalar values. The reader's REJECT verdict is therefore appropriate: the core theoretical contribution is not proven, and the empirical results, while suggestive, do not repair a proof gap in the main theorem. I am not objecting to the empirical finding or to the general idea that gradient penalties change generator behavior; the concern is internal to the proof. Since the reader already rejected on essentially this basis, my recommendation is UNCHANGED (maintain REJECT).","tokens_in":33659,"tokens_out":8371,"duration_ms":75374,"concrete_test":"One analytical check: test Lemma 3.6 under the paper's own notation by taking a = \\|\\nabla_x D(G_{\\theta_t}(z))\\| = 0.1, g_0 = 1, and \\|\\epsilon\\| = 0.3. The claimed ordering |a - g_0| < a < a + \\|\\epsilon\\| becomes 0.9 < 0.1 < 0.4, which is false. If the authors instead intend squared penalties, the same example gives (a - g_0)^2 = 0.81 > a^2 = 0.01, so the ordering fails there as well. This single check falsifies Lemma 3.6 unless an additional lower bound a \\geq g_0/2 is proved from the CFG dynamics, and no such bound appears in the paper.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Theorem 3.7 is not established because Proposition 3.5 replaces the actual latent N-size r of Definition 3.1 by an upper bound, and Lemma 3.6 then orders those bounds. Definition 3.1 defines r = \\hat{\\epsilon}(2 inf_z L(z))^{-1} for the actual generator-discrepancy term L(z). In the proof of Proposition 3.5 (Eq. C22), the authors prove L(z) \\leq U(z), where U includes 2\\|G_{\\theta_t}(z_1)-G_{\\theta_t}(z)\\| plus sums of norms involving \\nabla_x D and R, and then define the r in Proposition 3.5 using U instead of L. Since r is inversely proportional to the infimized quantity, ordering U_{R1} < U_{R0} < U_{R\\epsilon} gives only an ordering of these upper-bound-based expressions; the actual radii r_{R1}, r_{R0}, r_{R\\epsilon} are not ordered, because the tightness of each upper bound can vary independently. In addition, Eq. C19 asserts G_{\\theta_{t+1}}(z) = G_{\\theta_t}(z) + \\sum \\eta_m \\delta(x) \\nabla_x D = G_{\\theta_t}(z) + \\sum \\eta_m \\delta(x)(\\nabla_x D + R), adding the scalar penalty R directly to the gradient vector. In Eq. (8), R is a scalar loss term added to the discriminator objective, not a perturbation of the generator's functional gradient; the object \\|\\nabla_x D + R\\| is therefore undefined without further specification. Lemma 3.6's norm ordering is also false in general: if a = \\|\\nabla_x D\\| and a < g_0/2, then |a - g_0| > a, so \\|R_1\\| > \\|R_0\\|, reversing the claimed inequality. Thus the central radius ordering is unsupported.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes Li-CFG, a variant of Composite Functional Gradient (CFG) GAN training that adds a Lipschitz-style gradient penalty to the discriminator. It defines a 'latent N-size' r (Definition 3.1), claims that increasing the discriminator gradient norm reduces r and hence increases sample diversity, introduces an ε-centered gradient penalty (Eq. (9)), and states Theorem 3.7 that the three gradient penalties yield radii satisfying r_R1 > r_R0 > r_Rε. The empirical sections compare Li-CFG with CFG and other GAN baselines on ring/grid, MNIST, CIFAR-10, LSUN, and ImageNet, and also apply the ε-centered penalty to BigGAN and DDGAN.","tokens_in":34128,"tokens_out":6667,"duration_ms":64053,"significance":"The empirical scope is a strength: the authors evaluate across many datasets and architectures, report FID, IS, and precision/recall, and show consistent improvements when adding their ε-centered penalty. If Theorem 3.7 were established, the paper would provide a useful new perspective connecting discriminator gradient penalties to latent-space diversity. However, the central proof is not sound: the derivation rests on an ill-defined update, replaces the quantity in Definition 3.1 by an upper bound, and uses a norm inequality that is false in general. As it stands, the theoretical claims are not established.","major_comments":[{"comment":"The update in Eq. (C19) writes Gθ_{t+1}(z) = Gθ_t(z) + Σ η_m δ(x)(∇_x D_m(Gθ_t(z)) + R), but R in Eq. (8) is a scalar regularization term in the discriminator objective (an expectation of squared norms). Adding this scalar directly to the gradient vector ∇_x D is undefined, and no derivation is provided to show that the gradient penalty enters the CFG generator update in this additive form. Since Proposition 3.5 and all subsequent norm comparisons use ∥∇_x D + R∥, the formula for r with a gradient penalty is not meaningful without a corrected derivation.","section":"Appendix C.1, Eq. (C19)"},{"comment":"Definition 3.1 defines r = ε̂(2 inf_z L(z))^{-1} with L(z) the actual ratio of generator discrepancies. In Eq. (C22) the proof shows L(z) ≤ U(z), and Proposition 3.5 then restates r using U(z) in place of L(z). Since r is inversely proportional to the infimum, replacing L by an upper bound gives r_P = ε̂(2 inf U)^{-1} ≤ r, not equality. The later ordering r_R1 > r_R0 > r_Rε compares these upper-bound expressions; the actual radii are not ordered because the tightness of each upper bound can differ independently. Theorem 3.7 therefore does not follow from the proof as written.","section":"Proposition 3.5 / Appendix C.1, Eq. (C22)"},{"comment":"The claimed ordering ∥R1∥ < ∥R0∥ < ∥Rε∥ is false in general. With a = ∥∇_x D∥ and g0 = 1, if a < 1/2 then |a − 1| > a, so ∥R1∥ = ∥ |a − 1| ∥ > ∥R0∥ = ∥a∥, reversing the first inequality. The condition ∇_x D ≤ 0 used elsewhere does not prevent this, because the gradient norm can still be small. Since Lemma 3.6 is the bridge from penalty norms to radii in Theorem 3.7, this is a load-bearing gap.","section":"Lemma 3.6, Eq. (C30)"},{"comment":"The proof claims that g_m(x) ≤ 0 implies ∇_x D(x) ≤ 0. In the CFG formulation, g_m is a vector-valued function on image space, so 'g_m(x) ≤ 0' is undefined, and a scalar inequality cannot be inferred for each component of ∇_x D. The subsequent argument that the ε-centered penalty enlarges ∥∇_x D∥ relies on this unproved sign condition, so the comparison supporting the ε-centered penalty is not established.","section":"Appendix C.2, proof of ∇_x D(x) ≤ 0"}],"minor_comments":[{"comment":"The symbol ε is used both as a vector (∥ε∥_2 = ε′) and as a scalar inside ε′ = sqrt(C N^2 ε^2); please introduce separate symbols for the vector and the scalar norm target.","section":"Eq. (9) and Appendix D.1"},{"comment":"In the LSUN B block, the 'ours(ε-centered)' row lists only a single FID value (8.78) instead of the IS/FID pair used in all other rows; please fix the formatting.","section":"Table 3"},{"comment":"References [4] and [6] appear to reference the same work by Johnson and Zhang; they should be merged to avoid duplicate citation.","section":"References"},{"comment":"The main text says z2 is distracted 'from a mode', while Appendix C says 'to a mode'; the terminology should be made consistent.","section":"Definition 3.4"},{"comment":"The penalty expressions are inserted as (∥∇D∥−g0)^2, ∥∇D∥^2, and (∥∇D−ε∥)^2 with the expectation and γ/2 omitted, so the quantities compared are not the actual penalties defined in Eqs. (6)–(9); this should be clarified or corrected.","section":"Appendix C.3, Eqs. (C27)–(C29)"}],"recommendation":"reject","confidential_remarks":"The theoretical core of the paper is not sound, and I cannot recommend publication in its current form. The empirical study is reasonably broad, and if the authors can supply a correct derivation or substantially weaken the claims to what the experiments support, a resubmission might be considered."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: the paper's main theorem is not proven, and the proof has a dimensionally invalid step. The ε-centered penalty is a small but reasonable variant of existing gradient penalties, and the empirical tables are wide-ranging; that is the real value here. The theory around 'latent N-size' does not hold up.\n\nWhat is new: the ε-centered gradient penalty (Eq. 9) simply shifts the penalty center, and the authors study its effect on diversity across several architectures. The empirical sweep is broad: MNIST, CIFAR-10, LSUN, ImageNet, plus integration with BigGAN and DDGAN. That is a useful data point, and the claim that penalty center affects diversity is plausible.\n\nThe soft spots are serious. In the proof of Proposition 3.5 (Eq. C19), the generator update is written as G_{\\theta_{t+1}}(z) = G_{\\theta_t}(z) + \\sum \\eta_m \\delta(x)(\\nabla_x D_m(G_{\\theta_t}(z)) + R). Here R is a scalar penalty from Eq. (6)/(8). Adding a scalar to a gradient vector is not defined. The same issue appears in the definition of \\|\\nabla_x D_m(Y)\\| in Proposition 3.5. Second, the proof replaces the actual latent N-size r by an upper bound U in Eq. C22, then orders those upper bounds. Since r is inversely proportional to the infimum of the original expression, ordering U does not imply ordering r. Third, Lemma 3.6 states \\|R_1\\| < \\|R_0\\| < \\|R_ε\\|, but for \\|\\nabla_x D\\| < g_0/2, the 1-centered norm |a - g_0| is larger than a, so the claimed inequality is false in general. The corollary \\nabla_x D \\le 0 is derived from a specific sign convention and does not rescue it. These are load-bearing, not cosmetic, flaws.\n\nThe experiments themselves are suggestive but not convincing: no error bars, no code release, and the authors admit their FID numbers for the CFG baseline differ substantially from the original CFG paper. The diversity improvements are modest and uneven across datasets.\n\nWho this is for: someone interested in a cheap trick to try in GAN training might read the ablation tables. For theory, the paper cannot be used.\n\nRecommendation: desk reject, or major revision with a correct proof and released code. If the authors fix the dimensional issue and the upper-bound argument, the paper could be a modest contribution, but as written it doesn't clear the bar.","headline":"The empirical sweep is useful, but the main theorem is unproven due to a dimensionally invalid generator update and an upper-bound argument that does not order the actual radii.","tokens_in":34670,"tokens_out":3260,"would_cite":false,"duration_ms":31379,"reading_group":"no","serious_thinker":"no","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims GAN output diversity is governed by a latent-space neighborhood radius, and that a new ε-centered gradient penalty shrinks that radius below standard 0- and 1-centered penalties, yielding more diverse images.","keywords":["generative adversarial networks","functional gradient methods","Lipschitz constraint","gradient penalty","epsilon-centered gradient penalty","latent neighborhood size","mode collapse","synthesis diversity"],"falsifier":"Train one architecture under the 1-centered, 0-centered, and $\\epsilon$-centered penalties and measure the latent N-size $r$ directly from the trained generator's input-output map, as Definition 3.1 defines it, rather than from the upper-bound expression in Proposition 3.5; if the measured ordering is not $r_{R_1} > r_{R_0} > r_{R_\\epsilon}$, the theorem's conclusion is refuted. A cheaper check of the premise is to compare training with $R$ inserted inside the generator update (Eq. C19) against training with $R$ used only as a discriminator regularizer: if the two trajectories differ while the theory only describes the perturbed one, the derivation is ungrounded.","tokens_in":33452,"feed_emoji":"🎨","tokens_out":16334,"duration_ms":130563,"temperature":0.7,"pith_summary":"This paper tries to put a theoretical handle on mode collapse in GANs by tying the diversity of generated images to a quantity it calls the latent N-size: the radius of the neighborhood in latent space whose points are pulled toward the same image mode. The claim is that shrinking this radius, by making the discriminator gradient norm $\\|\\nabla_x D\\|$ large, pushes the generator's outputs to spread over more modes, and so produces more diverse samples. To enlarge the gradient norm in a controllable way, the paper introduces an $\\epsilon$-centered gradient penalty that steers $\\|\\nabla_x D\\|$ away from zero by a tunable amount $\\epsilon'$ rather than toward 0 or 1. The central result is the ordering of latent N-sizes $r_{R_1} > r_{R_0} > r_{R_\\epsilon}$, with the $\\epsilon$-centered penalty giving the smallest neighborhood and, the paper argues, the best diversity. A sympathetic reader would care because this offers a parameterized, more stable alternative to composite functional gradient (CFG) GAN training, with reported FID and recall gains on MNIST, CIFAR-10, LSUN, and ImageNet.","feed_headline":"New ε penalty shrinks GAN latent neighborhoods, raising diversity","feed_subtitle":"The ε-shifted penalty tightens each mode's latent neighborhood, spreading generated images across more modes.","key_machinery":"The load-bearing object is the latent N-size, the paper's term for the neighborhood radius $r = \\hat{\\epsilon}\\,\\bigl(2\\inf_z(\\cdots)\\bigr)^{-1}$ of Definition 3.1, whose reciprocal measures how much the generator's output moves between consecutive functional-gradient steps; a small $r$ means nearby latent vectors get pushed toward distinct modes. Proposition 3.5 rewrites that radius in terms of the discriminator gradient norm $\\|\\nabla_x D\\|$ plus a penalty term $R$ inserted into the composite functional gradient (CFG) generator update, and the $\\epsilon$-centered gradient penalty $R_\\epsilon(\\theta,\\psi) = \\frac{\\gamma}{2}\\,\\mathbb{E}\\big[\\|\\nabla_x D_\\psi(\\hat{x}) - \\epsilon\\|^2\\big]$ with $\\|\\epsilon\\| = \\epsilon'$ is the control knob that enlarges the norm. Lemma 3.6 compares the three penalties' contributions to the radius expression, concluding $\\|R_1\\| < \\|R_0\\| < \\|R_\\epsilon\\|$, and Theorem 3.7 turns that into the radius ordering $r_{R_1} > r_{R_0} > r_{R_\\epsilon}$. The direction of the argument relies on the corollary $\\nabla_x D(x) \\le 0$, which makes subtracting $\\epsilon$ the move that increases $\\|\\nabla_x D\\|$.","core_discovery":"The paper's central claim is that the Lipschitz constraint on the discriminator and the diversity of generated samples are linked through the magnitude of the discriminator gradient. Under the CFG update rule, the latent N-size radius $r$ from Definition 3.1 is inversely controlled by $\\|\\nabla_x D\\|$: larger gradient norms shrink the neighborhood that collapses onto a single mode. Because the CFG derivation gives $\\nabla_x D(x) \\le 0$, subtracting a fixed offset vector $\\epsilon$ in the penalty term enlarges the effective norm by $\\|\\epsilon\\|$, and the paper shows through Proposition 3.5, Lemma 3.6, and Theorem 3.7 that the three common penalties order their latent N-sizes as $r_{R_1} > r_{R_0} > r_{R_\\epsilon}$. The theorem is offered as the theoretical foundation for why their Li-CFG method, which adds the $\\epsilon$-centered gradient penalty on top of CFG, generates more diverse images while training more stably than unpenalized CFG.","pith_inferences":["Because Lemma 3.6 compares upper-bound expressions rather than the actual radii $r$, a direct measurement of the neighborhood radius from a trained generator's input-output map would be the real test of the ordering; the bound comparison alone cannot rule out that the actual radii order differently.","The paper's own limitations section concedes that the relationship between CFG's $\\delta(x)$ and the penalty scale $\\epsilon'$ is not analyzed theoretically; a natural next step would be a joint condition on $(\\delta(x), \\epsilon', \\gamma)$ that keeps the latent N-size inside the convergent range.","The empirical tables show recall improving with the $\\epsilon$-centered penalty while precision sometimes drops slightly, suggesting the mechanism trades a little fidelity for broader mode coverage; that trade-off is a testable prediction of the neighborhood picture.","The authors report that their reproduced CFG FID scores run noticeably higher than the numbers in the original CFG paper, so the absolute FID comparisons reported here should be read as internally consistent rather than directly comparable across papers."],"forward_implications":["If the ordering $r_{R_1} > r_{R_0} > r_{R_\\epsilon}$ is correct, the $\\epsilon$-centered penalty yields the smallest latent N-size of the three, and the predicted diversity benefit should reproduce across architectures and datasets.","The hyper-parameter $\\epsilon'$ becomes a diversity dial: the paper's ablation on MNIST and LSUN Bedroom shows $\\epsilon' = 0.3$ giving the best FID and $\\epsilon' = 5$ failing to converge, which it reads as a too-small neighborhood.","Because the penalty is added to the loss rather than to the network, it can be combined with common GAN objectives; the paper reports improved FID and recall for WGAN, LSGAN, HingeGAN, BigGAN, and DDGAN when the $\\epsilon$-centered term is added.","The theory predicts a diversity-trainability trade-off: driving $\\|\\nabla_x D\\|$ too high makes the latent N-size too small, and the generator collapses to noise instead of learning modes.","Stabilizing $\\|\\nabla_x D\\|$ should stabilize FID across the CFG hyper-parameter $\\delta(x)$; the paper shows CFG's FID swings widely with $\\delta(x)$, while Li-CFG's stays flat."],"supporting_citations":[{"why":"Supplies the Composite Functional Gradient (CFG) formulation, including the generator update rule and analytic discriminator solution that Proposition 3.5 extends.","marker":"[4]"},{"why":"The journal version of the CFG framework whose continuous-time dynamics the paper uses to derive the functional gradient $g_m(x) = \\delta(x)\\nabla_x D(x)$.","marker":"[6]"},{"why":"Provides the 1-centered gradient penalty and the line-segment interpolation sampling scheme used by all three penalties in the comparison.","marker":"[3]"},{"why":"Provides the 0-centered gradient penalty and the local-convergence analysis that motivates adding a Lipschitz constraint to CFG.","marker":"[5]"},{"why":"Co-introduces the 0-centered gradient penalty formulation used as one of the two comparison penalties.","marker":"[7]"},{"why":"Supplies the dynamic-theory analysis used in Appendix B to equate CFG's generator updates with common GAN training near equilibrium.","marker":"[16]"},{"why":"Defines the WGAN baseline and the Wasserstein Lipschitz-constraint line that the paper compares against in every benchmark table.","marker":"[2]"},{"why":"The adversarial-latent-space diversity method used as a baseline in the generalization comparison, which the paper distinguishes from its neighborhood theory.","marker":"[27]"},{"why":"The mode-seeking diversity baseline whose objective the paper contrasts with its own latent N-size mechanism.","marker":"[28]"}],"fun_headline_variants":["Epsilon shift in GAN penalty boosts diversity and stability","New ε-penalty tightens GAN latent zones for richer samples","Lipschitz GAN method uses ε penalty to widen sample diversity","ε-centered gradient penalty shrinks GAN mode collapse risks","Li-CFG: ε penalty reduces latent neighborhoods, improves GANs"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the scalar gradient penalty $R$ can be inserted as a vector term inside the generator's update step, as written in Proposition 3.5 and Appendix C.1, Eq.~(C19), $G_{\\theta_{t+1}}(z) = G_{\\theta_t}(z) + \\sum_m \\eta_m \\delta(x)\\big(\\nabla_x D_m(G_{\\theta_t}(z)) + R\\big)$, even though $R$ is a scalar loss term attached to the discriminator objective; if that insertion is not valid, the latent N-size formula and the ordering $r_{R_1} > r_{R_0} > r_{R_\\epsilon}$ do not follow.","fun_headline_variants_meta":{"raw":{"variants":["Epsilon shift in GAN penalty boosts diversity and stability","New ε-penalty tightens GAN latent zones for richer samples","Lipschitz GAN method uses ε penalty to widen sample diversity","ε-centered gradient penalty shrinks GAN mode collapse risks","Li-CFG: ε penalty reduces latent neighborhoods, improves GANs"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000214,"raw_usage":{"total_tokens":1453,"prompt_tokens":999,"completion_tokens":454,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":615,"completion_tokens_details":{"reasoning_tokens":364}},"tokens_in":615,"tokens_out":454,"duration_ms":4002,"temperature":1.0,"reasoning_tokens":364,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T18:30:07.309378+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train one architecture under the 1-centered, 0-centered, and $\\epsilon$-centered penalties and measure the latent N-size $r$ directly from the trained generator's input-output map, as Definition 3.1 defines it, rather than from the upper-bound expression in Proposition 3.5; if the measured ordering is not $r_{R_1} > r_{R_0} > r_{R_\\epsilon}$, the theorem's conclusion is refuted. A cheaper check of the premise is to compare training with $R$ inserted inside the generator update (Eq. C19) against training with $R$ used only as a discriminator regularizer: if the two trajectories differ while the theory only describes the perturbed one, the derivation is ungrounded.","supporting_citations":[{"cited_title":"In: International Conference on Machine Learning, pp","cited_arxiv_id":null,"evidence_quote":"Supplies the Composite Functional Gradient (CFG) formulation, including the generator update rule and analytic discriminator solution that Proposition 3.5 extends."},{"cited_title":"IEEE Transactions on Pattern Analysis and Machine Intelligence 43(1), 17–32 (2019)","cited_arxiv_id":null,"evidence_quote":"The journal version of the CFG framework whose continuous-time dynamics the paper uses to derive the functional gradient $g_m(x) = \\delta(x)\\nabla_x D(x)$."},{"cited_title":"3481– 3490 (2018)","cited_arxiv_id":null,"evidence_quote":"Provides the 0-centered gradient penalty and the local-convergence analysis that motivates adding a Lipschitz constraint to CFG."},{"cited_title":"In: International Conference on Machine Learning, pp","cited_arxiv_id":null,"evidence_quote":"Defines the WGAN baseline and the Wasserstein Lipschitz-constraint line that the paper compares against in every benchmark table."},{"cited_title":"Conference on Neural Information Processing Systems 35, 8868–8881 (2022)","cited_arxiv_id":null,"evidence_quote":"The adversarial-latent-space diversity method used as a baseline in the generalization comparison, which the paper distinguishes from its neighborhood theory."},{"cited_title":"In: IEEE/CVF Computer Vision and Pattern Recognition Conference, pp","cited_arxiv_id":null,"evidence_quote":"The mode-seeking diversity baseline whose objective the paper contrasts with its own latent N-size mechanism."}],"review_version":1}