{"id":"78ecbc5d-1b9d-4fee-8c43-78d03ec26e58","arxiv_id":"2411.10368","paper_version":1,"verdict":"REJECT","confidence":"HIGH","novelty_score":3.0,"correctness_risk":"high","formal_verification":"none","parameter_count":0,"one_line_summary":"Adversarial training alone can mimic an autoencoder under strong capacity assumptions, which the paper proposes as the mechanism behind GAN-only image-to-image translation.","lead":"This paper claims that a bare GAN, without extra cycle or identity losses, can perform image-to-image translation because adversarial training behaves like an autoencoder when the generator and discriminator are both strong enough. The authors offer this as a mechanism for why simple GAN-only models work, and they show qualitative translation examples for faces, animals, and artwork.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Even granting perfect discrimination, Eq 6 is a distribution-matching objective, not a per-sample reconstruction loss; a dataset-permuting generator attains zero adversarial loss with large reconstruction error, so the proof's conclusion does not follow.","rationale":"The reader correctly identifies the perfect per-feature discrimination assumption in the transition from Eq 2 to Eq 3. My concern is more fundamental: even if that assumption is granted, the subsequent move to Eq 6 converts a paired sample-wise loss into an unpaired marginal distribution divergence. The conclusion 'G(x) will be pushed to x' would require the loss to be minimized pointwise for the same x, but adversarial training only minimizes an expectation over the generator's output distribution. This is not an exotic failure: a permutation of the dataset is a measure-preserving map, so any GAN loss based solely on distribution matching (WGAN, standard GAN, etc.) cannot distinguish identity from permutation. Hence the central theoretical claim collapses regardless of discriminator capacity. The paper's experiments actually show a persistent gap between GAN and autoencoder reconstruction losses, and the text admits the GAN cannot make outputs identical to inputs without an explicit reconstruction loss. That admission is consistent with my analysis. Because the proof's core implication is invalid and the experiments do not close the gap, rejection remains appropriate. My concern is distinct from but compatible with the reader's; I partially agree because perfect discrimination is not the fundamental blocker—the marginal expectation is.","tokens_in":10098,"tokens_out":4654,"duration_ms":47220,"concrete_test":"Analytic test: on a finite dataset X={x1,...,xm} with uniform pdata, consider any bijection π on X that is not the identity, and set G(x)=π(x). Compute Eq 6: since π is measure-preserving, E_{x∼pdata}[RD(π(x))]=E_{x∼pdata}[RD(x)] for every RD, so L=0. If this computation is verified, the adversarial objective has a global minimizer with maximal reconstruction error, invalidating the claim. Experimental complement: train a minimal GAN on a two-image dataset with a generator initialized to swap the images; if adversarial loss converges to the same value as identity while reconstruction error remains large, the paper's conclusion fails.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The proof in Section III.A.1 fails at the step 'From (2), we know that G(x) will be push to x when minimizing the loss function.' Equation (6) is L = E_{x∼pdata}[RD(x)] − E_{x∼pg}[RD(x)], a distribution-matching objective over the marginal pg. Even if D is perfectly able to distinguish every feature discrepancy—the reader's identified assumption—this objective does not compare G(x) with its own input x. It only compares the distribution of all generated outputs with the distribution of all real inputs. Consequently, any generator that preserves the marginal distribution exactly achieves the same loss as the identity map. For example, on a two-image dataset {A,B} with uniform probabilities, the swap map G(A)=B, G(B)=A yields pg=pdata and hence zero adversarial loss for any D, yet ||A−G(A)||>0. Thus Eq 6 cannot imply per-sample reconstruction; the central claim that 'adversarial training for autoencoder models yields results similar to traditional autoencoder models' is not supported by the derivation. Notably, the paper's own experimental text (Section IV.A, Fig. 6) concedes 'noticeable differences between the input and output images' and that the GAN 'cannot make them identical without incorporating a reconstruction loss,' which is consistent with this concern.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper argues that an image-to-image translation network using only an adversarial loss can behave like an autoencoder under two conditions: the generator can reconstruct the input, and the discriminator can perfectly distinguish real from generated images on every feature. The authors derive an algebraic equivalence between adversarial training and autoencoder reconstruction in Section III.A.1, give a geometric interpretation, and present qualitative experiments on face, animal, and art datasets. The central claim is that adversarial training alone implicitly enforces autoencoder-like per-sample reconstruction, and that the discriminator's role in translation is to preserve features common to both source and target datasets while replacing features unique to the source.","tokens_in":10400,"tokens_out":3025,"duration_ms":30071,"significance":"If the central claim were correct, the paper would offer a useful simplification of image-to-image translation: a bare GAN loss that preserves identity features without reconstruction or cycle losses, together with an explanation of what content and style mean in a dataset-driven sense. The paper does have some strengths: it clearly identifies the conditions it believes are sufficient, it is explicit about its constraints in Section IV.C, and it acknowledges in Section IV.A that the GAN cannot make images identical to the input without an explicit reconstruction loss. However, the theoretical derivation contains a load-bearing logical gap, the final loss in Eq. (6) is a distribution-matching objective rather than a per-sample reconstruction objective, and the experimental evaluation is qualitative only. As a result, the claimed equivalence between adversarial training and autoencoders is not established, and the contribution is substantially weakened. The paper does not ship machine-checked proofs, reproducible code, or quantitative comparisons; its value currently rests on a plausible but unproven intuition.","major_comments":[{"comment":"The transition from Eq. (2) to Eq. (3) is not valid without an additional, unstated assumption. Eq. (2) is a sum of L1 norms, so it equals sum_i |D_i(x) - D_i(G(x))|. To rewrite it as sum_i [D_i(x) - D_i(G(x))], one must assume D_i(x) > D_i(G(x)) for every i and every x. The paper asserts that perfect discrimination implies this ordering, but perfect discrimination only implies that the two values differ, not which is larger. A discriminator can perfectly separate the sets while having D_i(x) < D_i(G(x)) for some components, or while the sign varies across the sample. Without the sign condition, the derivation collapses, and the conclusion that minimizing the loss pushes G(x) toward x does not follow.","section":"Section III.A.1, Eqs. (2)-(3)"},{"comment":"Even granting the sign assumption, the final loss in Eq. (6) is a distribution-matching objective over the marginals pdata and pg, not a per-sample comparison of G(x) with its own input x. The statement \"From (2), we know that G(x) will be pushed to x when minimizing the loss function\" is therefore unsupported. A generator that computes the identity mapping and a generator that permutes the dataset (e.g., G(A)=B and G(B)=A on a two-image uniform dataset) produce exactly the same value of the expectation in Eq. (6), since pg = pdata in both cases, yet the reconstruction error ||x - G(x)|| differs dramatically. Thus Eq. (6) cannot imply per-sample reconstruction, and the paper's central claim that adversarial training yields autoencoder-like results is not established by the derivation.","section":"Section III.A.1, Eqs. (4)-(6) and the statement after Eq. (6)"},{"comment":"The experimental section does not substantiate the central claim. The paper reports only qualitative samples and a single reconstruction-loss curve; there is no quantitative comparison with an autoencoder on held-out data, no measure of perceptual similarity, and no statistical test. Moreover, the text accompanying Fig. 6 concedes that \"the GAN is capable of bringing G(x) close to x, but it cannot make them identical without incorporating a reconstruction loss.\" This admission is consistent with the theoretical gap: the marginal distribution-matching loss does not force identity. The reader is asked to accept similarity from a few favorable examples, which is insufficient evidence for the strong equivalence claim.","section":"Section IV.A and Fig. 6"},{"comment":"The key condition stated in Section IV.C—that the discriminator must be able to perfectly distinguish real from fake images whenever there is a discrepancy—is effectively the conclusion of the theorem, not a supporting assumption. If a discriminator can detect any discrepancy between G(x) and x, then the only way for G to fool it is to reproduce x exactly; this is a restatement of the desired result rather than a proof of it. The geometric interpretation in Section III.A.2 similarly assumes that alternating training brings each paired G(x) toward its own x, but the red arrows show G(x)'s moving toward the boundary in feature space, which does not imply each G(x) moves toward its own paired x rather than toward some other real point. The argument is therefore circular in a way that a referee cannot overlook.","section":"Section IV.C and Section III.A.2"}],"minor_comments":[{"comment":"The abstract and introduction contain grammatical issues that obscure the meaning, e.g., \"We show that adversarial for GAN models yields results comparable to those of existing methods\" and \"We investigate the relationship between GANs and autoencoders and provide an explanation for the efficacy of employing only the GAN component.\" These sentences should be rewritten for clarity.","section":"Abstract and Section I"},{"comment":"The notation in Eq. (3) is ambiguous: it writes a double sum but the outer sum is over a set I while the inner sum uses index i, and the paper alternates between using i as a sample index and as a feature index. Please define indices explicitly.","section":"Section III.A.1, Eq. (3)"},{"comment":"The phrase \"if the capacity of the network is not insufficient\" should read \"if the capacity of the network is insufficient.\" Also, \"make the information more campact\" is a typo for \"compact.\"","section":"Section IV.C.2"},{"comment":"The paper claims \"we used the mean square error between original image and generated image to evaluate the performance,\" but Figure 4's caption and the text describe reconstruction loss curves without reporting the actual values or a comparison table. Please provide numerical results and standard deviations.","section":"Section IV"},{"comment":"The related work section would benefit from a discussion of conditional GANs for image-to-image translation (e.g., Isola et al. [15] is listed but not discussed in the body), and from explicit statements about how the proposed method differs from a standard unpaired translation approach such as CycleGAN.","section":"References"}],"recommendation":"reject","confidential_remarks":"The paper's core theoretical claim is not supported because the derivation in Section III.A.1 contains a sign-assumption error and then reduces to a distribution-matching objective that does not enforce per-sample reconstruction. The experiments are qualitative and even concede the point. This is not a case where a local fix would suffice; a new theoretical framework and a properly controlled quantitative evaluation would be needed. I would not encourage resubmission without a fundamentally revised argument. The paper does honestly state its limitations, but honesty about the gap does not close it."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know up front. The paper is honest and its architecture is clean, but the central theoretical claim doesn't hold. The derivation in Section III.A.1 converts a per-sample reconstruction objective into a distribution-matching objective (Eq. 6), and that objective does not push G(x) toward x. A generator that permutes the dataset—swap A and B on a two-image set—gets zero adversarial loss while failing at reconstruction. That is not a pedantic edge case; it shows the proof's conclusion doesn't follow from its own equations.\n\nWhat's genuinely useful: the paper demonstrates that a GAN trained with only an adversarial loss can do some image-to-image translation, and the geometric story (global features first, then local) is a reasonable heuristic. The authors also state their two assumptions clearly and test them by varying dataset size and bottleneck width. The observation that smaller datasets make the discriminator memorize and push toward one-to-one mapping is worth noting. And they honestly concede in Section IV.A that the GAN cannot make outputs identical without a reconstruction loss—a concession that undercuts the abstract's stronger phrasing.\n\nThe soft spots are load-bearing. The step from Eq. 2 to Eq. 3 requires the unproved assumption that D_i(x) > D_i(G(x)) for every feature whenever they differ; even granting that, Eq. 6 only matches marginals, not per-sample pairs. The experiments are qualitative: no comparison against CycleGAN, no quantitative metrics beyond reconstruction loss on one dataset, no code or data. The novelty is also thinner than claimed, since reference [4] from the same group already applied a GAN-only loss to translation; what's new here is supposed to be the explanation, and that explanation isn't a proof.\n\nThis paper is for readers who want a suggestive hypothesis about why GAN-only losses work, not for readers who need a rigorous mechanism. If the authors reframe the contribution as a conjecture and back it with proper baselines and quantitative evaluation, it could be a workshop-level contribution. As it stands, the central claim is unsupported, and I wouldn't send it to referees.","headline":"The proof that a GAN-only loss reconstructs inputs fails at Eq. 6, and the experiments are qualitative, so the central claim is unsupported despite an honest writeup.","tokens_in":10863,"tokens_out":3036,"would_cite":false,"duration_ms":29351,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims that a GAN trained only with an adversarial loss can perform image-to-image translation because, when the discriminator is powerful enough, adversarial training becomes equivalent to autoencoder reconstruction.","keywords":["generative adversarial networks","image-to-image translation","autoencoder","adversarial loss","content and style decomposition","reconstruction without explicit loss","WGAN-like objective"],"falsifier":"Train the same pure-GAN model on a dataset whose images have two separable feature groups, one where the discriminator is deliberately weakened, for instance features carried only in high-frequency detail. If reconstruction error on the weak-feature group stays high while the easy group converges, the per-feature perfect-discrimination assumption is false and the claimed autoencoder equivalence fails.","tokens_in":9898,"feed_emoji":"🖼️","tokens_out":7726,"duration_ms":61815,"temperature":0.7,"pith_summary":"The paper tries to establish that a GAN using only its adversarial loss—no cycle loss, identity loss, or feature penalties—can perform image-to-image translation because adversarial training secretly behaves like an autoencoder. The argument is that a discriminator strong enough to detect any difference between a real image and its reconstruction lets the min-max loss be rewritten as a sum over per-feature differences, whose minimum pushes the generator output onto the input. This matters because it gives a mechanism for why simple GANs work, and it reinterprets 'content' and 'style' as features common to both datasets versus features present only in the target dataset. The paper supports the claim with experiments showing that GAN and autoencoder reconstruction losses track each other, and that pure adversarial translation preserves pose and global topology while retexturing.","feed_headline":"Pure adversarial loss matches autoencoder behavior","feed_subtitle":"A discriminator that catches every difference pushes generated images toward inputs, making extra losses optional.","key_machinery":"The load-bearing object is the min-max objective $\\min_G \\max_D \\frac{1}{m}\\sum_{x\\in I}\\|D(x)-D(G(x))\\|$, where $D$ maps images to a real-valued feature vector rather than to a binary label. Under the assumption that $D$ can perfectly distinguish $x$ and $G(x)$ on every feature, the loss separates into a sum over per-feature differences and then into a difference of expectations, $\\mathbb{E}_{x\\sim p_{\\text{data}}}[\\hat D(x)] - \\mathbb{E}_{x\\sim p_g}[\\hat D(G(x))]$ with $\\hat D(x)=\\sum_i D_i(x)$, a WGAN-like form. Minimizing this expectation pushes $G(x)$ toward $x$ without paired data, which is the mechanism that makes a pure GAN behave like an autoencoder.","core_discovery":"With a discriminator that can perfectly separate real from generated images feature by feature, adversarial training minimizes the same kind of difference as an autoencoder's reconstruction loss. Consequently, a GAN trained only with the adversarial loss makes G(x) approach x when the source and target datasets are the same; when the datasets differ, it preserves features common to both datasets and replaces features found only in the target dataset. The paper identifies the common features as 'content' and the target-only features as 'style', giving an operational, dataset-relative definition of this distinction.","pith_inferences":["A direct testable consequence the paper does not run: if you reduce the discriminator's sensitivity to one feature group, for example by blurring high frequencies in its input, reconstruction of that group should degrade while other groups stay intact.","The same reasoning suggests why auxiliary losses such as cycle consistency help in practice: they compensate for imperfect discriminators, and when perfect per-feature separation fails, extra reconstruction-style constraints fill the gap.","The framework yields an operational definition of style and content measurable from data alone, so style losses could in principle be derived from dataset statistics rather than chosen by hand."],"forward_implications":["When the source and target datasets coincide, a GAN trained with only the adversarial loss should reconstruct its input, matching autoencoder behavior without any explicit reconstruction term.","For different source and target datasets, the same single-loss GAN should preserve features common to both datasets and replace features present only in the target dataset.","Smaller datasets make the discriminator's perfect-distinction condition easier to satisfy, so pure-GAN translation should converge closer to a one-to-one mapping on small datasets.","Insufficient encoder capacity breaks the reconstruction condition, and the network then keeps only coarse pose and angle while losing fine identity details.","The content/style split in style transfer can be identified with common versus dataset-specific features rather than with semantic categories."],"supporting_citations":[{"why":"Supplies the GAN training setup and the adversarial min-max game that the paper strips down to a single GAN loss.","marker":"[11]"},{"why":"Defines the autoencoder/VAE reconstruction objective that the paper compares with adversarial training.","marker":"[12]"},{"why":"The WGAN loss whose per-example expectation form the paper's derived objective resembles, linking adversarial loss to reconstruction push.","marker":"[36]"},{"why":"Provides the StyleGAN-v2 generator and discriminator architecture used in all experiments.","marker":"[18]"},{"why":"Earlier pure-GAN food-image translation model that motivates the claim that extra penalties are unnecessary.","marker":"[4]"},{"why":"CycleGAN is the main comparison model and source of photo-to-painting datasets, representing the cycle-loss complexity the paper argues is unnecessary.","marker":"[6]"},{"why":"The style/content separation approach the paper reinterprets in terms of common versus target-only features.","marker":"[21]"}],"fun_headline_variants":["Pure adversarial loss matches autoencoder behavior","GAN translation works with only adversarial loss","No extra losses needed: GANs explained","Adversarial loss alone drives image translation"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The entire argument rests on the discriminator being able to perfectly tell a generated image from a real one on every individual feature whenever the two differ at all.","fun_headline_variants_meta":{"raw":{"variants":["Pure adversarial loss matches autoencoder behavior","GAN translation works with only adversarial loss","No extra losses needed: GANs explained","Adversarial loss alone drives image translation"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000502,"raw_usage":{"total_tokens":2352,"prompt_tokens":741,"completion_tokens":1611,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":357,"completion_tokens_details":{"reasoning_tokens":1557}},"tokens_in":357,"tokens_out":1611,"duration_ms":12795,"temperature":1.0,"reasoning_tokens":1557,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T19:41:02.073221+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train the same pure-GAN model on a dataset whose images have two separable feature groups, one where the discriminator is deliberately weakened, for instance features carried only in high-frequency detail. If reconstruction error on the weak-feature group stays high while the easy group converges, the per-feature perfect-discrimination assumption is false and the claimed autoencoder equivalence fails.","supporting_citations":[{"cited_title":"Generative adversarial nets,","cited_arxiv_id":null,"evidence_quote":"Supplies the GAN training setup and the adversarial min-max game that the paper strips down to a single GAN loss."},{"cited_title":"Auto-encoding variational bayes,","cited_arxiv_id":null,"evidence_quote":"Defines the autoencoder/VAE reconstruction objective that the paper compares with adversarial training."},{"cited_title":"Wasserstein generative adver- sarial networks,","cited_arxiv_id":null,"evidence_quote":"The WGAN loss whose per-example expectation form the paper's derived objective resembles, linking adversarial loss to reconstruction push."},{"cited_title":"A style-based generator architecture for generative adversarial networks,","cited_arxiv_id":null,"evidence_quote":"Provides the StyleGAN-v2 generator and discriminator architecture used in all experiments."},{"cited_title":"Shape-preserving generation of food images for automatic dietary assessment,","cited_arxiv_id":null,"evidence_quote":"Earlier pure-GAN food-image translation model that motivates the claim that extra penalties are unnecessary."},{"cited_title":"Unpaired image-to-image translation using cycle-consistent adversarial networks,","cited_arxiv_id":null,"evidence_quote":"CycleGAN is the main comparison model and source of photo-to-painting datasets, representing the cycle-loss complexity the paper argues is unnecessary."}],"review_version":1}