{"id":"7ac406fc-239f-4810-ae27-3bbc5d2193de","arxiv_id":"1908.07269","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"RelGAN conditions a GAN on relative attribute vectors, the difference between target and source attributes, to achieve continuous facial attribute editing with better preservation of unchanged attributes than StarGAN and AttGAN.","lead":"This paper proposes RelGAN, an image-to-image translation method that takes the desired change in facial attributes as input rather than the full target attributes, enabling continuous editing while preserving unaffected traits. It reports better visual quality and smoother interpolation than StarGAN and AttGAN on facial attribute tasks.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The conditioning vector v = â - a is only interpretable if the generator can recover a from x, but a is never supplied and no experiment isolates this implicit inference; the central edit guarantee is therefore not established.","rationale":"I agree with the reader's weakest_assumption and make it the primary stress-test. The central claim has two parts—reliable attribute editing and preservation of unedited attributes—and both depend on interpreting v relative to the source state. Since a is not an input, the generator must estimate it, and the only supervision for that estimation is indirect (the matching discriminator). The paper's Table 2 and FID results are consistent with the model having learned a good implicit estimate, but they do not isolate its reliability: no error bars, no stratification by source-attribute ambiguity, and no test of out-of-distribution v. The reconstruction losses do not close the gap, because cycle consistency is compatible with spurious changes in zero-delta attributes as long as -v reverses them. The interpolation-metric issue (standard deviation of SSIM rewards a no-op) and the missing GANimation baseline are real secondary weaknesses in the evaluation of smoothness, but they do not attack the core mechanism as directly as the source-inference problem. The concrete test would settle whether the implicit inference is actually a bottleneck; until that is run, the correct disposition is the same conditional verdict the reader gave. No change to the verdict is needed.","tokens_in":12419,"tokens_out":12985,"duration_ms":143269,"concrete_test":"Using the authors' released code (or a retraining with the paper's stated hyperparameters), take the CelebA-HQ test set and single-attribute edit directions v derived from ground-truth labels. For each output G(x,v), compute per-attribute target accuracy with the same ResNet-18 classifier as in Section 4.4, checking whether the predicted attribute equals a+v rather than reporting only the Table 2 average. Stratify these per-attribute accuracies by the classifier's confidence on the relevant source attributes of x (e.g., predicted probability in [0.3,0.7] counts as low confidence). If target accuracy is markedly lower on low-confidence source images than on high-confidence ones, the model's success depends on an implicit source-attribute inference that is not robust; if the gap is small, the concern does not land.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 3.1 defines v = â - a, and the generator input is (x, v) (Figure 2, Section 4.2); neither G nor DMatch ever receives a. For the central claim of 'changing particular attributes of interest while preserving the other attributes' to hold, G must infer the source attribute vector a from x alone, because the same v corresponds to different target attributes for different source states: notably v_i = 0 means 'keep smile present' when a_i = 1 but 'keep smile absent' when a_i = 0. This inference is never supervised directly; it is only learned indirectly through DMatch's real/wrong-triplet loss. The experiments build v from ground-truth labels (Table 2), which is a correct oracle for the intended target but does not measure whether G's implicit source estimate is reliable, nor whether the learned mapping generalizes to v vectors that are inconsistent with the source (e.g., v_i = -1 on an image that already lacks attribute i). If source inference is wrong, the output silently realizes a different edit than the one requested, so the central claim fails. This is the least secure link in the argument: the reconstruction losses (Eqs. 4 and 5) ensure G(x,0)=x and round-trip invertibility, but neither prevents an edit from changing zero-delta attributes as long as applying -v reverses them in the cycle.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes RelGAN, a GAN-based method for multi-domain facial attribute editing that conditions the generator on a relative attribute vector v = â − a instead of the target attribute vector â. The framework consists of one generator and three discriminators: a realism discriminator, a matching-aware discriminator that classifies real, fake, and wrong triplets (x, v, x′), and an interpolation discriminator that estimates the interpolation coefficient. The paper evaluates attribute transfer, reconstruction, interpolation, and user preference against StarGAN and AttGAN, reporting lower FID, higher classification accuracy, better reconstruction, and lower SSIM-standard-deviation interpolation scores.","tokens_in":12733,"tokens_out":8854,"duration_ms":89734,"significance":"The relative-attribute reparameterization is a genuine interface improvement for attribute editing, and the matching-aware discriminator with its four wrong-triplet types is a correct and clearly specified design. The cycle and self reconstruction losses are principled, and the cross-dataset generalization experiment (CelebA-HQ to FFHQ) is a strong point. If the central mechanism—the generator's implicit inference of the source attribute vector—is confirmed, the method would be a solid contribution to multi-domain image translation. The current evidence does not yet close that gap, and the interpolation metric is degenerate, which tempers the novelty claim.","major_comments":[{"comment":"The formulation defines v = â − a, but the generator receives only (x, v) and never the source attribute vector a; therefore G must infer a from x in order for v to be interpretable. This inference is load-bearing because a zero entry v_i must mean \"keep the attribute present\" when x has the attribute and \"keep the attribute absent\" when it does not, and the same ambiguity applies to nonzero entries (e.g., v_i = −1 is only meaningful if a_i = 1). The experiments in Table 2 construct v from ground-truth labels, so they provide an oracle for the intended target but do not measure whether G's implicit estimate of a is reliable, nor do they probe generalization to v vectors that are inconsistent with the source image. The reconstruction losses (Eqs. 4 and 5) guarantee G(x, 0) ≈ x and G(G(x, v), −v) ≈ x, but they do not prevent zero-delta attributes from being altered during a nonzero edit. I recommend an explicit diagnostic: measure the classifier-confirmed attribute change for zero-delta entries as a function of the source attribute value, and test with intentionally wrong source attribute vectors to reveal what G actually does.","section":"Section 3.1, Eq. (1); Section 4.4, Table 2"},{"comment":"The interpolation metric, the standard deviation of SSIM between consecutive frames, is degenerate: a model that returns the same image for every α (a no-op generator) achieves σ = 0, the best possible score, because all consecutive SSIM values are identically 1. Consequently, Table 5 does not by itself support the claim that RelGAN improves interpolation quality. In addition, the interpolation discriminator (Section 3.5) is explicitly trained to make interpolated frames look like the endpoints, so the reported metric partly measures the training objective rather than an independent property. The paper should supplement Eq. (11) with a measure that the edited attribute actually moves monotonically along the interpolation path (e.g., classifier output for the target attribute as a function of α), and it should include a no-op baseline in the interpolation comparison.","section":"Section 4.6, Eq. (11)"},{"comment":"The empirical comparison is restricted to binary-attribute, target-attribute methods (StarGAN and AttGAN). The conclusion's claim of \"superior performance over the state-of-the-art methods\" is therefore broader than the evidence, especially because GANimation, which the paper cites and which is designed for continuous facial expression interpolation, is not included in any comparison. A comparison on the overlapping smile/expression task, or a suitably scoped claim about binary target-attribute baselines, would be needed to support the state-of-the-art statement.","section":"Section 4.3, Baselines; Section 5, Conclusion"}],"minor_comments":[{"comment":"The text says \"the first term aims at recovering ˆa from G(x, α v)\"; the symbol should be ˆα (the interpolation coefficient), since DInterp outputs a scalar degree of interpolation and not an attribute vector.","section":"Section 3.5, after Eq. (6)"},{"comment":"The baseline interpolation formula \"G(x, α a + (1−α)â)\" appears to have a and â reversed if α = 0 is meant to correspond to the input image in Figure 5; please check and correct the ordering.","section":"Section 4.6"},{"comment":"FID and classification results are reported as point estimates from presumably a single run; the paper should report means and standard deviations over multiple training runs, or at least state how many runs were used.","section":"Section 4.4, Tables 1 and 2"},{"comment":"The ablation study is presented only with qualitative images; adding quantitative metrics (e.g., FID and classification accuracy for each ablation row) would make the contribution of each loss term more convincing.","section":"Table 3"},{"comment":"The supplementary interpolation figures are visually informative, but no quantitative measure of attribute monotonicity along α is provided; such a measure would directly address the central claim of fine-grained continuous control.","section":"Section 4.6, Figure 14-17"}],"recommendation":"major_revision","confidential_remarks":"The paper is within the scope of the journal. The main risk is the unmeasured source-attribute inference mechanism, which is the weakest link in the central claim; the authors should be pushed to provide the diagnostic rather than just more qualitative images. The interpolation metric also needs to be replaced or augmented, since a no-op model trivially optimizes it. If the authors address these two points, the paper could be a solid contribution."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The one thing to know: this is a solid, incremental GAN paper whose key idea - condition on the change vector v = a_hat - a instead of the full target attribute vector - is simple and works. The matching-aware discriminator with four wrong-triplet types is cleanly constructed, and the interpolation discriminator is a reasonable add-on. The FID and classification gains over StarGAN and AttGAN are consistent, and the user study supports the qualitative claim that RelGAN preserves unchanged attributes better than the target-attribute baselines.\n\nThe soft spots are mostly in the evaluation. The interpolation metric, standard deviation of SSIM across adjacent frames, would be perfectly scored by a no-op model: if all outputs are identical, every SSIM is 1 and the std is 0. That is a real flaw in the quantitative interpolation claim, and the qualitative figures can't fully carry that weight. Second, there are no error bars anywhere, so the FID differences - which look large - could be smaller than run-to-run variance. Third, GANimation is cited but never compared, even though it is the obvious baseline for continuous attribute interpolation. Fourth, the classification accuracy on generated images is not broken down by source attribute state, so it doesn't directly verify the requested change magnitude was achieved, only that the output is classified as having the target attribute.\n\nOn the stress-test concern: the generator never receives the source attribute vector, so it must infer the attribute state from the image alone. The paper doesn't isolate this inference, and the model isn't tested on inconsistent v vectors (e.g., v_i = -1 on an image that already lacks attribute i). But I wouldn't call this fatal. The self-reconstruction loss and the strong preservation results in the hair-color transfer figures suggest the model learns something about source state. It's a missing robustness experiment, not a demonstrated failure of the central claim.\n\nThis paper is for anyone working on multi-domain image translation or facial attribute editing. It deserves a serious referee - the relative-attribute formulation is a legitimately new conditioning scheme that others will build on - but I would send it back for revision to fix the interpolation metric, add error bars, and include GANimation. The central idea is sound.","headline":"RelGAN's relative-attribute conditioning is a genuinely useful twist; the core formulation holds up, but the evaluation has soft spots that are repairable, not fatal.","tokens_in":666,"tokens_out":1407,"would_cite":true,"duration_ms":43630,"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":"The paper claims that conditioning facial image translation on relative attribute vectors—the difference between desired and current attributes—enables continuous editing of only selected attributes.","keywords":["image-to-image translation","relative attributes","facial attribute editing","generative adversarial networks","attribute interpolation","multi-domain translation","CelebA-HQ"],"falsifier":"Run the trained model on an input whose source attributes are intentionally misclassified by a standard attribute classifier, such as an androgynous face for gender, and pass a non-zero relative attribute for that label; if the output does not move in the intended direction, the premise that $G$ can interpret $v$ without the source vector fails.","tokens_in":12235,"feed_emoji":"🎨","tokens_out":6203,"duration_ms":59757,"temperature":0.7,"pith_summary":"RelGAN claims that multi-domain image translation can be driven by relative attributes, the difference $v = \\hat{a} - a$ between the target attribute vector and the source attribute vector, rather than by target attributes alone. The paper argues this frees the user from knowing the current attribute values of an input image and from specifying attributes that should not change, since zero entries in $v$ mark the untouched attributes. It further claims that scaling $v$ by $\\alpha$ produces smooth, realistic interpolation, giving fine-grained control over attribute strength. On facial attribute transfer and interpolation, the paper reports lower FID scores, higher attribute-classification accuracy on generated images, lower standard deviation of SSIM across interpolation steps, and majority user preference over target-attribute baselines.","feed_headline":"RelGAN edits faces by desired change, not target state","feed_subtitle":"Change vectors, not target labels, let the model keep untouched traits and interpolate smoothly.","key_machinery":"The load-bearing object is the relative attribute vector $v = \\hat{a} - a$. It turns a user's editing intent into a compact conditioning signal whose zero entries encode do-not-touch and whose magnitude encodes how much to change. Three discriminators carry the training: $D_{\\text{Real}}$ enforces realism; $D_{\\text{Match}}$, inspired by matching-aware discriminators, takes triplets and distinguishes matched real pairs from fake or wrong pairs; $D_{\\text{Interp}}$ regresses the interpolation coefficient $\\alpha$ on $G(x, \\alpha v)$, pushing interpolated frames to look like non-interpolated outputs. Cycle and self-reconstruction losses tie the generator to identity preservation.","core_discovery":"The central claim is that the translation function $G(x, v)$, conditioned on the desired change, can outperform the target-conditioned $G(x, \\hat{a})$ at preserving untouched attributes while enabling continuous control. The paper introduces $v = \\hat{a} - a$ and trains $G$ with an adversarial loss for realism, a matching-aware discriminator that classifies triplets $(x, v, x')$ as matched, mismatched, or fake, and an interpolation discriminator that predicts the interpolation degree $\\alpha$ from $G(x, \\alpha v)$. With cycle and self-reconstruction losses, the resulting model is claimed to change only the requested attributes and to interpolate smoothly. The evidence includes quantitative comparisons on CelebA, CelebA-HQ, and FFHQ and a large user study.","pith_inferences":["Beyond the paper, if the relative-attribute formulation is sound, it should generalize to any multi-domain translation task with annotated attributes, such as clothing, scene, or object attributes, since nothing in the design is face-specific.","The paper leaves implicit that the method's ceiling is set by source-attribute inference; a natural extension is to provide an explicit source-attribute estimate or a classifier-consistency loss, which would likely make the matching discriminator's job easier.","The interpolation discriminator's alpha-regression objective could be reused as an unsupervised disentanglement regularizer: any latent path whose interpolations fool the regressor is declared semantically meaningful.","The wrong-triplet construction suggests a testable extension: use continuous attribute regressors instead of binary labels, allowing $v$ to take graded values in training and thereby improving fine-grained control further."],"forward_implications":["A user can edit any subset of attributes without supplying the remaining attribute values; zero entries in $v$ suffice.","Interpolation becomes a scalar control: $G(x, \\alpha v)$ sweeps from the original to the fully edited image as $\\alpha$ goes from 0 to 1.","Because training uses real-valued relative attributes, the model can express and realize graded changes such as partial hair-color transfer or smile strength.","The trained model transfers to a new dataset, FFHQ, without retraining and still maintains its editing behavior, as measured by FID."],"supporting_citations":[{"why":"provides the single-generator multi-domain baseline and the generator backbone RelGAN adapts.","marker":"[7]"},{"why":"serves as the target-attribute baseline in FID, classification, interpolation, and user-study comparisons.","marker":"[8]"},{"why":"supplies the cycle-consistency idea used in the cycle-reconstruction loss.","marker":"[5]"},{"why":"introduces the matching-aware discriminator concept that $D_{\\text{Match}}$ adapts to triplets.","marker":"[10]"},{"why":"inspires the interpolation discriminator's adversarial regularizer for intermediate images.","marker":"[16]"},{"why":"provides the CelebA dataset with 40 binary attribute labels used for training and evaluation.","marker":"[18]"},{"why":"supplies the CelebA-HQ dataset used for training and most evaluations.","marker":"[19]"},{"why":"supplies FFHQ, the held-out dataset for testing generalization.","marker":"[20]"},{"why":"defines FID, the visual-quality metric used for quantitative comparison.","marker":"[24]"}],"fun_headline_variants":["RelGAN: change attributes, not targets","Face editing by relative change, not target labels","Continuous face edits via relative attribute vectors","RelGAN: interpolate and preserve with change vectors","Relative attributes enable smoother face edits"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The generator must be able to infer each input image's current attribute values from the pixels alone, because only the requested change, not the source values, is supplied as conditioning.","fun_headline_variants_meta":{"raw":{"variants":["RelGAN: change attributes, not targets","Face editing by relative change, not target labels","Continuous face edits via relative attribute vectors","RelGAN: interpolate and preserve with change vectors","Relative attributes enable smoother face edits"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00024,"raw_usage":{"total_tokens":1465,"prompt_tokens":840,"completion_tokens":625,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":456,"completion_tokens_details":{"reasoning_tokens":559}},"tokens_in":456,"tokens_out":625,"duration_ms":6299,"temperature":1.0,"reasoning_tokens":559,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T12:20:46.898138+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the trained model on an input whose source attributes are intentionally misclassified by a standard attribute classifier, such as an androgynous face for gender, and pass a non-zero relative attribute for that label; if the output does not move in the intended direction, the premise that $G$ can interpret $v$ without the source vector fails.","supporting_citations":[{"cited_title":"StarGAN: Unified generative adversarial networks for multi-domain image-to-image translation,","cited_arxiv_id":null,"evidence_quote":"provides the single-generator multi-domain baseline and the generator backbone RelGAN adapts."},{"cited_title":"AttGAN: Facial attribute editing by only changing what you want,","cited_arxiv_id":null,"evidence_quote":"serves as the target-attribute baseline in FID, classification, interpolation, and user-study comparisons."},{"cited_title":"Unpaired image-to-image translation using cycle-consistent adversar- ial networks,","cited_arxiv_id":null,"evidence_quote":"supplies the cycle-consistency idea used in the cycle-reconstruction loss."},{"cited_title":"Generative adversarial text to image synthesis,","cited_arxiv_id":null,"evidence_quote":"introduces the matching-aware discriminator concept that $D_{\\text{Match}}$ adapts to triplets."},{"cited_title":"Under- standing and improving interpolation in autoencoders via an adversarial regularizer,","cited_arxiv_id":null,"evidence_quote":"inspires the interpolation discriminator's adversarial regularizer for intermediate images."},{"cited_title":"Deep learning face attributes in the wild,","cited_arxiv_id":null,"evidence_quote":"provides the CelebA dataset with 40 binary attribute labels used for training and evaluation."},{"cited_title":"Progressive growing of GANs for improved quality, stability, and varia- tion,","cited_arxiv_id":null,"evidence_quote":"supplies the CelebA-HQ dataset used for training and most evaluations."},{"cited_title":"A style-based generator architecture for generative adversarial networks,","cited_arxiv_id":null,"evidence_quote":"supplies FFHQ, the held-out dataset for testing generalization."},{"cited_title":"GANs trained by a two time-scale update rule converge to a local Nash equilibrium,","cited_arxiv_id":null,"evidence_quote":"defines FID, the visual-quality metric used for quantitative comparison."}],"review_version":1}