{"id":"bcb92686-f50d-48e6-9e23-381adcce0cf6","arxiv_id":"2603.16551","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A new conditioning network lets diffusion models generate fair medical images for demographic intersections never seen in training, beating reweighting baselines on chest X-ray and fundus data.","lead":"This paper introduces CompDiff, a diffusion-model engine that encodes a patient's age, sex, and race as separate building blocks and recombines them to create medical images for demographic groups with no training examples. Its chest X-ray and eye-scan experiments show better image quality, fairer subgroup performance, and stronger downstream classifiers than fairness-through-reweighting baselines.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Zero-shot evidence in Table 3 is computed on tiny held-out subgroups with no reported sample sizes or error bars, so the 21% FID improvement may not be statistically separable.","rationale":"The paper's strongest claim is that CompDiff's representation-level factorization yields zero-shot generalization to demographic intersections absent from training, with Table 3 as the direct evidence. The reader's weakest assumption identifies exactly the load-bearing weakness: Table 3 lacks the standard deviations promised in §3.3 and likely uses tiny real reference sets. I reviewed the full text and find no other objection that is more central. The ablations in §3.4 and Table 5 support the mechanism, though they contain a minor numeric inconsistency (e.g., 104.2 vs 97.1 for strong L_comp) and rely on a strong flat-vs-hierarchical contrast; however, those do not directly test the zero-shot claim. The downstream TSTR results and three-seed protocol elsewhere give partial independent support, so this is not a rejection but a request for statistical substantiation. If the requested confidence intervals show non-overlap, the claim is credible; if not, the verdict should be reconsidered. Hence I keep the reader's CONDITIONAL verdict unchanged.","tokens_in":8265,"tokens_out":3680,"duration_ms":39939,"concrete_test":"Obtain the per-subgroup real test set sizes for Table 3 from the authors, then recompute FID for all three methods using bootstrapped 95% confidence intervals (e.g., 1000 resamples of the generated and real sets, with matched reference size). If the CIs for CompDiff overlap with FairDiffusion or baseline on any held-out subgroup, the zero-shot superiority claim in §3.3 is not statistically supported; also report the three-seed mean±std as promised.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim—that hierarchical composition enables zero-shot generalization to unseen intersections—rests on Table 3. The paper states in §3.3 that 'All results are computed on held-out generated test sets using three different generation seeds; we report mean and standard deviation,' yet Table 3 reports only point estimates for five held-out intersections. The rarest training groups in Table 2 are ~0.5% of the training split; the held-out groups are chosen by rarity, so the real test reference sets are likely tens of images. Subgroup FID at this size is unstable and sensitive to the number of real images used; without per-subgroup n, standard deviations, or confidence intervals, the claimed up-to-21% FID improvement cannot be distinguished from sampling noise. This matters because FairDiffusion's failure on 80+ Female Asian (247.2 vs baseline 210.7) is used as evidence that optimization-level reweighting cannot help when samples are absent; that comparison is subject to the same noise. The modeling premise (pairwise interactions and Lcomp anchor compose correctly for a triple) is plausible and is partially supported by Table 1 and Table 5, but the only direct evidence for the zero-shot claim is Table 3, and that evidence is statistically under-specified. The manuscript's own limitation statement (performance degrades relative to well-represented groups) is honest but does not address whether the zero-shot gains are real.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"CompDiff proposes a representation-level solution to demographic imbalance in medical image diffusion models. Instead of reweighting losses, it introduces a Hierarchical Conditioner Network (HCN) that decomposes demographic conditioning into single-attribute embeddings, pairwise interactions, and a composed token, which is concatenated with CLIP text embeddings as cross-attention context. The training objective combines the diffusion loss with a compositional-consistency loss, a KL term on the demographic latent, and an auxiliary classification loss on the final conditioning token. The paper reports improved FID, ES-FID, zero-shot FID on five held-out demographic intersections, and downstream classifier fairness on MIMIC-CXR and FairGenMed, with ablations supporting the hierarchical design and the auxiliary loss placement.","tokens_in":8547,"tokens_out":7726,"duration_ms":80844,"significance":"The central idea is timely and, if correct, significant: it reframes fairness interventions from optimization-level reweighting to representational factorization and provides initial evidence that unseen demographic intersections can be generated in a zero-shot manner. The experimental scaffolding is strong in several respects: three-seed reporting, held-out generated test sets, external pretrained classifiers, TSTR downstream evaluation, and systematic ablations of both architecture and loss terms. The released code is a further strength. The main reservation is not the architecture but the statistical basis of the headline zero-shot claim, which currently rests on a single table without uncertainty estimates. With that evidence properly grounded, the paper would be a solid contribution to fair medical image generation.","major_comments":[{"comment":"The protocol sentence in §3.3 states that all results are computed on held-out generated test sets using three generation seeds and that mean and standard deviation are reported, but Table 3 reports only point estimates for each method and intersection. No per-subgroup real-sample size is given. The held-out intersections were chosen by rarity; the rarest rows in Table 2 are 0.5% of training, so the real reference sets are likely on the order of tens of images. Subgroup FID at this scale is high-variance and depends on the real-image count. As a result, the claimed up-to-21% improvement (e.g., 18–40 M/A: 161.3→127.6) and the FairDiffusion failures (247.2 vs 210.7 for 80+ F/A; 265.5 vs 208.1 for 80+ M/A) cannot be separated from noise as reported. This is the only direct evidence for the zero-shot claim. Please provide per-subgroup n, mean(std) over seeds, and ideally bootstrap confidence","section":"§3.3, Table 3"},{"comment":"The auxiliary classification loss Laux is applied to the final token c using true age/sex/race labels. Consequently, the high sex/race accuracy and low age RMSE reported in Tables 1 and 5 are partly supervised properties of the conditioning pathway, not independent evidence that HCN's compositional factorization yields demographic control. The ablation HCN(no aux) versus CompDiff shows that the loss is necessary, and the flat-versus-hierarchical comparison under identical supervision is informative. However, the paper should state explicitly that demographic controllability is partly constructed by Eq. (5), and should ideally validate the semantics of c with an independent probing task not trained jointly with the generator before using demographic accuracy as evidence for the representation-level claim.","section":"§2.3, Eq. (5), Tables 1/5"},{"comment":"The zero-shot hypothesis is that pairwise interactions (h_a,s, h_a,r, h_s,r) and the Lcomp anchor compose correctly for an unseen triple. Table 3 lists only the removed triples; it does not report whether the constituent pairs were present in training or how often. For example, 80+ F/A has pairs (80+, F), (80+, A), (F, A), and if any pair is also absent or extremely rare, the pairwise-to-triple composition story is not the mechanism actually tested. Please report per-held-out-group pair coverage (counts or percentages) and, if possible, include a control where seen triples of matched rarity are removed to confirm that the method's advantage is specific to true zero-shot composition rather than to rarity in general.","section":"§2.2, §3.3"}],"minor_comments":[{"comment":"The ES-FID formula is ambiguous as typeset. If the intended expression is FID * (1 + (1/(|Ai|*FID)) * sum |FID - FID_j|) = FID + mean absolute deviation, please rewrite it to remove the apparent extra FID factor. As printed, the units are not dimensionally consistent.","section":"Eq. (6)"},{"comment":"The text states that increasing the compositional-consistency weight degrades FID to 104.2, but Table 5's 'Strong Lcomp' row reports FID 97.1 for λcomp=0.5. Either the 104.2 value corresponds to a different λcomp that is omitted from the table, or the text and table do not agree. Please reconcile.","section":"§3.4, Table 5"},{"comment":"The fundus baseline disease AUROC is reported without a standard deviation (0.94 with no std), while all other cells in the table have mean(std). Please add the missing value.","section":"Table 1"},{"comment":"The selection of the best checkpoint based on 'validation performance across the four dimensions' is not specified. State how the four metrics were aggregated into a single criterion and confirm that no hyperparameter or checkpoint selection was performed on the test set.","section":"§3.3"},{"comment":"The zero-shot experiment is reported only for the chest X-ray modality. If the method is presented as a general framework for medical images, either add a zero-shot evaluation on the fundus data or explicitly restrict the zero-shot claim to the modality tested.","section":"§3.1, §3.3"},{"comment":"The values of λcomp, λaux, and λKL are not given in the main text. Please report them and state whether they were tuned on the validation set; this is needed for reproducibility of the ablations.","section":"§2.3"}],"recommendation":"major_revision","confidential_remarks":"The paper fits the journal's scope and the core idea is promising. The main revision need is statistical transparency: Table 3 is the only direct evidence for the zero-shot claim and it lacks standard deviations and sample sizes. I also recommend that the authors clearly frame the auxiliary loss as constructing demographic controllability rather than as independent evidence. No concerns about novelty or fit; I would support acceptance after these issues are addressed."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe thing to know about CompDiff: the core idea is representation-level conditioning—decomposing demographics into single, pairwise, and full composition and feeding a dedicated token to cross-attention—and that is actually different from FairDiffusion's loss reweighting. The paper shows it improves FID and ES-FID on two modalities, and the TSTR downstream gains look real. The three-seed protocol, held-out intersections, external classifiers, and ablations are better than the subfield average.\n\nWhat's new and good: HCN hierarchy, the zero-shot protocol that removes five intersections from training, and the ablation showing flat MLP fails to recover demographic control while hierarchical succeeds under identical supervision. That is a nice piece of evidence for the structural inductive bias claim. The compositional consistency loss as a soft anchor is sensible.\n\nSoft spots, in proportion: The load-bearing zero-shot claim rests on Table 3, which reports only point estimates for five held-out groups. The table gives no standard deviations, no subgroup sample sizes, and the text promises mean (std) in §3.3. Given those groups are the rarest, real FID reference sets are likely in the tens of images. At that size, a 21% FID difference can be noise. FairDiffusion's failure on 80+ Female Asian is cited as confirming that optimization-level methods can't help, but that comparison has the same statistical fragility. This is fixable by releasing per-subgroup n and error bars or confidence intervals. Also, §3.4 text says strong L_comp degrades FID to 104.2, but Table 5 reports 97.1 for the same variant. Maybe one is on a different set, but the paper doesn't say so. And the two code URLs—anonymous repo in the body, named GitHub in the abstract—should be reconciled. On circularity: Laux supervises the final token, so sex accuracy 0.99 is partly constructed; the paper doesn't hide that and it doesn't undermine the FID/ES-FID fairness claims.\n\nBottom line: the central mechanism is plausible and the experiments are mostly honest. The zero-shot evidence needs tightening before I'd fully trust the headline claim, but this deserves a serious referee. I'd send it to review with a request for the missing statistics. For a reading group, worth a look if you care about fair generation.","headline":"Representation-level compositional conditioning is a genuinely new idea with solid scaffolding, but the zero-shot evidence is statistically under-specified and one ablation has an internal inconsistency.","tokens_in":9081,"tokens_out":2006,"would_cite":true,"duration_ms":20120,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"CompDiff shows demographic conditioning can be factorized hierarchically so unseen intersections compose from seen attributes.","keywords":["compositional demographic conditioning","fair medical image synthesis","intersectional bias","zero-shot generalization","diffusion models","representation learning","chest X-ray generation","fundus image generation"],"falsifier":"Generate or collect a test set where each demographic intersection has hundreds of real images, remove several intersections from training, and compare CompDiff against a flat encoder and a reweighting baseline with confidence intervals on subgroup FID. If the held-out intersections do not show a statistically significant FID advantage—or if an ablation replacing the hierarchy with a flat MLP of equal capacity matches CompDiff—the compositional-generalization claim is falsified. A second probe: train on synthetic data with known ground-truth attribute factors and check whether generated images","tokens_in":8075,"feed_emoji":"🩻","tokens_out":6191,"duration_ms":53123,"temperature":0.7,"pith_summary":"In medical image generation, models trained on imbalanced data produce worse images for rare demographic subgroups, and no loss-reweighting scheme can help with intersections that never appear in training. This paper argues the fix should come at the representation level: decompose demographic identity into single attributes, pairwise interactions, and a full composition, and feed that structured token into the diffusion model. On chest X-ray and fundus image benchmarks, the proposed CompDiff framework improves overall fidelity (FID 64.3 vs 75.1), reduces subgroup quality gaps, and generates plausible images for demographic intersections held out entirely from training. The paper also shows that classifiers trained on CompDiff-generated data achieve higher AUROC and lower demographic bias, suggesting that how demographics are represented matters for downstream fairness.","feed_headline":"CompDiff composes demographics, not losses, for fair image generation","feed_subtitle":"Factorized age/sex/race conditioning improves FID by up to 21% on groups never seen in training.","key_machinery":"The central mechanism is the Hierarchical Conditioner Network (HCN), which decomposes demographic conditioning into single-attribute embeddings ('grandparents'), pairwise interaction MLPs ('parents'), and a full-composition MLP ('child') that yields the demographic token c. The token is concatenated with CLIP text embeddings to form the cross-attention context for a Stable Diffusion UNet. The design is regularized by a compositional consistency term Lcomp = 1 - cos(hdemo, e_age + e_sex + e_race), a KL term on the variational latent, and auxiliary classification applied to the projected token. The HCN's structured factorization encourages parameter sharing across subgroups and is what allows","core_discovery":"This paper claims that demographic identity in medical images is compositional: a diffusion model can generalize to unseen demographic intersections if conditioning is factored hierarchically. CompDiff's Hierarchical Conditioner Network (HCN) embeds age, sex, race separately, learns pairwise interactions, and fuses them into a demographic token appended to CLIP text embeddings for cross-attention. Auxiliary classification on the projected token preserves demographic information; a variational latent and a cosine anchor to the additive baseline regularize training. On five held-out intersections with zero training examples, CompDiff improves FID by up to 21%, whereas FairDiffusion, which rewe","pith_inferences":["Editorial: If composition works for demographics, the same hierarchical factorization could be applied to other structured clinical attributes (e.g., pathology, severity, body region), potentially improving generation for rare disease combinations.","Editorial: A direct test would compare CompDiff and flat encoders on a synthetic dataset with known generative factors to see whether the hierarchy genuinely recovers compositionality or simply provides extra capacity.","Editorial: The paper's reliance on pairwise interactions raises a question: will triple or higher-order interactions degrade substantially when all three attributes are unseen together? Future graph-based interaction models might capture these better.","Editorial: If zero-shot compositional generation is confirmed, it reduces the need to gather rare demographic data for training, but clinical expert evaluation (not just automated metrics) would be needed before synthetic images are used in practice."],"forward_implications":["If CompDiff's central claim holds, loss-reweighting methods like FairDiffusion cannot be the complete answer to the imbalanced generator problem, because they cannot synthesize training signal for combinations absent from the data.","Rare demographic intersections can be generated with acceptable quality even when no training examples exist, which opens a path to augmenting clinical datasets for underrepresented groups.","The architecture of demographic conditioning—not just the loss—affects downstream classifier fairness; classifiers trained on CompDiff data show higher AUROC and lower underdiagnosis or equalized-odds differences.","Because HCN adds only 0.19% trainable parameters and requires no change to sampling, the approach is a cheap drop-in replacement for the conditioning head of a latent diffusion model."],"fun_headline_variants":["CompDiff: Factored demographics beat loss reweighting","No reweighting needed: CompDiff composes demographics for rare groups","Zero-shot intersections: CompDiff's hierarchical conditioning wins","Fair medical images: CompDiff factors demographics, not losses","CompDiff: Compositional diffusion for unseen demographic mixes"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"Zero-shot generalization is measured on five held-out intersections whose real test samples number only tens of images, so the subgroup FID values may not be stable enough to reliably separate methods; underneath that sits the premise that pairwise interactions trained on seen pairs compose correctly to unseen triples.","fun_headline_variants_meta":{"raw":{"variants":["CompDiff: Factored demographics beat loss reweighting","No reweighting needed: CompDiff composes demographics for rare groups","Zero-shot intersections: CompDiff's hierarchical conditioning wins","Fair medical images: CompDiff factors demographics, not losses","CompDiff: Compositional diffusion for unseen demographic mixes"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000259,"raw_usage":{"total_tokens":1450,"prompt_tokens":798,"completion_tokens":652,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":542,"completion_tokens_details":{"reasoning_tokens":570}},"tokens_in":542,"tokens_out":652,"duration_ms":6572,"temperature":1.0,"reasoning_tokens":570,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-02T18:01:21.959763+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Generate or collect a test set where each demographic intersection has hundreds of real images, remove several intersections from training, and compare CompDiff against a flat encoder and a reweighting baseline with confidence intervals on subgroup FID. If the held-out intersections do not show a statistically significant FID advantage—or if an ablation replacing the hierarchy with a flat MLP of equal capacity matches CompDiff—the compositional-generalization claim is falsified. A second probe: train on synthetic data with known ground-truth attribute factors and check whether generated images","supporting_citations":[],"review_version":2}