{"id":"adde0e07-7d0d-450d-9004-dc6401a502f3","arxiv_id":"2505.04306","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"MoDE improves occluded face recognition by generating multiple diffusion-repainted versions of a face and adaptively weighting their recognition scores with a trained identity gate.","lead":"This paper proposes MoDE, a module that restores occluded face images with several diffusion models and then uses a trained gate to combine the recognition scores of the restored images. The authors report accuracy gains on masked and otherwise obscured face datasets, though the evaluation has gaps that temper the headline claim.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The paper's own tables contradict 'MoDE outperforms all other methods': EMD alone beats MoDE-only on Occ CelebA (32.2 vs 31.6) and OVF (90.9 vs 87.5).","rationale":"The reader's weakest assumption (mask availability at test time) is an important limitation, but the more immediately load-bearing problem is that the paper's own reported numbers refute the headline claim. The mask issue affects real-world applicability and is a valid reason for requesting clarification; however, the internal contradiction in Tables I and II requires no external assumption to be recognized. I agree with the reader's conditional verdict, because the core method still shows consistent gains over its own baselines and the combined EMD+MoDE pipeline achieves the best results in most rows. The condition should explicitly require correcting the overstatement in Section IV-B1 and providing the mask-annotation protocol for OVF/WWCF. These are revisions rather than grounds for outright rejection, so the CONDITIONAL verdict remains appropriate.","tokens_in":12056,"tokens_out":6355,"duration_ms":64594,"concrete_test":"Re-run the OVF and Occ CelebA evaluations using the exact protocol of Section IV-A, with the same probe/gallery split, the same backbone for both MoDE-only and EMD-only configurations, and no other re-ranking. Report Top-1 accuracy with paired bootstrap confidence intervals over the probe set. If MoDE-only Top-1 remains below EMD-only Top-1 in either dataset, as the published tables indicate, then the assertion 'MoDE outperforms all the competing methods' must be removed or narrowed to 'MoDE+EMD outperforms all competing methods.'","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim in Section IV-B1, 'MoDE outperforms all the competing methods,' is directly contradicted by the paper's own experimental tables. In Table I (Occ CelebA), the DeepFace-EMD row reports EMD Top-1 accuracy of 32.2 versus MoDE's 31.6; in Table II (OVF), the same row reports EMD Top-1 of 90.9 versus MoDE's 87.5. These are not noise-level differences, especially the 3.4-point gap on OVF. Because the paper positions MoDE as a plug-and-play module that can be attached to any face recognition model, the comparison 'MoDE' in each row should be the same backbone as the 'EMD' baseline. Under that reading, the re-ranking alone outperforms MoDE alone in two of the five evaluated settings. The strongest claim therefore fails as stated; the paper can at most claim that MoDE consistently improves its own backbone and that the combination MoDE+EMD achieves the best results. This is a correctness issue, not a matter of external consensus or missing comparison, because it relies only on numbers inside the manuscript.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes MoDE (Mixture of Diffusion Experts), a plug-and-play module for occluded face recognition. Given an occluded face image and a binary mask m, the method runs several diffusion-based inpainting experts to produce n repainted images, keeps the original occluded image as expert E0, extracts features from all n+1 images with a fixed face recognition backbone, and trains an ID-Gate network to assign weights to the corresponding similarity vectors before the final argmax classification. Experiments are reported on synthetically occluded MS1M, LFW, and CelebA, plus two self-collected real datasets OVF and WWCF, using ArcFace, FaceNet, CosFace, FFR-Net, and DeepFace-EMD as backbones or baselines. The main claimed contributions are the identity-gated fusion in decision space and the use of diffusion repainting for occluded face recognition.","tokens_in":12318,"tokens_out":4973,"duration_ms":49899,"significance":"If the central claim were established, MoDE would be a useful and practical contribution: decision-space fusion of multiple inpainting hypotheses is a sensible way to exploit generative diversity, and the plug-and-play design is attractive for real deployments. The two self-collected datasets and the systematic comparison with several face recognition models are also valuable. However, the manuscript's own tables contradict the headline claim that 'MoDE outperforms all other methods,' and a core test-time assumption (the availability of an accurate occlusion mask) is never addressed for real-world data. These issues must be resolved before the contribution is credible. The paper does not provide code or released datasets, which further limits verification, but the core idea remains potentially sound.","major_comments":[{"comment":"The sentence 'MoDE outperforms all the competing methods' in Section IV-B1 and the Table I caption 'Our MoDE outperforms all other methods' are directly contradicted by the paper's own numbers. On Occ CelebA, the DeepFace-EMD row reports EMD Top-1 of 32.2 versus MoDE's 31.6; on OVF, EMD achieves 90.9 Top-1 versus MoDE's 87.5. In addition, on WWCF, the FaceNet row shows MoDE at 7.6 Top-1, below the FaceNet baseline of 8.6. Since MoDE is presented as a plug-and-play module, the fair comparison is against the same backbone, and in two settings the EMD re-ranking alone beats MoDE alone. The central claim should be revised to something like 'MoDE consistently improves its own backbone, and MoDE combined with DeepFace-EMD achieves the best results,' or the authors should provide a corrected comparison.","section":"Section IV-B1, Table I, Table II"},{"comment":"The repainting step conditions on a binary mask m: Eq. (4) splices known pixels sampled via Eq. (5) with unknown pixels sampled via Eq. (6). The paper never explains how m is obtained at test time for real-world images or for the OVF and WWCF datasets. For synthetic datasets, the text says masks are created using facial keypoint detection, but real deployment requires a mask estimation procedure. If the mask is inaccurate, the known-region sampling in Eq. (5) will mix corrupted pixels into every diffusion step, degrading the reconstruction. The authors should specify the mask estimation method or explicitly state that m is assumed to be given by an external oracle.","section":"Section III-A, Eqs. (4)-(6)"},{"comment":"The training and testing protocol for Occ MS1M is under-specified. The text states that during training 500 faces in Occ MS1M are used as probe and 2500 faces in MS1M as gallery, and during testing 500 faces in Occ MS1M are randomly selected as probe with 2500 MS1M faces as gallery. There is no statement that the training and test probes or galleries are disjoint. If the same identities or even the same images are reused, the reported Occ MS1M results could reflect ID-Gate memorization rather than generalization. The authors must clarify the exact split and, ideally, evaluate on a held-out set of identities.","section":"Section IV-A"},{"comment":"The discussion of the real-world results says 'MoDE performs better in dealing with face occlusion problems compared with most face recognition models.' This is too weak compared with the earlier 'outperforms all' claim, and it is also inconsistent with the FaceNet WWCF result where MoDE is worse than the baseline. The narrative should be aligned with the actual table entries, and the claims should be stated per backbone rather than as an unconditional superiority.","section":"Section IV-B2"}],"minor_comments":[{"comment":"The notation is inconsistent: Eq. (7) defines W(X), the weight vector for the whole set X, while Eq. (8) uses w(x_i) for individual experts. Please clarify the relationship between the vector W(X) and the scalar weights w(x_i).","section":"Section III-B, Eq. (8)"},{"comment":"The parameter 'jumpy size j' is listed in Implementation Details but is never defined or used in the equations. Define it in Section III-A or remove it from the text.","section":"Section IV-A"},{"comment":"There is a typo: 'aj iterative resampling step' should be 'an iterative resampling step.'","section":"Section III-A"},{"comment":"Several reference entries contain 'n.d.' placeholders for publisher, volume, and page information, for example references [4], [6], [7], [35], and [42]. The bibliography should be completed before publication.","section":"References"},{"comment":"The text says 'the second Noisy Top-K Gating is discussed in the appendix,' but no appendix is present in the manuscript. Either include the appendix or remove the reference.","section":"Section III-D"},{"comment":"The contribution bullet 'our work is the first to introduce the diffusion model to occluded face recognition' is a strong novelty claim and is not supported by an exhaustive survey. Please soften the claim or provide a more complete literature comparison.","section":"Section I and II"},{"comment":"The figures are referenced in the text (Figs. 2-7) but are not visible in the provided manuscript text; please ensure all figures are embedded and legible in the final submission.","section":"Throughout"}],"recommendation":"major_revision","confidential_remarks":"The paper's core idea is not circular and is potentially useful, but the manuscript as written overstates its results: the 'outperforms all' claim is refuted by the paper's own tables, and the mask-dependence of the method is unresolved for real-world use. These are fixable with revised claims, a clarified setup for mask acquisition, and a properly disjoint evaluation split. I would not reject, but major revision is needed. The authors may also be asked to release the self-collected OVF and WWCF datasets or at least describe their collection protocol in more detail, since the practical significance depends on them."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a plausible incremental application of known components—Repaint-style diffusion inpainting plus a learned gate that weights predictions from several repainted faces—and it does improve recognition over the raw occluded image in most tested settings. But the paper's central 'outperforms all methods' claim is false on its own numbers, and the method's load-bearing assumption about known occlusion masks is never explained for real data.\n\nWhat's genuinely new: I don't see a previous diffusion-based occluded-face-recognition method in the references, and gating over multiple repaintings in decision space is a sensible way to exploit diffusion diversity. The paper gives credit where it's due: the ablation in Table III shows ID-Gate beats simple averaging of repainted faces (+3.2 Top-1 on Occ CelebA), and the modular attachment to ArcFace/FaceNet/CosFace/FFR-Net shows consistent gains in most rows. That is a real empirical contribution, assuming results reproduce.\n\nSoft spots, in order of severity. First, the internal contradiction: Section IV-B1 says 'MoDE outperforms all the competing methods,' but in Occ CelebA DeepFace-EMD's EMD re-ranking gets 32.2 Top-1 versus 31.6 for MoDE, and on OVF EMD gets 90.9 versus 87.5. The paper can honestly claim that MoDE lifts each backbone and that MoDE+EMD beats everything, but not that MoDE alone outperforms all. Second, the known-mask assumption in Eqs. (4)-(6) is load-bearing. The paper doesn't say how masks are obtained for OVF/WWCF or at test time; with an imperfect mask, the known-region sampling injects corrupted pixels at every step and the whole approach degrades. This is not a minor implementation detail. Third, no error bars, code, or data release, and the real datasets are self-collected without detailed protocol. That makes the reported margins (e.g., +4.8 on CosFace Occ MS1M) hard to assess. Also missing are recent occlusion-robust methods and large-scale benchmarks with masks; the comparison set is thin.\n\nMy take: the architecture is coherent and the ablation supports the gating mechanism, but the claims need correction and the mask-handling story needs to be explicit before this can be trusted. I'd send it to a serious referee, because the core direction is worth engaging with, but I'd expect major revision. For my own work, I wouldn't cite it as-is.","headline":"A reasonable plug-and-play diffusion-repaint ensemble for occluded face recognition, but the paper's central 'outperforms all' claim is false on its own tables and the mask-availability assumption is left unspecified.","tokens_in":12843,"tokens_out":2228,"would_cite":false,"duration_ms":21543,"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":"This paper claims that adaptively weighting multiple diffusion-repainted versions of an occluded face, fused in the decision space by a learned identity gate, improves recognition over using the occluded image alone.","keywords":["occluded face recognition","Mixture of Diffusion Experts","diffusion repainting","identity gating","decision-space fusion","mask occlusion","plug-and-play face recognition"],"falsifier":"Take a fixed set of occluded probe images with known ground-truth masks, run MoDE with the exact mask and with masks corrupted by increasing amounts of mislabeled pixels, and compare Top-1 accuracy; if accuracy falls back to the no-repaint baseline as soon as the mask is slightly wrong, the central claim depends on perfect mask availability, which the paper does not establish for real images.","tokens_in":11857,"feed_emoji":"😷","tokens_out":10455,"duration_ms":89511,"temperature":0.7,"pith_summary":"MoDE (Mixture of Diffusion Experts) claims that occluded face recognition works better when a masked face is repainted several times by a diffusion model and the recognition scores of all repainted versions, together with the original occluded image, are combined by a learned identity gate in the decision space. The authors argue that this beats both recognizing the occluded image directly and averaging the repainted faces, because the gate can down-weight reconstructions whose random sampling drifted from the true identity. They position MoDE as a plug-and-play module that attaches to existing recognition backbones, and they report consistent gains on three public benchmarks and two self-collected real-world occlusion datasets. If the claim holds, masked and otherwise occluded face recognition can be improved without retraining the underlying recognition model.","feed_headline":"Weighing repainted faces boosts masked face recognition","feed_subtitle":"A learned gate fuses diffusion-repainted faces, lifting Top-1 accuracy on occluded benchmarks.","key_machinery":"The central mechanism is the combination of a mask-conditioned diffusion repainting expert and an identity-gating network (ID-Gate). Each expert runs a Denoising Diffusion Probabilistic Model with a repainting procedure that uses resampling: at denoising step $t$ the known region $m\\odot x^{\\mathrm{known}}_{t-1}$ is sampled from the input's noised distribution while the unknown region $(1-m)\\odot x^{\\mathrm{unknown}}_{t-1}$ comes from the denoiser, and the two are spliced with $r$ resampling passes to smooth boundaries. The ID-Gate is a multi-input softmax network that receives the extracted feature vectors of the occluded original $x_0$ and $n$ repainted faces $x_1,\\dots,x_n$ and outputs weights $w(x_i)$; the final prediction is the weighted sum of the individual similarity matrices, $S_X=\\sum_{i=0}^n w(x_i)s(x_i)$. The gate's job is to decide, per probe, which reconstructions carry trustworthy identity information and which should be down-weighted.","core_discovery":"The paper's central claim is that a stochastic generator plus a learned fusion rule beats any single reconstruction: the occluded image $x_0$ and $n$ diffusion-repainted faces are treated as $n+1$ experts, a shared recognizer extracts an embedding from each, and the ID-Gate network outputs weights $w(x_i)$ that combine the per-expert similarity matrices into $S_X=\\sum_{i=0}^n w(x_i)s(x_i)$, which becomes the recognition score. The authors report that this module raises Top-1 and Top-5 accuracy for every base recognizer on the synthetic mask benchmarks, and for most on the wild datasets, and that the gains stack with patch-wise EMD re-ranking. The underlying reasoning is that the diffusion model's sampling diversity supplies multiple plausible identities, and the gate learns which of those are trustworthy for a given probe.","pith_inferences":["The largest practical gap the paper leaves open is mask acquisition: the repainting equations require a precise occlusion map $m$, and a learned mask estimator or a gate that consumes mask uncertainty would make the module end-to-end.","Because the gate acts on similarity vectors rather than raw pixels, the same decision-space gating could attach to other stochastic reconstruction pipelines, such as super-resolution or 3D face reconstruction, where several plausible outputs must be reconciled for a downstream identity decision.","The plateau at four to five experts suggests a compute-accuracy trade-off: an adaptive policy could stop adding experts once the gate's weights stabilize."],"forward_implications":["MoDE transfers across recognition backbones: every baseline tested on the synthetic occluded benchmarks shows higher Top-1 accuracy when the module is attached, and most improve on the real-world datasets as well.","The identity gate beats simple equal averaging of repainted faces: the ablation on Occ CelebA shows that averaging repainted faces helps, but weighting them with ID-Gate helps more.","Recognition accuracy keeps rising with the number of diffusion experts up to about four or five, after which the gain plateaus.","MoDE and patch-wise EMD re-ranking are complementary: combining both produces the largest gains, such as a 19.6-point Top-1 improvement on Occ MS1M in the EMD+MoDE row."],"supporting_citations":[{"why":"Supplies the denoising diffusion probabilistic model that the repainting experts are built on.","marker":"[37]"},{"why":"Provides the mask-conditioned repainting procedure with resampling that MoDE adapts to face images.","marker":"[35]"},{"why":"Defines the ArcFace recognizer used as the main backbone and as a baseline for the reported gains.","marker":"[7]"},{"why":"Defines the FaceNet embedding baseline used to show that MoDE is plug-and-play across recognizers.","marker":"[4]"},{"why":"Supplies the feature-rectification occlusion-recovery baseline compared on synthetic and real occluded faces.","marker":"[41]"},{"why":"Provides the patch-wise Earth Mover's Distance re-ranking baseline that combines with MoDE for the largest gains.","marker":"[42]"}],"fun_headline_variants":["Diffusion repaint mixture with learned gate tops masked face ID","Gating multiple diffusion repaints improves occluded face recognition","Mixture of diffusion experts lifts accuracy on masked faces","Learned fusion of diffusion repaints wins at occluded face recognition"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method needs a correct map of which pixels are occluded for every test image, and the paper never says where that map comes from in practice.","fun_headline_variants_meta":{"raw":{"variants":["Diffusion repaint mixture with learned gate tops masked face ID","Gating multiple diffusion repaints improves occluded face recognition","Mixture of diffusion experts lifts accuracy on masked faces","Learned fusion of diffusion repaints wins at occluded face recognition"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000245,"raw_usage":{"total_tokens":1520,"prompt_tokens":912,"completion_tokens":608,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":528,"completion_tokens_details":{"reasoning_tokens":539}},"tokens_in":528,"tokens_out":608,"duration_ms":5650,"temperature":1.0,"reasoning_tokens":539,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T23:32:02.894763+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a fixed set of occluded probe images with known ground-truth masks, run MoDE with the exact mask and with masks corrupted by increasing amounts of mislabeled pixels, and compare Top-1 accuracy; if accuracy falls back to the no-repaint baseline as soon as the mask is slightly wrong, the central claim depends on perfect mask availability, which the paper does not establish for real images.","supporting_citations":[{"cited_title":"Repaint: Inpainting using denoising diffusion probabilistic models,","cited_arxiv_id":null,"evidence_quote":"Provides the mask-conditioned repainting procedure with resampling that MoDE adapts to face images."},{"cited_title":"Arcface: Additive angular margin loss for deep face recognition,","cited_arxiv_id":null,"evidence_quote":"Defines the ArcFace recognizer used as the main backbone and as a baseline for the reported gains."},{"cited_title":"Facenet: A unified embed- ding for face recognition and clustering,","cited_arxiv_id":null,"evidence_quote":"Defines the FaceNet embedding baseline used to show that MoDE is plug-and-play across recognizers."},{"cited_title":"A unified framework for masked and mask-free face recognition via feature rectification,","cited_arxiv_id":null,"evidence_quote":"Supplies the feature-rectification occlusion-recovery baseline compared on synthetic and real occluded faces."},{"cited_title":"Deepface-emd: Re-ranking using patch-wise earth mover’s distance improves out-of-distribution face identification,","cited_arxiv_id":null,"evidence_quote":"Provides the patch-wise Earth Mover's Distance re-ranking baseline that combines with MoDE for the largest gains."}],"review_version":1}