{"id":"9519d816-9e5f-46a1-86a6-02c7dee5cc98","arxiv_id":"2412.02627","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"A convex hull based experience replay method reduces catastrophic forgetting in continually updating personalized face GANs, with a new five-celebrity benchmark.","lead":"This paper defines a new learning problem: updating a personalized 2D or 3D face generator as new photo batches of the same person arrive over time, without forgetting earlier looks. It shows that a replay buffer selected as a convex hull in StyleGAN's latent space forgets less than random replay when storage is small.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"ER-Hull's claimed ~20% forgetting reduction is not yet causally tied to the convex-hull criterion: the reported gap lacks variance estimates and no ablation separates geometric selection from timestamp-coverage or random noise.","rationale":"The reader's weakest assumption correctly identifies the unproven geometric proxy as the central risk. My stress-test sharpens this into a falsifiable claim: even if the ER-Hull numbers are reproduced, nothing in the paper shows the improvement is caused by the convex-hull geometry. The absence of error bars and the small number of identities means the headline gap is not yet statistically grounded, and the algorithm's other differences from ER-Rand provide plausible alternative explanations. Because these concerns are addressable with additional experiments and analysis, the CONDITIONAL verdict remains appropriate; no change to the reader's recommended decision is needed. However, if the proposed correlation test failed, the central contribution would be reduced to an unverified heuristic and the verdict should move toward REJECT.","tokens_in":13292,"tokens_out":4200,"duration_ms":49679,"concrete_test":"On a fixed timestamp transition with buffer size 3, compute the Eq. 4 score for a diverse set of candidate buffers; then, starting from the same checkpoint, train the model once per candidate buffer and measure actual forgetting. Compute the Spearman rank correlation between Eq. 4 scores and measured forgetting across candidates. Also compare ER-Hull against a coverage-only baseline that greedily maximizes the number of timestamps represented in the buffer without any convex-hull distance computation. If the correlation is not significantly negative, or if the coverage-only baseline matches ER-Hull, the convex-hull criterion is not the cause of the claimed improvement.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim is that ER-Hull's convex-hull buffer selection reduces forgetting by roughly 20% at buffer size 3. Two conditions are required for this claim to hold: (a) the measured gap is real rather than run-to-run noise, and (b) the gap is caused by the convex-hull objective rather than by another difference between ER-Hull and ER-Rand. Condition (a) is insecure: every reported number is a single-run average over only 5 identities (4 for 3D), with no standard errors, random seeds, or repeated runs. In Table 3, the average LPIPS forgetting gap (0.98 vs 0.78) is heavily driven by one identity, Michael B. Jordan (1.22 vs 0.73), while IU shows almost no gap (0.67 vs 0.65). With this sample size, the headline improvement could be within plausible noise. Condition (b) is also unaddressed. ER-Hull differs from ER-Rand not only in geometric scoring but also in the candidate-buffer constraints and the timestamp-count normalization in Eq. 4; the paper itself labels the convex-hull proxy a hypothesis in Section 3.4 and never demonstrates that the Eq. 4 score correlates with actual forgetting under the training protocol. Without such a link, a random or coverage-based baseline might reproduce the same results, and the proposed mechanism would be unsupported.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces a continual learning setting for personalized generative face models: a StyleGAN (or EG3D) generator is fine-tuned sequentially on batches of a single person's photos that differ in style, lighting, and pose across timestamps. To combat catastrophic forgetting, the authors propose two experience replay strategies. ER-Rand is a balanced reservoir sampler. ER-Hull selects the replay buffer by minimizing the average distance from all currently available latent codes to the convex hull of the buffer's latent codes (Eqs. 3 and 4), using a RANSAC approximation. The paper evaluates these methods on a new dataset of five celebrities with ten timestamps, for 2D (MyStyle) and 3D (My3DGen) personalization, using reconstruction (LPIPS, ID) and synthesis (FID, ID) metrics. The central empirical claim is that at a buffer size of 3, ER-Hull reduces forgetting by roughly 20% relative to ER-Rand, while at larger buffer sizes both methods approach the upper bound.","tokens_in":13643,"tokens_out":6176,"duration_ms":56685,"significance":"The paper identifies a practical and understudied problem: continual personalization of generative face models in an open-world setting. The new dataset, evaluation protocol, and the two replay baselines are valuable assets for future research. If the reported advantage of ER-Hull is real, it provides a simple, geometric memory-selection rule that could be useful for small-storage continual learning. The paper is also honest about its limitations, including the remaining gap to the upper bound and the computational costs. However, the significance currently rests on a small set of single-run experiments whose statistical robustness is not demonstrated, and the proposed selection rule is not causally validated.","major_comments":[{"comment":"All reported numbers are single-run averages over five identities (four for 3D) with no error bars, no random seeds, and no repeated runs. The headline 3D LPIPS forgetting improvement (0.78 vs 0.98 in Table 3) is driven largely by one identity, Michael B. Jordan (0.73 vs 1.22), while IU shows almost no gap (0.65 vs 0.67). With this sample size, the claimed ~20% reduction could be within run-to-run noise. Please provide results over multiple seeds or at least standard errors/confidence intervals for the buffer-3 comparisons, which carry the paper's main claim.","section":"§4.1 and Tables 1–4"},{"comment":"The convex-hull selection rule is introduced as a hypothesis ('We hypothesize that the best image to discard from the replay buffer is the one whose latent code is closest to the convex hull of the remaining latent anchors') and is never connected to actual forgetting dynamics. The comparison between ER-Hull and ER-Rand also differs in the candidate-buffer constraints and the timestamp-count normalization in Eq. (4), so an observed improvement cannot be attributed specifically to the geometric objective. Please include an ablation that uses the same constraints and normalization as ER-Hull but selects buffers randomly, or report a correlation between the Eq. (4) objective and measured forgetting across timestamps.","section":"§3.4, Eq. (4)"},{"comment":"The sentence describing the buffer constraints appears self-contradictory: 'we additionally constrain the algorithm to only consider possible replay buffers that contain at least one sample from each previously seen batch of data when k < t and no more than one example from each batch when k ≥ t.' For k < t, it is impossible for a buffer of size k to contain at least one sample from each of the t previously seen batches. The conditions are likely inverted, and this error defines the search space of the proposed method. Please correct it and make the exact constraint unambiguous.","section":"§3.4, ER-Hull constraints"},{"comment":"The text states that ER-Hull is 'slightly better than ER-Rand for all 4 celebrities for both reconstruction and synthesis.' Table 4 shows that for Margot Robbie, ER-Hull has synthesis forgetting FID of 31.2 while ER-Rand has 30.5, so ER-Hull is worse on that metric. This factual mismatch should be corrected. Additionally, the 20-timestamp experiment (Table 5) uses a single identity and only reports synthesis forgetting with no variance estimates; this is too limited to support the claim that the method scales to longer deployments.","section":"§4.3 and Table 4"}],"minor_comments":[{"comment":"The text says 'Our next sampling algorithm ER-Random (Fig. 3)' but the algorithm described and shown in Fig. 3 is ER-Hull; please correct the name.","section":"§3.4, first paragraph of ER-Hull"},{"comment":"The text refers to 'ER-Hill' (in the paragraph on buffer size 10); this is a typo for 'ER-Hull'.","section":"§4.2"},{"comment":"The notation d(Xj, Hull(Rt)) is used both for the full batch from timestamp j and for the subset of its samples that are available in Xt ∪ Rt−1. Please clarify the definition for j < t, where only the retained samples in the replay buffer are available.","section":"Eq. (3)"},{"comment":"The captions say the metrics are 'scaled by ×10' without specifying which columns; for example, LPIPS values in Table 3 (e.g., 1.17 for the lower bound) suggest that LPIPS is also scaled. Please state the scaling explicitly for each metric.","section":"Table captions"},{"comment":"The synthesis task says latent codes are sampled from 'the convex hull of Xt'; please clarify that this means the convex hull of the latent codes corresponding to the images in Xt.","section":"§3.5, Synthesis evaluation"}],"recommendation":"major_revision","confidential_remarks":"The paper addresses a useful and timely problem, and the proposed method is simple enough to be a practical baseline if validated. However, the empirical foundation as it stands is thin: no variance estimates, a single 20-timestamp run, and a claim about 'all 4 celebrities' that is contradicted by Table 4. The constraint description in §3.4 is also internally inconsistent. These issues are fixable within revision, but they currently prevent confidence in the central claim."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nQuick take: this paper's real contribution is the problem formulation — continual learning for personalized unconditional face generators as image batches arrive in stylistically coherent chunks — and a new buffer-selection heuristic based on convex hulls in StyleGAN latent space. Both are genuinely new relative to the class-conditional GAN continual learning works they cite. The evaluation framework is thoughtful: per-timestamp held-out sets, AIP and forgetting separated, reconstruction and synthesis, 2D and 3D, and a K-Means replay baseline. They are also honest about limits, labeling the hull proxy a hypothesis and including a limitations paragraph.\n\nThe soft spots are real but mostly fixable. No error bars, no seeds, no repeated runs. With five identities (four for 3D), the headline ~20% forgetting reduction is a point estimate. The stress-test note says the Table 3 average is driven by one identity; looking at the table, all four 3D identities show ER-Hull better on LPIPS forgetting, but IU's margin is tiny, so the direction is consistent while the magnitude is not secure. More important, there is no ablation separating the convex-hull objective from the timestamp-coverage constraints and normalization in Eq. 4. Nothing shows that the geometric score, rather than coverage, is what buys the gain. Calling Eq. 4 \"optimal\" over a RANSAC sample is loose wording. The upper bound is trained with a different iteration budget (1000 total for MyStyle versus 10x1000 for ER), so \"upper bound\" is a weak reference point. And no code or data is released, which matters because the datasets are part of the contribution. The citation pattern looks fine; the My3DGen self-citation is on point.\n\nBottom line: plausible idea, consistent but under-powered evidence. It deserves a serious referee, not a desk rejection. A revision needs seeds/error bars, an ablation of the hull criterion, a stronger upper bound, and released data/code. I would bring it to reading group as a maybe, mainly to debate whether the hull proxy is a useful prior or just a roundabout coverage heuristic.\n\nRecommendation: send to peer review; expect heavy revision.","headline":"A useful new continual-learning formulation and a plausible hull-based buffer heuristic, but the headline forgetting-reduction margin is a single-run point estimate over very few identities.","tokens_in":14087,"tokens_out":3944,"would_cite":true,"duration_ms":41496,"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 a replay buffer selected by minimizing the average distance from available StyleGAN latent codes to the convex hull of the buffer's latent codes reduces catastrophic forgetting in personalized 2D and 3D face models…","keywords":["continual learning","experience replay","catastrophic forgetting","personalized generative models","StyleGAN latent space","convex hull","face generation","3D generative models"],"falsifier":"Run the same continual fine-tuning protocol with ER-Hull's selected buffers replaced by random buffers that satisfy the same timestamp-coverage constraint, repeated over many seeds; if forgetting on held-out test images from all past timestamps matches ER-Hull's, then the convex-hull rule is not the cause of the reported gain.","tokens_in":13127,"feed_emoji":"🧠","tokens_out":8071,"duration_ms":78484,"temperature":0.7,"pith_summary":"The paper introduces a continual learning setting in which a personalized generative face model is updated as new batches of a person's photos arrive over time, and it shows that naive sequential fine-tuning forgets earlier appearances. It reports that with a relatively large replay buffer, simple random experience replay already suppresses most forgetting, but at a 30% buffer-to-timestamps ratio random replay still forgets. The proposed algorithm, ER-Hull, chooses which past photos to store by minimizing the average distance from all available latent codes to the convex hull of the stored latent codes in StyleGAN's latent space, and the paper reports about a 20% reduction in forgetting versus random replay for both 2D and 3D personalized face models. The paper also contributes a dataset of five celebrities across ten timestamps and an evaluation protocol based on reconstruction and synthesis metrics.","feed_headline":"Hull-shaped replay buffers cut face-model forgetting by ~20%","feed_subtitle":"With tiny buffers, replaying photos whose latent codes sit closest to a convex hull forgets about 20% less.","key_machinery":"The central object is the convex hull of the latent codes of the images kept in the replay buffer, defined inside the $W+$ latent space of a pretrained StyleGAN. ER-Hull selects the buffer $R_t$ that minimizes the sum over available data batches of the squared distances from each batch's latent codes to $\\mathrm{Hull}(R_t)$, with a normalization for how many distinct timestamps are covered, and it searches over candidate buffers with RANSAC. The hypothesis doing the work is that an image whose latent code lies closest to the hull of the remaining buffer is the safest to discard, because the hull can interpolate that appearance later. This geometric proxy is what converts memory selection into a covering problem in latent space.","core_discovery":"On the paper's own terms, the central discovery is that the informativeness of a replay buffer for a personalized face model admits a geometric measure: keep the buffer $R_t$ whose latent anchors form a convex hull minimizing the average squared distance to every latent code currently available, normalized by the number of unique timestamps represented (Eqs. 3 and 4). With a buffer of 3 images across 10 timestamps, ER-Hull outperforms random sampling and a k-means baseline on both inversion and synthesis, cutting forgetting by roughly 20% in LPIPS and FID while also improving average incremental performance for 2D MyStyle and 3D My3DGen models. At a 1:1 buffer-to-timestamp ratio, both replay methods nearly tie the off-line upper bound, making the selection criterion matter mainly in the low-storage regime. The paper frames the contribution as the first domain-incremental continual learning study for unconditional generative models.","pith_inferences":["Beyond the paper, the convex-hull objective can be read as picking a buffer that spans the observed appearance manifold; a testable next step would be to ablate the hull against simpler summaries such as the centroid or k-medoids of the latent codes to isolate which geometric property carries the memory benefit.","Beyond the paper, the storage arithmetic implies that deployments with hundreds of timestamps would need buffer ratios far below 30%, where random replay is known to fail; this suggests combining hull-based selection with compressed latent anchors instead of raw images.","Beyond the paper, the method is demonstrated only on face GANs, so a natural extension is to test whether the hull criterion transfers to non-face generative models whose latent spaces are less disentangled, where the geometric proxy would be weaker.","Beyond the paper, ER-Hull's RANSAC search costs about 40 minutes per timestamp, so an incremental or closed-form update of the hull buffer as new timestamps arrive is a practical direction the paper leaves open."],"forward_implications":["If ER-Hull's gain is real, then at small buffer-to-timestamp ratios the content of the replay buffer, not just its size, determines how much a personalized face model forgets.","At larger ratios, experience replay becomes easy enough that random selection already approaches the offline upper bound, so geometry-aware selection is only needed for storage-constrained deployment.","The same benefit transfers to 3D personalized face models, where forgetting is more severe, indicating the geometric memory policy is not specific to 2D StyleGAN tuning.","Extending from 10 to 20 timestamps at a 30% buffer ratio keeps the advantage (FID forgetting 29.8 for ER-Hull vs 51.0 for ER-Rand on one identity), which points toward usefulness in long-term deployments.","The storage cost of the method is just the replay buffer itself, so the hull selection buys better retention without requiring all past images to be kept."],"supporting_citations":[{"why":"Supplies the personalized 2D generative prior and its fine-tuning protocol that the continual learning setup extends.","marker":"[28]"},{"why":"Supplies the personalized 3D generative model and its My3DGen training/evaluation protocol.","marker":"[33]"},{"why":"Underlies My3DGen as the 3D GAN backbone, making the 3D experiments possible.","marker":"[6]"},{"why":"Defines the StyleGAN latent space whose convex hulls ER-Hull optimizes.","marker":"[20]"},{"why":"Supplies the random experience replay baseline and the k-means episodic memory variant ER-Hull is compared against.","marker":"[7]"},{"why":"Provides the pretrained encoder used to invert each photo into a StyleGAN latent code before hull selection.","marker":"[35]"}],"fun_headline_variants":["Convex hull replay shrinks face-model forgetting by 20%","Geometric replay cuts face-model forgetting by ~20%","Hull-based replay buffers trim face forgetting by 20%","Convex-hull replay: 20% less forgetting for face models","Replay hulls cut face-model forgetting ~20%"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that a replay buffer whose compressed internal coordinates form a convex hull closest to the other available coordinates is the one that best preserves the model's memory of past appearances; this geometric proxy is stated as a hypothesis and never derived from how fine-tuning actually forgets.","fun_headline_variants_meta":{"raw":{"variants":["Convex hull replay shrinks face-model forgetting by 20%","Geometric replay cuts face-model forgetting by ~20%","Hull-based replay buffers trim face forgetting by 20%","Convex-hull replay: 20% less forgetting for face models","Replay hulls cut face-model forgetting ~20%"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000509,"raw_usage":{"total_tokens":2451,"prompt_tokens":892,"completion_tokens":1559,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":508,"completion_tokens_details":{"reasoning_tokens":1472}},"tokens_in":508,"tokens_out":1559,"duration_ms":12000,"temperature":1.0,"reasoning_tokens":1472,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T23:14:23.230834+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same continual fine-tuning protocol with ER-Hull's selected buffers replaced by random buffers that satisfy the same timestamp-coverage constraint, repeated over many seeds; if forgetting on held-out test images from all past timestamps matches ER-Hull's, then the convex-hull rule is not the cause of the reported gain.","supporting_citations":[{"cited_title":"Mystyle: A personalized generative prior","cited_arxiv_id":null,"evidence_quote":"Supplies the personalized 2D generative prior and its fine-tuning protocol that the continual learning setup extends."},{"cited_title":"Wang, Shengze Wang, and Roni Sengupta","cited_arxiv_id":null,"evidence_quote":"Supplies the personalized 3D generative model and its My3DGen training/evaluation protocol."},{"cited_title":"Efficient geometry-aware 3d generative adversarial networks","cited_arxiv_id":null,"evidence_quote":"Underlies My3DGen as the 3D GAN backbone, making the 3D experiments possible."},{"cited_title":"Analyzing and improv- ing the image quality of stylegan","cited_arxiv_id":null,"evidence_quote":"Defines the StyleGAN latent space whose convex hulls ER-Hull optimizes."}],"review_version":1}