{"id":"fef358fa-f940-476b-869e-976131762353","arxiv_id":"2411.14521","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A personalized facial age transformation method that uses an adapter network on top of the SAM global aging model, trained with 10 to 50 photos of one person, to produce re-aged images that resemble that person's actual appearance at the target age.","lead":"MyTimeMachine trains a small adapter on 10 to 50 personal photos spanning decades, adjusting a global facial aging model so re-aged faces resemble the person's actual appearance at the target age. The same idea extends to video by re-aging one keyframe and face-swapping it across frames.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The reported ID_sim metric is the exact max-ArcFace similarity that the personalized aging loss optimizes (Eq. 5 vs. Eq. 8), and Appendix E concedes ArcFace can be gamed; the headline identity advantage may reflect metric alignment rather than genuine re-aging.","rationale":"The reader's verdict is CONDITIONAL, with the weakest assumption about biased personal photo collections (metadata, GFPGAN restoration, retouching). I agree that data bias is real, but I do not think it is the most load-bearing issue: the paper's non-celebrity experiment (Table 3) partially addresses transfer beyond curated celebrity images, and the limitations section is candid about known artifacts. The sharper problem is that the quantitative identity metric is the same function the adapter is trained to maximize. The paper does not mention this overlap in Sec. 4.1.4; in fact, Appendix E shows that a competing personalization baseline can achieve a high ID_sim while being an 'adversarial example' for ArcFace, which undermines the metric's validity for exactly the claim being made. This does not prove MyTM is bad; the adapter design, the three losses, and the ablation study are coherent, and age accuracy measured by FP-Age is a somewhat independent signal. But it means the central claim 'identity-preserving re-aging that resembles the person at the target age' is not established by Tables 1-2 as reported. The single check that would settle this is an independent identity metric, which would separate genuine identity preservation from optimization of the evaluation metric. Before that, plus code/data release, the CONDITIONAL verdict is appropriate; I do not see a reason to change it to REJECT because the qualitative results and non-celebrity numbers provide partial support.","tokens_in":23165,"tokens_out":6725,"duration_ms":66197,"concrete_test":"Re-evaluate Tables 1 and 2 using an identity measure never used in training: e.g., a different face-recognition backbone trained on a disjoint dataset (not ArcFace) and mean (not max) cosine similarity over reference images. If MyTM's ID_sim advantage over FADING and FADING+Dreambooth shrinks or reverses, the identity-preservation claim rests on metric alignment. Also score the 'SAM Pers. f.t.+MyStyle' outputs from Fig. 11 with this independent metric; if its high ID_sim drops, that confirms the original ArcFace score was inflated.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing weakness is not primarily data authenticity; it is that the headline identity metric is the training objective. The personalized aging loss (Eq. 5) is 1 - max_j <R(y_p_tgt), R(x_j)> over reference images within ±3 years of the target age. The reported identity metric, ID_sim (Eq. 8), is max_j <R(y_tgt), R(x_j)> over the same type of reference window, averaged over target ages. The only distinction stated is that evaluation reference images are 'not used in training,' but they come from the same curated celebrity collection and the same GFPGAN restoration/alignment pipeline (Appendix B), so optimizing Eq. 5 on that distribution transfers directly to Eq. 8. Baselines like SAM and FADING are not trained with this objective, which biases the reported 11.7% ID_sim advantage. Appendix E itself warns that SAM Pers. f.t. + MyStyle achieves high ID_sim while producing 'adversarial examples for arcface' with visibly poor quality, demonstrating that ID_sim can be inflated by non-identity factors. The user study (Fig. 4) is small and reports no significance tests, so it does not independently resolve the issue. The architectural idea and ablations are plausible, but Table 1/Table 2 identity numbers cannot carry the central claim as currently reported.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes MyTimeMachine (MyTM), a personalized facial age transformation method built on the global aging encoder SAM and a fixed StyleGAN2 decoder. For a given person, MyTM trains a per-identity adapter network that predicts an offset to SAM's W+ latent code conditioned on the target age, using three losses: a personalized aging loss that maximizes ArcFace similarity to reference images of the same person near the target age (Eq. 5), an extrapolation regularization that keeps outputs close to SAM's global prediction outside the training age range (Eq. 6), and an adaptive w-norm regularizer (Eq. 7). The method is evaluated for age regression (de-aging) and progression on a curated set of 12 celebrities and 5 non-celebrities, with an additional video re-aging extension based on face-swapping. The paper claims that with 10-50 personal images spanning 20-40 years, MyTM produces re-aged images that are both age-accurate and identity-preserving, outperforming global and naive-personalization baselines.","tokens_in":23428,"tokens_out":7248,"duration_ms":67101,"significance":"Practical significance is potentially high: it addresses an applied problem (VFX, forensic aging) where personal photo collections are available, and the proposed adapter-plus-regularization recipe is simple and clearly specified. The ablation in Fig. 8 and the dataset-size study in Fig. 7 are useful and support the claim that each component contributes and that around 50 images is a reasonable operating point. The video extension via face-swapping is pragmatic, and the user-study preference data, though small, is directionally favorable. However, the identity-preservation evidence is currently undermined by the fact that the headline metric ID_sim (Eq. 8) is the same quantity that the personalized aging loss (Eq. 5) optimizes, and the quantitative evaluation rests on aggregate means over 8-12 subjects without error bars. These issues must be resolved before the central claims can be considered established.","major_comments":[{"comment":"The personalized aging loss is L_pers-age = 1 - max_j <R(y^p_tgt), R(x_j)> over a reference set D_tgt defined by a_tgt +/- 3 years, and the reported identity metric is ID_sim = max_j <R(y_tgt), R(x_j)> over the same kind of reference window, averaged over target ages. The only stated distinction is that the evaluation reference images are not used in training, but they come from the same curated collection and the same GFPGAN restoration/alignment pipeline (Appendix B). Consequently, MyTM is directly trained to optimize the very quantity on which its identity-preservation advantage is reported (e.g., Table 1: 0.67 vs 0.60; Table 2: 0.70 vs 0.62), while the baselines are not trained with this objective. This is load-bearing because identity preservation is one of the two headline claims. Please re-evaluate with an identity metric that is not part of the training objective (for example, a different face-recognition backbone or human 'same identity' judgments), and report per-celebrity results with variance. The max over the reference set can also be inflated by a single outlier reference, so a mean or median over the reference set should be reported alongside it.","section":"§3.3.1, Eq. (5); §4.1.4, Eq. (8)"},{"comment":"The quantitative claims are supported by aggregate means over 12 celebrities (8-10 per task) with no standard deviations, confidence intervals, significance tests, or per-subject breakdowns, and the dataset is not released. The ablation in Fig. 8 is additionally performed on a single celebrity (Al Pacino). The preprocessing pipeline restores low-quality and grayscale images with GFPGAN, which can alter aging texture (wrinkles, sagging), and ages are inferred from photo metadata where available; both steps can bias the personalized aging trajectory. Please report per-celebrity numbers with error bars, clarify how age labels were obtained and verified, and make the dataset or a reproducible subset available. Without these, the general claim that 'as few as 50 images' suffices is not established beyond anecdote.","section":"§4.1.1, Appendix B"},{"comment":"The paper's own comparison shows that SAM Pers. f.t. + MyStyle achieves a high ID_sim (0.60) while producing images described as 'adversarial examples for arcface' with visibly poor quality. This is internal evidence that ID_sim can be inflated by non-identity factors, so ID_sim alone cannot carry the identity-preservation claim. The user study in Fig. 4 asks participants to match the reference images while preserving style, which conflates target-age resemblance with identity preservation, and it reports no significance tests. Please add a human evaluation that isolates identity (e.g., forced-choice 'same person?' at matched apparent age) and report the number of participants, per-condition counts, and confidence intervals.","section":"Appendix E, Fig. 11"},{"comment":"The reported aggregates mix interpolation and extrapolation targets. For age regression, Ours (30~70) evaluated at a_tgt <= 70 includes target ages within the training range (30-70) and extrapolation ages (0-20); for age progression, targets 40-100 include the boundary age 40 and extrapolation beyond it. Since extrapolation is a central claimed advantage, please report interpolation and extrapolation performance separately for both tasks, in addition to the pooled numbers.","section":"§4.2.1, Tables 1-2"}],"minor_comments":[{"comment":"The section title 'Loss Funtions' should be 'Loss Functions'.","section":"§3.3"},{"comment":"The caption says 'The input test images match those in Fig. 10 (top)' but should refer to Fig. 9 for the pre-trained baseline comparison.","section":"Fig. 10 caption"},{"comment":"The adaptive w-norm schedule is written as lambda_reg(Delta_age) = 1 - <pi * Delta_age / 100>; please specify whether <·> denotes cosine, state the intended domain of Delta_age, and confirm the sign and range of lambda_reg, since the text describes it as a cosine function.","section":"Eq. (7)"},{"comment":"The abstract says 'as few as 50 images' while the introduction says 'as few as 10, ideally 50'; please align these statements with the dataset-size results in Fig. 7.","section":"Abstract and Section 1"},{"comment":"Age ranges such as '01~76' and '03~96' should be written without leading zeros (e.g., '1~76', '3~96').","section":"Table 5"},{"comment":"For many celebrity photos, capture-time metadata is unavailable; state how age labels were obtained or verified beyond metadata.","section":"Appendix B"}],"recommendation":"major_revision","confidential_remarks":"The paper fits ACM TOG and the core idea is worth pursuing. My main concern is the metric/training-objective overlap; I would like the revision to re-run the identity evaluation with an independent metric and report per-subject variability. I also hope the authors will release the celebrity dataset or a sufficiently large anonymized subset, since the current evaluation cannot be reproduced otherwise. The related-work coverage is adequate and I see no attribution problems."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague, here's my read on 2411.14521.\n\nThe core idea is genuinely new: a lightweight MLP adapter that adds a personalized offset to SAM's global W+ latent code, trained on a few dozen personal photos, with three losses—personalized aging loss (max ArcFace similarity to reference images within ±3 years of the target age), extrapolation replay to SAM, and an adaptive w-norm cosine schedule. The ablations are clean and each piece earns its place: the extrapolation loss alone cuts Age_MAE from 17.9 to 11.0. The video extension via face-swapping is pragmatic and they don't oversell it. The limitations section is honest (glasses, hair color, red-eye).\n\nThe load-bearing problem is the identity metric. Eq. 5 is 1 - max ArcFace similarity to reference images near the target age; Eq. 8, the reported ID_sim, is exactly that max similarity. Evaluation reference images are said not to be used in training, but they come from the same curated celebrity collection, through the same GFPGAN restoration and alignment pipeline. So the model is being scored on the same objective it was trained on, and baselines are not trained with that objective. That inflates the reported 11.7% ID_sim advantage. The paper itself shows the danger in Appendix E: SAM Pers. f.t. + MyStyle gets high ID_sim while producing adversarial examples with poor visual quality. The user study is small (10 pairs per task, ~24-29 responses per comparison) and reports no significance tests, so it doesn't independently settle identity preservation.\n\nOther concerns: 12 celebrities, no error bars, dataset not released, code not provided. These are addressable but they limit what we can verify.\n\nI don't think this is fatal. The architecture is sensible and the losses are well-motivated. But the headline quantitative claim—that MyTM has high identity preservation—is not supported by the current numbers. The metric needs to be decoupled from the training loss, e.g., a different face recognizer trained on a different distribution, or reference images from outside the curated set. Until then, treat the identity numbers as evidence of objective alignment, not of genuine re-aging.\n\nWho's the paper for? Anyone working on StyleGAN-based face editing or personalized generation. The adapter idea will likely survive the metric fix. I'd send it to peer review—it deserves a serious referee—but my own verdict would be a conditional accept at best, with the metric redesign as the main revision. I wouldn't cite the current numbers in my own work until they fix the evaluation.","headline":"Novel adapter-based personalization for face aging, but the headline identity metric is the training objective, so the central quantitative claim doesn't hold as reported.","tokens_in":24004,"tokens_out":2789,"would_cite":false,"duration_ms":25867,"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":"A small personal photo album—as few as 50 images spanning decades—is enough to synthesize how a specific face ages, both into the past and the future.","keywords":["facial age transformation","personalization","age progression","age regression","adapter network","StyleGAN2","latent space editing","identity preservation"],"falsifier":"Hold out a person's childhood photos, train MyTM only on photos from ages 50–70, de-age a 70-year-old test photo to age 10, and compare the output side-by-side with the withheld childhood photo; if a substantial fraction of subjects produce outputs that do not resemble the withheld early photos—measured by ArcFace similarity against the actual childhood reference set falling below the paper's reported interpolation levels—then the extrapolation claim is falsified.","tokens_in":22941,"feed_emoji":"⏳","tokens_out":7246,"duration_ms":63033,"temperature":0.7,"pith_summary":"This paper argues that accurate facial aging for a specific person does not require a giant dataset—only a few dozen photos of that person taken over 20 to 40 years. The method, MyTimeMachine (MyTM), starts from a global aging model that knows how faces age on average and adds a lightweight person-specific adapter network that corrects that average trajectory using the personal photo collection. With as few as 50 images, the resulting re-aged faces both land at the requested target age and resemble how that individual actually looked at that age, and the same personalization extends to video. This matters because movie de-aging, missing-person aging, and personal time-machine apps could then work from an ordinary photo album instead of generic templates that produce plausible but wrong faces.","feed_headline":"A 50-photo album yields realistic person-specific face aging","feed_subtitle":"Using as few as 50 images across decades, the method de-ages or ages a person in a way that preserves identity.","key_machinery":"The load-bearing object is the Age Adapter Network, a family of 18 style-wise MLPs (plus a shared global MLP and an aging MLP) that reads SAM's W+ latent code and the target age and emits a per-style latent offset ΔW+. The offset is added to the global code before the frozen StyleGAN2 decoder, so personalization never retrains the generator or the age encoder. The argument is carried equally by the three losses: the personalized aging loss uses an ArcFace reference set near the target age to anchor identity without relying on a possibly biased global age classifier; the extrapolation regularization replays SAM's output for out-of-range target ages to prevent overfitting to the training age boundaries; and the adaptive w-norm regularization, with a cosine schedule on the age gap, keeps latent codes near the data manifold for editing while staying close enough to SAM's inversion for fidelity.","core_discovery":"The central claim is that a personalized facial age transformation can be learned from a small longitudinal photo collection (ideally 50 images spanning 20–40 years) by combining two sources of knowledge: a global aging prior learned from many identities, and a personal adapter that corrects that prior for one individual. MyTM builds on SAM, a StyleGAN2-based encoder that maps an input face and a target age into the W+ latent space. A new Age Adapter Network—a set of MLPs—takes SAM's predicted latent code and the target age and outputs a latent offset; adding that offset to the global code and decoding with the frozen StyleGAN2 generator produces the personalized re-aged face. Three losses drive personalization: a personalized aging loss that pulls the output's identity features toward the person's own photos near the target age (using maximum ArcFace cosine similarity over a reference set), an extrapolation regularization that keeps outputs close to the global model when the target age lies outside the training range (experience replay), and an adaptive w-norm regularization that schedules proximity to the mean latent code as a cosine function of the input-to-target age gap to balance inversion accuracy and editability. On a curated 12-celebrity longitudinal dataset, MyTM reports age errors of 7.7 (regression) and 6.3 (progression) MAE with identity similarities around 0.76–0.78, outperforming SAM, CUSP, AgeTransGAN, FADING, and naive personalization baselines such as DreamBooth fine-tuning.","pith_inferences":["The adapter-plus-reference-set pattern suggests a general recipe for personalizing other attribute editors (e.g., weight, hairstyle, or skin health) with a handful of exemplars; the paper demonstrates it for age, but the machinery is attribute-agnostic.","The adaptive w-norm cosine schedule implies that a continuous trade-off curve between inversion fidelity and editability can be navigated per-example; a testable extension would be to replace the cosine with a learned schedule or apply the same schedule to other latent-editing tasks.","The video result suggests temporal consistency can be decoupled from the aging model itself by delegating it to a face-swapper, and if face-swapping improves, the same pipeline could handle other identity-preserving edits such as expression or lighting changes.","A practical caution follows from the weakest assumption: the method's reliability on real photo albums depends on restoration not erasing wrinkles or sagging; a robustness test would re-run the experiments with different restoration settings to quantify this sensitivity."],"forward_implications":["With roughly 50 personal photos spanning decades, one can generate de-aged actors for VFX without generic aging templates, and age people forward for missing-person or forensic use.","The same trained model extrapolates beyond the photo collection's age range (e.g., trained on ages 20–40, aged to 100) with a reported age MAE of 6.3, so the personalization is not limited to interpolation.","Personalization transfers to video by re-aging a single keyframe and face-swapping into the rest of the clip, giving temporally consistent results in under 5 minutes versus over 3 hours for direct video editing baselines.","Naive personalization of global models (DreamBooth-style fine-tuning) overfits to the training age range, whereas MyTM's combination of personal and global priors avoids that failure.","The improvement holds for non-celebrity YouTube subjects, not just curated celebrity photos, indicating the approach is not limited to professionally captured faces."],"supporting_citations":[{"why":"Supplies SAM, the global age encoder and StyleGAN2 latent-space prior that MyTM personalizes.","marker":"[Alaluf et al. 2021]"},{"why":"Provides the FFHQ dataset and StyleGAN2 backbone whose frozen decoder renders the re-aged faces.","marker":"[Karras et al. 2021]"},{"why":"ArcFace identity features anchor the personalized aging loss and the ID_sim evaluation metric.","marker":"[Deng et al. 2022]"},{"why":"The e4e encoder defines the inversion-editability trade-off that the adaptive w-norm regularization addresses.","marker":"[Tov et al. 2021]"},{"why":"DreamBooth is the naive personalization baseline that overfits to the training age range, motivating the adapter design.","marker":"[Ruiz et al. 2023]"},{"why":"FADING is the diffusion-based state-of-the-art baseline MyTM must beat on age accuracy and identity preservation.","marker":"[Chen and Lathuilière 2023]"},{"why":"Source of the w-norm regularization idea that MyTM adapts with a cosine schedule on the age gap.","marker":"[Richardson et al. 2021]"},{"why":"GFPGAN restores old low-quality personal photos during preprocessing, which shapes the aging cues the adapter learns.","marker":"[Wang et al. 2021]"}],"fun_headline_variants":["50 photos personalize face aging","Personalized aging from a 50-photo album","MyTimeMachine turns 50 photos into custom aging","A tiny photo set enables personalized face aging"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The personal photo collection must genuinely represent how the person aged: ages come from metadata and old photos are restored with an external enhancer (GFPGAN), so if the collection is retouched, professionally lit, or biased in pose or lighting, the adapter may learn artifacts rather than the true aging trajectory, and the claimed resemblance at the target age fails.","fun_headline_variants_meta":{"raw":{"variants":["50 photos personalize face aging","Personalized aging from a 50-photo album","MyTimeMachine turns 50 photos into custom aging","A tiny photo set enables personalized face aging"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000977,"raw_usage":{"total_tokens":4228,"prompt_tokens":1100,"completion_tokens":3128,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":716,"completion_tokens_details":{"reasoning_tokens":3084}},"tokens_in":716,"tokens_out":3128,"duration_ms":20276,"temperature":1.0,"reasoning_tokens":3084,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T15:12:17.007576+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Hold out a person's childhood photos, train MyTM only on photos from ages 50–70, de-age a 70-year-old test photo to age 10, and compare the output side-by-side with the withheld childhood photo; if a substantial fraction of subjects produce outputs that do not resemble the withheld early photos—measured by ArcFace similarity against the actual childhood reference set falling below the paper's reported interpolation levels—then the extrapolation claim is falsified.","supporting_citations":[{"cited_title":"Designing an encoder for StyleGAN image manipulation","cited_arxiv_id":null,"evidence_quote":"The e4e encoder defines the inversion-editability trade-off that the adaptive w-norm regularization addresses."},{"cited_title":"Encoding in Style : a StyleGAN Encoder for Image -to- Image Translation","cited_arxiv_id":null,"evidence_quote":"Source of the w-norm regularization idea that MyTM adapts with a cosine schedule on the age gap."}],"review_version":1}