{"id":"0df24922-2cc9-4891-ae72-3fdd0491491b","arxiv_id":"2505.09256","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"Pose-TTA improves pre-trained face recognition at inference by generating matching side-profile views with a portrait animator and aggregating real and synthetic embeddings with fixed weights.","lead":"This paper introduces a test-time method that improves face recognition by generating matching side-profile images with a portrait animator, avoiding distorted frontal views. A weighted mix of real and synthetic face embeddings boosts accuracy on pose-heavy benchmarks without retraining the recognition model.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Identity preservation of the generated side-profile images is asserted but never measured; if LivePortrait leaks driving identity, the weighted aggregation cannot support the claimed gains.","rationale":"I agree with the reader's weakest_assumption: the method's validity hinges on whether LivePortrait-generated images preserve source identity. The paper never quantifies this. The qualitative figure (Fig. 2) is insufficient because a face that looks reasonable to a human can still contain appearance cues from the driving identity that distort the embedding space. This concern is load-bearing because the entire contribution is the use of synthetic side-profile views; if those views are not faithful to the source identity, the weighted aggregation in Eq. (2) is not aligning two views of the same person but blending identities. The fact that reported gains are positive on pose-heavy benchmarks does not rule out leakage, since the fixed weights and the flip-based baseline could mask or even exploit the leakage. The proposed test—measuring embedding similarity margins between source, generated, and driving images—would settle whether the mechanism operates as described. I do not see an internal inconsistency in the method; the paper is transparent about frontal-heavy datasets showing marginal or negative changes. The main weaknesses are the unmeasured identity-preservation assumption, the target-benchmark hyperparameter choice, and the lack of released code/significance estimates, all of which support a conditional verdict. Therefore I recommend no change to the reader's conditional verdict.","tokens_in":11456,"tokens_out":7997,"duration_ms":78471,"concrete_test":"Take a random subset of 1,000 pairs from CPLFW and, for each, compute the source image Is, the driving image Id, and the generated image Î using the paper's pipeline. Using each pre-trained backbone (e.g., AdaFace ResNet101), compute cosine similarity between embeddings: s_src_gen = cos(Θ(Is), Θ(Î)), s_src_drv = cos(Θ(Is), Θ(Id)), and s_src_self = cos(Θ(Is), Θ(Is)). If the average (s_src_gen - s_src_drv) is not positive, or if s_src_gen is much closer to s_src_drv than to s_src_self, identity is not preserved. Report these numbers for same-identity and different-identity pairs separately; if the margin is negative on different-identity pairs, the synthetic augmentation is actively harmful and the reported net gains need re-evaluation.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Pose-TTA's central mechanism is that LivePortrait, after the optional horizontal flip in Section II-A, renders the source identity in the driving pose, and that this generated image behaves like another view of the same person. This is the premise that makes the weighted aggregation in Eq. (2) meaningful. The paper provides only qualitative evidence (Fig. 2) and no quantitative measure of identity retention. If the animator transfers appearance cues from the driving image (notably when the unseen side of the face must be hallucinated), the generated feature for a source image becomes contaminated with the driving identity. In verification, this can inflate similarity on impostor pairs; the chosen weights (wsyn=0.25) may be compensating for such leakage rather than exploiting a clean pose-aligned view. The claim in Section II-A that the method 'reduces identity information loss' is therefore unsupported. The reported gains on CPLFW and CFP-FP could still arise from the flip-based alignment alone or from weighted averaging, but without an identity-preservation measurement the paper does not establish that its defining component contributes the hypothesized mechanism.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces Pose-TTA, a test-time augmentation method for face recognition. For a pair of face images, a Face Selector uses an off-the-shelf yaw estimator to designate the more frontal image as the source and the more profile image as the driving image; LivePortrait animates the source into the driving pose, with an optional horizontal flip of the source when the yaw directions disagree. The pre-trained recognition model's embeddings for the original, flipped, synthetic, and flipped-synthetic images are aggregated with fixed weights wreal=0.75, wsyn=0.25 and L2-normalized before verification. Experiments cover five benchmarks (CPLFW, CFP-FP, LFW, CALFW, AgeDB), several backbones/losses, and training sets including the synthetic DCFace set. The paper claims consistent improvements, particularly on pose-heavy benchmarks, and compares against frontalisation and unweighted averaging.","tokens_in":11591,"tokens_out":7644,"duration_ms":72920,"significance":"The idea of generating matching side-profile views rather than frontalising is interesting and practically attractive because it requires no retraining. The evaluation across six training sets and five architectures is a strength, and Table II provides a useful comparison against frontalisation. However, the central claim is only partially supported: Table I contains multiple negative deltas, the gains are concentrated on CPLFW and CFP-FP, and no statistical significance is reported. The method also depends on two hyperparameters tuned on the same test benchmarks, and the identity-preservation premise of the portrait animator is not measured. If these issues are addressed, the contribution could be a useful incremental result for inference-time pose robustness.","major_comments":[{"comment":"The weights wreal=0.75 and wsyn=0.25 are selected by evaluating the five test benchmarks in Table III and then used to report the final results on those same benchmarks in Tables I and II. This is test-set tuning; it makes the reported gains optimistic and weakens the claim that the method is a fixed, training-free procedure. Please either select weights on a validation split or via cross-validation, or report results across the weight grid and clearly state that the reported configuration is the best on the test sets.","section":"Section III-C, Table III"},{"comment":"The paper's core premise is that LivePortrait preserves the source identity while transferring pose, but no quantitative identity-preservation measurement is provided; Figure 2 is qualitative. If the animator transfers identity cues from the driving image, the synthetic embedding used in Eq. (2) can inflate impostor similarity, and the observed gains may not reflect pose alignment. Please add a quantitative identity-retention evaluation, for example embedding similarity of the generated image to the source versus to the driving image, or verification accuracy on synthetic source/driving pairs, using the same pre-trained recognition model.","section":"Section II-A, Eq. (2), Fig. 2"},{"comment":"The statement that Pose-TTA 'consistently improves inference performance' is contradicted by Table I, which contains negative deltas on CFP-FP, LFW, CALFW, and AgeDB (for example AdaFace ResNet50 WebFace4M has -0.10 on CFP-FP, and multiple LFW and AgeDB entries are negative). The later text acknowledges marginal drops, which should be reconciled with the abstract. Moreover, no confidence intervals or significance tests are provided, so gains of +0.05 or +0.09 cannot be distinguished from noise. Please revise the claims to match the evidence and add appropriate statistical evaluation.","section":"Abstract, Section IV-A, Table I"}],"minor_comments":[{"comment":"The statement that the head pose estimator 'is not used during the pose augmentation process, thereby eliminating dependence on its performance' is inaccurate because the Face Selector uses yaw values to choose the source/driving images and to decide the horizontal flip; please revise the wording.","section":"Section II-A"},{"comment":"The caption says that baseline results 'without TTA' use the aggregation of original and flipped embeddings, but that is itself a form of test-time augmentation; please clarify that the baseline is flip-only TTA rather than 'without TTA'.","section":"Table I caption"},{"comment":"There are several proofreading issues: the average accuracy '94,97' uses a comma as a decimal separator, Fig. 1 contains the typo 'recongition', and some delta entries in Table I do not match the difference between the printed baseline and TTA accuracy columns (for example AgeDB row 1, where 98.00 to 97.87 is -0.13 but the table prints -0.03).","section":"Table I, Fig. 1"},{"comment":"The Ethical Impact Statement claims that the authors analysed Pose-TTA's performance across demographic groups, but no such subgroup analysis appears in the experiments; please either add the analysis or remove the claim.","section":"Ethical Impact Statement"},{"comment":"The section is titled 'Test Time Adaptation' although the method is test-time augmentation; please use the terminology consistently.","section":"Section II-B"},{"comment":"Since the method adds a portrait animation step at inference, reporting per-pair runtime or computational cost would help assess practical utility; no timing or cost comparison is currently provided.","section":"Section III-A"}],"recommendation":"major_revision","confidential_remarks":"The paper is within scope for FG and the idea is interesting, but the empirical support for the central claim is currently overstated and the weight selection on the test benchmarks is a concern. I recommend major revision rather than rejection because the issues can be addressed with additional experiments and a more careful presentation of the results."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's the short version: this paper does something new—using a portrait animator to synthesize matching side-profile views at test time instead of frontalizing—and tests it broadly. It's an incremental, practical contribution, not a conceptual shift. The experiments cover five backbones and six training sets, and the results on pose-heavy benchmarks (CPLFW, CFP-FP) are positive across nearly every configuration. The comparison with frontalization in Table II is a fair way to show the method's premise. Credit where it's due: the writing is clear, the method is simple, and the weighted aggregation ablation in Table III honestly shows that synthetic-only features hurt.\n\nThe soft spots are real but not fatal. First, the weights wreal=0.75 and wsyn=0.25 are selected in Table III on the same benchmarks used for the headline results, so there's test-set tuning. No held-out validation, no error bars. Given that many gains are under a percentage point and several frontal benchmarks drop (LFW, CALFW, AgeDB all have negative deltas in Table I), the abstract's 'consistently improves inference performance' is an overstatement. The honest summary is: consistent gains on pose-heavy sets, neutral to slightly negative elsewhere.\n\nSecond, the stress-test concern lands. The paper's mechanism assumes LivePortrait preserves the source identity when generating the driving pose, but identity retention is never measured—only qualitative examples in Fig. 2. The flip-based alignment is a sensible heuristic, but if the animator leaks driving-identity cues, the weighted aggregation could be compensating for contamination rather than exploiting a clean pose-aligned view. The downweighting to 0.25 for synthetic features limits the damage, but the paper should either report an identity-retention metric (e.g., verification accuracy of generated images against the source identity) or temper the 'reduces identity information loss' claim.\n\nNo code, which doesn't help reproducibility. The ethics statement about restricted access is a legitimate choice, but for a training-free inference method, code would make the contribution much easier to adopt.\n\nBottom line: this is worth a real referee. The idea is plausible, the experiments are broad, and the concerns are addressable with a small amount of additional work—held-out weight selection, error bars, and a direct identity-preservation check. If those come back clean, the paper is a solid incremental result. I'd send it out for review rather than desk-reject, and I'd suggest the reviewer focus on the weight-selection protocol and identity measurement.","headline":"Pose-TTA is a modest, clearly explained inference-time plug-in that improves pose-heavy face verification with portrait animation; the broad experiments are useful, but the 'consistently improves' claim overstates the tables and identity preservation is never directly measured.","tokens_in":12156,"tokens_out":2732,"would_cite":false,"duration_ms":25450,"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":"Pose-TTA claims that a pre-trained face recognition model can be improved at inference, without retraining, by generating a matching side-profile view of the more frontal face and aggregating the two embeddings with fixed weights.","keywords":["test-time augmentation","face recognition","pose-invariant recognition","portrait animation","feature aggregation","head pose alignment","face verification","frontalisation"],"falsifier":"Take a pose-heavy benchmark such as CPLFW and replace every driving image's identity with a different person's face while keeping the yaw angle; if the reported gain persists, pose alignment is doing the work, whereas if it disappears, the animator is injecting driving-identity information into the augmented embeddings.","tokens_in":11194,"feed_emoji":"🪞","tokens_out":7682,"duration_ms":68927,"temperature":0.7,"pith_summary":"Pose-TTA claims that a frozen face recognition model can be made more accurate on pose-diverse image pairs at inference time, with no retraining or fine-tuning. The idea is to stop trying to frontalise side-profile faces and instead use a portrait animator to generate a matching side-profile view from the more frontal image, then compare the two aligned views. A small face selector chooses which image supplies identity and which supplies head pose, and a weighted average of the original and synthetic embeddings (weight 0.75 on real, 0.25 on synthetic) keeps generative artifacts from dominating. On pose-heavy benchmarks the method reports average gains of about 0.72 percentage points on CPLFW and 0.19 on CFP-FP across several pre-trained backbones and training sets, while leaving near-frontal benchmarks roughly unchanged. If correct, this gives a plug-in inference wrapper that improves pose robustness without dataset-specific training.","feed_headline":"Synthetic profiles boost pose-heavy face recognition, no retraining","feed_subtitle":"Frozen recognition models gain about 0.72 point on CPLFW and 0.19 on CFP-FP by comparing matched side profiles.","key_machinery":"The central mechanism is the source/driving split with symmetry-guided flipping. The face with the smaller absolute yaw is treated as the identity source; the face with the larger absolute yaw contributes only its head pose. When the yaw directions disagree, the source is horizontally flipped before being fed to the portrait animator, so the two faces point roughly the same way and the generator does not have to hallucinate the occluded half of the face. The second load-bearing piece is the fixed-weight aggregation: synthetic embeddings are down-weighted ($w_{syn}=0.25$, $w_{real}=0.75$) relative to the real image, and the combined vector is normalised along the channel axis before verification. Together these pieces convert test-time augmentation from a generic averaging step into a pose-alignment step that preserves identity information.","core_discovery":"On its own terms, the paper's claim is that pose-invariant face recognition can be decomposed at test time into two steps: generate, rather than reconstruct. Given a pair of faces, the face selector uses an off-the-shelf head pose estimator to label the image with smaller absolute yaw as the source (identity) and the image with larger absolute yaw as the driving image (pose only). If the two yaw signs differ, the source is horizontally flipped so the animator transfers pose without having to invent the unseen side of the face. The animator copies only head pose, not expression, so identity-related details such as eye and mouth shape are not biased toward the driving person. Both the original and the generated images, plus their horizontal flips, are embedded by the frozen recognition model and combined with fixed weights $\\hat{\\mathbf{y}}_{tta} = \\frac{1}{|T|}\\sum_i w_T \\Theta_{target}(\\tau_i(x))$, where $w_{real}=0.75$ and $w_{syn}=0.25$, followed by channel-wise normalisation. The reported result is that this pipeline consistently improves accuracy on pose-heavy benchmarks while barely changing frontal-dominant ones.","pith_inferences":["A natural test of the mechanism is to scramble the driving image's identity: if pose alignment alone drives the gain, the reported improvement should survive; if it collapses, the animator is leaking identity from the driving face and the aggregation is mainly masking that leakage.","The fixed 0.75/0.25 weights are tuned on these benchmarks; an input-dependent reliability estimate (for instance, from the animator's own stitching confidence) could recover part of the small losses seen on frontal-heavy sets such as AgeDB-30 (-0.04).","The wrapper nature of the method means it could be stacked with future encoders or pose estimators; the only moving parts that need updating are the animator and the two weights.","A direct identity-preservation audit of the generated profiles (does the synthetic face retrieve the right person in a large gallery?) would tell whether the gains are pose alignment or synthetic-data artifacts."],"forward_implications":["A pre-trained face recognition model improves on pose-diverse pairs (average +0.72 on CPLFW, +0.19 on CFP-FP) without any training or fine-tuning.","Frontalisation at test time is the wrong comparison: frontalising the profile hurts pose-heavy accuracy (about -0.12 average), while matching side profiles gains about +0.37 average.","The weighting is doing real work: equal weights (0.5/0.5) underperform 0.75/0.25, and relying entirely on synthetic images drops CPLFW accuracy from 87.87 to 85.77.","The horizontal-flip alignment contributes a small but consistent gain (about +0.04 average), so facial symmetry is part of the method, not decoration.","Because the recognition model is frozen, the wrapper can be attached to different backbones (ResNet18/50/101, ViT) and training losses (AdaFace, ArcFace) without per-dataset retraining."],"supporting_citations":[{"why":"Supplies the portrait animator that transfers source identity into the driving head pose; the core generation step of Pose-TTA.","marker":"[18]"},{"why":"Provides the off-the-shelf head pose estimator used by the Face Selector to assign source/driving roles and to decide whether to flip the source.","marker":"[41]"},{"why":"Provides the observation that averaging TTA predictions can hurt accuracy, motivating the weighted aggregation scheme.","marker":"[45]"},{"why":"Supplies the AdaFace pre-trained recognition model used as a frozen target in the main experiments.","marker":"[27]"},{"why":"Provides the CPLFW benchmark, the pose-heavy evaluation set where the largest average gain (0.72) is reported.","marker":"[64]"},{"why":"Provides the CFP-FP frontal-to-profile benchmark where the method reports an average gain of 0.19.","marker":"[43]"}],"fun_headline_variants":["Test-time pose augmentation boosts face recognition without retraining","Synthetic side profiles sharpen pose-heavy face recognition","Pose-TTA: generate, don't frontalize, for pose-invariant face ID","No retraining needed: test-time pose augmentation improves face matching"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method assumes that the portrait animator, after the optional horizontal flip, preserves identity closely enough that the generated side-profile embedding behaves like another view of the same person.","fun_headline_variants_meta":{"raw":{"variants":["Test-time pose augmentation boosts face recognition without retraining","Synthetic side profiles sharpen pose-heavy face recognition","Pose-TTA: generate, don't frontalize, for pose-invariant face ID","No retraining needed: test-time pose augmentation improves face matching"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000216,"raw_usage":{"total_tokens":1450,"prompt_tokens":981,"completion_tokens":469,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":597,"completion_tokens_details":{"reasoning_tokens":397}},"tokens_in":597,"tokens_out":469,"duration_ms":4714,"temperature":1.0,"reasoning_tokens":397,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T21:35:21.513448+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a pose-heavy benchmark such as CPLFW and replace every driving image's identity with a different person's face while keeping the yaw angle; if the reported gain persists, pose alignment is doing the work, whereas if it disappears, the animator is injecting driving-identity information into the augmented embeddings.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the off-the-shelf head pose estimator used by the Face Selector to assign source/driving roles and to decide whether to flip the source."},{"cited_title":"Shanmugam, D","cited_arxiv_id":null,"evidence_quote":"Provides the observation that averaging TTA predictions can hurt accuracy, motivating the weighted aggregation scheme."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the AdaFace pre-trained recognition model used as a frozen target in the main experiments."},{"cited_title":"Zheng and W","cited_arxiv_id":null,"evidence_quote":"Provides the CPLFW benchmark, the pose-heavy evaluation set where the largest average gain (0.72) is reported."},{"cited_title":"Sengupta, J.-C","cited_arxiv_id":null,"evidence_quote":"Provides the CFP-FP frontal-to-profile benchmark where the method reports an average gain of 0.19."}],"review_version":1}