{"id":"967697aa-fb48-4f6a-9fd5-99f3f88397e3","arxiv_id":"2608.10346","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"A single face-landmark model can train on multiple differently annotated datasets and output arbitrary queryable landmarks by representing each landmark as a normalized position along a face part contour.","lead":"This paper introduces a way to represent any facial landmark as a number between 0 and 1 along the outline of a face part, such as an eyebrow or lip. Using this representation, a single model can be trained on several differently annotated face datasets at once and can output any requested landmarks at runtime.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Table 7's held-out landmark test may not actually withhold the tested FPALPs: with fused AFLW19+300W+WFLW training, many 'held-out' 300W landmarks are still supervised through WFLW (and vice versa), so the reported edge over spline could be memorization unless a global FPALP blacklist was used.","rationale":"The reader's weakest assumption concerned the semantic consistency of the unified face template; my concern is a different but adjacent weak spot in the same dynamic-generalization story. The strongest claim includes both unified training and arbitrary dynamic queries, and Table 7 is the paper's only quantitative comparison of dynamic queries against geometric interpolation. If the masking protocol is ambiguous, the quantitative support for the dynamic contribution is unverified, even though Tables 2-4 and the WFLWE result in Table 6b remain largely intact. This does not change the overall CONDITIONAL verdict, but it should be added as an explicit acceptance condition: the authors must clarify or correct the held-out protocol. The concern is concrete and testable, and it is not an attack on the authors: it is a request for a missing experimental detail that determines whether a key comparison is valid. A single-dataset training setup or a global FPALP blacklist would fully resolve it, and the WFLWE zero-shot result provides a fallback line of evidence for dynamic generalization.","tokens_in":22029,"tokens_out":13934,"duration_ms":130159,"concrete_test":"Rerun the three reduced-supervision settings of Table 7 with an explicit global FPALP blacklist: for every dataset in the training fusion, remove all training samples whose annotated FPALP value falls in the held-out set, so a withheld 300W contour landmark is also withheld when it appears in WFLW. Then remeasure direct FPALP queries versus cubic-spline NME. If the gap disappears or shrinks materially, the dynamic-generalization claim must be weakened; if the gap persists, the concern is resolved. The authors should also state in the table caption whether the reduced-supervision model was trained on a single dataset or on the full fusion.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central dynamic-FLD evidence is the controlled held-out experiment in §A.6/Table 7, which is explicitly used to argue that the model 'learns more than a geometric interpolation rule.' The text says a subset of native 300W/WFLW landmarks is withheld during training, making the queried FPALPs 'unseen,' but it does not specify whether the reduced-supervision model is trained on the full fusion (AFLW19+300W+WFLW) or on a single dataset. Under the fused protocol, withholding a 300W landmark does not make its FPALP unseen: for example, every 300W face-contour FPALP (0/32, 2/32, ..., 32/32) is also present in WFLW's 33-point contour (Tables 12 and 13), and 300W's outer/inner lip positions coincide exactly with WFLW's outer/inner lip positions. Conversely, when WFLW landmarks are withheld, roughly half of the common 68-landmark positions are still supervised through 300W. Thus the reported 19.0%/16.3%/15.7% improvements over cubic splines could be inflated by the model having seen those exact FPALP-to-coordinate mappings in another dataset. If the authors instead trained those rows on a single dataset only, that needs to be stated; as written, the dynamic-generalization claim is not cleanly established by Table 7. This does not invalidate the WFLWE zero-shot result in Table 6b, where the 28 queried points are genuinely absent from 300W training, but Table 7 is the only head-to-head quantitative comparison against interpolation and should not rest on an ambiguous masking protocol.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Face Part-Anchored Landmark Positions (FPALPs), in which each facial landmark is represented as a normalized progression value along a user-defined face-part contour, together with a query-based transformer decoder that takes FPALP-plus-text embeddings as input and predicts the corresponding landmark coordinates. The authors claim two system-level benefits: unified training, in which one model can be trained end-to-end on a fusion of differently annotated face-landmark datasets (AFLW19, 300W, WFLW), and dynamic prediction, in which landmarks can be queried at arbitrary FPALPs at inference time without retraining. Experiments compare the method with state-of-the-art detectors on WFLW, 300W, AFLW-19, and cross-dataset targets (COFW, COFW68, WFLW68), ablate the training data, text encoder, image encoder, and decoder depth, and include a held-out landmark experiment against cubic-spline interpolation.","tokens_in":22376,"tokens_out":9143,"duration_ms":81197,"significance":"If the claims hold, this is a practically useful step toward a single face-landmark model that is not bound to one N-point annotation protocol: it removes the need for per-dataset heads, supports query-driven landmark selection, and makes no use of 3D priors. The paper is unusually transparent about its limitations, including template-alignment imprecision, the inability to evaluate interpolated density benchmarks, and language bias in the text encoder. The cross-dataset zero-shot evaluation on WFLWE (Table 6b) is a valuable falsifiable design, and the dataset-ablation study in Table 4 directly probes the unified-training hypothesis. However, the head-to-head evidence against geometric interpolation in Table 7 is currently ambiguous because the masking protocol may leak FPALPs across datasets, and all performance claims rest on single-run point estimates without released code or checkpoints.","major_comments":[{"comment":"The controlled held-out landmark experiment does not state whether the reduced-supervision model is trained on the fused AFLW19+300W+WFLW set or on a single dataset. Under fused training, the withheld FPALPs are not necessarily unseen: Tables 12 and 13 show that every 300W face-contour FPALP (0/32, 2/32, …, 32/32) and all 300W outer/inner-lip FPALPs also occur in WFLW, and most 300W eyebrow, nose-bridge, and nose-boundary FPALPs coincide as well. A model trained on the fusion could therefore memorize the exact FPALP-to-coordinate mapping from another dataset, and the reported 19.0%, 16.3%, and 15.7% improvements over cubic splines would be inflated. The authors should explicitly state the training protocol for the reduced-supervision rows, and if fused training was used, re-run the experiment with a global FPALP blacklist across all datasets, or use single-dataset training for this comparison.","section":"§A.6, Table 7; Tables 12–13"},{"comment":"All quantitative results are single-run point estimates without standard deviations, confidence intervals, or significance tests, and the SOTA comparison in Table 2 cites previously reported numbers without shared evaluation infrastructure or code release. Since several claimed advantages are small (for example, 2.80 vs. 2.78 NME on 300W, and 4.05 vs. 4.02 on WFLW), the paper should report multiple seeds, provide error bars, or explicitly temper the claim of consistent superiority over prior methods; otherwise the comparative claims are not fully supported.","section":"Table 2; Section 4.1"}],"minor_comments":[{"comment":"The sentence before Tables 10–13 says the mapping is tabulated for “the AFLW, COFW, 300W, and 300W datasets,” but the fourth dataset should be WFLW, and “Tab. Table 10-13” is ungrammatical.","section":"§A.16"},{"comment":"The column headers in Table 1 run together (“TrainingCycles InferenceCalculation StorageParameterLandmarkThroughput”); they need spacing to be readable.","section":"Table 1"},{"comment":"Equation (8) should define the terms “RelativePosition” and “index” formally and state explicitly that the denominator Np in FPALP = pos/(Np−1) is the number of landmarks in the unified face-part template Tp, since the current notation is ambiguous about whether Np refers to the dataset-specific or unified template.","section":"§A.9, Eq. (8)"},{"comment":"The description “alternating landmark indices within the face-part definitions” should specify whether the alternation is applied to the original dataset ID ordering or to the sorted FPALP ordering; these two orderings differ for non-monotone contours such as closed eye and lip curves.","section":"§A.6, Table 7"},{"comment":"The set of 28 WFLW landmarks defining WFLWE should be enumerated or otherwise unambiguously specified, because some WFLW positions share an FPALP value with a 300W position (for example, the eye point at 4/8 equals the 300W eye point at 3/6), and it is not clear whether those shared-value points are included in the “absent/undefined” set.","section":"Table 6b and §A.6"},{"comment":"The paper would benefit from releasing the FPALP mapping tables, dataset-adapter configurations, and trained checkpoints; given that the method depends on the manual face-part and start/end ordering choices, code or detailed mapping files are important for reproducing Table 2 and Table 7.","section":"General"}],"recommendation":"major_revision","confidential_remarks":"The deciding issue is the Table 7 masking protocol. If the reduced-supervision rows were trained on the fused dataset without a global FPALP blacklist, then the head-to-head comparison against interpolation is not valid as written, and the dynamic-generalization claim loses its strongest quantitative support. The zero-shot WFLWE result is better controlled, but the ambiguity in Table 7 must be resolved before publication. I would also encourage the editor to weigh the absence of code and single-run metrics when judging the strength of the SOTA-comparative claims."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's my take on arXiv:2608.10346. The paper's real contribution is FPALP, a simple 1D normalization of landmarks along face-part contours, plus a text-conditioned query decoder that lets one model train on multiple N-point landmark datasets and query arbitrary landmarks at inference. That combination is genuinely new relative to LAB, LDDMM-Face, CLD, and the authors are appropriately measured about it being a system-level contribution rather than a SOTA-chasing one. The fused-training results on WFLW, 300W, and AFLW-19 are competitive, and the adapter-separated comparison is good practice.\n\nThe experiments are reasonably comprehensive: cross-dataset evaluation, dataset ablations, backbone ablations, and a controlled held-out experiment against cubic splines. The paper is honest about limitations—template alignment is approximate, and they explicitly decline to evaluate on interpolation-based high-density benchmarks because evenly spaced FPALPs may not match the interpolation sampling. That kind of self-awareness makes me trust the rest more.\n\nNow the soft spots. The biggest one is the held-out landmark experiment in §A.6/Table 7. The text doesn't specify whether the reduced-supervision model was trained on the fused set or on a single dataset. If it's fused, many 'held-out' 300W landmarks are still supervised through WFLW's denser 98-point template, and vice versa—e.g., every even-indexed face contour FPALP in 300W also appears in WFLW's contour. The reported 15–19% edge over spline could then be memorization, not generalization to unseen FPALPs. The authors need to state the training protocol and, if fused, blacklist all occurrences of the held-out FPALPs globally or use a single-dataset setting. Table 6b's WFLWE zero-shot is cleaner and helps, but Table 7 is the only head-to-head versus interpolation and needs to be fixed.\n\nAlso minor: no code or checkpoints, and all metrics are single-run point estimates. The template alignment algorithm is described only loosely; the 2.22-pixel intra-cluster average helps, but the clustering initialization and manual start/end ordering are not fully specified.\n\nOverall, the central claim—that FPALPs enable useful unified and dynamic FLD—holds up in shape, but the strongest evidence for dynamic generalization is currently ambiguous. This deserves a serious referee and a request for revision, not a desk reject.\n\nRecommendation: send to peer review, ask for clarity on the masking protocol, release code/checkpoints, and add error bars.","headline":"A solid incremental contribution to FLD unification; the central dynamic-query evidence is currently undermined by an ambiguous held-out protocol that needs clarification.","tokens_in":22924,"tokens_out":2736,"would_cite":true,"duration_ms":22882,"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 single model can be trained on any mixture of face landmark datasets and answer arbitrary landmark queries at runtime.","keywords":["face landmark detection","Face Part-Anchored Landmark Positions","unified face template","dynamic landmark prediction","cross-dataset generalization","transformer decoder","landmark queries","face alignment"],"falsifier":"Take a fourth landmark dataset whose contour ordering or point spacing differs from the three used here, align it into the unified template, and measure the average intra-cluster distance; if it climbs well above the reported 2.22 pixels, or if a model trained on that fusion loses its edge over per-dataset models on cross-template evaluations, the semantic-alignment premise fails.","tokens_in":21819,"feed_emoji":"📍","tokens_out":7900,"duration_ms":66628,"temperature":0.7,"pith_summary":"Face landmark datasets disagree on how many points define a face and what each point means, so today one model must be trained per dataset and it can only emit that dataset's fixed set of points. This paper tries to remove both restrictions with a single representation: every landmark is described as a progression value between 0 and 1 along the contour of a face part (eye, lip, jawline, etc.), a value the authors call FPALP. Because any landmark from any dataset can be written this way, the datasets can be merged and one network trained on the union; at runtime, the user supplies FPALP-based queries for exactly the landmarks wanted, and the same network predicts them. The paper shows this unified dynamic model matches or beats prior specialist models on WFLW, 300W, and AFLW-19 while also predicting landmarks that were never seen during training, such as denser or partial face-part layouts. A reader should care because this collapses many per-dataset models into one trainable, queryable detector and points toward a common interface for face geometry.","feed_headline":"One model learns every face-landmark dataset at once","feed_subtitle":"Landmarks become positions on face-part curves, so one network outputs any requested points at runtime.","key_machinery":"The load-bearing object is the Face Part-Anchored Landmark Position (FPALP), a normalized coordinate that locates any facial landmark as a value in [0,1] representing its fractional distance from the start to the end of a user-defined face-part contour (eyes, brows, lips, nose, face contour). To build it across datasets, the authors first make a unified face template T_U by clustering the landmarks of AFLW-19, 300W, and WFLW, yielding tight clusters with average intra-cluster distance 2.22 pixels; each cluster becomes a position on a face-part curve. Closed contours duplicate their start landmark as the end so that fraction 0 and 1 meet. That FPALP value, added to a text embedding of the face part name, forms an image-agnostic query; a cross-modality decoder, combining self-attention, deformable image cross-attention, and text cross-attention, refines the query and its coordinate prediction over several layers. The FPALP is what lets different annotation schemes share one parameter set, and the text-plus-progression query is what lets the number of output landmarks be chosen freely at runtime.","core_discovery":"The central claim is that a single end-to-end model can be trained on the fusion of multiple N-point face landmark datasets and then, at inference, answer arbitrary landmark queries, producing any number of landmarks the user asks for. The discovery supporting this is that the landmark definitions of AFLW-19, 300W, and WFLW, though formally different, are semantically aligned through shared face parts: the same physical points recur across datasets and are roughly evenly spaced along each part's contour. The paper encodes each landmark as an FPALP, a fraction between 0 and 1 of the way along its containing face-part curve, and uses the FPALP plus the face part's text embedding as a query. A decoder refines the query against the image and outputs coordinates. On the full benchmarks the model matches or outperforms prior per-dataset specialists (e.g., 4.05 NME on WFLW without dataset adapters, 2.47 on 300W common, 1.02 NMEdiag on AFLW-19), and with lightweight per-dataset adapters it improves further, while a 300W-only model transfers to unseen templates such as WFLW68 and even to the 28 WFLW landmarks absent from 300W.","pith_inferences":["If the FPALP alignment stays semantically consistent as more datasets are added, the unified template could keep absorbing new annotation schemes without retraining from scratch, a scale-up the paper only demonstrates for three datasets.","The same query mechanism points toward open-vocabulary face parts: a stronger multilingual text encoder might let a user name a novel facial region and receive landmarks for it, which the paper lists as future work.","A sharp stress test would be to fuse a dataset whose contour ordering runs opposite to the three used here; if the average intra-cluster distance grows well beyond 2.22 pixels, unified training should lose its edge over per-dataset models.","Because the paper left COFW out of training after observing degradation, the method's tolerance for annotation noise is not yet characterized; auditing label quality before fusion may be essential in practice."],"forward_implications":["One set of weights trains jointly on any number of N-point landmark datasets, so adding a dataset no longer means training a new backbone or regression head.","At inference the same network can emit sparse or dense layouts, from four points per face part to high granularity multipliers, by loading different FPALP queries, without retraining.","Joint training on varied templates regularizes the model: fused-dataset training generally beats single-dataset training and transfers to unseen templates in near zero-shot fashion.","With small per-dataset adapters, the unified model's numbers surpass the reported specialist SOTA on WFLW, 300W, and AFLW-19, showing that the shared representation retains dataset-specific precision.","Because each query is a face-part phrase plus a progression value, landmark detection becomes addressable by language-like instructions, the interface the paper says it aims to enable."],"supporting_citations":[{"why":"Supplies the AFLW 19-point face template and its evaluation protocol.","marker":"[48]"},{"why":"Supplies the 300W 68-point face template and the common/challenging test splits.","marker":"[31]"},{"why":"Supplies the WFLW 98-point face template and motivates the boundary-line, evenly spaced landmark view.","marker":"[39]"},{"why":"State-of-the-art specialist baseline whose separate per-dataset training motivates the unified approach.","marker":"[40]"},{"why":"Baseline used in single- and cross-dataset comparisons of landmark accuracy.","marker":"[13]"},{"why":"Cascaded-transformer baseline used in cross-dataset comparisons.","marker":"[18]"},{"why":"PossLoss supervises the attention map that generates the initial landmark queries and coordinate predictions.","marker":"[47]"},{"why":"Wing loss supervises the intermediate and final coordinate predictions during decoder refinement.","marker":"[7]"},{"why":"Pretrained text encoder that produces the face-part embeddings combined with FPALPs in landmark queries.","marker":"[30]"},{"why":"Deformable attention layer that refines each query against image features at the predicted location.","marker":"[49]"}],"fun_headline_variants":["One face model answers any landmark query","Unified model predicts any face landmarks on demand","Landmarks as fractions: one net, all datasets","Single network handles every N-point face dataset","Arbitrary face landmarks from one trained model"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method's success rests on the assumption that aligning the landmark templates of different datasets into one shared template keeps each landmark's meaning intact; if landmarks that should be the same point land far apart in the merged template, the unified training signal becomes contradictory.","fun_headline_variants_meta":{"raw":{"variants":["One face model answers any landmark query","Unified model predicts any face landmarks on demand","Landmarks as fractions: one net, all datasets","Single network handles every N-point face dataset","Arbitrary face landmarks from one trained model"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000183,"raw_usage":{"total_tokens":1362,"prompt_tokens":1044,"completion_tokens":318,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":660,"completion_tokens_details":{"reasoning_tokens":249}},"tokens_in":660,"tokens_out":318,"duration_ms":3410,"temperature":1.0,"reasoning_tokens":249,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T14:22:00.278240+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a fourth landmark dataset whose contour ordering or point spacing differs from the three used here, align it into the unified template, and measure the average intra-cluster distance; if it climbs well above the reported 2.22 pixels, or if a model trained on that fusion loses its edge over per-dataset models on cross-template evaluations, the semantic-alignment premise fails.","supporting_citations":[{"cited_title":"Face alignment by coarse-to-fine shape searching","cited_arxiv_id":null,"evidence_quote":"Supplies the AFLW 19-point face template and its evaluation protocol."},{"cited_title":"300 faces in-the-wild challenge: The first facial landmark localization challenge","cited_arxiv_id":null,"evidence_quote":"Supplies the 300W 68-point face template and the common/challenging test splits."},{"cited_title":"Look at boundary: A boundary-aware face alignment algorithm","cited_arxiv_id":null,"evidence_quote":"Supplies the WFLW 98-point face template and motivates the boundary-line, evenly spaced landmark view."},{"cited_title":"Sparse local patch transformer for robust face alignment and landmarks inherent relation learning","cited_arxiv_id":null,"evidence_quote":"State-of-the-art specialist baseline whose separate per-dataset training motivates the unified approach."},{"cited_title":"Pixel-in-pixel net: Towards efficient facial landmark detection in the wild.IJCV, 2021","cited_arxiv_id":null,"evidence_quote":"Baseline used in single- and cross-dataset comparisons of landmark accuracy."},{"cited_title":"Towards accurate facial landmark detection via cascaded transformers","cited_arxiv_id":null,"evidence_quote":"Cascaded-transformer baseline used in cross-dataset comparisons."},{"cited_title":"Possloss: A reliable and sensitive facial landmark detection loss function","cited_arxiv_id":null,"evidence_quote":"PossLoss supervises the attention map that generates the initial landmark queries and coordinate predictions."},{"cited_title":"Wing loss for robust facial landmark localisation with convolutional neural networks","cited_arxiv_id":null,"evidence_quote":"Wing loss supervises the intermediate and final coordinate predictions during decoder refinement."},{"cited_title":"Sentence-bert: Sentence embeddings using siamese bert- networks","cited_arxiv_id":null,"evidence_quote":"Pretrained text encoder that produces the face-part embeddings combined with FPALPs in landmark queries."},{"cited_title":"smiling girl with party wig","cited_arxiv_id":null,"evidence_quote":"Deformable attention layer that refines each query against image features at the predicted location."}],"review_version":1}