{"id":"973a80be-a47f-4233-8783-9817543a41ae","arxiv_id":"2608.00903","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"A training-free inference framework combining target-aware reference expansion, soft top-k palette voting, and cycle-gated temporal fusion improves segment-matching colourisation on animation videos.","lead":"PeCA adds a test-time inference layer to paint-bucket colourisation for animation, using context from extra reference views and neighbouring frames without any retraining. Across five backbones and several benchmark settings, it raises segment-level accuracy by roughly 3 to 14 percentage points, with the largest gains on frozen foundation models.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Unverified comparison target: 'DACoN 1.1 + PeCA' may be tested against a simplified top-1 base rather than the official trained DACoN 1.1 inference, so the headline gain on a trained pipeline is not yet confirmed.","rationale":"I read the paper in good faith. The method is internally coherent, the ablations are consistent with the design, and the gains across frozen backbones and two reference protocols suggest that test-time aggregation and temporal fusion have real value. The reader's weakest-assumption analysis focused on Active Reference Expansion, but the ablation (Tab. 5) shows ARE alone contributes only about 1 point, so it is not the most load-bearing element. I instead see the most load-bearing concern as the fairness and verifiability of the 'trained DACoN 1.1 + PeCA' comparison. PeCA's probability aggregation is defined as a direct replacement for the Eq. (2) top-1 retrieval rule, and the paper's implementation text describes building a 'Base inference' with exactly that rule. If the baseline DACoN 1.1 numbers are from that simplified base rather than the official trained pipeline, then the central claim about improving a task-trained pipeline is not established. This is not an accusation of misconduct; it is an ambiguity that the manuscript does not resolve and that code release or an explicit statement would settle. The reader already flagged missing code and validation-set concerns; this is a more specific instance of the same verifiability issue, so I mark partial agreement. The verdict remains CONDITIONAL: the paper should be accepted only after the official-baseline comparison is confirmed or the claim is restricted to frozen backbones.","tokens_in":26262,"tokens_out":7574,"duration_ms":87282,"concrete_test":"Obtain the official DACoN 1.1 code and checkpoint. Reproduce the Table 1 row 'DACoN 1.1' on PBC-3D one-shot design-sheet colourisation using the official default inference; verify the reported 68.01 segment Acc. Then insert PeCA strictly as a post-processing step on the official pipeline's region features/outputs and recompute. If (a) the official baseline reproduces 68.01 and (b) +PeCA again gives 72.04 Acc (and the analogous PBC-Real first-frame gains in Table 3), the central claim is confirmed. If the official baseline differs from the Eq. (2) base, or if +PeCA no longer improves over the official pipeline, the headline gain is an artifact of comparing against a simplified base.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The strongest claim is that PeCA improves the trained DACoN 1.1 pipeline (e.g., Table 1: 68.01→72.04 Acc; Table 3: 70.34→74.41 on PBC-3D). However, the manuscript's implementation description defines the inference base as the Eq. (2) top-1 retrieval rule using pooled region features, and says it 'build[s] a Base inference for SOTA [29] and frozen foundation models' (Sec. 4.1). It never states explicitly that the reported 'DACoN 1.1' numbers are produced by the official, unmodified DACoN 1.1 code/checkpoint, or that PeCA is applied to that official pipeline's outputs. If the DACoN 1.1 rows are instead a re-implemented top-1 baseline using DACoN features, then PeCA's probability aggregation (Sec. 3.4) is directly replacing the only inference rule of that baseline, and the large gains may partly reflect recovering the gap between top-1 retrieval and the actual DACoN inference, not an improvement over the trained pipeline the paper claims to improve. This is load-bearing because the 'trained DACoN 1.1' results are the headline evidence for the 'plug-and-play on task-trained models' part of the central claim. No code or checkpoints are released, so this cannot currently be checked from the manuscript.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces PeCA, a training-free, plug-and-play inference framework for paint-bucket colourisation in animation videos. PeCA builds on segment-level retrieval: region descriptors are obtained by pooling backbone features, and correspondences are computed by cosine similarity. The framework adds three test-time components: Active Reference Expansion (ARE), which augments reference design sheets and selects a target-conditioned support subset via a submodular facility-location objective; Probability Aggregation (PA), which soft-votes over top-k correspondences to obtain per-region palette probabilities; and Cyclic-gated Temporal Fusion (CT), which refines these probabilities along cycle-consistent links between adjacent frames. Experiments on PBC-3D, PBC-Real, and a newly constructed Anita-Pirate long-video test set report consistent gains over a top-1 retrieval baseline on the trained DACoN 1.1 pipeline and on several frozen foundation backbones (DINOv2, SAM2.1, SigLIPv2, DINOv3, CLIP).","tokens_in":26600,"tokens_out":6301,"duration_ms":71149,"significance":"If the reported gains are faithful, PeCA is a useful contribution: it is model-agnostic, requires no task-specific training, and the paper provides unusually thorough validation for a test-time inference method, including ablations of each component (Tables 5-6), hyperparameter sensitivity surfaces (Fig. 7), multiple backbones and reference protocols, and a new long-video stress test with honest failure-case analysis in App. H. The circularity concern raised in the stress-test note is only partially supported: ARE and PA do use the same similarity scores, but Table 5 shows ARE alone contributes about one point, so the method does not hinge on that assumption. The main unresolved issues are the verification of the DACoN 1.1 baseline and the absence of statistical uncertainty; these are load-bearing for the headline claim that PeCA improves a trained SOTA pipeline.","major_comments":[{"comment":"The claim that PeCA improves the trained DACoN 1.1 pipeline is not yet verifiable from the manuscript. The implementation details state that the authors 'build a Base inference for SOTA [29] and frozen foundation models', and the base rule is defined as the top-1 retrieval in Eq. (2). It is never stated explicitly that the rows labelled 'DACoN 1.1' are produced by the official DACoN 1.1 code/checkpoint and inference protocol, rather than by a re-implemented top-1 retrieval baseline using DACoN features. If the latter is the case, then PeCA's probability aggregation (Eq. 6) is replacing the only inference rule of that baseline, and a substantial fraction of the reported gain (e.g., 68.01 to 72.04 in Table 1) could be due to recovering the gap between simple top-1 retrieval and the actual DACoN inference. No code or checkpoints are released, so this cannot be checked. Please clarify exactl","section":"Sec. 4.1, Eq. (2), Tables 1-4"},{"comment":"All reported metrics are single-run percentages with no variance, confidence intervals, or significance tests. This matters because ARE is stochastic: App. B.3 and Table B2 sample flips, rotations, and affine transformations. A single draw cannot support claims of 'consistent improvements', especially for small gains such as Pix-B-MIoU improvements of 0.31-0.82 points (Tables 1-2). Please report mean and standard deviation over multiple runs or seeds, and ideally a paired significance test for the headline comparisons.","section":"Tables 1-6, App. B.3"},{"comment":"The fixed hyperparameters (top-k=64, tau=0.05, B=31, m=4) are said to be used across all experiments, but the manuscript does not describe any validation split or selection criterion used to choose these values. Without this information, it is unclear whether the defaults were selected on the test benchmarks, which would inflate the reported gains. Please specify how the hyperparameters were chosen, or evaluate sensitivity on a held-out validation set rather than on the final test sets.","section":"Sec. 4.1, Fig. 7"}],"minor_comments":[{"comment":"The sentence 'When k=1 and tau->0, it reduces to hard copying' is imprecise: for k=1 the softmax is degenerate for any tau. Consider reformulating as 'when only the top-1 candidate is retained'.","section":"Sec. 3.4, Eq. (5)"},{"comment":"The transform composition is unclear: is the chain T_affine * T_90 * T_vflip * T_hflip applied with each probability independently, or is a single composition sampled? Please clarify the sampling procedure.","section":"App. B.3, Eq. (vi) and Table B2"},{"comment":"The VIPSeg label-propagation experiment selects reference frames using ground-truth panoptic classes present in the target video. This is an oracle protocol; please state explicitly that these results are an upper-bound diagnostic rather than a deployable setting.","section":"App. F"},{"comment":"On Anita-Pirate, DACoN 1.1 + PeCA has the same Pix-Acc as DACoN 1.1 (94.29). The text says the framework provides 'consistent improvements'; please acknowledge tied metrics when they occur.","section":"Table 4"},{"comment":"The Anita-Pirate benchmark is constructed semi-automatically by inferring shadow/highlight boundaries from colour discontinuities and assigning majority colours. Please include a discussion of potential label noise and how it may affect conclusions when this benchmark is used as a stress test.","section":"App. A"},{"comment":"No code or official checkpoints are released, so exact reproduction is difficult; the project page appears to host only the new test set. A statement on planned code release would strengthen reproducibility.","section":"Overall"}],"recommendation":"major_revision","confidential_remarks":"The paper is potentially a solid contribution, but the central claim that PeCA improves the trained DACoN 1.1 pipeline depends on the exact definition of the reported DACoN 1.1 baseline. If the published numbers come from a re-implemented top-1 retrieval baseline rather than the official DACoN 1.1 checkpoint, the claim would be significantly weaker. The lack of error bars and the unspecified hyperparameter selection procedure add to this concern. I recommend major revision, not rejection, because these issues can be addressed by verification and additional reporting."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"PeCA is a genuinely useful idea: instead of training a new colourisation model, it wraps existing region-matching backbones with three test-time steps—target-conditioned reference expansion, soft top-k voting in palette space, and cycle-gated temporal fusion—and reports consistent gains on PBC-3D, PBC-Real, and a new long-video benchmark. The contribution is real, and the paper is well-built. Ablations isolate each component (ARE alone gives about a point; PA and CT carry most of the gain), hyperparameter sensitivity is shown over wide ranges, and they evaluate five backbones under three reference protocols. The submodular facility-location selection with greedy approximation is standard and correctly applied. The appendix also includes a useful analysis of probability quality (entropy, GT margin) that explains why soft voting helps. I believe the main claim—that this plug-in substantially improves frozen backbone colourisation—probably holds.\n\nThe soft spots are about verification, not the core idea. There are no error bars, no code or checkpoints, and the fixed hyperparameters look like they were chosen after sensitivity analysis on the same evaluation benchmarks, so the exact magnitudes (3–14 points) should be read as provisional. The stronger concern is the DACoN 1.1 comparison: the setup text says they \"build a Base inference\" for the SOTA, and never explicitly confirms the reported DACoN 1.1 numbers come from the official checkpoint and official inference rule. If the base is a top-1 retrieval re-implementation, the gain over the trained pipeline may partly be recovering the gap between top-1 retrieval and real DACoN inference. That is the weakest link and needs an explicit statement and, ideally, released code.\n\nThe ARE assumption—that maximum feature similarity under geometric augmentation proxies region identity—is reasonable but the ablation shows ARE alone is not load-bearing, so that concern is minor. The limitations section is honest: missing reference colours and line leakage are acknowledged.\n\nThis paper is for researchers working on animation colourisation and test-time inference. It is not a conceptual leap, but the engineering and evaluation are careful, and the plug-and-play framing makes it immediately useful. I would send it to a competent referee, and I would specifically ask about the DACoN base protocol and repeated-run variance. If those are clarified, it is a clear accept for a specialist venue.","headline":"A useful, carefully evaluated test-time plug-in for paint-bucket colourisation; the headline DACoN comparison needs an explicit protocol statement before I'd trust the trained-pipeline gains.","tokens_in":714,"tokens_out":1023,"would_cite":true,"duration_ms":37744,"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":"PeCA claims that paint-bucket colourisation improves at test time by pooling spatial context from expanded reference views, softening top-k correspondences into palette probabilities, and fusing those probabilities only along cycle-consiste","keywords":["animation colourisation","paint-bucket colourisation","region correspondence","test-time inference","foundation models","temporal consistency","video colourisation","submodular maximization"],"falsifier":"Take a character whose left and right sides have different colours, build a target shot in a pose where a horizontally flipped design sheet yields the highest feature similarity, and run PeCA with and without that flip in the candidate pool; if segment accuracy does not improve when the flipped view is selected, the similarity-proxy assumption behind Active Reference Expansion fails. A direct quantitative check is the GT-Margin of Eq. (10): if adding ARE-selected views does not raise the probability gap between the ground-truth colour and its strongest competitor, the expansion is supplying di","tokens_in":26127,"feed_emoji":"🎨","tokens_out":10269,"duration_ms":108996,"temperature":0.7,"pith_summary":"Hand-drawn animation colourisation assigns every line-enclosed region a colour from a fixed character palette, and current pipelines do this by matching each target region to a reference region. The paper argues that failures come less from missing matches than from ambiguous, noisy correspondences, and that context can resolve the ambiguity. It introduces PeCA, a training-free inference wrapper that expands the reference bank with target-aware augmented views, converts noisy top-k matches into per-region palette probabilities via soft voting, and refines those probabilities along bi-directionally consistent matches between adjacent frames. Across design-sheet and first-frame colourisation benchmarks, PeCA reports consistent segment-accuracy gains of roughly 3.9 to 12.1 points on one-shot reference tasks, with the largest gains on frozen foundation backbones that have never seen colourisation training. A sympathetic reader would take this as evidence that context engineering at test time is a viable alternative to training stronger colourisation models.","feed_headline":"Test-time context lifts animation colourisation by up to 12 points","feed_subtitle":"Reference views, soft palette voting, and temporal gating make frozen backbones better at paint-bucket colourising without retraining.","key_machinery":"The load-bearing object is the per-region palette probability vector: a distribution over the discrete colour palette obtained by soft top-k voting. It makes colour evidence from different views and frames commensurable, which is what lets temporal context act on the same label space. Three mechanisms move evidence into that space: Active Reference Expansion (Eqs. 3-4) scores augmented reference views by the maximum cosine similarity they offer to target regions and selects a budgeted subset via a submodular facility-location objective; Probability Aggregation (Eqs. 5-6) soft-votes top-k reference matches into palette probabilities; Cyclic-gated Temporal Fusion (Eqs. 7-8) keeps only bidirect","core_discovery":"The paper's central claim is that region-matching colourisation becomes more robust without task-specific training when evidence is pooled from three kinds of context. Active Reference Expansion augments design sheets with geometric transforms and greedily selects a budgeted subset that best covers target regions via a submodular facility-location objective. Probability Aggregation replaces hard top-1 copying with temperature-sharpened soft top-k voting into the shared colour-palette simplex, giving per-region colour probabilities comparable across frames. Cyclic-gated Temporal Fusion refines those probabilities only along bidirectional cycle-consistent matches between adjacent frames. The p","pith_inferences":["The dependence on geometric augmentation means PeCA reduces mild in-plane pose and layout gaps but not out-of-plane rotations or never-observed character poses; a natural extension is to pair the same facility-location selection with learned novel-view generation instead of handcrafted transforms.","The active reference selection objective could be inverted into an interactive production tool: instead of expanding views, the artist could be told which few frames or design-sheet views to colour first, minimising manual effort while maximising downstream automatic colourisation accuracy.","The palette-probability interface is task-agnostic; any discrete-label assignment problem with per-region descriptors and a fixed label set could adopt the same soft-voting plus cycle-gated temporal fusion, with the VIPSeg result as a first indication.","A stricter test of the temporal cue would be a video in which colours legitimately change over time, such as a lighting shift or palette swap; PeCA's cycle gate should refuse fusion there, and measuring that refusal would bound the method's applicability."],"forward_implications":["Frozen foundation models become usable paint-bucket colourisers out of the box: SAM2.1-Large's one-shot design-sheet segment accuracy rises from 34.54% to 46.65% when PeCA is wrapped around it.","The gains do not saturate when more design sheets are available: 5-shot and max-shot reference protocols still show improvements, contrasting with the earlier observation that extra references give marginal returns.","Longer videos are handled without accumulating error: on the new 206-frame Anita-Pirate benchmark, PeCA improves accuracy and raises a temporal-consistency F1 from about 91% to 97%.","The same inference wrapper transfers to non-colour label propagation: on VIPSeg superpixel semantic labelling, PeCA improves segment accuracy and mIoU over direct matching with two frozen backbones.","Because PeCA is plug-and-play, it can be combined with future region-matching backbones without retraining, so descriptor-quality improvements and test-time context improvements are additive."],"supporting_citations":[{"why":"Supplies the DACoN and DACoN 1.1 pipelines, the task-trained SOTA baseline that PeCA wraps, plus the key-frame colourisation protocol and first-frame reference protocol.","marker":"[29]"},{"why":"Provides BasicPBC-Ref, the design-sheet key-frame colourisation baseline and benchmark protocol that PeCA compares against.","marker":"[8]"},{"why":"Provides BasicPBC, the paint-bucket colourisation baseline, and the PBC-3D and PBC-Real benchmarks used in most experiments.","marker":"[9]"},{"why":"DINOv2 is the frozen foundation backbone used in the main training-free experiments, including the ablation study.","marker":"[31]"},{"why":"SAM 2.1 serves as a frozen segmentation backbone whose pooled region features PeCA improves in both one-shot and first-frame colourisation settings.","marker":"[38]"},{"why":"Provides the submodular maximization theory that justifies the greedy approximation guarantee for the facility-location view-selection objective in Active Reference Expansion.","marker":"[17]"},{"why":"Motivates test-time augmentation of reference views, the precursor idea that PeCA extends with budgeted, target-conditioned view selection.","marker":"[43]"},{"why":"Motivates using cycle consistency over time as a reliable temporal signal, the basis for Cyclic-gated Temporal Fusion.","marker":"[54]"}],"fun_headline_variants":["Training-free context boosts animation colorization by up to 12 points","Plug-and-play context improves animation colorization without retraining","Spatio-temporal context lifts animation colorization at test time","12-point gain in animation colorization from test-time context","No retraining: context alone lifts animation colorization up to 12 points"],"cache_read_input_tokens":2816,"weakest_assumption_plain":"The load-bearing premise is that the maximum cosine similarity between a target region's average-pooled descriptor and an augmented reference view's descriptors is a reliable proxy for the colour support that view provides; if geometric augmentation creates high-similarity distractors with wrong region identity, the selected views could add noise rather than evidence.","fun_headline_variants_meta":{"raw":{"variants":["Training-free context boosts animation colorization by up to 12 points","Plug-and-play context improves animation colorization without retraining","Spatio-temporal context lifts animation colorization at test time","12-point gain in animation colorization from test-time context","No retraining: context alone lifts animation colorization up to 12 points"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001231,"raw_usage":{"total_tokens":4829,"prompt_tokens":617,"completion_tokens":4212,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":361,"completion_tokens_details":{"reasoning_tokens":4124}},"tokens_in":361,"tokens_out":4212,"duration_ms":31483,"temperature":1.0,"reasoning_tokens":4124,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T00:40:37.160883+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a character whose left and right sides have different colours, build a target shot in a pose where a horizontally flipped design sheet yields the highest feature similarity, and run PeCA with and without that flip in the candidate pool; if segment accuracy does not improve when the flipped view is selected, the similarity-proxy assumption behind Active Reference Expansion fails. A direct quantitative check is the GT-Margin of Eq. (10): if adding ARE-selected views does not raise the probability gap between the ground-truth colour and its strongest competitor, the expansion is supplying di","supporting_citations":[{"cited_title":"In: Proceedings of the IEEE/CVF International Conference on Computer Vision","cited_arxiv_id":null,"evidence_quote":"Supplies the DACoN and DACoN 1.1 pipelines, the task-trained SOTA baseline that PeCA wraps, plus the key-frame colourisation protocol and first-frame reference protocol."},{"cited_title":"Paint Bucket Colorization Using Anime Character Color Design Sheets","cited_arxiv_id":"2410.19424","evidence_quote":"Provides BasicPBC-Ref, the design-sheet key-frame colourisation baseline and benchmark protocol that PeCA compares against."},{"cited_title":"In: Proceedings of the IEEE/CVF conference on Computer Vision and Pattern Recognition","cited_arxiv_id":null,"evidence_quote":"Provides BasicPBC, the paint-bucket colourisation baseline, and the PBC-3D and PBC-Real benchmarks used in most experiments."},{"cited_title":"Tractability3(71- 104), 3 (2014)","cited_arxiv_id":null,"evidence_quote":"Provides the submodular maximization theory that justifies the greedy approximation guarantee for the facility-location view-selection objective in Active Reference Expansion."},{"cited_title":"In: Proceedings of the IEEE/CVF International Confer- ence on Computer Vision","cited_arxiv_id":null,"evidence_quote":"Motivates test-time augmentation of reference views, the precursor idea that PeCA extends with budgeted, target-conditioned view selection."},{"cited_title":"In: Proceedings of the IEEE/CVF conference on Computer Vision and Pattern Recognition","cited_arxiv_id":null,"evidence_quote":"Motivates using cycle consistency over time as a reliable temporal signal, the basis for Cyclic-gated Temporal Fusion."}],"review_version":1}