{"id":"a1284b66-3785-4238-b122-b1af872c4226","arxiv_id":"2412.07744","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"StyleMaster improves reference-image video stylization by extracting global and local style separately, training on model-illusion pairs, and adding a motion adapter and gray tile ControlNet.","lead":"StyleMaster is a system that turns videos and images into a chosen artistic style using a reference style image and text prompts. It combines a new way to build training data, texture-preserving patch selection, and a motion adapter to make stylized videos smoother and more style-faithful.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Triplet loss treats same-style images from different pairs as negatives, so the global style projector may be trained to separate images of the same style.","rationale":"The reader's weakest assumption focuses on whether jigsaw rearrangement preserves human-perceived style within a positive pair. I agree that this is fragile, but I identify a more concrete and internal flaw: the triplet loss's negative sampling. The paper explicitly states that any image outside the pair is a negative, and since the dataset is generated by sampling one style and two objects per pair, many negatives share the anchor's style. This is a direct contradiction between the training objective and the stated goal of extreme style consistency. It is testable and, if it lands, it calls into question the entire global-style extraction contribution. I do not change the reader's CONDITIONAL verdict because the flaw is not necessarily fatal; it might be mitigated in an unstated implementation or the model could still learn useful style cues despite the label noise. However, the paper must release code/data and demonstrate that a corrected negative sampling does not alter the conclusions. This is a stronger, more specific reason for the conditional status than the reader's jigsaw-based concern, hence partial agreement.","tokens_in":15402,"tokens_out":5648,"duration_ms":57916,"concrete_test":"Request the illusion dataset and training code. Count the fraction of training triplets whose negative shares the style description of the anchor (expected about 1/65). Then retrain the global projection module with corrected negative sampling that guarantees a different style for each negative, keeping all hyperparameters and the jigsaw pair construction fixed. Compare CSD and ArtFID on Table 1 and CSD/UMT on Table 2. If the corrected model yields comparable or better style metrics, the current negative sampling is not the source of the reported gains; if the metrics drop significantly, the reported results depend on a flawed supervisory signal.","verdict_should_be":"UNCHANGED","load_bearing_attack":"In Sec. 3.1, each illusion pair is created by randomly selecting one style and two objects, producing an anchor and a positive. Sec. 3.2 defines the triplet loss with 'any image outside this pair as the negative sample.' Because pairs are not grouped by style description, roughly 1/65 of all other images in the dataset share the anchor's style. With 10K pairs and two images per pair, each anchor has about 300 same-style negatives among ~20K candidates, so random negative sampling yields ~1.5% mislabeled triplets. The loss then explicitly pulls the anchor away from same-style images, directly contradicting the goal of learning a style-invariant representation. These mislabeled negatives are initially close in CLIP space and will dominate gradients, forcing the projector to separate images of the same style. The reported CSD gains could then arise from a shortcut (e.g., object identity or low-level jigsaw artifacts) rather than from style abstraction. Thus the 'absolute style consistency' claimed for the dataset holds at the pixel level within a pair, but not at the supervision level of the triplet labels, undermining the central contribution of the global style extractor.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes StyleMaster, a DiT-based video stylization system that combines local texture selection from CLIP patch features, a global style projection trained on a self-generated 'model illusion' paired dataset, a motion adapter with a negative scale, and a gray-tile ControlNet for content guidance. The authors claim state-of-the-art style resemblance and temporal coherence in both stylized video generation and video/image style transfer, and introduce model-illusion-based paired style data as a novel resource for contrastive style learning. The manuscript includes quantitative comparisons against StyleID, InstantStyle, CSGO, VideoComposer, StyleCrafter, and DomoAI, plus ablations of the proposed components.","tokens_in":15632,"tokens_out":3373,"duration_ms":35060,"significance":"If the claims hold, the model-illusion dataset construction is a clever and low-cost alternative to manually curated style datasets, and the combined global-local style description with a negative-scale motion adapter is a practical recipe for reference-image video stylization. The paper also makes a useful attempt at image-to-video transfer by training a lightweight motion adapter on still videos, and the gray-tile ControlNet idea is simple and effective for style-transfer content control. However, the absence of released code or data, the small evaluation sets, and the lack of statistical significance testing leave the central claim of 'significant improvement' under-supported; the paper's own ablations and tables reveal several metric trade-offs that are explained post hoc rather than established.","major_comments":[{"comment":"The triplet loss treats 'any image outside this pair' as a negative, but the illusion dataset is generated from only 65 style descriptions (Supplement, Sec. 7.2). With 10K pairs, each anchor has roughly 300 other images that share the same style but belong to different pairs. Randomly sampled negatives will therefore include about 1.5% same-style images, and the loss explicitly pulls them apart, contradicting the goal of learning a style-invariant representation. The manuscript does not specify negative mining or grouping by style label. This is load-bearing for the global style extractor: if mislabeled negatives dominate gradients, the reported CSD gains may come from a shortcut unrelated to style. Please either group negatives by style description, track and exclude same-style negatives, or provide evidence (e.g., measuring same-style embedding distances before and after training) that the effect is negligible.","section":"Sec. 3.1 and Sec. 3.2, Eq. (1)"},{"comment":"The 'absolute style consistency' claim relies on the assumption that jigsaw-rearranged images are perceived as having the same style. Jigsaw scrambling also breaks composition, creates artificial boundaries, and may introduce low-level artifacts that a projector could exploit as a shortcut. The paper provides no human evaluation or style-classifier validation that these pairs actually share the intended artistic style. Please add a perceptual or classifier-based check on the illusion pairs, and ideally analyze what the global projector attends to, to rule out jigsaw-specific artifacts.","section":"Sec. 3.1, Fig. 4"},{"comment":"The evaluation is based on very small test sets (192 style-prompt pairs, 96 content-style pairs, and 4 videos for transfer) with no error bars, confidence intervals, or significance tests. The word 'significant' in the abstract is therefore not supported statistically. Moreover, Table 2 shows that on the main style metric (CSD-Score), StyleMaster scores 0.463, below VideoComposer's 0.680; the explanation that VideoComposer copies content is plausible but is offered post hoc and not backed by analysis. Please report per-pair distributions, pairwise significance tests, and a content-copying measure to substantiate the claim.","section":"Sec. 4, Tables 1 and 2"},{"comment":"The proposed method is worse than StyleID on LPIPS (0.61 vs 0.55) and much worse on CFSD (2.37 vs 1.06). The paper says it 'slightly underperforms' in content alignment, but without error bars one cannot tell whether this is significant, and CFSD is a direct content-preservation metric. Because the central claim includes 'both style resemblance and temporal coherence' plus implicit content preservation, the trade-off should be quantified and discussed in a falsifiable way, e.g., by reporting the Pareto frontier or a human preference study that weighs style fidelity against content retention.","section":"Sec. 4.1, Table 1"}],"minor_comments":[{"comment":"The title and abstract contain typos: 'Y our' in the header and 'casual V AE' in Sec. 4 should be 'Your' and 'causal VAE', respectively.","section":"Abstract"},{"comment":"The symbol α is used both as the margin in the triplet loss (Eq. 1) and as the motion-adapter scale in Eq. (2). Please use distinct symbols to avoid confusion.","section":"Sec. 3.4, Eq. (2)"},{"comment":"The main text says the video stylized-generation test set has 12 style images, but the supplement's video style transfer comparison uses 16 style images (Table 8). Please reconcile these numbers.","section":"Sec. 4, Dataset paragraph"},{"comment":"The narrative states that α = −0.3 'owns the best visual quality', but the table reports VisualQuality 2.370 at −0.3 versus 2.279 at 0; while plausible, no statistical significance is given, and the reader cannot tell whether this is within noise.","section":"Sec. 4.4.2, Table 4"},{"comment":"Several references are incomplete (e.g., StyleID lacks page numbers, and the StyleMaster project page appears as two different URLs: zixuan-ye.github.io/stylemaster in the abstract and style-master.github.io in the supplement). Please unify and complete.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The core idea of using model-illusion pairs for contrastive style learning is interesting and potentially useful, but the triplet negative-sampling issue is a genuine methodological weakness that could undermine the global-style extractor. The paper also makes very strong 'significant outperformance' claims on a small, non-released evaluation. I would like to see the authors clarify the negative-sampling protocol, add grouping or hard-negative filtering, and provide error bars or a human study. The novelty relative to CSGO and StyleTokenizer should be positioned more carefully; the model-illusion pairing is new, but the overall contrastive framework is not. No concerns about authorship or scope beyond the above."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper is a solid engineering contribution, but the evidence doesn't yet support the headline. The most interesting piece is the model-illusion dataset: generating style-consistent pairs by pixel rearrangement is genuinely creative and cheap. The patch-selection strategy and the negative motion adapter scale are also neat, and the gray tile ControlNet is a simple, sensible fix. The ablations are well-designed and show each component earns its place; the writing is clear and the related work is positioned accurately.\n\nThe soft spot that matters most is in the triplet supervision. The paper claims 'absolute style consistency' from the illusion pairs, but the triplet loss samples negatives from outside the pair without grouping by style. With 65 styles and 10K pairs, about 1.5% of the negatives share the anchor's style. That alone would be tolerable noise, but those same-style negatives are likely to be the hard negatives in CLIP space, so they can disproportionately shape the gradient. The projection layer may end up pushing same-style images apart, which directly contradicts the goal of a style-invariant extractor. The authors should either group negatives by style or explicitly exclude same-style images from the negative set. This is a fixable flaw, but it undercuts the central claim that the dataset guarantees style-consistent supervision.\n\nOther concerns are more conventional: no code, no data release, no error bars. The image transfer test set is 96 pairs (comparable to prior work, but still small), and the video transfer comparison uses only 4 videos, with a commercial app as a baseline. Some metric trade-offs—like accepting worse LPIPS while touting better ArtFID—are explained after the fact; that's defensible, but it needs more evidence. The limitation section only flags dynamic styles, not this supervision issue.\n\nOverall, the method is coherent and the work deserves a serious referee. I would send it to review with a request for code/data, a larger or independently sourced evaluation, and a direct discussion of the same-style negative problem. The central idea is worth engaging with, even if the current claims outrun the evidence.","headline":"A clever, coherent video stylization system with a genuinely novel dataset idea, but the triplet supervision has a same-style negative flaw and the experimental evidence is too thin to back the 'significant improvement' claim.","tokens_in":16147,"tokens_out":3422,"would_cite":true,"duration_ms":33824,"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":"StyleMaster claims that splitting style extraction into a global contrastive projection and prompt-filtered texture patches lets a diffusion model restyle videos to match a reference image without leaking the image's content.","keywords":["video stylization","style transfer","video generation","diffusion models","contrastive learning","model illusion","style extraction","motion adapter"],"falsifier":"A concrete check: take style categories with strong spatial composition (formal portraits, symmetrical architecture, scenes with a clear horizon) and have human raters or a style-similarity metric judge whether each illusion pair looks like the same style. If a substantial share of composition-heavy pairs are rated as different styles — or if a global projection trained on illusion pairs fails to beat one trained on the existing Style30K groupings when both are evaluated on real reference images outside the training distribution — then the claimed benefit of 'absolute style consistency' is not the source of the reported gains.","tokens_in":1897,"feed_emoji":"🎨","tokens_out":5459,"duration_ms":101902,"temperature":0.7,"pith_summary":"The paper argues that failures in video stylization — output far from the reference style, or content copied from it — trace back to the style-extraction stage, and that this stage needs both global style cues and local texture. To get global cues without content leakage, the authors train a projection on top of CLIP using contrastive pairs generated by 'model illusion': one image is a jigsaw rearrangement of the other, so the pair shares a style by construction while differing in content. For texture, they keep only the CLIP image patches least similar to the text prompt, which carry brushwork and material feel without carrying the subject. A lightweight motion adapter trained on still videos, used at a negative scale at inference, bridges the image-to-video gap and pushes outputs further from the real-world domain, strengthening the stylized look. The paper reports that the resulting system, StyleMaster, outperforms prior stylized video generators and transfer pipelines on style resemblance, text alignment, and motion quality.","feed_headline":"Jigsaw illusion pairs teach video AI to transfer style, not content","feed_subtitle":"StyleMaster splits a style image into global look and brushwork, then restyles video while keeping text meaning intact.","key_machinery":"Four mechanisms carry the argument: (1) the illusion dataset — pairs produced by running a text-to-image sampler in parallel on an image and a jigsaw-rearranged copy with two different object prompts, so the paired outputs are pixel rearrangements with identical style by construction; (2) the global projection — an MLP over the CLIP image embedding trained with triplet loss so that the anchor and its rearranged twin map close together while unrelated images map apart; (3) the local selection — keeping only the $k=15$ CLIP patch tokens with the lowest cosine similarity to the text prompt, so the Q-Former texture tokens carry strokes and materials rather than the subject; (4) the motion adapter — LoRA weights on the temporal-attention query, key, and value matrices trained on still videos and applied with a negative scale $\\alpha=-0.3$ at inference, which restores dynamics and, because the training data was real-world, exaggerates departure from the real-world domain, strengthening stylization. Style enters the denoising transformer through a dedicated style cross-attention that runs in parallel with text cross-attention, and a gray tile ControlNet supplies layout-preserving content guidance for transfer.","core_discovery":"On the paper's own terms, the central discovery is that style can be decomposed into two complementary signals and that both can be extracted without content leakage: a global style description from a contrastively trained MLP projection over the CLIP image embedding, and a local texture description from the roughly fifteen CLIP patches least similar to the text prompt, aggregated by a Q-Former. The enabling ingredient is a training dataset of pairs that are guaranteed to share a style: using the model-illusion trick, one image is generated as a jigsaw rearrangement of the other under a different object prompt, so the two pictures are pixel rearrangements of each other and 'absolute style consistency' holds within every pair. Contrastive triplet training on these pairs teaches the projection to encode style rather than content. Style is then injected through a separate style cross-attention alongside text cross-attention, and the same module handles both stylized generation and, with a grayscale tile ControlNet for content guidance, video-to-video style transfer.","pith_inferences":["Editorial extension: the illusion-pair recipe is a general substitute for manually grouped style datasets; any contrastive style or attribute extractor that suffers from noisy group labels could adopt pixel-rearrangement pairs, not just video stylizers.","Editorial extension: the negative-scale trick on a domain-trained adapter generalizes beyond motion — training an adapter on any unwanted domain and inverting its scale at inference is a cheap way to push generation away from that domain.","Editorial extension: a testable prediction of the paper is that the global projection's advantage grows as style categories become more compositional (e.g., portraits, scenes), where jigsaw rearrangement breaks layout; if the advantage instead vanishes there, the 'absolute style consistency' claim is doing less work than the contrastive objective alone."],"forward_implications":["Style extraction, not just style injection, is the bottleneck: contrastive training on guaranteed-consistent pairs plus prompt-filtered texture patches raises style resemblance while improving text alignment over methods that inject all reference tokens.","A single image-trained pipeline transfers to video by swapping in a motion adapter, so the same model serves stylized generation and video style transfer without per-video optimization.","The negative motion-adapter scale converts real-world-domain training data into an implicit stylization boost: as the scale moves from 0 to -1, the reported CSD style score rises from 0.443 to 0.465.","Removing color from the tile ControlNet condition prevents the content guide from contaminating the injected style, which is what enables clean video-to-video transfer.","Ablations show the two streams are complementary: global projection alone or texture selection alone underperforms their combination on both style resemblance and text-video alignment."],"supporting_citations":[{"why":"The contrastive style-learning baseline and its manually grouped Style30K dataset, whose inconsistent style labels motivate the illusion-generated pairs.","marker":"[24]"},{"why":"Supplies the model-illusion mechanism the paper adapts: generating images whose pixel rearrangement changes perceived content while keeping style.","marker":"[15]"},{"why":"The main stylized-video-generation competitor, and the source of the Q-Former aggregation idea and the test set the paper extends.","marker":"[28]"},{"why":"The main competitor that injects all reference tokens directly, used as the content-leakage failure case the method must beat.","marker":"[46]"},{"why":"Provides the dual cross-attention adapter mechanism through which the extracted style is injected alongside text.","marker":"[52]"},{"why":"Contributes the motion adapter trained on still videos that the paper reuses, now with a negative inference scale.","marker":"[4]"},{"why":"An image stylization competitor whose identified-layer style injection is contrasted with the paper's extractor design, and which forms a video-transfer baseline when combined with AnyV2V.","marker":"[44]"},{"why":"A competitor that supplies the tile-ControlNet content-guidance idea and the CSD metric used for style-resemblance evaluation.","marker":"[51]"}],"fun_headline_variants":["Jigsaw illusion pairs train video AI to separate style from content","StyleMaster learns video style from paired images sharing exact style","Video restyling: global style plus texture patches minus content leakage","StyleMaster pairs jigsaw images to teach video style transfer cleanly"],"cache_read_input_tokens":18304,"weakest_assumption_plain":"The load-bearing premise is that jigsaw-rearranging an image's pixels leaves its style unchanged in the way viewers actually perceive it, so pairs produced by the model-illusion trick really are two different subjects sharing one style; if the rearrangement destroys composition or texture, the contrastive projection learns a signal that has little to do with the style users want.","fun_headline_variants_meta":{"raw":{"variants":["Jigsaw illusion pairs train video AI to separate style from content","StyleMaster learns video style from paired images sharing exact style","Video restyling: global style plus texture patches minus content leakage","StyleMaster pairs jigsaw images to teach video style transfer cleanly"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000457,"raw_usage":{"total_tokens":2302,"prompt_tokens":962,"completion_tokens":1340,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":578,"completion_tokens_details":{"reasoning_tokens":1268}},"tokens_in":578,"tokens_out":1340,"duration_ms":9975,"temperature":1.0,"reasoning_tokens":1268,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T18:31:45.172208+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A concrete check: take style categories with strong spatial composition (formal portraits, symmetrical architecture, scenes with a clear horizon) and have human raters or a style-similarity metric judge whether each illusion pair looks like the same style. If a substantial share of composition-heavy pairs are rated as different styles — or if a global projection trained on illusion pairs fails to beat one trained on the existing Style30K groupings when both are evaluated on real reference images outside the training distribution — then the claimed benefit of 'absolute style consistency' is not the source of the reported gains.","supporting_citations":[{"cited_title":"Visual ana- grams: Generating multi-view optical illusions with diffu- sion models","cited_arxiv_id":null,"evidence_quote":"Supplies the model-illusion mechanism the paper adapts: generating images whose pixel rearrangement changes perceived content while keeping style."},{"cited_title":"Videocomposer: Compositional video synthesis with motion controllability","cited_arxiv_id":null,"evidence_quote":"The main competitor that injects all reference tokens directly, used as the content-leakage failure case the method must beat."}],"review_version":1}