{"id":"cfb6e08e-5084-47d4-af8d-01288a110e57","arxiv_id":"2504.13224","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":3.0,"correctness_risk":"high","formal_verification":"none","parameter_count":5,"one_line_summary":"ICAS is a style-transfer pipeline that freezes IP-Adapter's style path, lightly tunes its content path, and adds ControlNet structure conditioning to preserve multi-subject layouts.","lead":"ICAS combines IP-Adapter and ControlNet to transfer the style of a reference image onto pictures containing multiple subjects, while keeping the layout and identity of each subject. The paper claims better style consistency and subject clarity than existing adapters, based mostly on qualitative comparisons and a small user study.","discovery_kind":"incremental","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The evaluated system omits the proposed partial fine-tuning: Section IV-A says all experiments use a pre-trained IP-Adapter without further fine-tuning, so the claimed superiority of ICAS was never actually tested.","rationale":"Reading the paper in good faith, the strongest version of the claim is that a lightly fine-tuned content-injection branch, combined with frozen IP-Adapter style blocks, ControlNet, and cyclic multi-content embeddings, gives state-of-the-art multi-subject style transfer. What would have to be true is that the fine-tuned system was actually built and evaluated. The manuscript itself undercuts this: Section IV-A says all Section IV experiments use the pre-trained IP-Adapter without further fine-tuning. That makes the main contribution untested in the reported comparisons. The Section IV-C \"Content-Only\" ablation appears to describe the missing experiment, but it is not reconciled with Section IV-A's statement, and no trained checkpoint, parameter counts, training data, or quantitative subject-matching scores are supplied. Independent confirmation is absent. The absence of error bars or significance tests in Table 1 means even the user-study numbers should be treated as descriptive. This is a correctness and support problem, not a disagreement with community consensus; the proposed method might work, but the paper does not demonstrate it. The reader's weakest_assumption identifies the same mismatch, so I agree with the rejection. I recommend leaving the verdict unchanged.","tokens_in":11742,"tokens_out":3317,"duration_ms":34269,"concrete_test":"The decisive check is to release the checkpoint trained under the Section IV-C Content-Only protocol and rerun the same 150 stylization tasks with (1) No-Finetune, (2) Content-Only ICAS, and (3) full ICAS with trained gating and cyclic embeddings, scoring subject clarity and style fidelity with paired significance testing. If Content-Only does not significantly beat No-Finetune, the claimed contribution is unsupported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section III-A describes fine-tuning a small content injection sub-block as the core of ICAS, and Section IV-C claims Content-Only fine-tuning beats Full-Finetune and No-Finetune. Yet Section IV-A reports that both training settings produced similar results and then states, \"for convenience and computational efficiency, all subsequent experiments (Section IV) are conducted with the pre-trained IP-Adapter without further fine-tuning.\" If that is literal, Figures 2, 4, 6 and Table 1 never evaluate the proposed partial fine-tuning; they evaluate an off-the-shelf IP-Adapter plus ControlNet and cyclic embeddings. The central claim that ICAS outperforms baselines is therefore about a system that omits the paper's central contribution, and no evidence is shown that the gating parameters and content cross-attention layers were ever trained for the reported results. The only quantitative support is a 20-participant user study with no variance or significance measures, so one cannot infer that the tested system is superior, let alone that the untested fine-tuned system is.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes ICAS, a framework for multi-subject style transfer that combines a pre-trained IP-Adapter (for style injection) with a ControlNet branch (for structural conditioning), and introduces a gating mechanism plus a cyclic multi-content embedding strategy. The stated contribution is a lightweight partial fine-tuning scheme that updates only the content injection branch while freezing the style path. The authors claim superior performance over baselines such as InstantStyle, StyleShot, and IP-Adapter, supported by qualitative comparisons and a 20-participant user study.","tokens_in":11952,"tokens_out":4380,"duration_ms":45205,"significance":"If the method worked as described and were properly evaluated, the direction of combining IP-Adapter and ControlNet for multi-subject style transfer with minimal fine-tuning could be of interest to the community. The paper also addresses a practical need: efficient style transfer without inversion or large stylized datasets. However, the current manuscript does not validate these claims: the experiments do not implement the proposed partial fine-tuning, the reported quantitative evidence is a user study without statistical support, and the method description contains internal contradictions and an ill-defined gating equation. The potential significance is therefore not realized in the present form.","major_comments":[{"comment":"Section IV-A states that 'all subsequent experiments (Section IV) are conducted with the pre-trained IP-Adapter without further fine-tuning.' This directly contradicts the central contribution described in Section III-A and Section IV-C, which is partial fine-tuning of the content injection branch. Consequently, Figures 2, 4, 6 and Table 1 evaluate a system that omits the proposed fine-tuning. The 'Content-Only (Ours)' condition in the ablation study of Section IV-C is therefore never actually realized in the reported results; the 'No-Finetune' condition is the de facto evaluated system. This means the central claim that ICAS outperforms baselines is not supported by the experiments.","section":"Section IV-A"},{"comment":"The user study in Section IV-D reports only mean scores from 20 participants, with no standard deviations, confidence intervals, or significance tests. The differences between ICAS and InstantStyle on Style Fidelity are 4.32 versus 4.22, which is likely within inter-rater variability. Without per-participant or per-task variance, and without any statistical test (e.g., paired t-test or Wilcoxon), the conclusion that 'ICAS consistently outperforms' is not justified.","section":"Table 1"},{"comment":"Equation (4) defines F_sim = α·A_R + (1−α)·Q + g, where A_R and Q are N×d matrices and g is defined in Eq. (3) as a d-dimensional vector via g = σ(Wg(e_C·e_R) + bg). The addition of a d-vector to an N×d matrix is undefined without specifying broadcasting; if broadcast over the N dimension, every spatial token receives the identical bias, which is not an 'adaptive' per-token gate. If g is intended to be per-token, Eq. (3) cannot produce an N-dimensional vector because e_C·e_R is a global similarity. This makes the core gating mechanism ill-defined, and the claim that it provides 'fine-grained modulation of style strength' is not supported.","section":"Section III-B, Eq. (4)"},{"comment":"The paper is internally inconsistent about whether fine-tuning is required. Section III-A calls ICAS a 'lightweight and tuning-free method' but then immediately states that 'we only update a small content injection sub-block.' Section IV-C claims that the Content-Only partial fine-tuning strategy is superior to No-Finetune, yet Section IV-A says all experiments used the pre-trained adapter without fine-tuning. The reader cannot determine whether the proposed method requires training or not, and the experimental results are consistent only with the no-fine-tuning interpretation. This ambiguity undermines the reproducibility of the method.","section":"Section III-A and Section IV-C"},{"comment":"The ablation studies in Section IV-C repeatedly refer to quantitative measures such as 'subject-level matching score' and 'FID / style match / subject fidelity,' but no numerical values are reported anywhere in the paper. The conclusions of the ablations are based solely on qualitative figures (Figures 3 and 4) with no error bars or quantitative backing. Consequently, the claim of 'comprehensive ablation experiments' is overstated; the evidence is anecdotal.","section":"Section IV-C"}],"minor_comments":[{"comment":"The paper contains numerous typos and inconsistent terminology: 'criss-cross attention' should be 'cross-attention,' 'iPadAdapter' in the Figure 2 caption should be 'IP-Adapter,' 'Evol ving' in Section II-C should be 'Evolving,' 'Instantstyle' should be 'InstantStyle,' and 'Styleshot' should be 'StyleShot.'","section":"Throughout"},{"comment":"The phrase 'The Base Model Setup. The Base Model Setup.' is duplicated in the 'Base Model Setup' paragraph. This appears to be an editing error.","section":"Section IV-A"},{"comment":"Equation (4) introduces a scalar α that is described as 'user-controllable,' but the paper does not specify how α is set in the experiments or whether it is learned. The relationship between the gating parameters (Wg, bg) and α is not clarified.","section":"Section III-B"},{"comment":"The structure preservation module is described as a 'residual map' added to U-Net features with a scaling parameter γ. The paper does not define the spatial dimensions of R_S or how the projection ϕ is parameterized, making it difficult to reproduce the implementation.","section":"Section III-C"},{"comment":"In the comparison with state-of-the-art methods, the paper states that baselines' weights are 'manually adjusted to maintain consistent intensity sensitivity.' No details are given about how this adjustment was performed, which weakens the fairness of the comparison.","section":"Section IV-B"}],"recommendation":"reject","confidential_remarks":"The manuscript reads like an early draft. The central contribution is not evaluated: the experiments explicitly omit the proposed fine-tuning, and the gating mechanism is mathematically ill-defined. The only quantitative support is an underpowered user study. Even with a major revision requiring new experiments and a rewritten method section, the paper would need substantial additional work to establish a defensible claim. I recommend rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe short version: the paper's central claim does not survive contact with its own experimental section. ICAS is presented as a lightweight fine-tuning of only the content injection branch of IP-Adapter, with a gating term and cyclic multi-content embeddings. But Section IV-A states that all subsequent experiments use the pre-trained IP-Adapter without further fine-tuning, and Section III-A even calls the method 'tuning-free' before describing fine-tuning. So Figures 2, 4, 6 and Table 1 appear to evaluate off-the-shelf IP-Adapter plus ControlNet, not the method as described. If that is literal, the claimed superiority of ICAS was never tested.\n\nThere is something genuine to work with here. Combining IP-Adapter's style path with ControlNet's structural conditioning is a sensible recipe for multi-subject style transfer, and the cyclic multi-content embedding is a reasonable heuristic for preserving distinct subjects. The gating mechanism is a small addition—a sigmoid on the content-style similarity—and is not deeply motivated, but it is not harmful. Those pieces could make a useful empirical note if they were evaluated honestly.\n\nThe soft spots are large, though. The only quantitative evidence is a 20-participant user study with mean scores and no variance, no significance tests, and no per-participant detail. The ablations are qualitative figure comparisons; the claimed quantitative metrics in Section IV-C ('FID / style match / subject fidelity') are asserted but never reported. There is no code, no data, and no reproducibility package. The abstract's 'new paradigm' language is not supported by anything in the paper.\n\nOn citation pattern: the paper cites the relevant adapters and style transfer literature, and I don't see missing critical references, though the self-positioning against InstantStyle and StyleShot is brief and mostly qualitative.\n\nMy recommendation: this deserves a desk reject in its current form, not because the idea is worthless but because the paper does not evaluate what it proposes. If the author actually runs the partial fine-tuning, reports the comparison between trained and untrained content branches with proper metrics and error bars, and releases code, it could become a minor but citable contribution. As is, I would not send it to referees.","headline":"The paper's headline contribution—partial fine-tuning of the content branch—is never actually tested in its own experiments, which fall back to a pre-trained IP-Adapter plus ControlNet; the remaining evidence is qualitative figures and a 20-person user study without statistics.","tokens_in":12472,"tokens_out":3245,"would_cite":false,"duration_ms":30990,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"ICAS claims that combining a frozen IP-Adapter style path, ControlNet structure control, and cyclically injected content embeddings preserves multiple subjects' identities during style transfer without large stylized datasets or inversion.","keywords":["style transfer","multi-subject generation","diffusion models","IP-Adapter","ControlNet","content embedding","gating mechanism","structure preservation"],"falsifier":"Run ICAS twice on the same held-out multi-subject set, once with the content cross-attention branch updated and once with it frozen, and compare per-subject identity preservation. If the two outputs are indistinguishable, partial fine-tuning is not load-bearing; if the frozen version loses or merges subjects, then the paper's claim that all reported experiments used the pre-trained adapter without fine-tuning is inconsistent with the method it proposes.","tokens_in":11521,"feed_emoji":"🎨","tokens_out":8700,"duration_ms":81837,"temperature":0.7,"pith_summary":"This paper claims that multi-subject style transfer can be done without retraining the diffusion model or running expensive inversion, by combining a frozen IP-Adapter style-injection path with a ControlNet structure-preservation branch and a cyclic multi-content embedding mechanism. The proposed framework, ICAS, adds a gating term to the style cross-attention so style strength adapts to content, injects structure residuals at scale $\\gamma=0.7$, and optionally fine-tunes only the content injection sub-block. The author argues this preserves each subject's identity and the global layout while transferring color, texture, and atmosphere, and that it beats adapter-based and data-intensive baselines in structure preservation, style consistency, and inference efficiency. The stakes are practical: a low-compute recipe for multi-subject stylization in real-world settings with limited data.","feed_headline":"Frozen adapters plus ControlNet keep every subject in style transfer","feed_subtitle":"Gated style injection, structure control, and cyclic content embeddings preserve identity without retraining.","key_machinery":"The central mechanism is the two-branch injection into U-Net attention: the Style Injection Module reuses IP-Adapter's pre-trained cross-attention with a gating term to blend style keys and values into the content query, and the Structure Preservation Module adds a ControlNet-derived residual map $R_S = \\phi(F_S)$ scaled by $\\gamma$ to U-Net features. A third piece, the cyclic multi-subject content embedding list, repeatedly injects several encodings of the content image so each subject receives its own attention signal. Together these carry the argument: style comes from the frozen adapter, layout from the ControlNet residual, and subject identity from multiple content embeddings.","core_discovery":"The central claim is that decoupling style and structure and injecting them through separate adapter branches lets a pre-trained diffusion model perform multi-subject style transfer with only minimal, selective parameter updates. The style injection output is $F_{sim} = \\alpha A_R + (1-\\alpha)Q + g$, where a sigmoid gate $g$ modulates style contribution based on content-style similarity, and the structure branch adds $\\gamma R_S$ to U-Net features at selected blocks. The paper reports that with $\\gamma=0.7$, ICAS preserves multiple subjects better than InstantStyle, StyleShot, and the original IP-Adapter, and that a 20-participant user study gives ICAS the highest mean scores on style fidelity, subject clarity, and overall aesthetic. It further states that freezing the style path and updating only content-related parameters achieves the best balance of quality and computational overhead.","pith_inferences":["If the statement that all experiments used the pre-trained IP-Adapter without further fine-tuning is taken literally, then the partial content-branch fine-tuning module is not actually tested by the reported results, and the observed gains would come entirely from the frozen style path, ControlNet at $\\gamma=0.7$, and cyclic content embeddings.","Because the ablation describes the gating term as possibly 'solely a scale parameter in the code,' the gating mechanism may be functionally equivalent to IP-Adapter's existing strength adjustment; a direct test would compare ICAS against IP-Adapter with matched per-subject scales on the same content set.","The cyclic multi-content embedding recipe could generalize to other adapter-based multi-subject generation tasks, such as multi-person personalization or story illustration, where encoding each subject multiple times through augmentation or segmentation and alternating injections may improve identity preservation.","The $\\gamma=0.7$ sweet spot suggests that structure scale and style strength interact, so a natural extension would be a joint calibration or a learned schedule for $\\gamma$ conditioned on the number of subjects and the complexity of the scene."],"forward_implications":["Multi-subject style transfer can be performed without inversion or large stylized datasets, using a frozen pre-trained style adapter plus a structure control branch.","Freezing the style path while lightly updating only the content cross-attention preserves pre-trained style knowledge and adapts to multi-subject content at minimal parameter cost.","Cyclically injecting multiple content embeddings into the content path keeps individual subjects from being merged, occluded, or lost in the background.","A ControlNet conditioning scale around $\\gamma=0.7$ balances geometry preservation and style richness across the tested multi-subject scenes.","In the reported small user study, ICAS scores higher than original IP-Adapter and InstantStyle on style fidelity, subject clarity, and overall aesthetics."],"supporting_citations":[{"why":"Supplies the IP-Adapter cross-attention style injection mechanism that ICAS freezes and reuses as the Style Injection Module.","marker":"[26]"},{"why":"In-text source for the ControlNet-style structural conditioning branch used by the Structure Preservation Module.","marker":"[24]"},{"why":"The adapter-based baseline most similar to ICAS; the paper compares against it on multi-subject style fidelity.","marker":"[10]"},{"why":"The data-intensive fine-tuning baseline ICAS claims to surpass without requiring large stylized corpora.","marker":"[13]"},{"why":"Representative inversion-based approach used to motivate ICAS's lower computational cost.","marker":"[6]"},{"why":"Feature-injection baseline representative of inversion-style pipelines the paper aims to avoid.","marker":"[7]"},{"why":"Training-free style-injection method that motivates the attention-injection design used in ICAS.","marker":"[8]"}],"fun_headline_variants":["ICAS: decouple style and structure for multi-subject transfer","Frozen style branch plus ControlNet: efficient multi-subject transfer","Multi-subject style transfer with selective fine-tuning, not full retraining","Keep every subject's identity: ICAS style transfer with gate control"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The central claim assumes that a pre-trained IP-Adapter's style-only path, combined with ControlNet at scale 0.7 and cyclically injected content embeddings, generalizes to arbitrary multi-subject inputs; if the partial fine-tuning of the content branch is genuinely required, then the reported experiments, which say they used no fine-tuning, do not test the proposed method.","fun_headline_variants_meta":{"raw":{"variants":["ICAS: decouple style and structure for multi-subject transfer","Frozen style branch plus ControlNet: efficient multi-subject transfer","Multi-subject style transfer with selective fine-tuning, not full retraining","Keep every subject's identity: ICAS style transfer with gate control"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000552,"raw_usage":{"total_tokens":2640,"prompt_tokens":958,"completion_tokens":1682,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":574,"completion_tokens_details":{"reasoning_tokens":1606}},"tokens_in":574,"tokens_out":1682,"duration_ms":14593,"temperature":1.0,"reasoning_tokens":1606,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T12:20:24.287068+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run ICAS twice on the same held-out multi-subject set, once with the content cross-attention branch updated and once with it frozen, and compare per-subject identity preservation. If the two outputs are indistinguishable, partial fine-tuning is not load-bearing; if the frozen version loses or merges subjects, then the paper's claim that all reported experiments used the pre-trained adapter without fine-tuning is inconsistent with the method it proposes.","supporting_citations":[{"cited_title":"Zhang, N","cited_arxiv_id":null,"evidence_quote":"In-text source for the ControlNet-style structural conditioning branch used by the Structure Preservation Module."},{"cited_title":"Tumanyan, M","cited_arxiv_id":null,"evidence_quote":"Feature-injection baseline representative of inversion-style pipelines the paper aims to avoid."},{"cited_title":"Chung, S","cited_arxiv_id":null,"evidence_quote":"Training-free style-injection method that motivates the attention-injection design used in ICAS."}],"review_version":1}