{"id":"a02428ac-dbf9-483d-b559-2021ea973ec7","arxiv_id":"2607.25622","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":8,"one_line_summary":"A dual-branch FLUX-based generator with dynamic temporal gating and region-aware losses improves the face-vs-appearance trade-off in personalized person image generation.","lead":"This paper combines an existing face-identity branch with an appearance branch in a single diffusion model, and adds a training procedure that dynamically balances the two signals. It also introduces a 100-image benchmark for evaluating holistic person identity preservation.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The central claim of a 'better trade-off' is not operationalized: DBS does not Pareto-dominate open-source baselines (e.g., Qwen-Image), and no holistic metric or preference study supports the superiority claim.","rationale":"The reader's weakest_assumption focused on the theoretical sufficiency of scalar gates to repair representation-space mismatches. That is a meaningful concern about the mechanism, but the central claim is empirical: DBS achieves a better trade-off than open-source baselines. For that claim to hold, 'better trade-off' must be measurable and DBS must actually win under that measure. The paper does neither. Table 1 shows DBS is not Pareto-dominant even among open-source methods, and no composite metric or user study defines 'better.' This makes the headline assertion unfalsifiable, which is a more direct threat to the central contribution than the gating-mechanism question. The reader's rationale did mention missing significance tests and the self-built benchmark, but did not make the conceptual point that the trade-off claim lacks an operational definition. I agree with the CONDITIONAL verdict because the empirical work is suggestive but not decisive; additional validation (preference study and defined holistic metric) could settle it. No change in verdict is needed, but the condition should explicitly require an operationalized trade-off measure and statistical support.","tokens_in":9629,"tokens_out":9864,"duration_ms":105450,"concrete_test":"Run a pairwise human preference study on Pexels-100 comparing DBS against Qwen-Image, InfiniteYou, and InstantCharacter: for each of 100 generated images, raters choose which better preserves holistic identity. Report win rates and a binomial significance test. Additionally, pre-register a holistic score, e.g., a weighted combination of Face and ReID (or DINOv2), and verify whether DBS outperforms open-source baselines across a range of weights. If the preference win rate does not significantly exceed 50% or if no weight region yields DBS as the best, the central claim is unsupported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The abstract and conclusion state that DBS achieves a better trade-off between facial fidelity and appearance consistency than existing open-source baselines. However, 'better trade-off' is never defined. Table 1 shows DBS (Face 0.7875, ReID 0.9122, DINOv2 0.3202, CLIP-T 0.2008) is not Pareto-dominant among open-source methods: Qwen-Image has higher ReID (0.9230) and CLIP-T (0.2279) at lower Face (0.4640) and DINOv2 (0.2486). Whether DBS's point on the Pareto frontier is preferable is a value judgment, not something the paper establishes. The paper presents no combined holistic metric, no preference study, and no trade-off curve. Without such a criterion, the central claim is unfalsifiable as stated. This is more load-bearing than the mechanism question: even if the gating is theoretically sound, the empirical assertion is not testable. The small improvements over NDB in Table 2 (Face +0.013, ReID +0.002, DINOv2 +0.018) further underscore that noise could explain the reported gains, since no error bars or significance tests are provided.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper addresses holistic identity preservation in personalized person image generation, proposing a naive dual-branch (NDB) baseline that combines a global appearance branch (InstantCharacter-style) and a local face branch (InfiniteYou-style) on the FLUX.1-dev backbone. It then introduces Dynamic Balancing Scaling (DBS), a fine-tuning strategy with two components: Adaptive Temporal Gating (ATG), which applies time-dependent scalar gates to the face and appearance branch residuals (Eqs. 2-5), and Region-Aware Optimization (RAO), which decomposes the denoising loss into face, appearance, and global terms with a Pareto-style soft aggregation (Eqs. 6-10). The paper also introduces a new benchmark, Pexels-100, and reports quantitative metrics (Face, ReID, DINOv2, CLIP-T) and qualitative comparisons. The central claim is that DBS achieves a better trade-off between facial fidelity and appearance consistency than existing open-source baselines.","tokens_in":9935,"tokens_out":2796,"duration_ms":33487,"significance":"If validated, the work offers a lightweight, backbone-frozen fine-tuning recipe for coordinating heterogeneous identity branches in a diffusion transformer, which could be practically useful and generalizable beyond person generation. The introduction of a dedicated holistic-consistency benchmark is also a potentially valuable community resource. However, the empirical evidence in the current manuscript is substantially weaker than the claims. All quantitative results come from a single run with no error bars or significance tests; the Pexels-100 benchmark is introduced but not released; and the central 'better trade-off' claim is not operationalized with any holistic criterion, preference study, or Pareto analysis. The method's mechanism is plausible but rests on the assumption that scalar, time-dependent gates can fix branch imbalance, which is asserted rather than demonstrated. The paper ships no code, no benchmark, and no machine-checked artifacts; its current value is mainly conceptual.","major_comments":[{"comment":"The central claim 'DBS achieves a better trade-off' is not operationalized and, as stated, is unfalsifiable. In Table 1, DBS does not Pareto-dominate open-source baselines: Qwen-Image has higher ReID (0.9230 vs 0.9122) and CLIP-T (0.2279 vs 0.2008) but lower Face (0.4640 vs 0.7875); InfiniteYou has lower ReID but higher Face. Whether DBS's point is preferable is a value judgment. The paper needs a defined trade-off criterion, a combined holistic metric, a preference study, or an explicit Pareto-frontier analysis. Without this, the main abstract and conclusion claims are not testable.","section":"Abstract / Table 1"},{"comment":"All quantitative results are single-run with no error bars, seeds, or significance tests. In Table 2, the full model differs from NDB by Face +0.013, ReID +0.002, DINOv2 +0.018, and CLIP-T +0.002; these differences are within typical run-to-run noise. The same issue affects Table 3 comparisons with MSE, LS, and FairGrad. The ablations therefore do not currently support the claim that ATG and RAO contribute meaningfully. Report means and standard deviations over at least three independent runs and, if appropriate, paired significance tests.","section":"Tables 2 and 3, Sections 4.3-4.4"},{"comment":"The evaluation is built on a self-constructed benchmark that is not released, and the construction is described in only one sentence: 'We used ChatGPT to generate 100 text prompts and paired them with these images.' No details are given on the number of identities, the prompt-image pairing procedure, filtering criteria, or diversity. The ReID metric (TransMatcher) is co-authored by a co-author, creating a conflict-of-interest concern that should be addressed with independent metrics or human evaluation. The paper should release the benchmark (or a clear description with statistics) and include additional holistic metrics (e.g., CLIP appearance similarity, human preference study).","section":"Section 4.2 / Pexels-100"},{"comment":"The motivation for ATG rests on the perturbation analysis in Figure 2, which reports an average difference in noise-prediction change when removing each branch. No details are given about how this average is computed, across how many images, or with what variance. Moreover, the proposed scalar gates (Eqs. 2-3) only rescale branch residual magnitudes; if the instability arises from incompatible representation spaces (cross-attention residuals vs ControlNet residuals), such gating cannot address it. The paper should provide quantitative evidence that the gates actually rebalance branch contributions over the denoising trajectory, e.g., by plotting the learned gate values and the resulting perturbation magnitudes before/after training.","section":"Section 3.3 and Figure 2"}],"minor_comments":[{"comment":"Typo: 'the denoising trajectory' should be 'along the denoising trajectory'. Also, 'at different identity information' should be 'at different identity granularities' or similar.","section":"Section 3.4 (first paragraph)"},{"comment":"Abbreviation inconsistency: the caption defines 'RAB: Region Aware Optimization' but the text and equations use 'RAO'. Please unify.","section":"Figure 3 caption"},{"comment":"The row label 'Upcoming Ours' is awkward and perhaps an artifact; use 'Ours'. Also, the names 'Nano banana' and 'Nano banana2' should be capitalized consistently.","section":"Table 1"},{"comment":"The face similarity threshold of 0.75 for filtering PPR10K is not attributed to a specific face recognition model. Specify which model and which image pairs were used to compute similarity.","section":"Section 4.2"},{"comment":"Only Face scores are shown per generated image. Since holistic consistency is the focus, include ReID or DINOv2 scores for each qualitative example, or at least state how the Face scores were computed.","section":"Figure 4"}],"recommendation":"major_revision","confidential_remarks":"The proposed method is simple and plausible, and the benchmark idea is useful, but the evaluation falls short of supporting the central 'better trade-off' claim. I am particularly concerned about the absence of any uncertainty quantification, the non-released self-built benchmark, and the use of a co-authored ReID metric. The authors should be asked to either release the benchmark and code, provide multi-run statistics and a defined trade-off criterion, or add an independent human evaluation. If these cannot be provided, the claims should be substantially softened."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Worth a look if you work on diffusion-based personalization: the authors take two existing adapters (InstantCharacter appearance, InfiniteYou face), bolt them onto FLUX.1-dev, observe that the face branch dominates, and fix it with timestep-dependent gates plus a region-aware loss. The diagnosis is the best part: they measure how much removing each branch perturbs noise prediction and show the face branch dominates, then address it directly. That is real, reproducible reasoning.\n\nWhat's new is the specific combination and the gating/loss scheme, not a new theory. The Pexels-100 benchmark is a modest addition, but calling it the 'first benchmark' for holistic consistency is not defended; it is self-built, not released, and the pairing of 100 Pexels images with ChatGPT prompts is described in a sentence. The PPR10K training-set construction (10,444 pairs, 850 identities) is also under-specified.\n\nSoft spots are where the evaluation does not support the abstract. 'Better trade-off' is never defined. Table 1 doesn't show Pareto dominance: Qwen-Image has higher ReID and CLIP-T with lower face and DINOv2; whether DBS is preferable is a value judgment. No holistic metric, no preference study, no error bars, no significance test. Many deltas in Tables 2 and 3 are tiny (e.g., ReID +0.002, CLIP-T +0.002). One run per configuration could easily explain these. The ReID metric, TransMatcher, is co-authored by one of the paper's authors; not circular by itself, but combined with the self-built benchmark, it argues for caution.\n\nThe mechanism itself is plausible. I don't see an internal contradiction in the equations. The weakest assumption is that scalar gates on frozen residuals can repair what is fundamentally a cross-representation imbalance; the perturbation analysis suggests the imbalance is real, but the fix is tested only on the authors' own tables.\n\nWho benefits: practitioners who want a lightweight way to combine off-the-shelf identity adapters and need a starting point. It deserves a serious referee, because the combination is useful and the weak points are clear. But I'd want code, benchmark release, repeated runs, and a stated decision criterion for 'better trade-off' before treating the central claim as established.","headline":"Useful engineering note on combining a face branch and an appearance branch for person personalization; the reported gains are small and the 'better trade-off' claim needs a defined criterion and open artifacts before it convinces me.","tokens_in":10455,"tokens_out":4274,"would_cite":false,"duration_ms":47514,"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":"The paper claims that holistic identity preservation in personalized person image generation reduces to coordinating a face branch and an appearance branch with time-aware scalar gates and region-aware supervision, without retraining the di","keywords":["personalized image generation","identity preservation","face consistency","appearance consistency","diffusion transformer","temporal gating","region-aware optimization","holistic identity"],"falsifier":"Concretely, measure the perturbation gap after removing each branch, as the paper does, and then reimplement the face branch with the same injection mechanism and feature dimensions as the appearance branch while retraining only the gates; if face over-dominance remains, relative signal strength is not the cause and the gate hypothesis fails.","tokens_in":9476,"feed_emoji":"🖼️","tokens_out":5052,"duration_ms":51071,"temperature":0.7,"pith_summary":"The paper tries to establish that the usual trade-off in personalized person image generation—keeping either the face or the overall appearance consistent, but not both—is not a hard limit of the generation model but a coordination problem between two conditioning pathways. It proposes a simple two-branch setup, one branch carrying global appearance cues and one carrying facial identity, and then shows that instability in that setup comes from the face branch dominating the prediction. The proposed fix, DBS, uses timestep-dependent scalar gates to rebalance the branches during denoising and a region-aware loss that protects facial identity while letting appearance guidance strengthen. If correct, this makes holistic identity preservation a lightweight fine-tuning problem on a frozen diffusion transformer, and the new Pexels-100 benchmark gives the community a shared way to measure it.","feed_headline":"Lightweight gates balance face and clothes identity in AI portraits","feed_subtitle":"Tuning only gate parameters coordinates face and appearance branches on frozen diffusion transformers.","key_machinery":"Adaptive Temporal Gating (ATG): per-layer scalar gates that are functions of the diffusion timestep; the face gate is 1 + delta*tanh(...) to stay close to its pretrained regime, and the appearance gate is a time-dependent multiple of the face gate with a progressive relaxation coefficient that starts near zero and grows with training. These gates multiply the branch residuals before injection into the transformer, letting the model shift weight from face to appearance as denoising proceeds. Region-Aware Optimization (RAO): the denoising error map is split into face-mask, person-mask, and global regions; each regional loss is normalized by an EMA reference, and a log-sum-exp aggregation with","core_discovery":"On the paper's own terms, the central claim is that face fidelity and appearance consistency can be jointly preserved by explicitly separating identity into two conditioning branches and then coordinating them with scalar, time-dependent gates plus spatially structured supervision. The face branch is treated as a stable anchor: its gate is bounded near unity, and the appearance branch's gate is initialized relative to the face gate and gradually relaxed during training. A Pareto-style region-aware loss then activates facial and appearance objectives only when they underperform relative to a running average, while the global objective continues to optimize image quality. The authors report th","pith_inferences":["If the rebalancing is truly achieved by scalar gates, the same recipe should transfer to other heterogeneous conditioning branches (pose, style, layout) on frozen diffusion transformers—a testable extension the paper does not run.","Because only gate parameters are trained, DBS-like tuning could be an inexpensive adapter layer for new backbone versions, letting identity preservation ride on top of improved generators.","The paper's diagnosis that face-branch over-dominance reflects injection-mechanism strength rather than inherent importance implies that aligning injection mechanisms (same pathway type) might achieve similar balance without gates—an alternative the authors do not compare.","The authors concede that complex poses and fine structures like hands still distort; whether gating can ever address structural defects is an open question beyond the paper's current scope."],"forward_implications":["Face fidelity and appearance consistency can be improved jointly by tuning only gate parameters, leaving the generative backbone frozen.","The naive dual-branch setup plus DBS provides a controllable basic framework: users can strengthen or weaken face versus appearance conditioning by adjusting gate scales.","The perturbation-gap measurement quantifies branch dominance and predicts where coordination is needed.","Region-aware Pareto-style supervision protects identity-critical regions without sacrificing global image quality.","The Pexels-100 benchmark offers a reusable protocol for comparing holistic identity rather than face-only fidelity."],"fun_headline_variants":["Gated branches preserve whole-person identity in AI portraits","Dynamic scaling coordinates face and body identity","Unified dual-branch control for holistic identity","Time-gated identity beats face-only consistency","Balance face and appearance with adaptive gates"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The load-bearing premise is that branch conflict is caused by relative signal strength, not by incompatible representation spaces or injection mechanisms; if conflict persists under equalized architectures, the scalar gates cannot fix it.","fun_headline_variants_meta":{"raw":{"variants":["Gated branches preserve whole-person identity in AI portraits","Dynamic scaling coordinates face and body identity","Unified dual-branch control for holistic identity","Time-gated identity beats face-only consistency","Balance face and appearance with adaptive gates"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000215,"raw_usage":{"total_tokens":1246,"prompt_tokens":707,"completion_tokens":539,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":451,"completion_tokens_details":{"reasoning_tokens":472}},"tokens_in":451,"tokens_out":539,"duration_ms":6550,"temperature":1.0,"reasoning_tokens":472,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-01T01:50:45.854362+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Concretely, measure the perturbation gap after removing each branch, as the paper does, and then reimplement the face branch with the same injection mechanism and feature dimensions as the appearance branch while retraining only the gates; if face over-dominance remains, relative signal strength is not the cause and the gate hypothesis fails.","supporting_citations":[],"review_version":1}