{"id":"efffb62a-49e1-49ee-b65b-111095a7f58a","arxiv_id":"2509.05659","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":3.0,"correctness_risk":"high","formal_verification":"none","parameter_count":4,"one_line_summary":"EditIDv2 fine-tunes only PerceiverAttention cross-attention weights on 3K images to inject editability into Flux-based ID customization, reporting selective gains on the self-proposed IBench benchmark.","lead":"EditIDv2 is a method for generating images of a person in complex scenes from long text prompts, while changing pose and expression instead of pasting the face. It fine-tunes a small attention module on about 3,000 images, and claims state-of-the-art editability on a benchmark the authors themselves proposed.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq. (9) defines an ID loss as a cosine similarity between a spatial diffusion latent and a face-recognition embedding, but no projection is specified, making the loss uncomputable as written and undermining the central training mechanism.","rationale":"The reader's weakest-assumption analysis targets Eq. (9), and I agree that it is the most load-bearing flaw. The paper's central contribution is a training scheme that injects editability under minimal data lubrication; the ID loss is one of the three named mechanisms, and the joint dynamic training depends on it. If Eq. (9) is not a well-defined operation, the method cannot be reproduced from the text, the ablation in Section 5.3.1 cannot be interpreted, and the claimed empirical support for the central claim collapses. This is an internal correctness problem, not a disagreement with field consensus. It is also concrete: the dimension mismatch is visible from the paper's own architecture description. The proposed test—making the projection explicit and re-running the ablation—would settle whether the issue is a notation gap or a substantive error. Because the current manuscript provides neither code nor the missing projection, rejection is appropriate.","tokens_in":10436,"tokens_out":4798,"duration_ms":56774,"concrete_test":"Analytical dimension test: instantiate Eq. (9) with the paper's actual module shapes. Let x_hat_0 be the predicted Flux clean latent with shape [B, C, H, W] (e.g., [B, 16, 64, 64]) and let e_ref be an Antelopev2 face embedding with shape [512]. Attempt to compute the cosine numerator x_hat_0 · e_ref; it should fail without a defined projection or reshape. Then run the Section 5.3.1 ablation with a corrected, explicitly specified loss—e.g., decode x_hat_0 through the Flux VAE, extract an Antelopev2 embedding from the decoded image, and compute cosine similarity against e_ref. If the claimed joint-loss editability gains persist under this corrected loss, the concern is resolved; if not, the central training mechanism is unsupported as written.","verdict_should_be":"REJECT","load_bearing_attack":"The central claim—that minimal data lubrication plus a decomposed PerceiverAttention, ID loss, joint dynamic training, and offline fusion achieves SOTA editability in IBench—depends critically on Eq. (9). As written, the ID loss is L_ID = 1 - (x_hat_0 · e_ref)/(||x_hat_0||_2 · ||e_ref||_2). In the Flux/DiT setup, x_hat_0 is the predicted clean latent representation from the diffusion model, with shape [B, C, H, W] (or equivalently a token sequence [B, N, d]). The reference ID embedding e_ref comes from a face-recognition model such as Antelopev2 and has a compact dimension, typically [d] with d=512. These are vectors in different spaces: one is a spatial image latent in VAE/flow-matching latent space, the other is a discriminative face embedding. The dot product x_hat_0 · e_ref is therefore not dimensionally defined unless a projection, pooling, or decoding-and-re-encoding operation is introduced. The paper never specifies such a projection. Section 3.3 claims the loss 'quantifies identity fidelity by calculating the cosine similarity between the generated image and the reference ID features,' and Section 4.2 repeats that training extracts the cosine similarity between the generated image and the reference ID features. If the intended computation is to decode x_hat_0 through the VAE and re-encode the resulting image with Antelopev2, that is a nontrivial and non-obvious step that must be stated; if instead a direct latent-embedding cosine is intended, the dimensions must be aligned. Neither is provided. Consequently, the joint dynamic training scheme in Eqs. (10)-(11) cannot be implemented as written, and the Section 5.3.1 ablation attributing improvements to the ID loss loses its experimental foundation. This is the most load-bearing weakness: even if the qualitative comparisons are visually plausible, the paper's core technical contribution rests on an undefined mathematical operation. Separate concerns—such as Table 1 not strictly supporting the 'SOTA' label (Infi","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes EditIDv2, a method for identity-preserving text-to-image customization aimed at high-complexity narrative scenes and long prompts. It builds on the authors' prior EditID framework and the Flux/PuLID base model. The main ideas are: decomposing the PerceiverAttention cross-attention module, introducing an ID loss based on cosine similarity, jointly training the ID loss with the flow-matching diffusion loss, using a dynamic integration strength schedule, and merging multiple fine-tuned attention weight sets offline via a weighted fusion. Only about 3K labeled images are used to fine-tune the cross-attention weights. The paper claims state-of-the-art editability results on the IBench benchmark, particularly on pose, landmark, and expression diversity metrics.","tokens_in":10920,"tokens_out":3517,"duration_ms":36627,"significance":"If the central claims held, the contribution would be notable: a minimal-data, partial fine-tuning scheme that improves editability in long-prompt narrative scenes while preserving identity, and a modular decomposition of the ID integration process that could transfer to other ID customization systems. The paper also gives a clear practical emphasis on business-relevant scenarios. However, the technical foundation is undermined by an undefined ID loss, and the reported table does not support the claimed SOTA status. The benchmark and primary baseline come from the authors' own prior work, and no error bars or protocol details are given, so the empirical contribution is not currently established.","major_comments":[{"comment":"The ID loss is defined as a cosine similarity between x̂₀, the predicted clean latent of the Flux DiT (a spatial latent or token sequence), and e_ref, a compact face-recognition embedding from Antelopev2 (typically 512-d). These vectors live in different spaces; the dot product is dimensionally undefined unless a projection, pooling, or decode-and-re-encode operation is specified. No such operation appears in §3.3 or §4.2, although §4.2 refers to the 'cosine similarity between the generated image and the reference ID features.' This makes the joint training objective uncomputable as written and undermines the experimental foundation.","section":"§3.3, Eq. (9)"},{"comment":"The abstract and introduction claim SOTA results in the editability indicators of IBench, but Table 1 does not support this. InfiniteYou (sim) reports higher Yaw (25.02 vs 18.17), Roll (12.30 vs 11.39), and Exprdiv (0.635 vs 0.611) than EditIDv2. Moreover, FID, Dino, and Fgis are missing for EditIDv2, so generation quality and structural fidelity cannot be compared; the text criticizes InfiniteYou's high FID but reports no FID for EditIDv2. The table therefore does not substantiate the central claim.","section":"§5.2, Table 1"},{"comment":"The evaluation uses IBench, a benchmark proposed in the authors' own prior work EditID [11], and the primary baseline EditID is also their own. The offline fusion coefficients w_i in Eq. (12) are 'automatically adjusted based on validation set performance,' i.e., tuned to the target benchmark. No error bars, statistical significance tests, or cross-validation are provided. These factors make the SOTA claim fragile and difficult to verify independently.","section":"§5.1 and §3.4, Eq. (12)"},{"comment":"The ablations for the ID loss and offline fusion strategies are presented only with qualitative examples (Figs. 5 and 6) and no quantitative metric tables. Given that the paper's central claims concern the contribution of these components to editability, the absence of IBench metrics for the ablations prevents the reader from confirming which component is responsible for the reported improvements.","section":"§5.3"}],"minor_comments":[{"comment":"The roles of query, key, and value in Eq. (6) are unusual: Q is computed from the ID features and K/V from the generation branch. In standard cross-attention, the query usually comes from the generation branch and keys/values from the conditioning branch. Please clarify the intended attention direction and how this relates to the PerceiverAttention design.","section":"§3.2, Eq. (6)"},{"comment":"The paper repeatedly calls the approach 'tuning-free' while fine-tuning the cross-attention weights on about 3K images. This terminology is misleading; a term such as 'partial fine-tuning' or 'minimal tuning' would be more accurate.","section":"Abstract and §4.2"},{"comment":"The dashes for FID, Dino, and Fgis for EditIDv2 are not explained. If these metrics were not computed, state so explicitly; if they were, report them. The same applies to the missing Dino/Fgis for InfiniteYou.","section":"Table 1"},{"comment":"The training-data section says data come from MyStyle and 'publicly crawled data from the internet,' with no details on licenses, filters, or preprocessing. For reproducibility, include a more precise description of the data collection and labeling pipeline.","section":"§4.1"}],"recommendation":"reject","confidential_remarks":"The paper's central SOTA claim is contradicted by its own Table 1, and the main training objective in Eq. (9) is undefined as written. Since the evaluation is based on the authors' own IBench benchmark and the primary baseline is their own EditID, the empirical evidence is not sufficiently independent. These are load-bearing issues rather than presentation problems, and I do not see how they can be resolved within the current scope without re-running the core experiments."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a real attempt to fix a real problem — editability vs. consistency in ID customization under long prompts — and the specific recipe (fine-tune only PerceiverAttention weights, add an ID loss, dynamic integration strength, offline fusion of several weight sets) is modestly new, even though each ingredient is known. The qualitative figures look plausible, and the 'data lubrication' idea — using a few thousand images of the same ID with different attributes to inject editability — is worth taking seriously.\n\nThe soft spots are in the details. The biggest one is Eq. (9). The ID loss is written as a cosine similarity between x_hat_0, a spatial diffusion latent, and e_ref, a 512-d face-recognition embedding. Those live in different vector spaces and the dot product is not defined as written. No projection, pooling, or decode/re-encode step is specified. Since that loss is the centerpiece of the joint dynamic training in Eqs. (10)-(11) and the Section 5.3.1 ablation, the core training mechanism is unimplementable as the paper stands. This is not a minor typo; it needs a concrete correction or the central claims lose their experimental foundation.\n\nThe 'SOTA in IBench' claim also outruns Table 1. EditIDv2 is missing FID, Dino, and Fgis, and InfiniteYou has higher Posediv on all three Euler angles. The method is best on Aesthetic, Landmarkdiff, and ClipI, which is real, but it's not universally SOTA. Also, IBench comes from the authors' own EditID paper, and the offline fusion coefficients are tuned on the validation set, so there's some circularity. 'Tuning-free' is used in an unusual way — the method fine-tunes attention weights on a dataset, which is fine, but the label invites confusion.\n\nGiven all that, the paper deserves a serious referee, but as submitted it needs major revision: define Eq. (9) properly, report missing metrics and error bars, and soften the universal SOTA claim. If those are addressed, the combination could be a useful data point for the subfield.","headline":"Modest new combination for editability in ID customization, but the core ID loss in Eq. (9) is undefined as written and the 'SOTA' claim outruns the evidence.","tokens_in":11416,"tokens_out":4101,"would_cite":false,"duration_ms":44606,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Fine-tuning only a small cross-attention module with about 3,000 images lets character-ID models follow long narrative prompts while keeping the face recognizable.","keywords":["text-to-image generation","identity customization","editability","PerceiverAttention","data lubrication","flow matching","long-prompt narrative scenes","ID loss"],"falsifier":"Check the dimensions in Eq. (9): if the generated latent x_hat_0 is a spatial feature map and e_ref is a 512-dimensional face embedding, compute the loss as written and see whether training runs at all or needs an unspecified projection. Separately, run a controlled ablation that fixes the 3K-image data budget and varies only the offline fusion weights; if the editability gains disappear when the consistency-biased variant is removed, the claimed stepwise control mechanism is not what drives the result.","tokens_in":10376,"feed_emoji":"🎭","tokens_out":3579,"duration_ms":39109,"temperature":0.7,"pith_summary":"This paper tries to establish that character identity can be made editable in a text-to-image diffusion model by fine-tuning only a small cross-attention bridge, using about 3,000 images as a data lubricant. The claim is that this preserves identity similarity while giving long, multi-clause prompts real control over pose, expression, and scene, where prior methods either paste the reference face rigidly or lose editing fidelity. The evidence is the EditIDv2 pipeline built on the Flux/PuLID base, evaluated with the IBench suite of editability metrics. If correct, character customization no longer needs large-scale retraining to be both consistent and controllable in complex narrative scenes.","feed_headline":"3,000 images can make character-ID edits follow long story prompts","feed_subtitle":"EditIDv2 fine-tunes only PerceiverAttention weights to lift pose and expression diversity while keeping identity similarity—no full model tu","key_machinery":"The central mechanism is the decomposed PerceiverAttention module used as the ID feature integration bridge. It separates the computation into Q from the ID feature, K and V from the generation branch, modulates the query path with alpha(t) = alpha_0 * (1 - t/T), defines an ID loss as cosine similarity between the generated latent and the reference ID embedding, and jointly trains with the flow-matching diffusion loss under cosine-decayed ID guidance. Offline weighted fusion of multiple fine-tuned attention weight variants, one favoring consistency and one favoring editability, gives stepwise control. This set of mechanisms is what the paper claims carries the editability injection.","core_discovery":"EditIDv2 claims that the bottleneck for editable ID customization is not the text-to-image backbone or the face encoder but the ID feature integration module: the PerceiverAttention cross-attention that injects identity into the generation branch. By decomposing that module into separate query, key, and value paths, training only its cross-attention weights with an ID cosine-similarity loss jointly with the flow-matching diffusion loss, and then fusing several fine-tuned weight sets offline, the method reports large gains in pose diversity, landmark diversity, and expression diversity on long narrative prompts while holding identity similarity at 0.659 Facesim versus 0.735 for the rigid PuLI","pith_inferences":["The paper leaves implicit that data lubrication is about attribute diversity per identity rather than identity count; a direct extension would vary the number of identities versus images per identity to map where editability gains saturate.","If the mechanism is as general as claimed, applying the same fine-tune-only-cross-attention recipe to other ID adapters with different face encoders or different DiT bases should produce similar editability gains; this is testable by measuring pose and landmark variance without the full IBench suite.","The cosine-similarity ID loss in Eq. (9) is written between a spatial image latent and a compact face embedding; a reader should expect either a projection or a pooled latent to make the dot product well-defined, and the paper's silence on this point is the most likely implementation gap.","The reported trade-off implies a predictable relationship: as the initial integration strength alpha_0 decreases, editability metrics should rise while Facesim falls; plotting that curve would give practitioners a direct calibration tool."],"forward_implications":["Only the PerceiverAttention cross-attention weights need to be trainable, so adapting an existing ID pipeline costs a small fine-tune rather than full model training.","Long-prompt narrative scenes, with actions, expressions, and environmental changes, become addressable by the same ID customization model instead of only simple single-subject prompts.","A modest drop in face similarity, about 0.08 Facesim, is traded for large gains in pose diversity and landmark diversity, making the consistency-editability trade-off explicit and tunable.","The offline fusion idea allows deployment to choose between consistency-biased and editability-biased weights per task without retraining.","The paper states the framework can extend to other ID customization algorithms equipped with character feature branches."],"supporting_citations":[{"why":"PuLID supplies the Flux-based base model whose feature combination EditIDv2 retains and fine-tunes.","marker":"[8]"},{"why":"EditID defines the decoupling of the character feature branch and the IBench evaluation framework that EditIDv2 adopts.","marker":"[11]"},{"why":"Scaling rectified flow transformers supplies the flow-matching training objective used in the joint dynamic training.","marker":"[6]"},{"why":"DiT provides the Transformer backbone architecture whose attention structure motivates the PerceiverAttention decomposition.","marker":"[16]"},{"why":"ArcFace/Antelopev2 is the face recognition model providing ID features used for the ID loss and similarity evaluation.","marker":"[5]"},{"why":"MyStyle is a source of the small labeled training data used for data lubrication.","marker":"[15]"},{"why":"InfiniteYou serves as a comparative baseline whose instability, reflected in high FID, supports the claimed advantage of the fusion strategy.","marker":"[9]"}],"fun_headline_variants":["Tune only cross-attention to keep faces in long-text image edits","EditIDv2: data-lubricated ID integration for complex scenes","Character edits follow long prompts with 3K images and targeted tuning","Identity stays, edits deepen: EditIDv2's targeted attention fix","From rigid IDs to expressive edits: EditIDv2's Perceiver trick"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The ID loss in Eq. (9) is written as a cosine similarity between a generated spatial image latent and a compact face-recognition embedding, but the paper does not specify the projection that puts these two quantities in the same vector space; if that operation is not meaningful, the joint training scheme cannot be implemented as written.","fun_headline_variants_meta":{"raw":{"variants":["Tune only cross-attention to keep faces in long-text image edits","EditIDv2: data-lubricated ID integration for complex scenes","Character edits follow long prompts with 3K images and targeted tuning","Identity stays, edits deepen: EditIDv2's targeted attention fix","From rigid IDs to expressive edits: EditIDv2's Perceiver trick"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000594,"raw_usage":{"total_tokens":2609,"prompt_tokens":726,"completion_tokens":1883,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":470,"completion_tokens_details":{"reasoning_tokens":1786}},"tokens_in":470,"tokens_out":1883,"duration_ms":12450,"temperature":1.0,"reasoning_tokens":1786,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T05:15:03.193757+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Check the dimensions in Eq. (9): if the generated latent x_hat_0 is a spatial feature map and e_ref is a 512-dimensional face embedding, compute the loss as written and see whether training runs at all or needs an unspecified projection. Separately, run a controlled ablation that fixes the 3K-image data budget and varies only the offline fusion weights; if the editability gains disappear when the consistency-biased variant is removed, the claimed stepwise control mechanism is not what drives the result.","supporting_citations":[{"cited_title":"Pulid: Pure and lightning id customization via contrastive alignment.Advances in neural information pro- cessing systems, 37:36777–36804, 2024","cited_arxiv_id":null,"evidence_quote":"PuLID supplies the Flux-based base model whose feature combination EditIDv2 retains and fine-tunes."},{"cited_title":"Mystyle: A personalized generative prior","cited_arxiv_id":null,"evidence_quote":"MyStyle is a source of the small labeled training data used for data lubrication."}],"review_version":1}