{"id":"00b2eaa1-17a7-49be-b2cb-4299943cfbca","arxiv_id":"2412.16832","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"RealisID uses two complementary control branches to deliver scale-robust identity fidelity, fine facial control, and zero-shot multi-person customization for text-to-image generation.","lead":"RealisID is a face-customization method that combines a face-cropped local branch with a location-guided global branch to keep identity fidelity for small faces and to control face location, pose, and expression. The same framework extends to multi-person images without training on multi-person data, which makes it potentially useful for production portrait editing and personalized image generation.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Multi-person customization relies on an untrained linear combination of branch outputs; the only evidence is a self-comparison that does not test overlap or cross-identity leakage.","rationale":"The reader's weakest_assumption identifies the same load-bearing concern: multi-person inference assumes a linear combination of local-branch sums and global-branch averages with no training or derivation. I agree that this is the point on which the paper's most ambitious claim depends. I considered whether the asymmetric baseline prompting for small faces or the lack of error bars might be more fundamental, but those affect the strength of the comparison rather than the validity of a headline capability. The multi-person superposition, by contrast, is asserted without evidence; if it fails under overlap or identity-confusion stress, the claim of 'easily extended to handle multi-person customization, even only trained on single-person datasets' is directly invalidated. The paper's own Limitation section candidly says per-person clothing and hairstyle control is not solved, which is consistent with a conditional acceptance, but it does not address the possibility of identity cross-contamination in the summed injections. The proposed concrete test would settle whether the superposition holds by measuring both true-pair identity retention and cross-pair leakage under overlapping and non-overlapping conditions. Because the reader already assigned CONDITIONAL, and my analysis reinforces rather than shifts that judgment, the appropriate verdict remains CONDITIONAL; hence UNCHANGED.","tokens_in":15249,"tokens_out":4669,"duration_ms":44575,"concrete_test":"Build a two-person evaluation using 20 identity pairs from CelebA-HQ. For each pair (A, B), generate images with (1) non-overlapping face boxes and (2) boxes with roughly 30% overlap. For each generated image, run face detection and FaceNet verification of the left face against A and B, and the right face against A and B, reporting both true-pair and cross-pair similarities. Compare with single-person generation at the same box sizes. If the true-pair FaceNet similarity drops by more than 0.05 relative to the single-person baseline, or if the cross-pair similarity comes within 0.05 of the true-pair, the linear superposition is cross-contaminating. Additionally, swap the order of the two reference images while keeping location masks fixed; the identity-to-location mapping should swap exactly, and any asymmetry reveals an order-dependent flaw in the summation.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's multi-person extension, described under 'Inference for Multi-Person Customization' and Fig. 2(c), sums local-branch injections and averages global-branch injections. This rule is not derived, not trained, and not analyzed for cross-identity interference. Each local branch was trained on single-person data to map a face crop and an ID embedding to facial detail features; nothing in the training objective in Eq. (7) enforces that adding two such injection maps in the same noise latent preserves each identity. The problem is most acute when face regions overlap: local branches operate on overlapping crops of z_t, and the relocated injections are superimposed at the same spatial positions, so the summed feature can conflate the two identities. The only quantitative support is Table 9, which compares FaceNet for single-person vs multi-person generation on the same 40 identities. That comparison is a self-comparison with no baseline, does not separate overlapping from non-overlapping layouts, and does not measure identity confusion (e.g., whether face A inherits features of face B). Since the headline contribution includes zero-shot multi-person customization from single-person training, this untested superposition is the most load-bearing assumption. If it fails, the multi-person claim collapses even though the single-person branch design may be sound.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes RealisID, a zero-shot identity customization method built on SDXL and two ControlNet-style branches. A local branch crops and up-samples the face region of the noisy latent and injects identity, pose, and expression information, while a global branch consumes a face-location mask and injects layout information. The two branches are trained jointly on single-person data with the loss in Eq. (7), which adds a face-region-weighted denoising term. At inference, multi-person customization is performed by summing the local-branch injections and averaging the global-branch injections across identities. Experiments compare RealisID with IP-Adapter-face-plus, FlashFace, PhotoMaker, InstantID, and PuLID on regular and small-face settings, and include ablations, fine-control demonstrations, and a single-vs-multi-person self-comparison.","tokens_in":15588,"tokens_out":6745,"duration_ms":59714,"significance":"If the reported results are reliable, RealisID would be a useful contribution: it offers a simple branch decomposition that jointly provides small-face identity fidelity, fine location/pose/expression control, and a zero-shot route to multi-person customization from single-person training. The paper's strengths are a clearly specified architecture, an explicit training objective with a face-mask term, and quantitative comparisons against five published baselines. The ablation evidence for branch complementarity is also informative. However, the multi-person extension rests on an untested linear-superposition assumption, the quantitative evidence lacks error bars and significance tests, and the small-face evaluation protocol is asymmetric between methods. These issues directly affect the headline claims, so the manuscript needs additional evidence before it can be accepted.","major_comments":[{"comment":"The multi-person customization claim is supported only by an untrained integration rule. Summing the local-branch injections and averaging the global-branch injections across identities is not motivated by the training objective in Eq. (7), which only ever sees one identity, and no analysis shows that the summed injections preserve identity separation. Table 9 is a self-comparison of RealisID single vs. multi on the same 40 identities; it does not measure cross-identity leakage (e.g., whether the generated face of person A contains features of person B), does not separate overlapping from non-overlapping layouts, and includes no baseline. Because multi-person customization is one of the three headline requirements, this needs direct evidence: e.g., identity confusion rates, separate results for overlapping and non-overlapping face boxes, and a comparison against a method trained for multi-ID generation or against an oracle combination.","section":"Inference for Multi-Person Customization; Fig. 2(c); Table 9"},{"comment":"The central quantitative claim of small-face superiority is not backed by variance or significance information. Table 1 reports single-point means over 2800 images per method; differences such as FaceNet 0.767 vs. 0.693 for InstantID and 0.767 vs. 0.497 for PuLID are presented without standard deviations, confidence intervals, or tests, so the reader cannot tell whether the ordering is stable. In addition, the small-face protocol is asymmetric: methods that cannot control face size receive the extra text prompt \"a full-body people image,\" while RealisID instead receives a location guidance mask. This changes the conditioning between arms and may affect all four metrics; the prompt list in Table 5 does not include this extra phrase. Please report per-method error bars and a same-prompt control (e.g., give all methods the same text prompt while RealisID additionally uses location guidance, or evaluate PhotoMaker and PuLID with the identical unmodified prompt as well).","section":"Quantitative Comparisons; Table 1"},{"comment":"The scale-alignment mechanism is not isolated. Table 2 and Fig. 4 compare the full model with and without the entire local branch, so they show that the local branch matters overall, but they do not show that the proposed cropping/up-sampling and relocation operation (Eqs. (4)-(5)) are what confer scale robustness. An ablated local branch that consumes the full latent z_t without cropping/up-sampling, or one that omits relocation, would directly test the paper's stated mechanism. Without such an ablation, the \"scale-robust\" property could plausibly come from the face-mask loss in Eq. (7) or from the global location branch rather than from the up-sampling operation.","section":"Ablation Studies, Scale Robustness; Table 2; Fig. 5"}],"minor_comments":[{"comment":"The caption says \"the odd rows display the quantitative results,\" but the context indicates these are qualitative results; please correct the wording.","section":"Figure 3 caption"},{"comment":"The first sentence of the Regular Case paragraph says \"the odd rows display the quantitative results\" and should read \"qualitative results.\"","section":"Section 'Qualitative Comparisons'"},{"comment":"The notation around Eq. (5) is confusing: Eq. (3) already defines i_t^l, Eq. (4) then defines \\hat{i}_t^l = R(i_t^l, c_loc), and Eq. (5) redefines i_t^l using \\hat{z}_t. Please reorder or rename the variables so the dependency chain is unambiguous.","section":"Eqs. (3)-(5)"},{"comment":"There is a typo: \"desnoier\" should be \"denoiser.\"","section":"Text after Eq. (5)"},{"comment":"The paper does not state whether code or trained models will be released; given that the architecture is the main contribution, a code-release statement would improve reproducibility.","section":"Reproducibility"}],"recommendation":"major_revision","confidential_remarks":"The paper is competently written and the local/global branch idea is interesting. The single-person results are plausible, but the multi-person claim is the weakest point: the superposition rule is not validated for cross-identity leakage, and Table 9 does not answer the question. I would like to see the requested multi-person diagnostics and error bars before acceptance. I do not see grounds for rejection if these experiments are added."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"RealisID is a worthwhile single-person identity customization method with a genuinely novel local-branch design; the multi-person part is a nice story backed by thin evidence. The local branch crops the face latent, up-samples it, runs a ControlNet, and relocates the injected features back to the original position. That is a concrete, sensible mechanism, and the ablations (Table 2) do show the local branch contributes identity fidelity in small-face settings while the global branch helps layout. The reported gains over InstantID and PuLID on FaceNet for small faces are consistent, and the qualitative examples support the claim. The global location mask is also a clean way to control face position. This is a real contribution.\n\nThe soft spots are real but mostly not fatal. There is no released code, no error bars or significance tests, and the small-face comparison uses an extra \"full-body people image\" prompt for baselines that cannot control face size. That is a reasonable workaround but still an asymmetry; the numbers should be read with that in mind. The fine-control experiments compare against InstantID only, and the pose/expression metrics are L1 distances to DECA parameters, which is fine but limited.\n\nThe biggest soft spot is the multi-person customization. The method sums local-branch injections and averages global-branch injections at inference, trained only on single-person data. The only quantitative evidence is Table 9, a self-comparison of FaceNet between single- and multi-person generation, which cannot detect cross-identity leakage. The qualitative results, including overlapping faces, look plausible, but there is no baseline comparison and no metric measuring whether face A has inherited features from face B. Given that \"zero-shot multi-person from single-person training\" is a headline claim, this needs more than a self-comparison. I suspect it works reasonably for two non-overlapping faces and degrades with overlap, but the paper does not show that.\n\nThe authors do state a candid limitation: they cannot precisely control clothing or hairstyle per person in multi-person scenes. That honesty keeps the single-person claims credible.\n\nBottom line: this is a serious paper with a believable single-person contribution. The multi-person generalization is a promising heuristic, not an established capability. The paper is for anyone working on identity customization or controllable portrait generation; they will get something from the architecture. It deserves a real peer review, and the reviewer should push for code, error bars, and a proper multi-person evaluation with identity-confusion metrics.","headline":"A solid single-person identity customization paper with a clever local-branch design; the zero-shot multi-person claim is an unproven heuristic that needs much stronger evidence.","tokens_in":16075,"tokens_out":2410,"would_cite":true,"duration_ms":20517,"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":"RealisID claims one zero-shot adapter can preserve small-face identity, control pose and location, and compose multiple people by pairing a face-cropped local branch with a location-guided global branch.","keywords":["identity customization","text-to-image diffusion","small face identity fidelity","ControlNet","face location control","pose and expression control","multi-person customization","zero-shot personalization"],"falsifier":"Take two reference identities, generate a two-person image with heavily overlapping location masks, and measure each person's FaceNet similarity against its reference. If per-identity similarity falls well below the same method's non-overlapping score, or identities appear swapped between the two boxes, the linear-combination multi-person claim fails. A second test: run the same recipe with three or four identities and check whether global averaging still preserves each face at the specified location.","tokens_in":15068,"feed_emoji":"🎭","tokens_out":9978,"duration_ms":76809,"temperature":0.7,"pith_summary":"RealisID is an identity-customization method for text-to-image diffusion models. It claims that a single zero-shot adapter—no per-identity fine-tuning—can keep a face recognizable even when it is small in the frame, can place that face at a chosen location with a chosen pose and expression, and can put several specified identities into one image. Existing adapters, the paper argues, handle some of these but not all at once. RealisID's answer is to split the job: a local branch works only on a cropped and up-sampled face latent, while a global branch reads a location mask and the whole latent. If this works as claimed, one trained model replaces separate tools for portraits, small-face ID, and group customization.","feed_headline":"One face model handles small faces, fine control, and groups","feed_subtitle":"Local face-crop branch plus global location branch gives small-face fidelity, pose control, and multi-person images.","key_machinery":"The load-bearing objects are two ControlNet variants (trainable conditional copies of the diffusion U-Net that inject signals through zero convolutions) and the data flow between them. The local branch receives a cropped, up-sampled face latent $\\hat{z}_t$ and facial-landmark coordinates $c_{p\\&e}$; its output is down-sampled and relocated into a zero tensor according to the location mask $c_{loc}$. The global branch receives the full latent $z_t$ plus the mask $c_{loc}$. Both are conditioned on a projected CLIP identity embedding $p_{id}$. The training loss adds a face-mask-weighted noise-reconstruction term $\\lambda \\|(\\epsilon_\\theta(z_t,p_{text},t)-\\epsilon)\\odot c_{loc}\\|^2$ to the standard diffusion loss, concentrating detail learning inside the face box. At multi-person inference, the local injections are summed and the global injections averaged, on the assumption that local face signals do not interfere.","core_discovery":"The central claim is that the complementarity of two ControlNet-style branches is enough to satisfy all three requirements simultaneously. The local branch crops the face region from the noisy latent, up-samples it, and injects identity plus facial-landmark pose/expression conditions; because the face is normalized in scale, identity details survive even when the target face is a small fraction of the image. A relocation operation then places the local features back into the full latent at the position given by the location mask. The global branch takes the same identity embedding and the mask itself as conditions, providing layout, body/background placement, and overall coherence. The authors further claim that because each branch is a ControlNet variant, multi-person inference can be done without multi-person training by summing local-branch outputs and averaging global-branch outputs for the different identities. The paper's stated boundary is that this per-person control covers face-region factors only; per-person clothing and hairstyle in group images remain governed by the text prompt.","pith_inferences":["We infer the 'sum local, average global' recipe is a linear-superposition assumption; it should be stress-tested by increasing overlap between location masks and by going beyond three identities, since the paper only self-compares FaceNet on two/three-person examples.","We infer that the local branch's robustness hinges on face-detection quality: if the detector misses or mis-sizes a face, the up-sampling normalization will encode the wrong region, so detector noise is likely the main practical failure mode.","The relocation operation is a hard spatial copy, so identity features may not blend with surrounding content; the paper's own ablation shows the global branch is needed for harmony, and we suspect the same boundary will appear for partial occlusions.","Because training data are single-person and style is a global signal, the method likely transfers to other backbones but will inherit each backbone's text-encoder limitations on spatial language."],"forward_implications":["Small faces keep identity: because the local branch sees a scale-normalized face crop, FaceNet/CLIP-I scores stay high for face sizes down to 1/7 of the image edge.","Fine control becomes mask-driven: editing the binary mask changes location and size, and swapping facial landmarks changes pose and expression, with no text engineering.","Multi-person generation emerges from single-person training: summing local ControlNet outputs and averaging global outputs creates two-person and three-person composites without multi-person data.","The framework plugs into a pretrained SDXL backbone and inherits its prompt and style handling.","The method's own scope note: per-person clothing and hairstyle in group images are not precisely controllable."],"supporting_citations":[{"why":"Supplies ControlNet, the conditional-injection architecture used for both branches.","marker":"Zhang, Rao, and Agrawala 2023"},{"why":"IP-Adapter; supplies the ID-embedding projection/injection scheme and the random prompt-drop training strategy.","marker":"Ye et al. 2023"},{"why":"SDXL; the frozen text-to-image backbone RealisID builds on.","marker":"Podell et al. 2023"},{"why":"CosmicMan; the single-person image-text dataset used for training both branches.","marker":"Li et al. 2024a"},{"why":"InstantID; the main zero-shot baseline and source of the identity-conditioned classifier-free guidance recipe.","marker":"Wang et al. 2024a"},{"why":"PhotoMaker; baseline and reference for stacked ID-embedding and evaluation metrics.","marker":"Li et al. 2024b"},{"why":"PuLID; baseline for identity fidelity comparison.","marker":"Guo et al. 2024"},{"why":"Provides the delayed-subject conditioning used at inference.","marker":"Xiao et al. 2023"},{"why":"FlashFace; baseline for fine facial control and small-face comparison.","marker":"Zhang et al. 2024"}],"fun_headline_variants":["Local-global branches give scale-robust identity and group control","Two-branch complementation solves small-face fidelity and multi-person","Scale-robust identity and fine control via local-global branches","Local crop and global layout: identity fidelity for small faces and groups","Multi-person identity from single-person training via branch complementation"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"Multi-person inference assumes that the control signals from different identities combine linearly—local branch outputs can simply be summed and global outputs averaged—and the paper gives no derivation or dedicated multi-person training to justify that superposition, so overlapping or crowded faces could cross-contaminate identities.","fun_headline_variants_meta":{"raw":{"variants":["Local-global branches give scale-robust identity and group control","Two-branch complementation solves small-face fidelity and multi-person","Scale-robust identity and fine control via local-global branches","Local crop and global layout: identity fidelity for small faces and groups","Multi-person identity from single-person training via branch complementation"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001131,"raw_usage":{"total_tokens":4724,"prompt_tokens":994,"completion_tokens":3730,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":610,"completion_tokens_details":{"reasoning_tokens":3642}},"tokens_in":610,"tokens_out":3730,"duration_ms":18523,"temperature":1.0,"reasoning_tokens":3642,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T10:15:04.538208+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take two reference identities, generate a two-person image with heavily overlapping location masks, and measure each person's FaceNet similarity against its reference. If per-identity similarity falls well below the same method's non-overlapping score, or identities appear swapped between the two boxes, the linear-combination multi-person claim fails. A second test: run the same recipe with three or four identities and check whether global averaging still preserves each face at the specified location.","supporting_citations":[],"review_version":1}