{"id":"2753e9ac-24b0-4bfb-b68b-2af7232b56d1","arxiv_id":"2607.15619","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"Using identifier-based structured context instead of plain-language instructions improves multi-reference image generation in semantic alignment and reference consistency.","lead":"StructGen encodes multiple reference images in a dictionary with identifiers and rewrites instructions to point at those identifiers, instead of relying on verbose natural-language descriptions. The method is tested on multi-reference image generation, with claimed gains in prompt following and identity/object consistency.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The headline comparisons change input format and training data simultaneously; the sole format ablation reports only ID on two subsets, so the claim that identifier anchors themselves drive the gains is not yet cleanly isolated.","rationale":"The reader's weakest assumption identifies exactly the same load-bearing concern: the empirical comparison changes both input format and training data, so the performance gap cannot be attributed to the structured context alone. My stress-test confirms this is the most serious threat to the central claim. The paper's own ablation (configuration (a) vs (e)) is the correct experiment, but it is too narrow: it reports only the ID metric on two challenging subsets, whereas the central claim spans prompt following and subject consistency as well. A single controlled experiment that varies only the instruction format on the identical training mixture and evaluates all three metrics across all benchmark subsets would settle whether the identifier anchors are doing the work. Because the paper already receives a CONDITIONAL verdict, this concern does not change the reader's verdict; it reinforces it. No ad hominem is intended; the issue is purely experimental isolation.","tokens_in":21774,"tokens_out":4463,"duration_ms":49708,"concrete_test":"Retrain the exact configuration (e) — same combined dataset (MICo-De&Re + curated 15,965 samples), same 13k steps, same mixed-modality sampling — with only one change: replace every identifier-based instruction with the corresponding natural-language instruction (same entity descriptions rendered as text, no <human_i>/<obj_i> placeholders, reference images still interleaved in the BAGEL-style context). Evaluate on all eight StructGen Bench subsets with PF, SC, and ID. If the natural-language variant matches or exceeds the structured variant on PF and SC, the central claim that explicit identifier anchors drive the semantic-alignment gains is not supported; if PF/SC drop substantially, the confound is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 3.1's central claim is that explicit identifier anchors, not the accompanying real-image dataset or the BAGEL-MICo initialization, are what improve multi-reference generation. The main evidence in Tables 1 and 2 does not isolate this. StructGen is fine-tuned from BAGEL-MICo on a combined dataset (10,665 MICo-De&Re samples plus 15,965 curated samples, per Table 4), while the principal baselines are the original BAGEL-MICo, Echo-4o, BAGEL, or OmniGen2; none is trained on the same combined data in the same format. On OmniContext, StructGen is given a constructed structured variant while baselines receive the original natural-language version, so format and training data are co-varied. Table 3 attempts an ablation: (a) natural-language format on the curated dataset versus (e) the full configuration, and (b) structured format on MICo data. This is the right idea, but it only reports the facial ID metric on two subsets (Single/Multi Human + Scene under Outfit-Conditioned). The headline claims also cover prompt following and subject consistency, which are not tested in the format ablation. If the structured format's benefit is mostly identity preservation, the PF/SC gains in Tables 1-2 could partly arise from the extra curated data or from the easier structured input variant rather than from the identifier mechanism itself.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes StructGen, a framework for multi-reference image generation that replaces free-form natural-language instructions with a structured, dictionary-like context in which each reference image is assigned an identifier (e.g., <human_1>, <obj_1>, <scene>) and the instruction refers to these identifiers. The authors argue that this structured format reduces ambiguity and improves attribute–subject association and generation consistency. To support the approach, they construct a real-image-based training dataset via a four-stage curation pipeline (collection, captioning, extraction, instruction synthesis), fine-tune BAGEL-MICo on this dataset combined with MICo data, and introduce a new benchmark (StructGen Bench) with eight subcategories and three metrics (prompt following, subject consistency, facial ID). Experiments on OmniContext and the proposed benchmark report consistent improvements over OmniGen2, BAGEL, BAGEL-MICo, and Echo-4o.","tokens_in":22131,"tokens_out":3034,"duration_ms":34478,"significance":"If the central claim is established, the paper would make a useful contribution to multi-reference image generation by showing that a structured input format with explicit identifiers can improve grounding and consistency, and by providing a real-image-derived training pipeline and a dedicated benchmark. The strengths of the paper include: a clearly described data curation pipeline that addresses a genuine data-scarcity problem; a new benchmark with eight controlled subcategories; the use of multiple complementary metrics; and the commitment to release code. The paper also includes an ablation study, which is the right methodological instinct. However, the central claim — that the structured format itself, rather than the additional curated training data or the different input variant used during evaluation, is responsible for the gains — is not cleanly supported by the current experiments.","major_comments":[{"comment":"The headline comparisons do not isolate the effect of the structured context. StructGen is fine-tuned from BAGEL-MICo on a combined dataset (10,665 MICo-De&Re samples plus 15,965 curated samples, per Table 4), while the principal baselines (BAGEL-MICo, Echo-4o, BAGEL, OmniGen2) are not trained on this combined data in the same format. Moreover, on OmniContext, StructGen receives a structured variant with identifiers while the baselines receive the natural-language version. Thus two variables — input format and training data — change simultaneously. The claim in Section 3.1 that identifier anchors themselves improve reference grounding and generation consistency is load-bearing; to support it, the authors should add a control that trains BAGEL-MICo (or a same-initialization model) on the same combined data with natural-language instructions, and evaluate StructGen's structured format agai","section":"Section 4.1, Tables 1 and 2, Table 4"},{"comment":"The ablation study is too narrow to support the paper's central claim. Table 3 reports only facial ID consistency on two subsets (single- and multi-human with scene reference under Outfit-Conditioned). The headline claims in Tables 1–2 also cover prompt following (PF) and subject consistency (SC), and the Section 3.1 argument is about attribute–subject association and generation consistency, not merely identity preservation. The authors should report PF and SC metrics for configurations (a)–(e), or justify why these are not needed. Without this, the conclusion that 'structured context modeling and our curated dataset each contribute independently' is only tested on one dimension.","section":"Section 4.3, Table 3"},{"comment":"The interpretation of configurations (a)–(d) as 'each isolates a single factor relative to the full configuration (e)' is not accurate. Configuration (a) is 'training on our curated dataset using natural language format' whereas (e) is the full configuration (curated + MICo data, structured format); so (a) versus (e) changes both the input format and the training data composition. Configuration (b) is 'training on MICo data while converted into our structured format,' which changes the dataset composition (MICo-only versus curated+MICo) as well. Thus, even on the ID metric, the individual contributions of 'structured format' and 'curated dataset' are confounded. The authors should add configurations such as: curated data + structured format, curated data + natural language, MICo data + structured format, and MICo data + natural language, all with matched training steps and data size, to","section":"Section 4.3, Table 3"},{"comment":"The benchmark uses GPT-4.1 both to synthesize the instructions (including the identifier-based variants) and as the VLM judge for prompt following and subject consistency. Since the method is designed around the identifier format, and the judge is from the same model family that produced the instructions, there is a risk of evaluation bias favoring the format that the judge itself can more easily parse. The authors should provide a human-validation study on a subset of StructGen Bench (e.g., 100–200 samples) to confirm that the GPT-4.1 judgments correlate with human judgments for both input formats, or at least report the agreement rate. This is particularly important because the structured-vs-natural-language comparison is the core of the paper.","section":"Section 3.3, StructGen Bench construction"}],"minor_comments":[{"comment":"Typo: 'planing out coherent spatial arrangement' should be 'planning out coherent spatial arrangement.'","section":"Abstract"},{"comment":"Typo: 'the croped view' should be 'the cropped view.'","section":"Section 3.2, Reference Extraction"},{"comment":"The table header 'ID: Single/Multi' is ambiguous; the caption should explicitly state that these are the single-human and multi-human subsets with scene reference under the Outfit-Conditioned task, as done in the text.","section":"Table 3"},{"comment":"No confidence intervals or statistical significance tests are reported for the PF/SC/ID differences. Given that the gains on some subsets are modest (e.g., Table 2, Multi Human + Scene ID), providing error bars or a significance test would strengthen the conclusions.","section":"Tables 1 and 2"},{"comment":"The notation D_ref = {id_i : z_ref^i} is informal: it is not a standard set-builder definition, since id_i is a token and z_ref^i is a sequence. Clarifying that this is a mapping from identifier tokens to token sequences would improve precision.","section":"Section 3.1, Eq. (3)"},{"comment":"The paper states that the mixed-sampling strategy samples each reference 'between either the image or text modality with equal probability,' but it is not clear whether this applies at training time only or also at inference. Clarifying this would help reproducibility.","section":"Section 4.1"}],"recommendation":"major_revision","confidential_remarks":"This is a solid systems paper with a clear direction and a useful benchmark. The main issue is experimental isolation: the central claim that the structured identifier format drives the improvements is confounded by training data and input format. This is fixable with additional control experiments (natural-language training on the same combined data, and a fuller ablation that reports PF/SC in addition to ID). I would like to see the authors add a human-evaluation or human-agreement check for the VLM judge, given that the benchmark and judge share the same model family. If these experiments are added and confirm the structured-format advantage, the paper could be a strong contribution to the multi-reference generation literature."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Juan, quick read of StructGen. The core idea is simple and I think right: instead of making the user write “the green and white patterned shirt from image 2,” give each reference a placeholder token (<human_1>, <obj_1>, <scene>) and let the model point into a dictionary. That is a clean, practical reformulation of multi-reference generation. The real-image data curation pipeline is also genuinely useful: parse a real photo into human/belonging/scene references via a VLM, extract clean entity crops with Qwen-Image-Edit, and post-filter with InsightFace/Hungarian matching. The benchmark is a real asset too—eight orthogonal subsets and separate PF/SC/ID metrics. They ship detailed training config, and the code link is provided. No fitted parameters or hidden tricks. The contributions are real.\n\nThe soft spot is exactly what the skeptic flagged. The headline comparisons in Tables 1 and 2 do not isolate the structured format. StructGen is fine-tuned from BAGEL-MICo on a combined dataset (10,665 MICo + ~16k curated samples), while the principal baselines are the original BAGEL-MICo, Echo-4o, BAGEL, and OmniGen2—none trained on the same combined data. On OmniContext, StructGen gets the structured identifier variant while the baselines get the natural-language version. So format and training data are co-varied, and the PF/SC gains could partly come from the extra curated real-image data or from an easier input variant rather than from identifier anchors themselves. Table 3 is the right ablative experiment, but it only reports facial ID on two subsets. The results are directionally consistent—full config 0.65/0.38 vs 0.51/0.30 for natural language on curated data and 0.37/0.23 for structured on MICo—but they leave PF and SC untested, which is exactly where the paper claims “semantic alignment” gains. Also mild, but worth noting: the benchmark is generated with GPT-4.1 and judged with GPT-4.1, a circularity common in this literature but a reason to treat the StructGen Bench numbers as relative, not absolute.\n\nIs the central claim still likely true? Yes. The mechanism is well motivated, the ablation is consistent for ID, and the qualitative examples look compelling. But “likely true” and “demonstrated” are different. A serious referee should ask for the same base model trained on the same combined data in both natural-language and structured formats, with PF and SC reported in the ablation. That is more compute, but not hard.\n\nMy call: solid paper, worth refereeing, with one load-bearing result that needs a cleaner isolation experiment before Section 3.1’s claim is taken as established. I’d cite it for the data pipeline and benchmark either way.","headline":"Clear, practical reformulation of multi-reference generation with a genuinely useful real-image dataset and benchmark; the identifier-format claim is plausible but the headline evaluation confounds format with extra training data.","tokens_in":22580,"tokens_out":2636,"would_cite":true,"duration_ms":26909,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Multi-reference image generation improves when references are tagged with identifiers and instructions point to those identifiers, rather than describing each reference in natural language.","keywords":["multi-reference image generation","structured context","identifier-based instruction","diffusion models","dataset curation","benchmark","human-centric generation","attribute-subject association"],"falsifier":"Fine-tune a strong baseline on the exact same combined training mixture (the 15,965 curated structured samples plus the prior multi-image training data, same number of training steps, same trainable layers) but in plain natural-language format, and evaluate on the same benchmark subsets; if scores match or exceed StructGen's, the structured format is not the source of the gains. Conversely, if ablating to natural language drops performance even with the full data mixture, the format claim is supported.","tokens_in":21674,"feed_emoji":"🖼️","tokens_out":7451,"duration_ms":76069,"temperature":0.7,"pith_summary":"This paper tries to establish that the main bottleneck in multi-reference image generation is the input format, not just model capacity. It argues that natural-language instructions describing several reference images quickly become verbose and ambiguous, so the paper proposes StructGen, which encodes references in a dictionary-like structure in which each image is assigned a unique identifier (for example, <human_1>, <obj_1>, <scene>) and the instruction composes those identifiers explicitly. This converts cross-reference association from implicit reasoning into explicit searching, which the paper claims yields more accurate attribute–subject grounding and more consistent generation. To support this, the paper builds a real-image-based data curation pipeline, a training framework that fine-tunes a unified multimodal model on the structured format, and a benchmark with eight human-centric subsets; experiments report consistent gains in prompt following, subject consistency, and facial identity preservation, particularly as the number of references grows.","feed_headline":"Identifiers beat verbose prompts for multi-reference image generation","feed_subtitle":"Structuring references as a dictionary with IDs cuts ambiguity and keeps faces, clothes, and scenes consistent.","key_machinery":"The load-bearing object is the structured context: a reference dictionary D_ref = {id_i : z_i} that maps each identifier (such as <human_1>, <obj_1>, <scene>) to either its encoded image tokens or its detailed caption, concatenated with an identifier-based instruction c_id. This replaces the fixed-order concatenation of image-then-text tokens that base models use, and it is what turns reference resolution into an explicit search over dictionary entries. Supporting it are a four-stage data pipeline (collection with face/human filtering, structured vision-language captioning, generative extraction with pose perturbation and identity-matched post-filtering, and LLM-based instruction synthesis)","core_discovery":"The central claim is that replacing free-form natural language with a structured, identifier-based context — a reference dictionary paired with an identifier-based instruction — improves multi-reference image generation. Identifiers act as explicit anchors: they avoid entangling content and relations in ambiguous language, turning cross-reference association from implicit understanding into explicit searching. The paper further claims that its curation pipeline, which parses real-world images into structured fields rather than distilling synthetic compositions, preserves spatial diversity and human–environment interactions that prior datasets lack, and that this combination yields higher fid","pith_inferences":["Implicit extension: the dictionary-plus-identifier pattern should generalize beyond human-centric data to any enumerable reference set (object composition, layout-controlled generation, multi-image editing), since nothing in the mechanism is specific to people or clothing; a larger-scale application would test this.","The paper's mechanistic claim predicts a testable artifact: in attention maps, identifier tokens should attend strongly to their dictionary entries during generation; measuring that would provide direct evidence for the 'explicit searching' story.","Because the evaluation changes input format and training data simultaneously relative to baselines, the cleanest inference is that the format and the data each contribute, but their independent contributions on the full data mixture are not yet isolated; a controlled data-matched comparison is the natural next step.","The dataset is small (~16k curated samples), so scaling the pipeline to more categories and broader scenes would show whether the advantage widens or saturates — a question the paper leaves open."],"forward_implications":["Adopting a dictionary-and-identifier context requires no architectural changes to interleaved multimodal models, so the format could transfer across model families as a drop-in training recipe.","Structured instructions are far more concise than natural-language equivalents (about 64 versus 101 tokens in the paper's example), reducing user effort and the model's reasoning load simultaneously.","Because references can be supplied as images or as captions interchangeably, the same framework covers both multi-image composition and text-only generation with precise attribute association.","Real-world image curation with identity post-filtering yields a training distribution that better reflects natural interactions and backgrounds, which likely explains gains in subject and scene consistency beyond the input-format effect.","The gains concentrate in the most complex settings — multi-human with scene references — where natural-language baselines nearly fail, so structured grounding appears most valuable precisely when association complexity is highest."],"fun_headline_variants":["Structured IDs beat verbose prompts for multi-reference image synthesis","Dictionary-style references sharpen multi-reference image generation","Identifiers as anchors: clearer multi-reference image synthesis","Explicit IDs cut ambiguity in multi-reference image generation","Structured context outperforms long prompts in multi-reference generation"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The evaluation attributes the performance gains to the structured context format, yet StructGen is trained on a larger, curated dataset than the baselines and receives identifier-tagged inputs where baselines receive natural language, so two variables change at once.","fun_headline_variants_meta":{"raw":{"variants":["Structured IDs beat verbose prompts for multi-reference image synthesis","Dictionary-style references sharpen multi-reference image generation","Identifiers as anchors: clearer multi-reference image synthesis","Explicit IDs cut ambiguity in multi-reference image generation","Structured context outperforms long prompts in multi-reference generation"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000133,"raw_usage":{"total_tokens":949,"prompt_tokens":699,"completion_tokens":250,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":443,"completion_tokens_details":{"reasoning_tokens":174}},"tokens_in":443,"tokens_out":250,"duration_ms":3506,"temperature":1.0,"reasoning_tokens":174,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-01T22:43:33.318483+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Fine-tune a strong baseline on the exact same combined training mixture (the 15,965 curated structured samples plus the prior multi-image training data, same number of training steps, same trainable layers) but in plain natural-language format, and evaluate on the same benchmark subsets; if scores match or exceed StructGen's, the structured format is not the source of the gains. Conversely, if ablating to natural language drops performance even with the full data mixture, the format claim is supported.","supporting_citations":[],"review_version":1}