{"id":"53636da8-33d3-401f-affa-4c92b0057030","arxiv_id":"2505.21488","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"A layout-free method that predicts subject regions from initial noise and refines them via guidance, improving multi-subject prompt alignment and layout diversity.","lead":"This paper proposes generating multi-subject images by reading the subject layout from the initial random noise of a diffusion model and refining it during generation, instead of using an external layout. This avoids conflicts with the model's own prior and improves prompt adherence while preserving composition diversity.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The soft-layout predictor is trained on only 1,500 simple synthetic prompts from 20 MSCOCO classes, yet the central claim depends on it discovering prompt-alignable layouts for arbitrary complex prompts; its generalization and robustness under guidance are asserted but never quantitatively measured.","rationale":"I read the paper in good faith. The method is well-motivated: it avoids external layout conflicts by predicting layouts from the diffusion model's own features, and the qualitative results (Figs. 6, 8, 10–12), user study (Table 1), and ablations (Table 3, Fig. 13) provide real evidence that the pipeline can produce prompt-aligned multi-subject images with natural compositions. The reader's weakest assumption—that initial noise encodes a prompt-alignable layout discoverable by a small network—is reasonable; prior work (Crystal Ball, InitNO) supports the noise-layout connection, and the paper's Figure 2 illustrates it. My concern narrows the 'discoverable' part: the discovery mechanism is a network trained on a small, narrow dataset, and the paper provides no quantitative measure of layout prediction quality on the evaluation distribution. This is not a disagreement with consensus; it is an internal-evidence gap. The training set uses only 20 MSCOCO classes, simple prompt templates, and no attribute-binding or visually-similar-class compositions, yet the central claim covers arbitrary multi-subject prompts. The paper's own conclusion admits the guidance can push latents off the prior, which changes the predictor's inputs in an unquantified way. These are concrete, testable risks. If the proposed IoU test passes, the concern is resolved and the conditional verdict stands; if it fails, the central mechanism is undermined and the paper would need substantial revision. Given the absence of this evidence, I do not change the reader's CONDITIONAL verdict—code release and quantitative layout-accuracy evaluation should be conditions.","tokens_in":15352,"tokens_out":10137,"duration_ms":107806,"concrete_test":"Run the trained soft-layout network on 100 held-out complex prompts (e.g., the user-study prompts and the qualitative examples in Figs. 8, 10–12). For each prompt, generate with the full method and with vanilla SDXL using the same seeds; at t=50 and t=40, extract the predicted hard-layout and compare it to the GroundedSAM segmentation of the final image using mean IoU over subject clusters. Report IoU separately for the training-distribution prompts and the held-out complex prompts, and for guided vs. unguided latents. If held-out IoU drops substantially (e.g., >0.2) relative to training-distribution IoU, or if guidance reduces IoU, the asserted generalization and the central noise-induced-layout mechanism are not supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim—that a small network can predict and refine a noise-induced layout at every denoising step—rests on the soft-layout predictor generalizing far beyond its training distribution. In §4.1, the network is trained on ~1,500 SDXL images synthesized from prompts using only 20 MSCOCO classes, 1–3 classes per prompt, quantities 1–10, and generic prefixes/postfixes (Appendix A.2). At inference, it is applied to arbitrary complex prompts such as \"a polar bear, a grizzly bear, a panda bear, and a koala bear\" (Fig. 8) and to latents that have been modified by decisive guidance (§4.3). The paper asserts generalization in Appendix A.2 (\"our soft-layout network generalizes well to unseen object classes, backgrounds, and prompt structures\") without any quantitative evaluation of layout prediction accuracy. If the predictor fails on such prompts, the hard-layouts from §4.2 are arbitrary K-Means splits, and the guidance losses in §4.3 will enforce incorrect masks, causing leakage or distortion rather than preventing it. Furthermore, the Conclusion concedes that guidance \"push[es] the optimized latent away from the prior distribution,\" so the predictor's input distribution at inference differs from the noisy-clean training latents; no analysis shows the predictor remains accurate under this shift. This is load-bearing because the entire method is built on the predicted layouts.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes \"Be Decisive\", an inference-time method for multi-subject text-to-image generation that removes the need for externally prescribed layouts. A small neural network predicts a time-varying \"soft-layout\" from SDXL's attention features at each denoising step; this soft-layout is clustered into a hard-layout used to bound attention (following Bounded Attention) and to define a \"decisiveness\" guidance loss that stabilizes subject boundaries across timesteps. The claimed contribution is that, by deriving the layout from the initial noise rather than imposing an external one, the method respects the model's prior, yielding diverse, prompt-aligned compositions. Experiments on T2I-CompBench, a user study, ablations, and a non-curated comparison support the qualitative and quantitative claims, although several load-bearing evaluations rely on the method's own predicted layouts or lack statistical rigor.","tokens_in":15665,"tokens_out":5299,"duration_ms":53767,"significance":"If the central claim holds, the method offers a practical way to achieve multi-subject alignment without user-provided layouts, addressing a known failure mode of text-to-image models. The idea of predicting a time-varying layout from the noise and then committing to it is elegant, and the paper includes several valuable components: a lightweight predictor trained on auto-generated data, an ablation study isolating the loss terms, a user study, and a non-curated comparison. The significance is conditional on the soft-layout predictor generalizing far beyond its narrow training distribution and on the layout-diversity metric actually measuring image diversity; both remain unestablished, so the contribution is promising but not yet fully demonstrated.","major_comments":[{"comment":"The load-bearing component of the method is the soft-layout predictor, yet its generalization is asserted rather than measured. The predictor is trained on ~1,500 SDXL outputs from 20 MSCOCO classes with 1–3 classes, quantities 1–10, and generic prefixes/postfixes, but at inference it is applied to arbitrary prompts (e.g., four bear species in Fig. 8) and to latents that have been altered by the decisive guidance described in §4.3; the Conclusion itself concedes that guidance pushes the latent away from the prior. Because the hard-layouts of §4.2 and all three guidance losses of §4.3 are computed from the predicted soft-layout, a distribution-shift failure of the predictor would turn the masks into arbitrary K-Means splits and the guidance into a source of artifacts. Please provide a quantitative evaluation of layout-prediction accuracy (e.g., mask IoU against GroundedSAM on unseen complex prompts) and an analysis of predictor behavior under guidance-modified latents.","section":"§4.1, §4.3, Appendix A.2, Conclusion"},{"comment":"The quantitative claims of 'consistently high scores across all metrics' rest on point estimates without error bars, significance tests, or multiple trials (Table 2). More importantly, the layout-diversity metric is computed from the method's own predicted hard-layouts (for Ours) or from LLM-generated layouts (for LLM+BA), not from the actual generated images; this makes the diversity comparison confounded with the layout-prediction/generation process and does not directly measure the variability of the final compositions. The paper should either measure diversity on the generated images (e.g., via segmentation of the outputs) or justify why predicted-layout diversity is a faithful proxy.","section":"§5.2, Table 2"},{"comment":"The non-curated comparison uses a single prompt and seven seeds. The claim that 'our method is able to correctly adhere to the prompt in each image' is made without a stated criterion for correctness, a blind evaluation, or an inter-rater agreement measure. Given that the user study (Table 1) is the primary evidence for semantic-leakage superiority, the single-prompt non-curated figure should be supported by a broader non-curated evaluation or by a quantitative correctness rate.","section":"§5.1, Figure 11"},{"comment":"The paper's central narrative assumes that the initial noise encodes a prompt-alignable layout that the predictor can discover. Figure 2 shows qualitative agreement, but there is no quantitative measure of how well the predicted layout at early timesteps matches the final image layout. Since the method's advantages are attributed to following the noise-induced layout rather than an external one, a quantitative test of this assumption (e.g., comparing predicted early layouts against final segmentation) would strengthen the causal claim and mitigate the risk that the predictor is merely segmenting attention without truly reading the noise-induced structure.","section":"§1, Figure 2"}],"minor_comments":[{"comment":"The captions contain a typo: 'noise-induce layouts' should be 'noise-induced layouts'.","section":"Captions of Fig. 6 and Fig. 10"},{"comment":"The text says 'We now turn to formally define the soft-layout' but gives no formal definition; the soft-layout is described only textually and via the triplet loss in Eq. (1). Please add an explicit definition or equation.","section":"§4.1"},{"comment":"The claim that 'our soft-layout network generalizes well to unseen object classes, backgrounds, and prompt structures' is not supported by any experiments in the appendix or main text; this is the same issue raised in the first major comment.","section":"Appendix A.2"},{"comment":"The 'w/o Ldecisive' row reports a prompt-alignment accuracy of 0.016, which is far below chance (0.5) in a forced-choice user study; please explain how participants could select the method's images so rarely and whether this row reflects the same number of comparisons as the other rows.","section":"Table 3"},{"comment":"The paper does not mention whether code or checkpoints will be released; given the dependency on a trained predictor and the high resource usage (36 GB VRAM), reproducibility would benefit from a release statement.","section":"§5.2"}],"recommendation":"major_revision","confidential_remarks":"The paper builds heavily on the authors' own Bounded Attention method and uses a training set generated by the same base model, SDXL. While this is not circular, the novelty is incremental in that the core contributions are the learned soft-layout predictor and the decisive guidance that replaces external layouts. The editor may want to weigh whether the technical depth is sufficient for a journal venue; the main risk is the unquantified generalization of the learned predictor, which the revision must address. The layout-diversity metric also needs rethinking. If the authors can provide the requested quantitative analyses, the paper could become a solid contribution."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The central idea here is worth taking seriously: instead of forcing an external layout onto a diffusion model, predict the layout latent in the initial noise, then hold to it through hard masks and guidance. That is genuinely new. The method builds on Bounded Attention (their own prior work, fairly cited) and uses a small Readout-Guidance-style network trained on roughly 1,500 synthetic prompts. The pipeline is coherent: soft-layout to K-Means hard layout, Hungarian labeling, then decisive guidance with cross, variance, and Dice losses.\n\nWhat the paper does well: the non-curated seed 0-6 comparison is unusually honest, showing the method succeeding 7/7 where baselines fail. The user study is meaningful, and the ablation user study (Table 3) shows the full method at 0.832, far above any ablated variant. The paper also openly states in Section 6 that guidance pushes the latent away from the prior, which is a real limitation.\n\nSoft spots, in proportion. Table 2 gives single point estimates with no error bars or significance tests; differences like 0.837 vs 0.772 may be real but are not established. Layout diversity is computed from the method's own predicted layouts, so it partly measures predictor variability rather than image composition diversity. No code is released. The training set is narrow: 20 MSCOCO classes, generic prefixes and postfixes. The stress-test concern about generalization to arbitrary prompts is fair, and the Appendix's assertion that the network \"generalizes well\" is not evidence. The non-curated qualitative results and CompBench performance on unseen prompts partially mitigate this, but there is no quantitative layout-prediction accuracy on unseen prompts, and the predictor's input distribution shifts once guidance is applied. That is the weakest load-bearing point. It does not break the paper, because the final outputs across many prompts suggest poor masks are not catastrophic, but it needs measurement or at least a targeted ablation.\n\nOverall, the central argument holds in its broad form. This is a solid advance in inference-time control for multi-subject generation, and the main caveat is evidence quality rather than a fatal flaw. The paper deserves serious refereeing; a good reviewer should push for error bars, code, and an explicit layout-prediction generalization evaluation.\n\nFor a reader working on attention-based control or multi-subject composition, this is worth a reading group slot and likely a citation after the evidence is cleaned up.","headline":"Noise-derived layout is a real idea, but the evidence needs error bars and code before I trust the generalization claims.","tokens_in":16184,"tokens_out":1995,"would_cite":true,"duration_ms":24232,"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-subject images get layouts read from the seed noise instead of imposed boxes","keywords":["multi-subject generation","text-to-image diffusion","subject leakage","noise-induced layout","soft-layout prediction","layout guidance","bounded attention","compositional generation"],"falsifier":"Decode a fixed set of initial noises once with the paper's full method and once with vanilla SDXL, segment both sets of outputs with the same segmentation model, and compare the overlap between the first-step hard layout and the final subject regions; if that overlap is no better than a random partition's overlap with those regions, the seed noise is not the carrier of the layout.","tokens_in":15154,"feed_emoji":"🎨","tokens_out":6342,"duration_ms":63001,"temperature":0.7,"pith_summary":"Text-to-image diffusion models tend to blur several requested subjects into one, losing counts and attributes. This paper argues that the cure is to stop imposing external layouts and instead read the subject layout already latent in the sampled initial noise. At each denoising step a small network predicts a soft-layout from the model's own attention features, turns it into subject masks, and guides the next step to stay true to those masks. The paper reports higher text-image alignment and more stable multi-subject generation than existing layout-guided methods, while keeping the diversity of the original model.","feed_headline":"Reading the layout out of seed noise fixes multi-subject images","feed_subtitle":"A small network predicts where each prompt subject will land, then keeps those boundaries stable through denoising.","key_machinery":"The soft-layout is the central object: a timestep-dependent feature map whose per-pixel vectors are close exactly when those pixels belong to the same subject. A lightweight readout network, built on the Readout Guidance head design, predicts it from attention features of the denoising model and was trained with a triplet loss on roughly 1,500 auto-segmented images. K-means turns the soft-layout into a hard-layout of k subject clusters plus background, labeled by cross-attention responses via the Hungarian algorithm; decisive guidance then optimizes the latent so the next soft-layout agrees with the current hard-layout, mixing a cross-attention term, a variance term that compacts clusters, and a Dice term that stabilizes boundaries across timesteps.","core_discovery":"On the paper's own terms, the central discovery is that the spatial arrangement of a multi-subject image is fixed early by the initial noise, and that a lightweight network can expose this arrangement well enough to use as a layout. The method then commits to that layout, refining it slightly rather than replacing it. Concretely, a soft-layout feature map is predicted from self-attention keys and cross-attention queries; K-means clusters it into subject and background segments; and a decisive-guidance loss keeps each cluster tight and its boundaries stable across timesteps while Bounded Attention suppresses inter-subject leakage. The reported result is that prompts with several classes, attributes, and quantities are rendered accurately, with layout diversity that matches the prior rather than the restricted diversity of LLM-generated layouts.","pith_inferences":["If the seed-noise layout premise generally holds, a similar readout could be trained once and reused across related tasks that need layout stability without user input, such as inpainting, editing with moving subjects, or consistent multi-subject video frames.","Because the layout is read from the model's own prior, the method inherits the model's blind spots: prompts requiring spatial relations the model never learned will still fail, as the paper itself notes. A testable extension is whether enlarging the small auto-segmented training set with spatial-relation prompts would push that ceiling.","The guidance step is the computational bottleneck; the paper's stated future direction of replacing iterative optimization with feature injection from a control map could be validated by checking whether the soft-layout network alone, without the loss, already yields the same final regions."],"forward_implications":["External layout inputs, whether user-drawn boxes or LLM-generated layouts, become unnecessary for multi-subject prompts.","Multi-class, multi-attribute, and quantity-sensitive prompts score higher on T2I-CompBench color, texture, single-class, and multi-class metrics than the compared layout-based and layout-free alternatives.","Sampling different seeds yields diverse, natural compositions because the denoising trajectory follows the seed's own noise-induced layout rather than an imposed one.","The method composes with personalization: integrating it with a subject-injection method turns single-person personalization into multi-person images.","The same Bounded Attention mask, when driven by a noise-derived layout, avoids the artifacts that appear when masks are prescribed externally."],"supporting_citations":[{"why":"Supplies Bounded Attention, the masking scheme and cross-attention loss used as the leakage-prevention backbone.","marker":"[Dahary et al. 2025]"},{"why":"Provides the Crystal Ball hypothesis that initial noise anticipates object positions, motivating the noise-induced layout.","marker":"[Ban et al. 2024]"},{"why":"Shows initial noise optimization shapes the final layout, supporting the claim that layout is determined early in denoising.","marker":"[Guo et al. 2024]"},{"why":"Readout Guidance, the architecture the soft-layout network is built on.","marker":"[Luo et al. 2024]"},{"why":"GroundedSAM, used to auto-segment the roughly 1,500 training images into subject segments for the triplet loss.","marker":"[Ren et al. 2024b]"},{"why":"T2I-CompBench, the benchmark used for quantitative color, texture, composition, and numeracy evaluation.","marker":"[Huang et al. 2023]"},{"why":"SDXL, the text-to-image backbone on which the method is implemented and evaluated.","marker":"[Podell et al. 2023]"}],"fun_headline_variants":["Seed noise dictates layout for multi-subject images","Predict layout from noise to fix multi-subject images","Noise-induced layout improves multi-subject generation","Reading layout from seed noise beats external layouts"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method rests on the premise that the initial noise of the diffusion model already contains a spatial layout for the final image that is prompt-alignable and that a small network can read it at every denoising step; if the noise does not carry that layout, the predicted hard masks will not track real subject regions and leakage will not be prevented.","fun_headline_variants_meta":{"raw":{"variants":["Seed noise dictates layout for multi-subject images","Predict layout from noise to fix multi-subject images","Noise-induced layout improves multi-subject generation","Reading layout from seed noise beats external layouts"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000172,"raw_usage":{"total_tokens":1251,"prompt_tokens":896,"completion_tokens":355,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":512,"completion_tokens_details":{"reasoning_tokens":296}},"tokens_in":512,"tokens_out":355,"duration_ms":4003,"temperature":1.0,"reasoning_tokens":296,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T13:26:29.358614+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Decode a fixed set of initial noises once with the paper's full method and once with vanilla SDXL, segment both sets of outputs with the same segmentation model, and compare the overlap between the first-step hard layout and the final subject regions; if that overlap is no better than a random partition's overlap with those regions, the seed noise is not the carrier of the layout.","supporting_citations":[],"review_version":1}