{"id":"1e01c7b0-00be-42f8-8ce3-67f26625edf6","arxiv_id":"2507.13388","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":3.0,"correctness_risk":"high","formal_verification":"none","parameter_count":2,"one_line_summary":"A softmax and spatial attention fusion of SDXL base and refiner latents yields an ImageNet FID drop of about 1 point, but the result is not statistically supported.","lead":"The paper fuses Stable Diffusion XL's base and refiner latents with learned attention maps, reporting small FID improvements on ImageNet. A generalist might see it as a cheap tweak to an existing pipeline, but the evidence is insufficient: no training details, no error bars, and only one baseline.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The reported FID gains depend on AGF/DSF convolution weights being trained, yet no training procedure, loss, or data split is described; if those weights are random, the proposed 'adaptive' fusion collapses to a fixed average.","rationale":"The reader's weakest assumption correctly identifies the missing training procedure as the central vulnerability. My analysis reinforces this: it is not merely a reproducibility concern, because with default initialization the proposed modules reduce to nearly uniform convex combinations of the base and refined latents. Under that reading, the architecture cannot perform adaptive fusion without training, so the reported FID/sFID/IS improvements are not attributable to the mechanism the paper claims to introduce. The absence of any training description in Sections 3.1-3.3 means there is no way to verify that the convolutional weights are optimized, and the anonymous repository link alone does not supply the missing specification. The reader's other points (no error bars, single baseline, notation inconsistencies such as the undefined Pspatial in DSF and the multi-view versus class-conditional mismatch) are valid but secondary. Because the central claim is unsupported in its current form, I agree with the REJECT verdict; no further adjustment is needed.","tokens_in":7390,"tokens_out":4543,"duration_ms":54457,"concrete_test":"Inspect the anonymous.4open.science repository for a training script or checkpoint for AGF/DSF. If none exists, rerun Table 1 using the paper's inference settings with (a) default random weight initialization for the fusion convolutions and (b) a fixed 0.5/0.5 average of Lb and Lr, over at least 3 independent seeds of 5,000 ImageNet samples. If the random-initialization or fixed-average FID is within statistical noise of the AGF/DSF values (or matches the SDXL baseline), the learned-fusion explanation fails. If a checkpoint is provided, instead fine-tune the modules with a documented loss and verify the exact reported numbers.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing premise is that the 7x7 convolutional layers in AGF (Section 2.2.1) and DSF (Section 2.2.2) have learned weights producing meaningful attention maps. The manuscript never specifies an optimizer, loss function, learning rate, number of epochs, data split, or any training stage; Section 3.1 describes only inference settings (50 DDIM steps for the base model, 15 for the refiner, CFG=5, and 5,000 generated images). If these convolution weights are untrained (for example, default PyTorch initialization), then the softmax/sigmoid outputs are near-constant, making Lf an approximately 50/50 per-pixel average of Lb and Lr. In that case, the reported improvements (FID 20.16 to 18.79 at 256x256, IS 219.74 to 232.04) would be a property of averaging the base and refiner latents, not of any learned adaptive mechanism. The paper also reports single-run numbers without error bars, but the deeper problem is that the mechanism asserted to cause the improvement is not specified. If a training stage exists, it must be documented; if it does not, the central claim is unsupported. The mismatch between the multi-view framing in Figure 1 and the class-conditional ImageNet experiments is a further concern, but the decisive gap is the missing training specification.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes DLSF, a post-hoc fusion procedure for SDXL that combines the base and refiner latents before decoding. Two fusion variants are introduced: Adaptive Global Fusion (AGF), which computes a channel-wise softmax-weighted sum of the base and refined latents, and Dynamic Spatial Fusion (DSF), which computes a spatial sigmoid mask from pooled features. The authors evaluate on class-conditional ImageNet at 256x256 and 512x512, reporting FID, sFID, IS, Precision, and Recall against an SDXL baseline, and they include an ablation of an additional refinement pass after fusion. The central claim is that the learned adaptive fusion improves generation fidelity and diversity relative to SDXL.","tokens_in":7743,"tokens_out":7236,"duration_ms":81919,"significance":"If the reported gains are real and reproducible, the contribution is a lightweight, inference-friendly integration of SDXL's base and refiner models; it requires only two small convolutional layers and could be applied without retraining the main diffusion model. The paper follows a standard evaluation protocol for conditional generation, reporting five metrics on 5,000 images across 1,000 classes, and it includes an ablation that tests a plausible design variant. The authors also provide an anonymous code link, which is a positive step toward reproducibility. However, the manuscript does not specify how the fusion layers are trained, does not compare against a trivial concatenation or averaging baseline, and gives no uncertainty quantification; these omissions currently prevent the central claim from being assessed. The idea is simple and potentially useful, but the evidence as presented is not sufficient.","major_comments":[{"comment":"The proposed AGF and DSF contain learnable convolutional layers, but the paper never specifies any training procedure: no loss function, optimizer, learning rate, number of epochs, data split, or validation protocol. Section 3.1 lists only inference hyperparameters (50 and 15 DDIM steps, CFG=5, and 5,000 generated images). If these convolution weights are randomly initialized, the softmax/sigmoid outputs are near-constant and the fusion reduces to an approximately fixed per-pixel average of Lb and Lr, so the reported FID improvements (e.g., 20.16 to 18.79 in Table 1) would not demonstrate an adaptive mechanism. This is the central load-bearing gap and must be resolved before the claims can be evaluated.","section":"Section 2.2.1, Section 2.2.2, Section 3.1"},{"comment":"The notation is internally inconsistent and directly affects reproducibility. In Section 2.2.1, the text states that Lconcat is processed by a '1×1 convolutional layer', but the equation immediately below uses Conv7×7. In Section 2.2.2, the equation for Mspatial uses Pspatial, which is never defined; the defined quantity is Pconcat. These errors must be corrected because the architecture is ambiguous as written.","section":"Section 2.2.1 and Section 2.2.2"},{"comment":"The comparison omits a simple concatenation or fixed-average baseline, so it is unclear whether the reported gains come from the adaptive weighting or merely from mixing the base and refiner latents. In addition, all FID/IS numbers are single runs with no standard errors or significance tests; on 5,000 samples, differences of roughly 1 to 1.4 FID points may be within sampling noise. The authors should report multiple seeds with confidence intervals and include a non-adaptive fusion control to isolate the effect of the learned weighting.","section":"Section 3.2, Tables 1 and 2"},{"comment":"The paper motivates 'multi-view image synthesis' in Figure 1 and Section 2.2, but the experiments in Section 3.2 evaluate class-conditional single-image generation on ImageNet and report no multi-view data, metrics, or comparisons. The stated task and the evaluation are mismatched, so the conclusions about multi-view synthesis are not supported by the presented evidence.","section":"Figure 1 and Section 3.2"}],"minor_comments":[{"comment":"Reference [16] is cited as a diffusion-based post-processing module of SDXL, but [16] is SDEdit, a separate image editing method; the citation does not support the claim as stated.","section":"Section 2.1"},{"comment":"The sentence claiming that 'AGF and DSF further enhance the balance between realism and coverage, raising Precision from 0.86 to 0.87 and Recall from 0.35 to 0.39' is not accurate at 512×512, where Table 2 shows Precision dropping to 0.85; the claim should be restricted to the 256×256 setting.","section":"Section 3.2"},{"comment":"The pooling asymmetry (average pooling on Lr and max pooling on Lb) is not justified; the authors should explain the choice or provide an ablation testing both pooling assignments.","section":"Section 2.2.2"},{"comment":"The text states that the decoder generates 1024×1024 images, while all quantitative experiments are at 256×256 and 512×512; the authors should clarify whether the fusion is resolution-specific or resolution-agnostic.","section":"Figure 1 and Section 3.1"},{"comment":"The manuscript contains numerous typos and formatting artifacts, including 'im-age', 'V AEs' with extra spaces, and duplicated table captions; a careful proofreading pass is needed.","section":"Throughout"}],"recommendation":"major_revision","confidential_remarks":"The manuscript reads as an incomplete draft: the central method is not fully defined because the training protocol for the fusion layers is absent, and the anonymous code link is not a substitute for the missing experimental details. The proposed idea is lightweight and could be useful, but the revision would need to be substantial, including a complete training description, a non-adaptive baseline, uncertainty quantification, and alignment between the stated multi-view task and the ImageNet evaluation."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe thing to know about arXiv:2507.13388 is that the headline result—about a one-point FID improvement on ImageNet from fusing SDXL base and refiner latents—depends on attention modules that are never actually trained, at least according to the paper. There is no optimizer, loss function, learning rate, number of epochs, or data split anywhere in the text. If those convolutional weights are left at default initialization, the softmax and sigmoid outputs are near-constant, and the fusion becomes a per-pixel average of the two latents. The reported gains might then be a property of averaging, not of any learned adaptive mechanism.\n\nWhat is worth a look: the input pairing is sensible. Base and refiner latents do carry complementary information, and combining them post hoc is a reasonable thing to test. The ablation showing that an extra refinement pass after fusion hurts (Table 3) is a genuinely interesting small observation. And the experiments use standard metrics at two resolutions, so the numbers are at least comparable.\n\nThe problems are not minor. In addition to the missing training specification, there are no error bars or multiple seeds, so a ~1 FID difference is not clearly significant. There is no baseline that simply averages Lb and Lr, which would have isolated the adaptive module's contribution. There are notation slips (1x1 conv written as Conv7x7 in Section 2.2.1, Pspatial used but not defined in Section 2.2.2). And the paper frames everything as 'multi-view image synthesis' while the experiments are class-conditional ImageNet with one output per prompt—that mismatch suggests the authors haven't fully aligned their story with their setup.\n\nThe stress-test note is right. The load-bearing premise is that the fusion weights are trained, and the manuscript provides no evidence for that. If the authors can document a training stage and add a simple averaging baseline with error bars, the paper becomes a modest workshop-level contribution. As it stands, the central claim is unsupported.\n\nMy recommendation: desk reject. It's not a question of importance—the idea is a routine extension of CBAM-style attention to a new input pair—but of incomplete evidence. Send it back with a clear list of what's missing; if the missing details show up, the paper could merit a light-peer-review venue.\n\nHope this helps.","headline":"Fusing SDXL base and refiner latents with attention modules is plausible, but the paper never shows the fusion weights are learned, so the reported FID gains may just be averaging.","tokens_in":8213,"tokens_out":3835,"would_cite":false,"duration_ms":44022,"reading_group":"no","serious_thinker":"yes","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims that fusing the base and refined latents of Stable Diffusion XL with either of two lightweight attention modules lowers FID and raises Inception Score on class-conditional ImageNet generation.","keywords":["diffusion models","latent fusion","image synthesis","Stable Diffusion XL","spatial attention","FID","Inception Score","feature aggregation"],"falsifier":"Regenerate the 256x256 ImageNet set with the fusion convolutions replaced by a fixed uniform weight (equivalent to averaging $L_b$ and $L_r$) and compare FID to the reported 18.79 and 18.89; if the average already gives the same FID, the learned attention maps are not doing the work.","tokens_in":7214,"feed_emoji":"🎨","tokens_out":6267,"duration_ms":62116,"temperature":0.7,"pith_summary":"This paper tries to show that the gap between the coarse base latent and the fine refined latent in Stable Diffusion XL can be bridged at decode time by an attention-weighted fusion, improving image quality. It proposes two modules: Adaptive Global Fusion (AGF), which softmax-weights the two latents per pixel, and Dynamic Spatial Fusion (DSF), which builds a spatial attention map from pooled features. Both are tested against the SDXL baseline on ImageNet at 256x256 and 512x512. If true, the result matters because it would provide a simple architectural change to existing two-stage diffusion pipelines.","feed_headline":"Fusing two SDXL latents cuts FID from 20.16 to 18.79","feed_subtitle":"Attention-weighted fusion of SDXL's two latents improves realism and diversity on ImageNet class-conditional generation.","key_machinery":"The load-bearing object is the fused latent $L_f = W_b \\odot L_b + W_r \\odot L_r$ for AGF, where $W_b,W_r$ come from a softmax over a 7x7 convolution on the channel-concatenated latents; for DSF, $L_f = M_{spatial} \\odot L_r + (1 - M_{spatial}) \\odot L_b$, where $M_{spatial}$ is a sigmoid spatial attention map built from average- and max-pooled channel features. These formulas turn two separate representations into one per-location weighted blend, which is what carries the reported quality gain.","core_discovery":"The central claim is that SDXL's two latents—the base latent $L_b$, which carries global structure, and the refined latent $L_r$, which carries detail—should not be decoded separately or simply concatenated, but fused with per-location attention weights. The paper reports that Adaptive Global Fusion (AGF), a softmax-weighted sum of the two latents, and Dynamic Spatial Fusion (DSF), a sigmoid-gated spatial mixture, both outperform the SDXL baseline on class-conditional ImageNet at 256x256 and 512x512, with FID falling from 20.16 to 18.79/18.89 at 256x256 and from 19.65 to 18.70 at 512x512, and Inception Score rising correspondingly. The authors interpret this as evidence that cross-latent communication preserves global coherence and local texture simultaneously.","pith_inferences":["Editorial extension: if the fusion convolutions are not trained, the reported gain may be largely an ensembling effect of averaging two latents; a direct test is to compare against plain averaging with the same compute budget.","Editorial extension: the same channel-concatenation-plus-spatial-attention recipe could transfer to other multi-stage generators, such as cascaded diffusion models, where a coarse and a refined representation are available.","Editorial extension: the paper evaluates only class-conditional ImageNet; a natural next step is text-to-image benchmarks where semantic alignment failures are more visible."],"forward_implications":["At 256x256, both AGF and DSF lower FID from SDXL's 20.16 to 18.79 and 18.89, and raise Inception Score from 219.74 to 230.43 and 232.04.","At 512x512, both modules reach FID 18.70 against SDXL's 19.65, and raise Inception Score to about 243.5 from 234.75.","Adding a further refinement pass after fusion (AGF/r, DSF/r) worsens FID and sFID and sharply lowers Inception Score, which the authors read as over-smoothing and diversity loss.","Precision and Recall at 256x256 improve, with Recall rising from 0.35 to 0.39, suggesting the fused outputs cover a broader distribution."],"supporting_citations":[{"why":"Defines the SDXL base and refiner models and the two-stage pipeline that DLSF modifies.","marker":"[6]"},{"why":"Introduces latent diffusion, the framework DLSF builds on.","marker":"[5]"},{"why":"Provides the ImageNet dataset and class labels used for evaluation.","marker":"[18]"},{"why":"Defines FID, the primary quality metric the paper reports.","marker":"[20]"},{"why":"Defines Inception Score, the secondary metric used to support the claim.","marker":"[22]"},{"why":"Supplies the DDIM sampling schedule used for inference.","marker":"[19]"},{"why":"Supplies classifier-free guidance, the conditioning method used in the experiments.","marker":"[13]"},{"why":"Provides the CBAM spatial attention pattern that DSF adapts.","marker":"[8]"}],"fun_headline_variants":["Fusing SDXL's two latents cuts FID to 18.79","Latent fusion improves SDXL fidelity on ImageNet","AGF and DSF fuse latents to sharpen synthesis","Dual-latent synergy reduces FID from 20.16 to 18.79","Attention-weighted latent fusion cuts FID by 1.37"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The reported gains depend on the assumption that the small convolutional layers inside AGF and DSF have usable weights; the paper never states how these layers are trained, what loss is used, or how the weights are initialized, so the improvements could come from the fusion formula alone or from chance.","fun_headline_variants_meta":{"raw":{"variants":["Fusing SDXL's two latents cuts FID to 18.79","Latent fusion improves SDXL fidelity on ImageNet","AGF and DSF fuse latents to sharpen synthesis","Dual-latent synergy reduces FID from 20.16 to 18.79","Attention-weighted latent fusion cuts FID by 1.37"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000769,"raw_usage":{"total_tokens":3387,"prompt_tokens":908,"completion_tokens":2479,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":524,"completion_tokens_details":{"reasoning_tokens":2385}},"tokens_in":524,"tokens_out":2479,"duration_ms":18299,"temperature":1.0,"reasoning_tokens":2385,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T16:59:20.876987+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Regenerate the 256x256 ImageNet set with the fusion convolutions replaced by a fixed uniform weight (equivalent to averaging $L_b$ and $L_r$) and compare FID to the reported 18.79 and 18.89; if the average already gives the same FID, the learned attention maps are not doing the work.","supporting_citations":[{"cited_title":"High-Resolution Image Synthesis with Latent Diffusion Models,","cited_arxiv_id":null,"evidence_quote":"Introduces latent diffusion, the framework DLSF builds on."},{"cited_title":"ImageNet: A Large -Scale Hierarchical Image Data- base,","cited_arxiv_id":null,"evidence_quote":"Provides the ImageNet dataset and class labels used for evaluation."},{"cited_title":"GANs Trained by a Two Time -Scale Up- date Rule Converge to a Local Nash Equilibrium,","cited_arxiv_id":null,"evidence_quote":"Defines FID, the primary quality metric the paper reports."},{"cited_title":"Improved Techniques for Training GANs,","cited_arxiv_id":null,"evidence_quote":"Defines Inception Score, the secondary metric used to support the claim."},{"cited_title":"CBAM: Convolutional Block Attention Module,","cited_arxiv_id":null,"evidence_quote":"Provides the CBAM spatial attention pattern that DSF adapts."}],"review_version":1}