{"id":"be66446b-7713-40c4-88ed-10822e0fffaa","arxiv_id":"2509.04126","paper_version":2,"verdict":"REJECT","confidence":"HIGH","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":5,"one_line_summary":"MEPG combines LLM-based spatial planning with a mixture of SDXL experts and staged local/global denoising, but the reported gains are modest and the main quality/diversity claims are not backed by experiments.","lead":"This paper proposes a text-to-image framework that uses a fine-tuned language model to plan object positions and a mixture of expert diffusion models to render each region in a different style. The authors report layout and style improvements, but the paper's own experiments only show a small GenEval gain, while the larger metric claims in the introduction are not measured anywhere.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Gating mechanism in §3.2 cannot implement spatial expert routing: gate input X is unspecified/usually scalar, so the central 'spatial-semantic' novelty is unsupported; promised FID/LPIPS evidence is also absent.","rationale":"The reader's weakest_assumption identifies the same mechanism, and I agree. The gate equation in §3.2 is the only description of expert selection; with G(X) 'usually a scalar' and no input definition, the method cannot implement the abstract's 'attention-based gating mechanisms' or 'selectively activated for each spatial partition.' This is an internal inconsistency, not just a disagreement with consensus. Even if one generously interprets G(X) as a vector over experts computed from some hidden features, the paper nowhere says those features include coordinates/masks; the cross-denoising fusion in §3.3 uses region prompts and expert outputs, but the gate itself is disconnected from the PSA coordinate stream. The central claim 'MEPG significantly outperforms ... in image quality and style diversity' is also unsupported because the promised FID/LPIPS results are absent, and the only numeric evidence (Table 1) shows a 0.03 total-score increase over SDXL with no error bars. I recommend keeping the reader's REJECT verdict (UNCHANGED), since both the method's defining mechanism and its headline evidence are unsupported as written.","tokens_in":11655,"tokens_out":4507,"duration_ms":41100,"concrete_test":"Obtain the actual SparseMoeBlock forward pass (or require the authors to specify G(X) and X). Then run a controlled gate-input ablation on GenEval: (A) X = [timestep, global prompt embedding]; (B) X = [timestep, per-region mask-pooled latent features/coordinates for each PSA box]. If (A) reproduces the reported 0.27 position score within noise, the method does not perform spatial expert routing and the central novelty collapses; if only (B) does, the paper's unspecified gate input is the missing load-bearing detail. Also report FID and LPIPS vs SDXL with multiple seeds to test the untested quality/diversity claim.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing assumption is that the MED module's gate performs spatially-aware expert routing. In §3.2 the gate is defined as w_i = σ(G(X))_i, and the text says G(X) is 'usually a scalar'; the input X is never specified, and no spatial coordinate, region mask, or attention feature is passed into the gate. A scalar gate output cannot index experts per spatial partition, and the equation is dimensionally incoherent as written (subscript i on a scalar). Because the paper's novelty over RAPHAEL/eDiffi is exactly 'spatial-semantic expert allocation', this gap is not cosmetic. If X is only the denoising timestep or a global text embedding, the method reduces to global MoE plus an LLM layout planner, and the claimed cross-region expert activation (and the 'precise spatial control' stated in the conclusion) has no mechanism. Moreover, the central quality/diversity claim is untested: the intro promises +12.3% FID and +18.7% LPIPS, but §5.1 reports only GenEval; no FID/LPIPS tables, no error bars, and the same-backbone total-score gain is 0.55→0.58 with position still 0.27. The most decisive flaw, however, is the undefined gate input: even with perfect planning, the 'multi-expert per spatial region' claim cannot hold unless the gate sees region information.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes MEPG (also called MEPT), a text-to-image framework that combines an LLM-based planning module (PSA) with a Mixture-of-Experts diffusion module (MED). The PSA module uses a LoRA-fine-tuned LLM to decompose complex prompts into spatial coordinates and per-region descriptions. The MED module is intended to route different expert models to different spatial regions during denoising, with a cross-denoising schedule that interleaves local and global generation. The authors claim improved compositional fidelity and style diversity, and report GenEval results showing Ours-MEPT at 0.58 overall vs. SDXL's 0.55, with qualitative style and edit examples.","tokens_in":11941,"tokens_out":4586,"duration_ms":43049,"significance":"If the central mechanism were sound, the idea of combining an LLM-based spatial planner with a lightweight portfolio of off-the-shelf SDXL experts would be practically useful, especially for editable layout control and style diversity. The use of public datasets (LAHR and LAION-SG) and open expert models is a strength, and the modular expert-swapping design is a worthwhile goal. However, the paper's load-bearing technical component—the spatially-aware gating mechanism—is not actually specified, and the quantitative evaluation does not measure the dimensions advertised in the introduction and abstract. The current manuscript does not provide a verifiable implementation of its core novelty, so the significance of the claimed results cannot be assessed.","major_comments":[{"comment":"The gating mechanism is the load-bearing component for the paper's central claim of 'spatial-semantic expert allocation' and 'cross-region generation'. However, the text states that G(X) is 'usually a scalar', while the formula indexes the output as σ(G(X))_i; a scalar cannot carry an index i, so the equation is dimensionally incoherent as written. More importantly, the input X to the gate is never specified. No spatial coordinates, region masks, or attention features are shown being passed into the gate. If X is only a global text embedding or denoising timestep, as in previous MoE text-to-image work, the expert selection is not spatially conditioned and the claimed per-region expert activation has no mechanism. The authors must define X explicitly, show how region information enters the gate, and verify that different spatial regions can select different experts.","section":"§3.2, gating formula w_i = σ(G(X))_i"},{"comment":"The Introduction promises '+12.3% FID improvement' and '+18.7% LPIPS gain', and the Abstract claims significant improvement in 'image quality and style diversity'. Yet the experiments section reports only GenEval scores; no FID or LPIPS values appear anywhere in the paper. Table 1 therefore cannot support the headline quantitative claims. The authors should either report the promised metrics or remove the unsupported claims from the abstract and introduction.","section":"§1 vs. §5.1, Table 1"},{"comment":"The reported GenEval gains are small and inconsistent. Ours-MEPT overall is 0.58 vs. SDXL's 0.55, with position improving from 0.15 to 0.27, but counting drops from 0.39 to 0.35 and color from 0.85 to 0.80. No error bars, variance, or significance tests are provided, so it is unclear whether the overall difference is meaningful. The comparison is also not 'same backbone' in a strict sense: the MED module uses four different SDXL-derived expert models (e.g., RealVisXL, nijiseven), not the original SDXL. The authors should clarify what the baseline actually is and provide statistical evaluation.","section":"Table 1, §5.1 and §5.3"}],"minor_comments":[{"comment":"The framework is called MEPG in the title and abstract but MEPT in several places (e.g., §1, §5.1, Table 1). Please use one consistent name.","section":"Throughout"},{"comment":"The text says the SparseMoeBlock consists of a gating function, a softmax function, and expert modules, but the formula uses a sigmoid function. Clarify which nonlinearity is actually used.","section":"§3.2"},{"comment":"The setup sets p1 = 0.7, while §5.1 states 'MEPT framework adopts p1 = 0.5'. Also, the fusion weights α_i^(t) in Eq. (2) are only described as 'empirically determined'. Please report the values and how they were chosen.","section":"§4.1 and §5.1"},{"comment":"The number of experts is denoted n in §3.2 and M in Eq. (2). Use a single notation.","section":"§3.2 and §3.3"},{"comment":"The conditions 'BaseModel', 'PSA+Global', 'PSA+MED', and 'prompt+MED' are not all defined in the text. Please define each ablation condition precisely.","section":"§5.3, Figure 5"},{"comment":"The table formatting is inconsistent (e.g., missing spaces in some entries) and no details are given for the baseline implementations. Also, the caption says 'Geneval' but the benchmark is 'GenEval'.","section":"Table 1"},{"comment":"The LoRA reference (Hu 2021) is missing author and title details. Please complete all references.","section":"References"}],"recommendation":"reject","confidential_remarks":"The paper appears not yet ready for peer review: the central gating mechanism is underspecified, the promised FID/LPIPS results are absent, and the reported GenEval gains are modest and not statistically supported. I would be willing to look at a revised version that defines the gate input, demonstrates spatial expert routing, and reports the advertised metrics."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: the paper bundles three plausible ideas—LLM-based layout planning, mixture-of-experts diffusion over SDXL variants, and a staged local/global denoising schedule—into one framework. The GenEval numbers are small but real, and the ablation suggests the MED module adds something. What the paper does not do is specify the mechanism that makes it new: the gate that supposedly does spatially-aware expert routing has no defined input.\n\nWhat is genuinely useful: the cross-denoising schedule in Eq. (1) is concrete and testable, and the SparseMoeBlock replacement of QKV/FF components without modifying base weights is a clean integration idea. The ablation in Fig. 5 (PSA only vs PSA+MED) shows a meaningful jump in total score, even though there are no error bars or sample counts.\n\nThe soft spots are serious. The abstract promises 'significantly outperforms' and the intro promises +12.3% FID and +18.7% LPIPS gains, but the experiments report only GenEval. The gate equation is written as wi = σ(G(X))i with G(X) 'usually a scalar'; X is never defined, and no spatial coordinate, mask, or attention feature is passed into it. That means the central claim of 'spatial-semantic expert allocation' is unsupported. If the gate sees only global text or timestep features, the method is a global MoE plus an LLM planner, and the 'precise spatial control' conclusion has no mechanism. The GenEval gains themselves are modest: total 0.55→0.58, position 0.15→0.27, and the model is worse than SDXL on counting. Hyperparameters are tuned per experiment (p1 = 0.7 vs 0.5), fusion weights are 'empirically determined,' and there is no code or checkpoint. The name inconsistency (MEPG vs MEPT) is minor but symptomatic.\n\nNone of this is circular: training on LAHR and LAION-SG and testing on GenEval is legitimate. The concern is not fabrication; it is under-specification. The paper still deserves a serious referee because the idea is worth exploring and the missing details can be fixed. A reviewer could ask for a specification of the gate input, FID/LPIPS tables, error bars, and code. That is a feasible revision path.\n\nIf you work on compositional T2I or MoE diffusion, this is worth a quick read as a direction to try, but I would not cite it in its current form. For a reading group, it could be a useful example of 'promising architecture, missing mechanism,' but not a model of rigorous experimentation.","headline":"Plausible framework, but the load-bearing spatial-routing gate has no specified input and the promised FID/LPIPS evidence is missing.","tokens_in":12553,"tokens_out":2786,"would_cite":false,"duration_ms":25671,"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":"The paper claims that decomposing a complex prompt into LLM-generated bounding boxes and routing each region through specialized diffusion experts improves compositional fidelity and style diversity over a single-model backbone.","keywords":["text-to-image generation","mixture of experts","spatial layout planning","LLM planning","LoRA fine-tuning","style diversity","compositional prompts","GenEval benchmark"],"falsifier":"Run a fixed prompt with only the bounding-box coordinates moved between regions (e.g., the cat box on the left vs the right) and log the gate weights; if the weights do not change, or if reassigning a region to a realism expert instead of an anime expert does not change the output style, then the per-region routing claim is false and the measured gains would come from the text-to-box planner.","tokens_in":11432,"feed_emoji":"🎨","tokens_out":10254,"duration_ms":88882,"temperature":0.7,"pith_summary":"The paper claims that text-to-image diffusion can follow densely packed, multi-object prompts much better if the prompt is first turned into a spatial plan and the image is then generated by a mixture of expert models, with each expert responsible for a region or style. It pairs a LoRA-finetuned language model that outputs bounding boxes and region descriptions with a mixture-of-experts diffusion block that routes each region to specialized style or realism experts, alternating local and global denoising to keep the composite coherent. The result, measured on the GenEval benchmark, is a total score of 0.58 against 0.55 for the same backbone and a jump in positional correctness from 0.15 to 0.27, plus qualitatively broader style diversity across experts. If this works as described, it points to a cheap upgrade path: expert modules can be added or swapped without retraining the base diffusion model, and users could edit layouts and region styles interactively.","feed_headline":"LLM-planned boxes raise image layout accuracy by 80 percent","feed_subtitle":"Per-region style experts handle complex prompts and boost GenEval position score from 0.15 to 0.27.","key_machinery":"The load-bearing mechanism is the Sparse-MoeBlock, which replaces the query/key/value and feed-forward modules of the diffusion backbone with weighted sums of corresponding modules from several expert models, the weights being wi = sigmoid(G(X))_i. Around it, cross-denoising schedules local generation first (p1 = 0.7 of steps) and global generation second, fusing expert outputs by xt = sum_i alpha_i(t) x_i(t), so regional detail and global coherence are both served. The PSA module supplies the spatial structure: a LoRA-tuned LLM converts text prompts into bounding-box coordinates and prompt-mask pairs.","core_discovery":"MEPG aims to establish that compositional image generation can be decomposed into a planning step and a regional generation step. The Position-Style-Aware module uses a LoRA-tuned LLM to turn a complex prompt into element lists, descriptions, and rectangular coordinates (x1,y1)-(x2,y2), yielding prompt-mask pairs. The Multi-Expert Diffusion module then routes each region: a gating function assigns weights to expert models and fuses their outputs, with local denoising dominating the first 70% of steps and a global pass consolidating the result. Evidence is the GenEval table, where the framework reaches 0.58 overall and 0.27 on position versus 0.55 and 0.15 for the SDXL backbone, plus qualitat","pith_inferences":["A direct test the paper does not run: sweep the bounding-box coordinates of one object while holding text fixed and inspect the gate weights; if they are unchanged, the 'spatial' expert routing is actually global style mixing, and the position gains come from the PSA planner.","The same planning-plus-routing decomposition could be wrapped around any diffusion backbone with replaceable qkv/feed-forward blocks; the paper only demonstrates it on four SDXL-style experts.","The per-region expert mixture, trained on style-sorted data, suggests an untested compositionality claim: arbitrary style checkpoints could be plugged in without retraining, making the gate a general-purpose style blender."],"forward_implications":["The base diffusion model's parameters stay frozen; only the gate and the LLM adapter are trained, so expert models can be swapped in or out without re-tuning the backbone.","Compositional accuracy improves on the tasks the framework targets: GenEval overall goes from 0.55 to 0.58 and position from 0.15 to 0.27 over SDXL, with two-object accuracy at 0.76.","Layout becomes an editable variable: the PSA module's coordinates can be changed at inference, and users can assign different style experts to different regions.","Style diversity is decoupled from a single model's aesthetic: mixing photorealistic, anime, and character-prone expert checkpoints yields distinct but globally coherent outputs."],"supporting_citations":[{"why":"Provides the GenEval benchmark whose object, attribute, and position scores are the paper's quantitative evidence.","marker":"Ghosh, Hajishirzi, and Schmidt 2023"},{"why":"LoRA is the adaptation method used to fine-tune the LLM in the PSA module.","marker":"Hu 2021"},{"why":"Supplies the coordinate-annotated high-resolution image dataset used to train the LoRA component.","marker":"Fang et al. 2025"},{"why":"Supplies the 540,000-image style dataset used to train only the gate in the MED module.","marker":"Li et al. 2024"},{"why":"Supplies the base LLM that the LoRA adapter modifies in the PSA module.","marker":"Yang et al. 2025"},{"why":"Establishes the mixture-of-experts routing idea that the MED module extends to spatial-semantic expert selection.","marker":"Jacobs et al. 1991"},{"why":"Represents the layout-conditioning approach that MEPG's regional prompt-mask design is intended to improve on.","marker":"Li et al. 2023"},{"why":"Represents prior text-feature-only MoE routing against which the paper positions its spatial routing claim.","marker":"Xue et al. 2023b"}],"fun_headline_variants":["Regional expert routing boosts compositional image generation","Multi-expert diffusion lifts prompt fidelity and layout accuracy","Planned regions, routed experts: better text-to-image","Style-aware LLM plans, experts render: 80% layout gain","One LLM plans boxes, many experts paint: layout up 80%"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The whole 'spatial expert routing' claim rests on the gate actually seeing region-specific information, but the paper defines the gate input only as 'usually a scalar' and never shows that changing a region box changes the weights.","fun_headline_variants_meta":{"raw":{"variants":["Regional expert routing boosts compositional image generation","Multi-expert diffusion lifts prompt fidelity and layout accuracy","Planned regions, routed experts: better text-to-image","Style-aware LLM plans, experts render: 80% layout gain","One LLM plans boxes, many experts paint: layout up 80%"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000473,"raw_usage":{"total_tokens":2211,"prompt_tokens":790,"completion_tokens":1421,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":534,"completion_tokens_details":{"reasoning_tokens":1338}},"tokens_in":534,"tokens_out":1421,"duration_ms":9705,"temperature":1.0,"reasoning_tokens":1338,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T10:20:57.417821+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run a fixed prompt with only the bounding-box coordinates moved between regions (e.g., the cat box on the left vs the right) and log the gate weights; if the weights do not change, or if reassigning a region to a realism expert instead of an anime expert does not change the output style, then the per-region routing claim is false and the measured gains would come from the text-to-box planner.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the GenEval benchmark whose object, attribute, and position scores are the paper's quantitative evidence."},{"cited_title":"A.; Jordan, M","cited_arxiv_id":null,"evidence_quote":"Establishes the mixture-of-experts routing idea that the MED module extends to spatial-semantic expert selection."}],"review_version":1}