{"id":"eeef1e50-57e6-47f2-84b6-558f6529ed38","arxiv_id":"2502.00826","paper_version":1,"verdict":"REJECT","confidence":"LOW","novelty_score":2.0,"correctness_risk":"high","formal_verification":"none","parameter_count":3,"one_line_summary":"A vague proposal for LLM-guided diffusion with dynamic KL weighting, backed by unsupported FID/IS tables.","lead":"This paper claims that combining a large language model with a diffusion model, plus a dynamic KL-weighting loss, improves text-to-image generation on COCO. The method is described only in broad strokes and the experimental evidence is not reproducible.","discovery_kind":"incremental","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Training loss in Eq. (6)-(7) is ill-posed: the KL term compares a forward transition over x_t with a reverse transition over x_{t-1}, so the proposed objective cannot be computed; all performance claims rest on an undefined loss.","rationale":"The reader identifies the unspecified α_t schedule as the weakest assumption. I agree that the absence of any formula or schedule for α_t is a serious reproducibility gap, but the more structural problem is that the KL term being weighted is mathematically undefined. The direction of the divergence in Eq. (6) is not the standard DDPM objective, and comparing forward and reverse transitions at different timesteps has no valid interpretation. This is a correctness flaw, not merely an omitted detail. I also note that the Inception Scores in Table I are not obviously implausible for COCO text-to-image benchmarks, so I would not base the rejection on that point. The paper has no independently checkable artifacts: no code, no full hyperparameter list, no human-evaluation protocol, and no formal verification. Because the central training objective cannot be evaluated as written, the reported performance improvements cannot be attributed to the proposed method. The reader's low-confidence REJECT verdict is therefore consistent with this analysis, though I reach it through a different, more fundamental technical failure.","tokens_in":8794,"tokens_out":4013,"duration_ms":42602,"concrete_test":"Analytically re-derive the ELBO for the generative model in Eqs. (1)-(4) and write out the KL term in Eq. (6) as an explicit integral with the same integration variable on both sides. If the left-hand distribution is over x_t and the right-hand distribution is over x_{t-1}, the term has no valid integral and Eq. (6) is not implementable; this one-page derivation settles the concern.","verdict_should_be":"REJECT","load_bearing_attack":"The central claim requires that the method in Section III-C actually optimizes a valid training objective. Equation (6) defines the loss as E_q[Σ_t D_KL(q(x_t|x_{t-1}) || p_θ(x_{t-1}|x_t, z_t))]. This KL divergence is not well defined: q(x_t|x_{t-1}) is a distribution over x_t conditioned on x_{t-1}, while p_θ(x_{t-1}|x_t, z_t) is a distribution over x_{t-1} conditioned on x_t, at different timesteps. There is no common measurable space on which to compute the divergence. The standard DDPM ELBO uses the tractable posterior q(x_{t-1}|x_t, x_0) on the left, not the forward transition q(x_t|x_{t-1}). Unless Eq. (6) is a typo, the described training procedure cannot be implemented, and α_t in Eq. (7) is weighting an undefined quantity. All reported improvements in Tables I-V therefore cannot be traced to the proposed objective. This is more fundamental than the missing α_t schedule: no specification of α_t can repair an ill-posed KL term.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a text-to-image generation method that combines an LLM-conditioned diffusion model with a 'dynamic KL-weighting strategy.' The method is described as a weak-to-strong guidance approach in which an LLM provides per-timestep text embeddings z_t and a time-dependent weight α_t is applied to each KL term in the training loss. The authors report quantitative experiments on COCO (FID and IS) plus human evaluations and ablation studies, claiming that the method outperforms DDPM, CLIP-guided diffusion, AttnGAN, and T2I-DA. The paper also claims improved efficiency, robustness to textual variability, and scalability to other datasets and modalities.","tokens_in":1741,"tokens_out":2161,"duration_ms":44273,"significance":"If the proposed method were fully specified and validated, a robust LLM-guided diffusion objective with a principled dynamic KL weighting could be a useful contribution to controllable text-to-image generation. The paper also introduces an 'alignment score' that, if properly defined, could be valuable for measuring semantic coherence between text and generated images. However, the significance is entirely contingent on the correctness of the training objective and the credibility of the experimental evidence. Since the central loss is ill-posed, α_t is unspecified, and the experimental support is incomplete, the claimed improvements cannot currently be attributed to the proposed method. The paper does not ship code, does not provide derivations, and reports results without error bars or evaluation details, so there is no independent way to verify any of the surface-level claims.","major_comments":[{"comment":"The training objective is ill-posed. Equation (6) defines the loss as E_q[Σ_t D_KL(q(x_t|x_{t-1}) || p_θ(x_{t-1}|x_t, z_t))]. The left argument is a distribution over x_t conditioned on x_{t-1}, while the right argument is a distribution over x_{t-1} conditioned on x_t and z_t. These are distributions on different measurable spaces at different timesteps, so the KL divergence is not computable. The standard DDPM ELBO uses the tractable posterior q(x_{t-1}|x_t, x_0) on the left, not the forward transition q(x_t|x_{t-1}). Unless Eq. (6) is a typo, the described training procedure cannot be implemented, and every quantitative result in Tables I-V rests on an undefined objective.","section":"Section III-C, Eq. (6)-(7)"},{"comment":"The dynamic weighting function α_t is never defined. The text states only that it is 'time-dependent' and 'designed to prioritize learning the early timesteps,' but no formula, schedule, or fitting procedure is given. Because α_t multiplies the ill-posed KL term in Eq. (7), no specification of α_t can fix the underlying problem; moreover, even if Eq. (6) were corrected, the choice of α_t would be an untested free parameter. The claimed benefit of the weighting strategy therefore cannot be evaluated.","section":"Section III-C, Eq. (7)"},{"comment":"The ablation study contains a row, 'No LLM and No KL Weighting,' with FID 42.1 and IS 4.2, which is exactly identical to the DDPM baseline row in Table I. This suggests either that the ablation is not actually a removal of both components from the proposed model, or that the table entries were copied. In either case, the result undermines the claim that both LLM guidance and dynamic KL weighting are 'critical to the success of our method,' because the ablated model is indistinguishable from the standard DDPM baseline.","section":"Table III"},{"comment":"The paper repeatedly promises a 'newly proposed alignment score' that 'quantifies the semantic coherence between text and image,' listing it as a key contribution and a motivation for the evaluation framework. However, this score is never defined, implemented, or reported anywhere in the experiments. The human evaluation tables (Tables II and V) ask participants to rate realism, relevance, and aesthetic quality, but these are not the promised quantitative alignment score. The absence of this central metric leaves the headline claim of improved text-image alignment without direct quantitative support.","section":"Introduction, Section I, and abstract"},{"comment":"The quantitative evaluation is not reproducible and the numbers are not contextualized. The paper does not report the image resolution, the number of generated samples, the specific COCO split (beyond '80-20'), the random seeds, or any error bars or confidence intervals for FID and IS. Without these details, the reported differences (e.g., FID 33.2 vs. 30.5) cannot be assessed for statistical significance. The manuscript also omits any comparison with modern text-to-image diffusion models such as Stable Diffusion or DALL-E, which limits the relevance of the claim that the method 'significantly outperforms existing approaches.'","section":"Section IV-B, Table I"}],"minor_comments":[{"comment":"The abstract contains grammatical errors, e.g., 'we presents a novel method' and 'a hybrid approach aimed at achieving both higher quality and efficiency in image synthesis from text descriptions' is a sentence fragment. The writing should be thoroughly revised.","section":"Abstract"},{"comment":"The cross-attention mechanism is described only as 'Attention(xt, zt)' with no formula, no definition of query/key/value projections, and no explanation of how the attention map at is used to modify the image features. This is insufficient for reproducing the method.","section":"Section III-B, Eq. (5)"},{"comment":"Much of the related-work section is off-topic. References such as [5] (InsectMamba), [15] (Cedille), [16] (Goldfish), [20] (LLMs in bioinformatics), and [21] (psycholinguistics) have no clear connection to text-to-image generation or LLM-guided diffusion, and the discussion of them distracts from the paper's focus.","section":"Section II"},{"comment":"The ablation study lacks a description of what 'No KL Weighting Strategy' means: is α_t fixed to 1, removed entirely, or something else? Without this specification, the reader cannot interpret the difference between the 'Full Model' and 'No KL Weighting Strategy' rows.","section":"Section IV-D and Table III"},{"comment":"The claim of 'approximately 20%' training-time reduction is presented with no timing measurements, no comparison of wall-clock time, no hardware configuration, and no convergence curves. The claim should either be removed or backed by actual measurements.","section":"Section IV-F"}],"recommendation":"reject","confidential_remarks":"This manuscript has several hallmarks that warrant editorial scrutiny beyond the technical deficiencies: the related-work references include many papers that are irrelevant to the topic (e.g., insect classification, French language models, psycholinguistics), the numerical results appear to contain a duplicated row between the main results and the ablation study, and the text has numerous grammatical errors. The editor may wish to verify that the submission is a genuine research contribution and not an artifact of a paper-generation pipeline. The technical issues alone justify rejection, but the integrity concerns may be worth a separate check."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The stress-test note is right, and it is the whole ballgame. Equation (6) defines a KL divergence between q(x_t|x_{t-1}), a distribution over x_t, and p_theta(x_{t-1}|x_t,z_t), a distribution over x_{t-1}. Those live on different spaces; the divergence is not computable. No amount of specifying alpha_t can repair that, and since the entire method section rests on this loss, the method as written cannot be trained. This is a load-bearing flaw, not a gap in exposition.\n\nWhat is the paper actually doing well? It names a real practical question: how to balance LLM conditioning against the diffusion model's own stochasticity over the course of generation. The weak-to-strong framing is a legitimate instinct, and the ablation study at least tries to isolate the contributions of LLM guidance and KL weighting. But that is where the credit ends. The central contribution, the dynamic weighting function alpha_t, is never given a formula or schedule. The promised 'alignment score' is never defined. There is no architecture detail, no hyperparameter list, no code, no data, no error bars.\n\nThe experimental section is worse than underspecified: it is internally suspicious. Inception Score around 4-5 on COCO is implausibly low; established text-to-image models score well above 20. The ablation row 'No LLM and No KL Weighting' is numerically identical to the DDPM baseline, which suggests the numbers were filled in rather than measured. The human evaluation reports an average rating of 4.6 for the proposed method versus 3.7 for the strongest baseline, with no participant demographics, no sample size per condition, no variance, no statistical test. The efficiency claims ('approximately 20% reduction in training time') and the scalability claims (OpenImages) appear out of thin air, with no supporting measurements.\n\nThe citation pattern is also sloppy: several references have nothing to do with the work (for example, an insect classification paper cited in the diffusion model overview), and the 'Weak to Strong' citation does not support the use made of it.\n\nThis paper is not ready for peer review. It has an invalid core objective, undefined key terms, and unverifiable empirical results. A serious referee would spend the entire time annotating fatal flaws. The right call is to desk reject. Nobody should build on this, and I would not cite it.","headline":"The core training loss in Eq. (6) is ill-posed, alpha_t is never defined, and the reported numbers are not credible; this paper should be desk rejected.","tokens_in":9507,"tokens_out":1395,"would_cite":false,"duration_ms":14204,"reading_group":"no","serious_thinker":"no","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"LLM-guided diffusion with dynamic KL weighting beats GAN and prior diffusion baselines on COCO text-to-image generation.","keywords":["text-to-image generation","diffusion models","large language models","dynamic KL weighting","weak-to-strong guidance","cross-attention conditioning","COCO dataset","human evaluation"],"falsifier":"Run the same LLM-guided diffusion pipeline with $\\alpha_t = 1$ for every timestep; if FID and IS stay at 30.5 and 5.4, the dynamic weighting is doing none of the work. Alternatively, any implementation that picks an explicit schedule and fails to reproduce Table I would show the published description is incomplete.","tokens_in":8586,"feed_emoji":"🖼️","tokens_out":6537,"duration_ms":61362,"temperature":0.7,"pith_summary":"This paper tries to establish that text-to-image generation improves when a large language model supplies a fresh semantic embedding at every denoising step and the diffusion loss is weighted differently at each timestep. The authors report the best numbers among the methods they compare with on COCO: FID 30.5, Inception Score 5.4, and a 4.6 out of 5 average human rating, beating DDPM, CLIP-guided diffusion, AttnGAN, and T2I-DA. They also report about 20% lower training time and robustness to paraphrased prompts. If true, the result would make LLMs a practical control signal inside the diffusion process, not just a front-end that rewrites prompts.","feed_headline":"LLM-guided diffusion beats GANs on text-to-image","feed_subtitle":"Per-step LLM embeddings plus dynamic KL weighting give FID 30.5 and top human ratings on COCO.","key_machinery":"The machinery is the LLM-conditioned reverse diffusion step $p_\\theta(x_{t-1}\\mid x_t, z_t)$, where $z_t$ is a per-timestep text embedding computed by the language model and injected through a cross-attention map, together with the weighted variational objective $L_{\\mathrm{weighted}} = \\mathbb{E}_q\\big[\\sum_{t=1}^T \\alpha_t D_{\\mathrm{KL}}(q(x_t\\mid x_{t-1}) \\,\\|\\, p_\\theta(x_{t-1}\\mid x_t, z_t))\\big]$. The time-dependent weight $\\alpha_t$ is meant to emphasize coarse structure early and fine detail later; a momentum-based fine-tuning step on high-confidence self-generated samples is the remaining component. The paper defines $\\alpha_t$ only symbolically and gives no schedule for it.","core_discovery":"The central claim is that a \"Weak to Strong\" training scheme, in which the LLM's influence is introduced progressively and the KL terms in the diffusion objective are reweighted by a time-dependent function, produces higher-quality, better-aligned images than existing text-to-image models. On the paper's reported numbers, the method has the lowest Fréchet Inception Distance (30.5) and the highest Inception Score (5.4) among the four baselines, and human raters prefer it by roughly a full point on a 1-5 scale. The ablation study attributes most of the gap to the LLM guidance and the dynamic KL weighting, and the authors additionally claim successful scaling to the OpenImages dataset. The paper does not provide the actual form of the weighting function.","pith_inferences":["The sharpest place to test the paper is the missing schedule for $\\alpha_t$: a re-run with any explicit schedule, compared with a constant-weight baseline, would separate the dynamic-weighting effect from the LLM conditioning effect.","The architecture as written is a cross-attention-conditioned diffusion model, so the distinguishing claim is really about the training curriculum; one could test whether ramping the LLM's influence with $t$ matters more than the embedding itself.","If the recipe transfers, per-timestep LLM conditioning should be tried in text-to-video and text-to-audio, where alignment must be maintained over time rather than in a single frame."],"forward_implications":["If the central claim holds, text-to-image systems can gain several FID points by conditioning every denoising step on an LLM embedding instead of treating the prompt as a one-shot input.","The reported 20% training-time reduction would make LLM-guided diffusion affordable for larger datasets and production deployments.","The reported generalization to novel prompts implies the LLM's semantic representation carries the model beyond memorized training captions.","The ablation ordering implies both the LLM guidance and the dynamic KL weighting are needed; removing either one degrades FID and IS by several points."],"supporting_citations":[{"why":"Supplies the weak-to-strong generalization idea that the paper's training scheme is named after and built on.","marker":"[1]"},{"why":"Supplies the multimodal-LLM-guided text-to-image generation pipeline that this work extends.","marker":"[2]"},{"why":"Provides the conditional-control formulation for text-to-image diffusion that the cross-attention conditioning step relies on.","marker":"[4]"},{"why":"States the diffusion forward/reverse process and variational objective underlying equations (1)-(7).","marker":"[6]"}],"fun_headline_variants":["Weak-to-strong LLM guidance boosts diffusion image quality","Dynamic KL weighting makes LLM-guided diffusion outperform GANs","LLM-guided diffusion with dynamic KL achieves lowest FID on COCO","Progressive LLM supervision sharpens diffusion text-to-image"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The entire reported improvement rests on the dynamic weighting function $\\alpha_t$, but the paper never specifies its formula or schedule, so the method cannot be reproduced or checked as written.","fun_headline_variants_meta":{"raw":{"variants":["Weak-to-strong LLM guidance boosts diffusion image quality","Dynamic KL weighting makes LLM-guided diffusion outperform GANs","LLM-guided diffusion with dynamic KL achieves lowest FID on COCO","Progressive LLM supervision sharpens diffusion text-to-image"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000971,"raw_usage":{"total_tokens":4092,"prompt_tokens":875,"completion_tokens":3217,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":491,"completion_tokens_details":{"reasoning_tokens":3146}},"tokens_in":491,"tokens_out":3217,"duration_ms":24042,"temperature":1.0,"reasoning_tokens":3146,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-09T17:34:03.220369+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same LLM-guided diffusion pipeline with $\\alpha_t = 1$ for every timestep; if FID and IS stay at 30.5 and 5.4, the dynamic weighting is doing none of the work. Alternatively, any implementation that picks an explicit schedule and fails to reproduce Table I would show the published description is incomplete.","supporting_citations":[{"cited_title":"Weak to strong generaliza tion for large language models with multi-capabilities,","cited_arxiv_id":null,"evidence_quote":"Supplies the weak-to-strong generalization idea that the paper's training scheme is named after and built on."},{"cited_title":"Master ing text-to-image diffusion: Recaptioning, planning, and gen erating with multimodal llms,","cited_arxiv_id":null,"evidence_quote":"Supplies the multimodal-LLM-guided text-to-image generation pipeline that this work extends."},{"cited_title":"Adding conditional co ntrol to text-to-image diffusion models,","cited_arxiv_id":null,"evidence_quote":"Provides the conditional-control formulation for text-to-image diffusion that the cross-attention conditioning step relies on."}],"review_version":1}