{"id":"c1004d39-f45e-4210-9fe4-b0e251ba6393","arxiv_id":"2508.02973","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"ANSWER replaces a fixed negative prompt with a per-step adaptive negative noise estimate from short diffusion-negative-sampling chains, improving prompt adherence in text-to-image diffusion models without training or external resources.","lead":"ANSWER is a new sampling recipe for text-to-image diffusion models that derives a fresh 'negative image' direction from the model's own noise predictions at each denoising step, so users do not need to write a negative prompt or run an external captioner. The authors report that adding ANSWER to Stable Diffusion and SDXL improves prompt adherence and wins human preference roughly 2:1 over classifier-free guidance and DNP.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The reported gains are attributed to per-step adaptive negative noise, but the paper never isolates that mechanism; a fixed-negative control using the same K-step DNS chain could show the adaptive recomputation is unnecessary.","rationale":"I agree with the reader's weakest-assumption identification. The method is clearly specified and testable, but the paper does not demonstrate that the K-step truncated DNS noise is a valid per-step negative direction; the only analytic handle is the K_t=1 degeneracy, and the rest is an empirical conjecture. The evaluation shows consistent mean improvements but lacks significance testing, error bars, code release, and a control that separates the adaptive recomputation from the extra compute and the DNS noise itself. These weaknesses support conditional acceptance rather than rejection: the proposed control experiment is cheap, decisive, and directly targets the central claim. No concern about author conduct is raised.","tokens_in":13058,"tokens_out":4229,"duration_ms":55669,"concrete_test":"Run Algorithm 1 with its K_t scheduler on the Attend&Excite, Pick-a-Pic, and DrawBench prompts, and compare against a fixed-negative control that computes one K-step DNS chain at t=T, normalizes the resulting noise, and reuses it in the NP update at every t>T/2 without recomputation, matching total denoiser evaluations, guidance scales, and random seeds. Report CLIP score, HPSv2, ImageReward, and a paired human preference study. If the fixed control is statistically indistinguishable from ANSWER, the adaptive per-step estimate is not the source of improvement; if ANSWER still wins by the claimed margin, the adaptivity concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing condition is in Section 3.3 and Algorithm 1: at each step t, the negative noise in the NP update, Eq. (5), is replaced by the noise produced by a K_t-step DNS chain run from the current latent z_t. The paper proves only the K_t=1 limit, where the construction reduces to CFG with rescaled guidance; for K_t>1 it asserts, without derivation or direct evidence, that this makes eps_n diverge from eps_phi and approximates the optimal negative direction n*(p,t). That is not a demonstrated property. The DNS chain follows a different, negatively guided trajectory, so its noise at time t-K_t need not be the best negative direction for the main chain's z_t. The scheduler and the mean/variance normalization are hand-designed, and no experiment isolates whether per-step recomputation is what drives the reported gains. A fixed-negative control that runs one DNS chain at t=T and reuses the normalized noise at every step would test exactly whether adaptivity is load-bearing; without it, the central 'adaptive' claim is unsupported.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes ANSWER (Adaptive Negative Sampling Without External Resources), a training-free modification of classifier-free guidance (CFG) sampling for text-to-image diffusion models. At each early denoising step, ANSWER runs a short K_t-step chain of Diffusion Negative Sampling (DNS), starting from the current latent, to produce a negative noise estimate that replaces the unconditional noise in the negative-prompting update of Eq. (5). The negative noise is normalized to match the statistics of the unconditional noise, and a scheduler decays K_t as t decreases. The authors argue that the optimal negative prompt n*(p,t) changes with the diffusion step, and that ANSWER approximates it directly in noise space, avoiding external captioning. Experiments on SDXL and SD compare ANSWER against CFG and DNP across Attend&Excite, Pick-a-Pic, DrawBench, PartiPrompts, and ImageNet using CLIP score, IS, FID, HPSv2, ImageReward, PickScore, plus an AMT human study. The paper reports that ANSWER outperforms the baselines on most metrics and is preferred by humans roughly two to one.","tokens_in":13251,"tokens_out":5633,"duration_ms":62575,"significance":"If the central claim is correct, ANSWER is a practically valuable drop-in sampler: it requires no training, no external captioner, and no user-specified negative prompt, and it can be applied to any model that supports CFG. The method is specified in full in Algorithm 1, which is a real strength: the procedure is reproducible in principle, and the ablations on K and on the normalization step are useful. The paper also addresses a genuine limitation of prior DNP work, namely that a single negative prompt estimated once cannot adapt to the changing latent along the diffusion chain. However, the significance is not yet established. The key mechanism—that a K_t-step DNS chain produces a noise estimate that acts as the optimal negative direction at step t—is asserted rather than derived or directly measured for K_t>1. The empirical evidence lacks error bars and significance tests, and one reported metric contradicts the summary text. The contribution is promising and within scope, but the load-bearing 'adaptive' claim needs a control experiment that isolates per-step recomputation.","major_comments":[{"comment":"The load-bearing step of ANSWER is the replacement of the negative noise term in Eq. (5) with the noise produced by a K_t-step DNS chain run from the current latent z_t. The paper only analyzes the K_t=1 limit, where the construction reduces to CFG with rescaled guidance, and for K_t>1 it asserts that 'ϵ_n* diverges from ϵ_ϕ' and approximates the optimal negative direction n*(p,t). This is not demonstrated. A DNS chain follows a different, negatively guided trajectory, so its noise at time t-K_t need not be the best negative direction for the main chain at time t. To support the central 'adaptive' claim, the authors should run a fixed-negative control: compute one DNS chain at t=T, normalize its noise, and reuse the same noise at every step. If the gains persist without per-step recomputation, the adaptivity is not the driver; if they vanish, the claim is supported. As written, the benefit attributed to adaptivity is not isolated.","section":null},{"comment":"No error bars, confidence intervals, or significance tests are reported for any of the quantitative metrics or win rates. The differences between methods are often small (e.g., CLIP scores of 32.97 vs. 33.62 in Table 1), and without measures of variance it is impossible to judge whether these differences are reliable. The win rates are point estimates with no indication of the number of images, seeds, or prompts. The authors should report standard errors or confidence intervals over multiple seeds or prompt subsets, and ideally a paired significance test for the main ANSWER versus CFG and ANSWER versus DNP comparisons.","section":null},{"comment":"The text states that 'SDXL+ANSWER outperforms SDXL (CFG) and SDXL+DNP across all metrics' for the ImageNet experiment and earlier claims similar broad superiority. This is contradicted by the Attend&Excite row in Table 1, where the IS value for ANSWER (11.15) is lower than that of DNP (11.65), as the preceding paragraph itself acknowledges. The summary sentence must be corrected to reflect the actual exception, and the discrepancy should be explained. A factual inconsistency in the main quantitative claim undermines confidence in the other reported results.","section":null},{"comment":"The AMT human evaluation is described only by aggregate percentages in Figure 7. The paper does not report the number of participants, the number of pairwise comparisons, the number of prompts used, or whether the reported rates (ANSWER 46–61%, DNP 21–32%, CFG 8–20%) are conditioned on excluding 'No Clear Winner' responses. Without this information, the abstract's claim that ANSWER is 'preferred by humans 2x more over the other methods' cannot be verified. Please provide the full experimental protocol and, if possible, a significance test or confidence interval for the preference rates.","section":null}],"minor_comments":[{"comment":"The sentence 'For all datasets, SDXL+ANSWER outperforms SDXL (CFG) and SDXL+DNP across all metrics' should be qualified to exclude the IS metric on Attend&Excite, as noted in the major comments.","section":null},{"comment":"The sampling update 'Sample(z, ϵ^t, t)' is not defined; the authors should specify whether they use DDIM, DDPM, or another sampler, since the DNS chain and the main chain must use the same sampler for the claimed equivalence to hold.","section":null},{"comment":"The scheduler constants in Schdlr(K,t) = K·((2t-T)/(2t-T+20))·((T+20)/T) are introduced without justification or sensitivity analysis; please state how these constants were chosen and discuss their effect on the results.","section":null},{"comment":"The claim that ANSWER has O(KT) complexity should be stated more precisely: each DNS step requires two model evaluations (ϵ_p and ϵ_ϕ), and the main NP step also requires two evaluations, so the total number of model evaluations is larger than the CFG baseline by a factor that depends on K and the scheduler. Reporting wall-clock time in Figure 5 would make the trade-off concrete.","section":null},{"comment":"Reference [6] is cited as 'Improving image synthesis with diffusion-negative sampling, 2024' without a venue; please provide the published reference or arXiv identifier if available.","section":null},{"comment":"The figure uses color to distinguish methods; adding textual labels or a colorblind-safe palette would improve accessibility, since the percentages are central to the human-preference claim.","section":null}],"recommendation":"major_revision","confidential_remarks":"The manuscript is within scope for a computer vision journal and addresses a timely problem. My main concern is not the method's novelty or the paper's framing but the gap between the claimed adaptive mechanism and the evidence: the K_t>1 behavior is unproven, and the empirical section lacks statistical grounding. The fixed-negative control I requested in the major comments is, in my view, essential for the central claim. With that experiment and proper statistical reporting, the paper could become publishable; without them, the 'adaptive' contribution is not established."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper's actual novelty is narrow but real: it replaces DNP's one-shot caption-based negative prompt with a per-step latent-space negative noise estimate, using a K-step truncated DNS chain, a scheduler, and a normalization step. The method is fully specified (Algorithm 1), training-free, and the empirical results are consistently positive across five benchmarks, with some genuinely nice qualitative wins on spatial, numeracy, and anatomy prompts.\n\nWhat it does well: the K_t=1 reduction to CFG with rescaled guidance is a clean sanity check; the normalization ablation (Fig. 9) directly shows why the noise-statistics mismatch matters; and the CNP variant is a useful conceptual bridge. The authors also correctly identify a real limitation of DNP—a single global caption can't be optimal across all diffusion steps—and they put their finger on the right intuition.\n\nThe soft spots are about evidence, not plausibility. First, no error bars, confidence intervals, or significance tests anywhere; the win-rate range in the text (41–52%) doesn't match the table (47–52%). Second, and more important, the paper never isolates the adaptive mechanism. A fixed-negative control—run one DNS chain at t=T, normalize, reuse at every step—would tell you whether per-step recomputation is load-bearing. The K_t=1 result shows the benefit is not inherent to the construction. The stress-test note is correct on this. Third, the theory leans heavily on [6]; the claim that the DNS chain's noise approximates the optimal negative direction at each step is asserted, not derived. Fourth, no code release, and no comparison with [2]'s token-level negative guidance, despite the 'first adaptive negative' claim. These aren't fatal—the method is clear and testable—but they limit what the current paper can conclude.\n\nThe authors themselves acknowledge the fixed-K protocol hides the hard-prompt advantage (Sec. 4.1), which is another sign the headline claim is overstated.\n\nWho gets value: anyone working on inference-time guidance for text-to-image models. It deserves a serious referee. I'd send it to review, with the request that the revision add the fixed-negative control, error bars, and a direct comparison with [2].","headline":"A clearly-specified training-free negative-sampling upgrade with real but unproven adaptivity; the evaluation is too thin to support the headline claim as-is.","tokens_in":13854,"tokens_out":2809,"would_cite":true,"duration_ms":30378,"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 proposes ANSWER, a training-free sampling procedure that recomputes a negative-guidance noise direction at every diffusion step inside the latent space, claiming it improves prompt adherence and human preference over…","keywords":["adaptive negative sampling","diffusion models","classifier-free guidance","negative prompting","training-free sampling","latent-space guidance","prompt adherence"],"falsifier":"Run ANSWER on a set of prompts while replacing the per-step negative noise with either the DNS noise computed once at the first step and reused everywhere, or with the normalized noise of an unrelated prompt; if prompt-adherence metrics stay roughly constant, then the adaptive component of ANSWER is not doing the work, and only the stronger guidance direction from the DNS chain matters. The paper's own $K$ ablation shows the trade-off, but it does not isolate adaptivity from the mere presence of a DNS-derived negative direction.","tokens_in":12758,"feed_emoji":"🖼️","tokens_out":6177,"duration_ms":68151,"temperature":0.7,"pith_summary":"The paper proposes ANSWER, a sampling procedure for text-to-image diffusion models that computes a negative-guidance noise direction anew at each denoising step, entirely inside the model's latent space. Its starting point is the observation that the optimal negative prompt for a given positive prompt changes as the noisy image evolves along the diffusion chain, so any single negative text prompt, whether hand-written or produced by a captioning model, is only optimal at one instant. ANSWER replaces the fixed negative-condition noise of negative prompting with the output of a short diffusion-negative-sampling chain started from the current latent, normalized to match the current noise statistics, and then applies the standard negative-prompting combination. The paper claims that this training-free, external-resource-free change improves prompt adherence, image quality, and human preference over classifier-free guidance and over diffusion-negative prompting on five benchmark suites. If correct, ANSWER is a drop-in improvement for any diffusion model that supports classifier-free guidance.","feed_headline":"Adaptive negative noise beats fixed negatives in diffusion sampling","feed_subtitle":"A training-free sampler recomputes the negative direction each step; humans preferred its outputs twice as often in tests.","key_machinery":"The central object is the diffusion-negative-sampling (DNS) chain from prior work, defined by $\\hat{\\epsilon} = \\hat{\\epsilon}_\\phi + s_n(\\hat{\\epsilon}_\\phi - \\hat{\\epsilon}_p)$, which generates \"negative images\" and, in DNP, is captioned to obtain a negative prompt. ANSWER repurposes this chain as a per-step oracle: instead of completing the chain and captioning it, it truncates the chain after $K_t$ steps and takes the noise $\\hat{\\epsilon}_{t_n}$ at time $t-K_t$ as a substitute for the negative-condition noise in the negative-prompting equation. Two auxiliary mechanisms carry the argument: a scheduler $K_t = K \\cdot \\frac{2t-T}{2t-T+20} \\cdot \\frac{T+20}{T}$ for $t \\in (T, ..., T/2)$ that smoothly transitions the chain from negative to unconditional guidance, and a normalization step that rescales $\\hat{\\epsilon}_{t_n}$ to the mean and variance of the unconditional noise $\\hat{\\epsilon}_\\phi$ at time $t$, which the ablations show removes blur introduced by the $K$-step gap between the two chains.","core_discovery":"On the paper's own terms, the central claim is that negative guidance can be made fully automatic and adaptive by replacing the fixed negative prompt in negative prompting with a negative noise estimate produced on the fly: at step $t$, run $K_t$ iterations of the diffusion-negative sampling chain, $\\hat{\\epsilon} = \\hat{\\epsilon}_\\phi + s_n(\\hat{\\epsilon}_\\phi - \\hat{\\epsilon}_p)$, initialized at the current latent $z_t$, and use the resulting noise at time $t-K_t$ as $\\hat{\\epsilon}_n$ in the negative-prompting equation $\\hat{\\epsilon} = \\hat{\\epsilon}_n + s(\\hat{\\epsilon}_p - \\hat{\\epsilon}_n)$. The paper motivates this by arguing that the ideal negative prompt $n^*(p,t)$ defined through the odds ratio $p_\\theta(p|z_t)/p_\\theta(n|z_t)$ varies with $t$, and that DNS provides the direction $-\\hat{\\epsilon}_p$ that maximizes this ratio. It then shows that a $K_t = 1$ version collapses to classifier-free guidance with a rescaled scale, so the method's benefit rests on $K_t > 1$, a hand-designed scheduler that decays $K_t$ over the first half of the chain, and a normalization that matches the mean and variance of the negative noise to the unconditional noise. The experiments report that SDXL + ANSWER outperforms SDXL with CFG and SDXL with DNP on CLIP score, ImageReward, HPSv2, and PickScore across the Attend&Excite, Pick-a-Pic, DrawBench, PartiPrompts, and ImageNet sets, with win rates around 50% and a 2x human preference over the baselines in a human evaluation study.","pith_inferences":["The paper fixes $K=5$ for all prompts; a prompt- and seed-dependent policy for $K$ or for the stopping time $t=T/2$ is a natural next step, since the paper's own ablation shows quality degrades when $K$ is too large and different prompts need different correction strength.","Because ANSWER operates entirely on noise estimates, it is composable with attention-based latent editing methods; the combination could fix spatial binding and anatomy simultaneously, though the paper does not test this.","The claim that the ideal negative prompt shifts along the chain could be tested directly by comparing DNS images at consecutive steps with caption similarity; if the negatives only change cosmetically, a single negative would suffice and ANSWER's adaptivity would be unnecessary.","An information-theoretic reading is that ANSWER distills a per-step negative text prompt into a noise vector, avoiding the quantization loss of captioning; this suggests the method may be most valuable for fine-grained attributes that captioners routinely miss, which the paper observes qualitatively."],"forward_implications":["Any diffusion model that already uses classifier-free guidance can switch to ANSWER without retraining or extra text inputs.","Per-step negative noise removes the two main costs of diffusion-negative prompting: the external captioning model and the fixed one-time negative caption.","For prompts where high CFG scale fails, such as spatial relations, numeracy, and imaginary scenes, ANSWER's odds-ratio boosting can force compliance, as the paper's qualitative examples show.","Human preference studies reported in the paper put ANSWER ahead of both CFG and DNP by roughly 2-to-1, with win rates of 41–52% on automated metrics.","The method's gains are largest in the early denoising steps; after the midpoint, CFG suffices, which bounds the extra inference cost to $O(KT)$ with small $K$."],"supporting_citations":[{"why":"Defines the DNS chain and the optimal negative prompt concept that ANSWER extends, and provides the DNP baseline it must beat.","marker":"[6]"},{"why":"Introduces classifier-free guidance, the base sampling procedure that ANSWER modifies and the primary comparison baseline.","marker":"[12]"},{"why":"Introduces the negative-prompting equation that ANSWER plugs its adaptive noise estimate into.","marker":"[18]"},{"why":"Supplies the energy-based contextual normalization argument used to justify the DNS noise as a negative condition.","marker":"[7]"},{"why":"Contributes the Attend&Excite dataset and comparison of prompt-compliance baselines used in the experiments.","marker":"[5]"},{"why":"Provides the Pick-a-Pic preference dataset and PickScore metric used to measure human-like preference gains.","marker":"[15]"}],"fun_headline_variants":["Training-free adaptive negative sampling sharpens prompt adherence","Adaptive negative sampling without external prompts beats fixed ones","No external prompts adaptive negative guidance improves diffusion","Adaptive negatives double human preference in diffusion sampling"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method assumes that a short DNS chain started from the current latent $z_t$ yields a noise estimate that faithfully represents the optimal negative direction $n^*(p,t)$ for that step, so that using it in the negative-prompting equation improves the odds of the positive prompt; the paper only proves the $K_t=1$ degenerate case and relies on empirical ablations for $K_t>1$.","fun_headline_variants_meta":{"raw":{"variants":["Training-free adaptive negative sampling sharpens prompt adherence","Adaptive negative sampling without external prompts beats fixed ones","No external prompts adaptive negative guidance improves diffusion","Adaptive negatives double human preference in diffusion sampling"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000785,"raw_usage":{"total_tokens":3546,"prompt_tokens":1105,"completion_tokens":2441,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":721,"completion_tokens_details":{"reasoning_tokens":2382}},"tokens_in":721,"tokens_out":2441,"duration_ms":20145,"temperature":1.0,"reasoning_tokens":2382,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T04:46:40.308798+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run ANSWER on a set of prompts while replacing the per-step negative noise with either the DNS noise computed once at the first step and reused everywhere, or with the normalized noise of an unrelated prompt; if prompt-adherence metrics stay roughly constant, then the adaptive component of ANSWER is not doing the work, and only the stronger guidance direction from the DNS chain matters. The paper's own $K$ ablation shows the trade-off, but it does not isolate adaptivity from the mere presence of a DNS-derived negative direction.","supporting_citations":[{"cited_title":"Improving image syn- thesis with diffusion-negative sampling, 2024","cited_arxiv_id":null,"evidence_quote":"Defines the DNS chain and the optimal negative prompt concept that ANSWER extends, and provides the DNP baseline it must beat."},{"cited_title":"Compositional visual generation with composable diffusion models","cited_arxiv_id":null,"evidence_quote":"Introduces the negative-prompting equation that ANSWER plugs its adaptive noise estimate into."},{"cited_title":"Compositional visual generation with energy based models","cited_arxiv_id":null,"evidence_quote":"Supplies the energy-based contextual normalization argument used to justify the DNS noise as a negative condition."},{"cited_title":"Attend-and-excite: Attention-based se- mantic guidance for text-to-image diffusion models, 2023","cited_arxiv_id":null,"evidence_quote":"Contributes the Attend&Excite dataset and comparison of prompt-compliance baselines used in the experiments."}],"review_version":1}