{"id":"e54518da-fecd-41f7-85b3-196890ffadb1","arxiv_id":"2505.21179","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"Normalized Attention Guidance (NAG) stabilizes attention-space extrapolation with L1 normalization and refinement, restoring negative prompting in few-step diffusion models across architectures and modalities.","lead":"A new inference-time method applies negative prompts to diffusion models by extrapolating in attention space with L1 normalization and blending. It aims to restore negative guidance in few-step image and video diffusion models where classifier-free guidance fails.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The reported metrics do not isolate negative-attribute suppression from generic quality enhancement, so the central 'negative guidance' claim is not yet verified.","rationale":"The reader's weakest assumption was that L1-based normalization and blending preserve directional guidance while preventing out-of-manifold drift. My concern is adjacent but more specific: the paper never directly measures whether the directional guidance is actually suppressing the negated attribute. The reported metrics are consistent with a generic quality-enhancement effect, which would undermine the 'negative guidance' part of the central claim even if the normalization/blending mechanism works as a stabilizer. This is a load-bearing gap because the paper's novelty and universality claims rest on semantic suppression, not just on improved FID or CLIP. I agree with the reader that the method is a reasonable heuristic with promising qualitative results and a solid ablation showing the stabilizers matter, so the verdict should remain conditional pending direct evidence. I did not find an internal inconsistency in the equations or a fatal flaw in the mathematical derivation; the concern is about the evidential link between the mechanism and the claimed effect. The proposed test would settle it by isolating the contribution of the negative direction, and it is straightforward to run with released code or a simple reimplementation.","tokens_in":19114,"tokens_out":5173,"duration_ms":62313,"concrete_test":"Run a controlled attribute-suppression experiment on Flux-Schnell (4-step) and SDXL-Lightning (8-step) with three negative prompts: 'glasses', 'tiger', and 'male'. For at least 512 images per condition, compute (a) CLIP similarity between the generated image and the negative text, (b) accuracy of an off-the-shelf attribute classifier (face detector with glasses attribute, object detector for tiger, gender classifier), and (c) the same metrics under a control that applies NAG with a positive-only quality prompt ('high resolution, sharp') in place of the negative prompt, keeping all other settings fixed. If NAG's reduction in negative-attribute presence is no larger than the control's, the negative direction is not the active ingredient and the central claim fails. Report full distributions, not only means, and include failure cases.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that NAG restores effective negative guidance, i.e., directional suppression of unwanted attributes, especially in few-step models. However, every quantitative headline in Table 1 and Section 5.1 is obtained with the universal negative prompt 'Low resolution, blurry,' which is primarily a quality phrase. CLIP score, FID, PFID, and ImageReward would all improve from any manipulation that sharpens output or increases contrast, even if it did not implement semantic negative guidance. The paper reports no direct measure of negative-attribute presence: no classifier accuracy for 'glasses', 'tiger', 'male', or 'static', and no CLIP similarity to the negative text. The user study in Section 5.5 asks participants to judge text alignment while considering the positive and negative prompt together, so it cannot separate suppression of the unwanted attribute from overall aesthetic improvement. The qualitative examples and ablation are suggestive, and the method's self-reported failure cases in Section 6 show it does not always suppress concepts, but none of this establishes that the guidance direction, rather than the normalization/blending regularizer, is what produces the gains. If NAG is simply a quality enhancer in attention space, the claim of universal negative guidance collapses to a weaker claim of generic feature regularization.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Normalized Attention Guidance (NAG), a training-free inference-time method for negative prompting in diffusion models. NAG computes attention outputs for positive and negative prompts, extrapolates in attention feature space as eZ = Z+ + φ(Z+ - Z-), applies L1-ratio clipping with threshold τ, and blends with the positive output using α. The authors claim that NAG restores effective negative guidance where CFG collapses, particularly in few-step models, and that it generalizes across UNet/DiT architectures, image/video modalities, and few-/multi-step settings. Experiments on seven few-step models, several multi-step CFG models, and one video model report CLIP/FID/PFID/ImageReward gains, supplemented by ablations, latency measurements, and a user study.","tokens_in":19366,"tokens_out":5122,"duration_ms":49929,"significance":"If the causal attribution to negative-attribute suppression is established, NAG would be a practical and broadly applicable plug-in for diffusion models. The method is simple and clearly specified (Algorithm 1, Table 5), and the paper provides unusually broad model coverage, including recent DiT models and a 14B video model, along with latency measurements and a component ablation. The authors are also candid about failure cases in Section 6. However, the current evaluation does not isolate negative-attribute suppression from generic quality enhancement, so the central claim is not yet verified. The paper also lacks comparisons with standard negative-prompt baselines and statistical uncertainty estimates, which are necessary to support the 'universal' and 'consistent improvements' claims.","major_comments":[{"comment":"The quantitative evaluation relies on the single negative prompt 'Low resolution, blurry' and on aggregate quality metrics (CLIP, FID, PFID, ImageReward) that would also improve from generic sharpening or contrast enhancement. The paper does not report any direct measure of the presence of the suppressed attribute (e.g., classifier accuracy for 'glasses', 'tiger', or 'male', or CLIP similarity to the negative text), and the user study in Section 5.5 instructs participants to judge text alignment while considering both positive and negative prompts, so it cannot separate suppression from aesthetic improvement. Consequently, the central claim that NAG implements directional negative guidance rather than acting as a generic feature regularizer is not yet supported. I recommend adding attribute-presence metrics on the generated samples for several negative concepts, or a controlled experiment in which the negative prompt targets a semantic attribute while the positive prompt does not mention it.","section":"§5.1, Table 1, §5.5"},{"comment":"There is no baseline against standard text-level negative prompting. For multi-step models, CFG with a negative prompt is the default practice; for few-step models, straightforward baselines such as prompt inversion or simply appending the negative prompt to the positive prompt could be used. Without such comparisons, the gains in Tables 1 and 2 could stem from the additional negative conditioning per se rather than from the attention-space mechanism of NAG. Please add these baselines to the evaluation.","section":"§5.1–5.3"},{"comment":"The hyperparameters (ϕ, τ, α) in Table 5 are chosen per model family, and the paper does not state whether they were tuned on the same COCO-5K evaluation set used in Tables 1, 2, and 6. Moreover, no error bars or repeated runs are reported; several FID and PFID differences are within typical noise (e.g., SD3.5-Large CFG FID +0.35 in Table 2, SD1.5 FID +0.14 and PFID +1.29 in Table 6). This weakens the 'consistent improvements' and 'universal' claims. Please report the tuning protocol and provide confidence intervals or at least per-prompt variance.","section":"Table 5, Table 2, Table 6"},{"comment":"The comparison with NASA is limited to UNet models, and the user study against NASA (Table 8) is conducted on only one model (DMD2-SDXL). Since the paper motivates NAG partly by NASA's instability on DiT architectures, a quantitative DiT comparison (even if it shows NASA collapse) would strengthen the claim. Also, the attribute-specific suppression of NAG versus NASA is not compared directly, so it remains unclear whether NAG's advantage is in guidance direction or in overall stability.","section":"§5.2, Appendix E"}],"minor_comments":[{"comment":"The computed z_nag is never assigned to hidden_states; the function returns the unmodified hidden_states. Replace 'return hidden_states' with 'return z_nag' or add an assignment before the return.","section":"Algorithm 1"},{"comment":"The symbol R is used both for the per-token ratio R[i] and for the vector in Eq. (9); please use a different symbol for the vector (e.g., lowercase r) and make the element-wise division explicit.","section":"Equations (8)–(9)"},{"comment":"Table 2 and the main text cite PAG as [12], but [12] is 'Rectified diffusion guidance' whereas Perturbed Attention Guidance is [59]; please correct the citation.","section":"References and Table 2"},{"comment":"There is a typo: 'the the limitations' should be 'the limitations,' and 'CFG ... fails' should agree in number with 'CFG'.","section":"Conclusion"},{"comment":"The terms 'Guidance Boundary' and 'Refinement Manifold' are introduced only in the caption; please define them in the text or make the figure self-contained.","section":"Figure 4"},{"comment":"The video evaluation is qualitative only; the contribution statement in Section 1 about improving 'motion characteristics' would benefit from a quantitative metric on the generated videos.","section":"Section 5.4"}],"recommendation":"major_revision","confidential_remarks":"The paper is timely and the method is simple enough to be a useful plug-in, but the experimental protocol must be strengthened before it can support the advertised claims. The main risk is that the reported gains are generic quality improvements rather than demonstrable negative-attribute suppression. Also, the citation inconsistency for PAG ([12] vs [59]) should be corrected."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things worth knowing. First, NAG is a simple, well-ablated fix to NASA's instability in attention-space guidance. The method is essentially L1-norm normalization plus alpha-blending on the extrapolated attention features, and the ablation shows both components matter. On the evidence here, it does restore usable negative prompting in few-step models where CFG collapses, across DiT and UNet, and the qualitative results in Figures 1 and 5 are often convincing. The breadth of models and modalities is a real asset.\n\nSecond, the stress-test note is right. Every quantitative headline uses 'Low resolution, blurry' as the negative prompt, which is a quality phrase. CLIP, FID, ImageReward, even PFID, would all improve if NAG merely sharpened outputs or increased contrast. There is no direct measure of attribute suppression: no classifier scores for 'glasses' or 'tiger', no CLIP similarity to the negative text. The user study asks about alignment with the positive prompt while showing the negative prompt, so it cannot separate removal of the unwanted attribute from aesthetic improvement. This matters because the abstract claims 'negative guidance' as the mechanism, not just feature regularization. The paper's own failure cases in Section 6 show the suppression is incomplete, but that doesn't close the gap: the quantitative gains could come from the normalization and refinement alone, without the semantic direction.\n\nOther soft spots are minor. Hyperparameters are tuned per model family (Table 5), presumably on the same COCO-5K set used for evaluation, and there are no error bars. FID degrades in a few configurations (SD3.5-Large CFG in Table 2, SD1.5 in Table 6). None of this kills the method; it is an empirical heuristic and presented as such.\n\nBottom line: this is a solid engineering contribution worth taking seriously. It deserves peer review, not desk rejection. A referee should ask for attribute-specific measurement and a comparison against a text-only negative prompt in a multi-step setting, plus error bars or per-seed variance. The paper would be stronger if the authors framed NAG as a stabilizing feature-space guidance technique, and presented the quality gains as quality gains, rather than leaning on the 'negative guidance' label.","headline":"A genuinely useful stabilization of attention-space negative guidance, but the paper's central claim that it provides semantic negative guidance is not yet isolated from generic quality enhancement.","tokens_in":19873,"tokens_out":2699,"would_cite":true,"duration_ms":28905,"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":"Attention-space trick revives negative guidance in few-step diffusion","keywords":["diffusion models","negative prompting","classifier-free guidance","attention guidance","few-step sampling","Normalized Attention Guidance","DiT","video generation"],"falsifier":"Reproduce Table 1 on 4-step Flux-Schnell with the universal negative prompt 'Low resolution, blurry'. The claim would be falsified if (a) the reported CLIP, FID, and ImageReward improvements do not reproduce, (b) ablations without normalization or without blending match or beat the full method, or (c) at a fixed scale the extrapolated attention features $\\hat{Z}$ land measurably far outside the distribution of unmodified $Z^{+}$ in a way that predicts the observed artifacts.","tokens_in":1947,"feed_emoji":"🎨","tokens_out":3824,"duration_ms":103713,"temperature":0.7,"pith_summary":"The paper claims that negative prompting—asking a diffusion model to suppress unwanted attributes such as \"blurry\" or \"glasses\"—can be made reliable in fast few-step samplers, where standard classifier-free guidance (CFG) breaks down because the positive and negative predictions diverge. It introduces Normalized Attention Guidance (NAG), a training-free plug-in that steers generation by extrapolating the cross-attention outputs away from the negative prompt, then applies L1-norm clipping and alpha-blending to keep those features near the valid data manifold. If true, this would restore a core controllability tool for high-efficiency diffusion models and extend negative prompting to new domains without retraining.","feed_headline":"Attention-space trick revives negative guidance in few-step diffusion","feed_subtitle":"Works across image and video models with no retraining, and adds little latency even in 1-step samplers.","key_machinery":"The central object is the NAG feature transform applied inside cross-attention layers: $\\tilde{Z} = Z^{+} + \\phi (Z^{+} - Z^{-})$, followed by token-wise L1-norm ratio clipping, $\\hat{Z}[i] = \\frac{\\min(R[i], \\tau)}{R[i]} \\tilde{Z}[i]$ with $R[i] = \\frac{\\| \\tilde{Z}[i] \\|_1}{\\| Z^{+}[i] \\|_1}$, and final refinement $Z_{\\mathrm{NAG}} = \\alpha \\hat{Z} + (1-\\alpha) Z^{+}$. This transform replaces output-space CFG extrapolation with a bounded, direction-preserving move in attention space; the two stabilization operations work as a magnitude guardrail and a manifold anchor, respectively, and are what let the guidance remain stable at high scale.","core_discovery":"The paper's central claim is that effective negative guidance can be recovered in few-step diffusion models by moving extrapolation out of the predicted-noise/output space and into the attention feature space, with two stabilizing operations that prior attention-manipulation methods like NASA lack. Specifically, NAG computes attention outputs $Z^{+}$ and $Z^{-}$ from the positive and negative prompts, forms an extrapolated feature $\\tilde{Z} = Z^{+} + \\phi (Z^{+} - Z^{-})$, then rescales each token's L1 norm to a bounded threshold $\\tau$ and blends the result with $Z^{+}$ by a factor $\\alpha$. The authors argue the L1 normalization preserves low-magnitude activations that carry subtle semantics while constraining extremes, and that the blending pulls features back toward the stable positive manifold; together these prevent the out-of-manifold drift and feature collapse they observe in prior methods. They claim this makes NAG a universal plug-in: it works where CFG collapses (few-step UNet and DiT models such as Flux-Schnell, SDXL-Lightning, and DMD2-SDXL), improves multi-step models when added on top of CFG and PAG, and transfers to video diffusion (Wan2.1) for both content suppression and motion enhancement.","pith_inferences":["The consistent behavior under a fixed L1 clip and blend across very different architectures hints that attention features in these models share a common scale structure; one could test this directly by measuring the distribution of $\\|Z\\|_1$ across layers and models and checking whether the default $\\tau = 2.5$ tracks a consistent quantile.","Because NAG operates on frozen attention features rather than on training, the same mechanism could be aimed at other steering objectives—concept erasure, style transfer, or prompt-based debiasing—provided a negative description can be written; the paper demonstrates the prompting use, not the full space of guidance tasks.","The early-stopping result suggests the semantic direction of a negative prompt is mostly determined in the first denoising steps; if bias suppression is applied only there, downstream steps remain untouched, which might make NAG-based debiasing less likely to degrade fine texture.","Nothing in the method requires the negative condition to be textual; using an image as the negative condition could test whether NAG's attention-space extrapolation works for reference-based editing."],"forward_implications":["Few-step models that ship without CFG support (Flux-Schnell, SANA-Sprint, SD3.5-Large-Turbo, NitroSD-Realism, DMD2-SDXL, SDXL-Lightning) gain a working negative-prompting interface without retraining.","When combined with CFG and PAG in 25-step models, NAG improves CLIP score and ImageReward while leaving FID and PFID roughly flat or slightly better, giving a way to add negative control on top of existing guidance.","Applying NAG only for the first fraction of denoising steps (e.g., $\\theta = 0.25$) keeps most of the measured benefit while cutting the added latency, since the guidance effect is concentrated in early timesteps.","NAG extends to video generation on Wan2.1, suppressing undesired content such as human faces or static motion and improving temporal dynamics.","The method requires no retraining and adds less latency than CFG on most model families (e.g., +13% on SANA and +12% on Wan2.1 versus +90% to +100% for CFG), making it a practical drop-in at inference time."],"supporting_citations":[{"why":"Classifier-Free Guidance, the output-space extrapolation baseline that NAG replaces and whose few-step failure motivates the method.","marker":"[9]"},{"why":"NASA, the prior attention-space guidance method that NAG stabilizes through normalization and refinement.","marker":"[13]"},{"why":"Flux-Schnell and Flux-Dev, the main DiT testbed for few-step and multi-step experiments.","marker":"[1]"},{"why":"SDXL-Lightning, a few-step UNet model used to establish coverage of the UNet architecture.","marker":"[7]"},{"why":"NitroFusion, the single-step UNet model in the few-step evaluation.","marker":"[8]"},{"why":"Wan2.1, the video model used to demonstrate cross-modal generalization.","marker":"[6]"},{"why":"PAG, the guidance method that NAG is integrated with in multi-step experiments.","marker":"[12]"}],"fun_headline_variants":["Attention-space extrapolation rescues negative guidance","Universal plug-in: attention-space negative guidance","Few-step diffusion: negative guidance via attention space","NAG: training-free negative guidance for diffusion","Attention-space fix revives negative guidance in few-step"],"cache_read_input_tokens":22016,"weakest_assumption_plain":"The whole method rests on the heuristic that L1-norm clipping plus alpha-blending keeps extrapolated attention features close enough to the true feature manifold, and that this holds across architectures and modalities without per-model tuning; the paper does not derive this from theory, only from the demonstrated experiments.","fun_headline_variants_meta":{"raw":{"variants":["Attention-space extrapolation rescues negative guidance","Universal plug-in: attention-space negative guidance","Few-step diffusion: negative guidance via attention space","NAG: training-free negative guidance for diffusion","Attention-space fix revives negative guidance in few-step"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001323,"raw_usage":{"total_tokens":5420,"prompt_tokens":1013,"completion_tokens":4407,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":629,"completion_tokens_details":{"reasoning_tokens":4337}},"tokens_in":629,"tokens_out":4407,"duration_ms":33907,"temperature":1.0,"reasoning_tokens":4337,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T13:32:14.444069+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Reproduce Table 1 on 4-step Flux-Schnell with the universal negative prompt 'Low resolution, blurry'. The claim would be falsified if (a) the reported CLIP, FID, and ImageReward improvements do not reproduce, (b) ablations without normalization or without blending match or beat the full method, or (c) at a fixed scale the extrapolated attention features $\\hat{Z}$ land measurably far outside the distribution of unmodified $Z^{+}$ in a way that predicts the observed artifacts.","supporting_citations":[{"cited_title":"Rectified diffusion guidance for conditional generation","cited_arxiv_id":null,"evidence_quote":"PAG, the guidance method that NAG is integrated with in multi-step experiments."}],"review_version":1}