{"id":"9f0c15d2-279b-4eee-b5d8-2dadbf575674","arxiv_id":"2506.19713","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"By boosting high-frequency guidance while keeping low-frequency guidance modest, frequency-decoupled guidance improves image quality and diversity metrics across several diffusion models.","lead":"This paper splits the guidance signal in diffusion image models into low-frequency and high-frequency parts and applies different strengths to each. The result is sharper images at low guidance settings, with better diversity than standard classifier-free guidance.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Figure 5's low- vs high-frequency comparison is not matched in total guidance strength; the conclusion that low-frequency guidance causes diversity loss may be an artifact of its larger norm.","rationale":"The reader's weakest assumption is that latent-space Laplacian bands correspond to perceptual structure and detail. That is a plausible concern for the interpretation of the method, but I see a more directly load-bearing problem in the isolation experiment used to establish the causal mechanism: the low- and high-frequency arms of Figure 5 are not matched in total guidance strength because low-frequency guidance has a much larger norm in these models. If this confound is real, the paper's central explanation for why FDG works is unsupported even if the empirical FID gains survive, since the motivating comparison for w_low<w_high would be an artifact of comparing different amounts of total guidance. This is a correctness risk in the experimental design rather than a question of whether the frequency transform maps sensibly from latent to image space. I therefore would keep the paper's CONDITIONAL verdict unchanged, but the condition should explicitly require norm-matched frequency ablations. The proposed test directly checks whether the diversity loss attributed to low frequencies persists when total guidance magnitude is equalized; if it does not, the authors would need to reframe FDG as an empirical guidance-reshaping method rather than evidence for a frequency-specific causal mechanism.","tokens_in":17872,"tokens_out":9370,"duration_ms":101806,"concrete_test":"Re-run the Figure 5 sweep with per-step normalized frequency bands: compute psi_low[Delta] and psi_high[Delta] at each sampler step, normalize each band to unit norm, apply w_low and w_high to the normalized bands, and sweep both scales while recording FID, recall, and precision. Equivalent formulation: choose w_low and w_high per step so that (w_low-1)||psi_low[Delta]|| = (w_high-1)||psi_high[Delta]||. If low-frequency-only no longer uniquely suppresses diversity, or if high-frequency-only reduces diversity once its correction magnitude is matched, then the paper's causal claim is confounded by total guidance strength rather than frequency content.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central mechanistic claim (Sec. 5.2, Fig. 5) is that low-frequency CFG components are the main cause of diversity loss and oversaturation, while high-frequency components enhance quality without affecting diversity. The supporting experiment compares the arms w_low=w, w_high=1 and w_low=1, w_high=w at the same scalar w. These arms do not apply equal total guidance: by Eqs. (5)-(6), the first injects (w-1)*psi_low[Delta] while the second injects (w-1)*psi_high[Delta], where Delta = D_c - D_u. Figure 7 itself reports that ||psi_low[Delta]|| is much larger than ||psi_high[Delta]|| over most of the sampling trajectory, with low-frequency norms dominating at early steps. Thus, at the same w, the low-frequency-only arm applies substantially stronger total correction than the high-frequency-only arm. Lower recall and higher saturation in the low-frequency-only arm can therefore be explained by total guidance strength alone, not by a frequency-specific effect. The same confound applies to the prompt-alignment analysis in Fig. 6. The empirical FID/recall gains of FDG in Table 1 may still hold, but the design principle w_low<w_high and the causal narrative that low-frequency guidance uniquely harms diversity are not established by the experiments as presented.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper analyzes classifier-free guidance (CFG) in the frequency domain and argues that low-frequency components of the guidance signal mainly govern global structure and condition alignment, while high-frequency components mainly contribute to visual fidelity. On this basis, the authors propose frequency-decoupled guidance (FDG), which applies separate CFG scales w_low and w_high to low- and high-frequency bands of the denoiser prediction. The central algebraic identity (Eq. 4), which follows from linearity of the frequency transform, is correct. The paper reports consistent FID and recall improvements over standard CFG across several models (EDM2, DiT-XL/2, Stable Diffusion 2.1, SDXL, SD3), plus improvements on human-preference and prompt-alignment metrics, and it claims FDG is a plug-and-play, no-retraining alternative to CFG.","tokens_in":18187,"tokens_out":4791,"duration_ms":50654,"significance":"If the empirical claims hold, FDG is a practically valuable contribution: it is simple, adds negligible sampling cost, works with pretrained models, and improves the quality-diversity trade-off relative to CFG. The paper also provides a plausible mechanistic story for why CFG hurts diversity and causes oversaturation, which is of independent interest. The main results are supported by experiments across multiple model families and datasets, and the central derivation is elementary and sound. However, the causal interpretation is currently under-supported by the ablations: the comparison in Fig. 5 does not control for total guidance strength, and the frequency decomposition is applied in latent space for the Stable Diffusion models without validating that latent-space bands correspond to perceptual image frequencies. These issues affect the paper's main explanatory claim, though the empirical gains of FDG may still be valid. The method is not circular: the derivation is a linearity identity, not a fitted quantity, and the per-model parameter choices are presented as design choices informed by sweeps.","major_comments":[{"comment":"The comparison between the arms (w_low=w, w_high=1) and (w_low=1, w_high=w) does not control for total guidance strength. From Eqs. (5)-(6), the injected guidance is (w-1)*psi_low[Delta] in the first arm and (w-1)*psi_high[Delta] in the second, where Delta = D_c - D_u. Figure 7 itself shows that ||psi_low[Delta]|| is much larger than ||psi_high[Delta]|| over most of the sampling trajectory. Lower recall, higher saturation, and worse FID in the low-frequency-only arm could therefore be explained by a larger L2 norm of the applied correction rather than by a frequency-specific effect. The same confound affects the prompt-alignment analysis in Figure 6. Please rerun the ablation with matched per-arm guidance norms (e.g., scaling w_high so that (w_high-1)||psi_high[Delta]|| = (w_low-1)||psi_low[Delta]|| at each step) and report recall, saturation, FID, and CLIP score as functions of the total injected norm; this is necessary to support the causal narrative that low-frequency guidance specifically harms diversity.","section":"Section 5.2, Figure 5"},{"comment":"The FDG parameters w_low and w_high are reported per model, but the selection procedure is not described: there is no validation split, no sensitivity analysis, and no error bars over seeds or prompt sets. Because FDG introduces two free parameters while CFG has one, a fair comparison requires a specified protocol for choosing (w_low, w_high) and an estimate of variance. The reported FID differences for SDXL (25.23 vs 24.60) and SD2.1 (24.99 vs 23.33) are small and could fall within run-to-run noise. Please add a validation protocol, confidence intervals or standard errors, and at least a one-dimensional sensitivity sweep around the chosen settings.","section":"Section 5.1, Table 8"},{"comment":"For the Stable Diffusion models, the Laplacian pyramid is applied directly to the latent-space predictions pred_cond and pred_uncond. The paper's mechanistic interpretation implicitly assumes that low- and high-frequency bands in the VAE latent space correspond to low- and high-frequency perceptual structure in the decoded image. This is not immediate because the latent space has a different spectral bias and the VAE decoder applies learned upsampling. Please validate the mapping, for example by decoding latent images whose frequency bands have been manipulated and measuring whether image-domain frequency content changes correspondingly, or by computing the same norm analysis on decoded predictions. Without this validation, the claim that 'low-frequency guidance governs global structure' is not established for latent diffusion models.","section":"Section 4, Algorithm 2"}],"minor_comments":[{"comment":"The pseudocode includes parallel_weights and a project() function that appear related to APG, but the main text does not describe when these are used in FDG. Please clarify whether they are part of FDG or optional add-ons, or remove them from the main algorithm.","section":"Algorithm 2"},{"comment":"The heading 'Guidane parameters' contains a typo; it should read 'Guidance parameters'.","section":"Table 8"},{"comment":"The PickScore column reports values such as 0.45 vs 0.55, which the appendix describes as win probabilities, but the table caption does not explain this. Please state explicitly in the caption that these are win rates against the CFG baseline, not raw PickScore values.","section":"Table 2"},{"comment":"The legend defines the three curves but does not explain what 'w' denotes for the CFG curve nor how the saturation metric is computed; please add these details to the caption.","section":"Figure 5"},{"comment":"Reference [62] applies frequency-aware guidance to diffusion models for image restoration; a brief sentence positioning FDG relative to this prior frequency-guided diffusion work would help the reader understand the novelty.","section":"Related work"}],"recommendation":"major_revision","confidential_remarks":"The paper is a reasonable candidate but needs major revision. The central algebraic identity is sound and the empirical gains, if reproducible, are useful. The main risk is that the mechanistic interpretation is supported by an ablation that does not match total guidance strength; if the authors can add a matched-norm experiment and a proper parameter-selection protocol, the contribution would be solid. The latent-space frequency mapping concern is also worth addressing, since it affects the explanatory claim for the Stable Diffusion models. I do not see circularity or any indication of fabricated results."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a solid empirical paper. FDG is a simple change to CFG — decompose the guidance update into low and high frequency bands and weight them separately — and it consistently improves FID and recall across EDM2, DiT, and the Stable Diffusion family. The gains are large. If the method holds up on replication, it's a genuinely useful plug-and-play alternative to CFG, and the frequency-domain lens is a nice way to think about guidance.\n\nWhat's actually new: the observation that CFG's uniform scale is suboptimal because low and high frequencies behave differently. The math is just linearity of the transform, but the point is the decoupling. The experiments are broad: class-conditional and text-to-image, multiple samplers, distilled models, and the combination with CADS and APG. They also show it works with different step counts. No retraining, negligible overhead. That's the good stuff.\n\nNow the soft spots. The biggest one is the mechanistic claim in Sec 5.2. The comparison in Fig. 5 between low-only and high-only guidance is not matched in total strength. Figure 7 shows that the low-frequency component of the guidance signal has much larger norm early in sampling, so at the same w the low-frequency arm injects a much stronger total correction. Lower recall in that arm could just be a strong-guidance effect, not a frequency-specific one. This confound also affects Fig. 6. The main FDG-vs-CFG result in Table 1 is not confounded in the same way — there they keep whigh at the CFG scale and reduce wlow, which is exactly the right test — but the design principle 'wlow should be below whigh' is motivated by the confounded sweep. They need to do an iso-norm comparison: scale each band so the injected correction has equal total norm, then re-test recall and saturation.\n\nSecond, the frequency decomposition is applied to latent tensors in the latent diffusion models. The paper doesn't justify that Laplacian bands in latent space correspond to perceptual high and low frequencies in image space. Maybe it works anyway, but the story is incomplete.\n\nThird, no error bars anywhere, and wlow/whigh are tuned per model on what appears to be the test set. For a method paper that's often tolerated, but given the confound, they should at least report variance and some validation split.\n\nNet: the method probably works as an empirical trick; the causal narrative needs more work. I'd send it to peer review and ask for the iso-norm experiment and error bars. It's a useful paper for practitioners and for anyone working on guidance mechanisms.","headline":"A useful plug-and-play tweak to CFG with a nice frequency-domain story, though the causal claim about which frequency hurts diversity is partly confounded by total guidance strength.","tokens_in":18668,"tokens_out":2673,"would_cite":true,"duration_ms":27140,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Classifier-free guidance works better when low frequencies are guided gently and high frequencies strongly, a scheme the paper calls frequency-decoupled guidance (FDG), which improves FID and recall at low guidance scales.","keywords":["classifier-free guidance","frequency-decoupled guidance","Laplacian pyramid","diffusion models","low guidance scales","diversity-quality trade-off","text-to-image generation","FID"],"falsifier":"Swap the scales: run FDG with $w_{\\mathrm{high}} < w_{\\mathrm{low}}$. The paper predicts oversaturation, reduced diversity, and worse FID; if that configuration matches or beats uniform CFG, the claimed division of labor between frequency bands is wrong.","tokens_in":17710,"feed_emoji":"🎨","tokens_out":6954,"duration_ms":62169,"temperature":0.7,"pith_summary":"Standard classifier-free guidance (CFG) applies one guidance strength to every frequency of the denoiser's prediction, which the paper argues is the root of the familiar trade-off: low scales give blurry, weakly aligned images, and high scales give oversaturated, less diverse ones. The paper's central claim is that low-frequency guidance controls global structure and condition alignment, while high-frequency guidance mostly controls fine visual detail, so the two should be scaled separately. It proposes frequency-decoupled guidance (FDG), which decomposes the CFG update with a Laplacian pyramid, applies a conservative scale to low frequencies and a stronger scale to high frequencies, and then reconstructs the guided prediction. Across class-conditional ImageNet models and Stable Diffusion text-to-image models, FDG reports better FID and recall than CFG at low guidance scales, with no retraining and nearly no extra sampling cost.","feed_headline":"Low and high frequencies get separate guidance, and images sharpen","feed_subtitle":"Splitting classifier-free guidance by frequency band improves FID and recall without retraining.","key_machinery":"The central object is the frequency decomposition of the denoiser's clean-image prediction under a linear, invertible transform $\\psi$, chosen as a single-level Laplacian pyramid: a coarse-to-fine decomposition that separates a signal into a low-frequency base and high-frequency detail. FDG computes $\\psi$ on both the conditional and unconditional predictions, applies the CFG interpolation separately with scales $w_{\\mathrm{low}}$ and $w_{\\mathrm{high}}$ to the two bands, and inverts the pyramid to obtain the guided prediction. This carries the argument because it converts the single global guidance scale of CFG into two per-band scales, making the paper's claim that low and high frequencies should be treated differently directly testable. The implementation also includes a projection step that keeps the guidance difference aligned with the conditional prediction, but the core mechanism is the per-band scaling.","core_discovery":"The discovery is an asymmetric frequency account of CFG. Viewing the CFG update as an interpolation in the frequency domain, the paper derives that a uniform scale $w$ acts identically on the low- and high-frequency components of the guided prediction, even though the two bands play different roles. Empirically, raising the low-frequency scale is what collapses diversity and causes oversaturation, while raising the high-frequency scale sharpens detail and improves quality without hurting diversity. FDG therefore sets $w_{\\mathrm{low}}$ close to or below the CFG scale and $w_{\\mathrm{high}}$ at or above it, and the paper reports consistent FID and recall improvements over CFG across EDM2, DiT-XL/2, Stable Diffusion 2.1, XL, and 3.","pith_inferences":["Editorial inference: if the frequency account is right, guidance schedules that increase $w_{\\mathrm{high}}$ over sampling time, rather than using a fixed split, should push the quality-diversity frontier further; the paper's norm analysis suggests the effective balance shifts as denoising progresses.","Editorial inference: the same per-band split could be applied to other guidance mechanisms and to non-image domains such as video or audio, where low-frequency structure and high-frequency texture have clear analogues.","Editorial inference: because the split is performed in latent space, a useful stress test is whether the gains persist when the same split is applied in decoded pixel space; that would separate a genuine perceptual-frequency effect from a latent-representation artifact."],"forward_implications":["At low guidance scales, FDG can replace uniform CFG as a drop-in sampling rule, improving FID and recall without retraining or extra compute.","Users can keep the diversity and natural color of low CFG scales while obtaining detail comparable to high CFG scales, by setting $w_{\\mathrm{low}}$ low and $w_{\\mathrm{high}}$ high.","Time-gated guidance methods, such as guidance interval, can be understood and tuned through the frequency norms of the guidance signal, since their practical benefit is an implicit emphasis on high-frequency guidance.","Distilled few-step models, where CFG often hurts, can still benefit from guidance through FDG, and text rendering in models like Stable Diffusion 3 improves because fine text detail lives in the high-frequency band."],"supporting_citations":[{"why":"Defines classifier-free guidance, the update rule FDG decomposes and modifies.","marker":"[17]"},{"why":"Supplies the Laplacian pyramid used as the frequency transform in FDG.","marker":"[8]"},{"why":"Provides the EDM2 class-conditional models and training-dynamics setup used as baselines.","marker":"[23]"},{"why":"Provides the DiT-XL/2 transformer baseline for class-conditional ImageNet generation.","marker":"[41]"},{"why":"Provides the Stable Diffusion 2.1 latent text-to-image model used as a baseline.","marker":"[44]"},{"why":"Provides the Stable Diffusion XL text-to-image model used as a baseline.","marker":"[42]"},{"why":"Provides the Stable Diffusion 3 rectified-flow transformer used as a baseline and for text-rendering tests.","marker":"[12]"},{"why":"Offers the guidance-interval baseline whose behavior FDG explains via frequency norms.","marker":"[28]"},{"why":"Introduces Autoguidance, whose guidance signal FDG analyzes in the frequency domain.","marker":"[24]"}],"fun_headline_variants":["Split CFG by frequency band: sharper images, no oversaturation","Separate low and high frequency guidance improves diffusion quality","FDG: split CFG into frequency bands to boost fidelity and diversity","High-frequency guidance sharpens; low-frequency keeps diversity","Frequency-split CFG: tune low and high bands separately for better images"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that low- and high-frequency bands in the model's latent prediction correspond to overall structure versus fine detail in the final image; if a Laplacian pyramid in latent space does not track perceptual frequency, the explanation of why FDG helps is not established.","fun_headline_variants_meta":{"raw":{"variants":["Split CFG by frequency band: sharper images, no oversaturation","Separate low and high frequency guidance improves diffusion quality","FDG: split CFG into frequency bands to boost fidelity and diversity","High-frequency guidance sharpens; low-frequency keeps diversity","Frequency-split CFG: tune low and high bands separately for better images"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000856,"raw_usage":{"total_tokens":3704,"prompt_tokens":914,"completion_tokens":2790,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":530,"completion_tokens_details":{"reasoning_tokens":2702}},"tokens_in":530,"tokens_out":2790,"duration_ms":22336,"temperature":1.0,"reasoning_tokens":2702,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T18:26:56.976916+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Swap the scales: run FDG with $w_{\\mathrm{high}} < w_{\\mathrm{low}}$. The paper predicts oversaturation, reduced diversity, and worse FID; if that configuration matches or beats uniform CFG, the claimed division of labor between frequency bands is wrong.","supporting_citations":[{"cited_title":"Guiding a diffusion model with a bad version of itself","cited_arxiv_id":null,"evidence_quote":"Introduces Autoguidance, whose guidance signal FDG analyzes in the frequency domain."}],"review_version":2}