Pith. sign in

REVIEW 3 major objections 6 minor 12 references

One-coefficient fix stops CFG oversaturation at zero extra cost

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · glm-5.2

2026-07-09 03:02 UTC pith:VWTBBVMK

load-bearing objection One-coefficient CFG repair derived from a clean terminal-exactness argument on a Gaussian calibration model; theory is solid but only proven on that model, with learned-checkpoint evidence that is suggestive but indirect. the 3 major comments →

arxiv 2607.07665 v1 pith:VWTBBVMK submitted 2026-07-08 cs.LG cs.NAmath.NA

Guidance Breaks the Fitted Operator: A Terminal-Fitted Repair for Classifier-Free Guidance

Shiheng Zhang This is my paper
classification cs.LG cs.NAmath.NA
keywords classifier-free guidancediffusion modelsDDIMnumerical analysisasymptotic-preserving schemesoversaturationfitted operatordiscretization error
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

Classifier-free guidance (CFG) is the standard technique for strengthening class-conditional signals in diffusion model sampling, but at high guidance scales it oversaturates and destabilizes. This paper argues that the root cause is a discretization mismatch: CFG changes the contraction rate of the sampler's discriminative subspace from exponent 1 to 1+w, which breaks the property that makes the standard DDIM step exact on the terminal (low-noise) stretch of sampling. On coarse step schedules, this causes the guided residual to diverge as the noise floor approaches zero. The paper proves a three-threshold barrier showing exactly when sign-flips, residual blow-up, and instability occur, and reads the familiar one-step oversaturation as the endpoint of this barrier—a solver artifact rather than a property of the continuous guided flow. The same analysis produces a repair: replace CFG's guidance coefficient w(r-1) with r^(1+w) - r, where r is the per-step noise ratio. This one-line, zero-extra-cost change is the unique coefficient that restores exactness on the terminal discriminative layer without requiring any spectral information. On the calibration Gaussian model, it removes the divergent blow-up and is first-order accurate against the exact guided flow. On learned CIFAR-10 checkpoints and in a Stable Diffusion 1.5 smoke test, it acts as a high-guidance stabilizer—cutting residual amplification and pixel saturation while preserving class-conditional accuracy—though it is not a universal image-quality win (KID can still favor vanilla CFG).

Core claim

The central object is the guided terminal exponent: on a commuting Gaussian calibration model, CFG re-stiffens the discriminative subspace (directions present in the marginal but collapsed in the class) to contraction exponent 1+w instead of the unguided exponent 1. This exits the class of fitted operators for which the DDIM step is layer-exact, causing a sigma_min-divergent residual blow-up on coarse meshes. The paper proves this is governed by three ordered step-size thresholds (sign preservation, residual amplification, absolute stability) and shows that the coefficient replacement w(r-1) -> r^(1+w) - r is the unique spectrum-free repair that restores terminal exactness, removes the blow-

What carries the argument

The guided terminal exponent mu_w = (1+w)A - wB, where A and B are per-coordinate contraction rates of the conditional and unconditional denoisers on a commuting Gaussian model. The three-threshold guided clock barrier (h_flat, h_sharp, h_infinity) in terms of step size h and guidance weight w. The fitted coefficient r^(1+w) - r, derived by a gauge change that rectifies the anomalous exponent back to 1, then transporting the sigma-clock Euler step through it.

Load-bearing premise

The entire theoretical analysis is carried out on a linear Gaussian model where the class and marginal covariances commute and the class is narrower in every direction. Real neural network checkpoints need not satisfy this, and the paper acknowledges the learned diagnostics are only 'consistent with' but do not verify the model's structural predictions.

What would settle it

If on real learned checkpoints the directional variance ordering between conditional and unconditional denoisers does not follow the class-subset pattern (a_i <= b_i), the exponent-1+w re-stiffening mechanism may not be the operative cause of oversaturation, and the fitted coefficient's improvements could stem from a different effect than the one the theory describes.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Practitioners using high guidance scales (g >= 7.5) can replace one coefficient in their DDIM+CFG sampler to reduce oversaturation and residual amplification at zero additional compute cost, as a drop-in one-line code change.
  • The three-threshold barrier provides a parameter-free prediction for the terminal edge of limited-interval guidance schedules, potentially replacing tuned guidance-interval cutoffs with a derived one.
  • The finding that oversaturation is a solver artifact (discretization mismatch) rather than a property of the continuous guided law suggests that other guidance pathologies at high w may also have discretization-induced components separable from the continuous-flow distortion.
  • The reverse-ordering case (class wider than marginal in some direction) creates a continuous-flow instability at exponent -w that no solver can repair, marking a hard boundary on where any discretization-level fix can help.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • If the discriminative-subspace re-stiffening mechanism holds on non-Gaussian or non-commuting geometries, the fitted coefficient could serve as a general-purpose terminal stabilizer for any flow-matching or diffusion sampler using CFG, not just DDIM.
  • The O(w^2) prefactor in the accuracy bound (not uniform in w) suggests there may be a w-dependent step-count requirement even with the repair, implying an optimal guidance-weight-to-step-count tradeoff curve that the paper does not fully characterize.
  • The FID/KID split (fitted wins FID, CFG wins KID) may reflect a fundamental tension between reducing discretization artifacts and preserving the distributional properties that KID measures—worth investigating whether the KID penalty is causally linked to the oversaturation removal or is an independent effect.
  • The class-subset hypothesis (class narrower than marginal in every direction) may be approximately testable on real checkpoints by estimating directional variances of conditional vs. unconditional denoisers, which would bridge the gap between the calibration model and the learned-checkpoint diagnostics.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 6 minor

Summary. This paper analyzes classifier-free guidance (CFG) through a numerical-analysis lens, building on a prior fitted-operator framework for unguided DDIM. On a commuting Gaussian calibration model (Hypothesis 1), the author shows that guidance re-stiffens the discriminative subspace to an anomalous terminal exponent 1+w, causing DDIM+CFG's terminal residual to diverge as sigma_min → 0 on coarse meshes. A guided clock barrier with three ordered step-size thresholds is proven (Theorem 1), and a one-coefficient, zero-extra-NFE repair—replacing w(r-1) by r^(1+w)-r on the guidance direction—is derived as the unique spectrum-free terminal-exact coefficient (Lemma 2). On the discriminative crossover, this repair removes the sigma_min-divergent blow-up and is first-order accurate against the exact guided flow (Theorem 2, Proposition 4). Empirical diagnostics on learned CIFAR-10 EDM checkpoints and a Stable Diffusion 1.5 smoke test show reduced residual amplification, clipping, and saturation, with 9/9 FID wins over CFG on the tested grid, though KID favors CFG and the dense-reference L2 diagnostic shows the fitted step is not a uniformly better integrator.

Significance. The paper provides a clean, parameter-free derivation of a one-coefficient CFG repair from terminal-exactness on a calibration model, with explicit constants and complete proofs (Appendix A). The barrier thresholds (Theorem 1) yield a falsifiable, parameter-free prediction for the terminal edge of limited-interval guidance (Remark 3). The empirical program is honestly scoped: the author reports the KID split, the dense-reference L2 failure (Appendix H), and the reverse-ordering instability boundary (Remark 4) alongside the positive results. The zero-extra-NFE, one-line implementation delta and the cross-domain SD 1.5 smoke test add practical value. The theory-to-practice gap—between the commuting Gaussian model and learned checkpoints—is the main limitation, and the author is transparent about it.

major comments (3)
  1. §4, Table 2 and Appendix H: The practical evidence is honestly scoped but the central tension between the theoretical mechanism and the empirical results is underdeveloped. The KID split (CFG wins every block) and the dense-reference L2 diagnostic (Table 5, ratios >1) show the fitted step is not a uniformly better integrator. Meanwhile, the theoretical motivation for the specific coefficient r^(1+w)-r rests on the commuting Gaussian model, and the learned-checkpoint diagnostics (Table 1) are described as 'consistent with' the model ordering without directly verifying that the guided terminal exponent on the discriminative subspace of a learned checkpoint approaches 1+w as sigma → 0 (Proposition 1's prediction). The paper would be substantially strengthened by a direct measurement attempting to estimate the effective terminal exponent on the discriminative subspace of the CIFAR-10 EDM den
  2. §3, Proposition 4 / Theorem 2: The first-order accuracy result is proven only for a single discriminative coordinate (a=0 < b) on the commuting model. The paper explicitly states: 'Extending this crossover accuracy to the tangential layers and learned geometry needs the spectrum-aware Gaussian factor (6) and is future work.' On the commuting model coordinates decouple, so this is less concerning there, but the gap to the multi-coordinate learned setting is not bridged even at the level of a multi-coordinate Gaussian model where the spectrum-aware factor (6) could be evaluated. A numerical experiment on a multi-coordinate commuting Gaussian (e.g., 2-3 coordinates with different b_i values) showing that the single-coordinate accuracy result does or does not extend would help clarify the scope of the theory.
  3. §2, Hypothesis 1 and Remark 4: The class-subset hypothesis (a_i ≤ b_i) is a load-bearing assumption, and the reverse ordering (b=0 < a) creates a continuous-flow instability. The paper acknowledges that 'real learned checkpoints need not obey it' but does not provide any diagnostic or argument for why the EDM CIFAR-10 checkpoints used in §4 might satisfy it, or how common the reverse ordering is in practice. Even an indirect argument (e.g., relating the class-subset property to the training setup where the unconditional model is a mixture of class-conditional models) would help readers assess the applicability of the theory.
minor comments (6)
  1. Figure 1: The caption says 'schematic' but the axes are unlabeled beyond the text annotations. Adding axis labels (e.g., sigma or lambda on the horizontal axis, exponent on the vertical) would improve clarity.
  2. §3, Eq. (15): The notation r^(1+w) could be confused with r raised to the power (1+w) in typeset form; the paper is consistent but a brief note on first use that this is exponentiation would help.
  3. Table 1: The 'clip_delta' column is defined as 'fitted CFG minus CFG for the median fraction of final-denoised pixels outside [-1,1]' but the negative values (improvement) are not immediately intuitive; a note that negative means improvement would help.
  4. §4, Table 3: The target accuracy for fitted CFG at w=8, N=16 jumps to 99.3%, notably higher than CFG's 93.8%. The paper notes 'the high fitted value at N=16 should not be read as fuller conditional fidelity, as it may partly reflect reduced diversity,' but this large jump deserves more discussion or investigation.
  5. Appendix F: The Stable Diffusion smoke test uses float16, which could introduce numerical artifacts at high guidance. A note on whether this was checked would be useful.
  6. Reference [12] is by the same author and is the foundational fitted-operator result this paper builds on. This dependency is clearly stated, but the paper could more explicitly delineate what is novel in this work versus [12].

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for a careful and constructive report. The recommendation is minor revision, and all three major comments are actionable. We agree with the substance of each: (1) the theory-to-practice gap should be bridged by a direct terminal-exponent measurement on learned checkpoints; (2) the single-coordinate accuracy result should be tested on a multi-coordinate Gaussian; and (3) the class-subset hypothesis needs an explicit argument connecting it to the EDM training setup. We commit to all three additions in the revision and explain below what each will contain, including honest accounting of what we can and cannot claim.

read point-by-point responses

  1. Referee: The central tension between the theoretical mechanism and the empirical results is underdeveloped. The KID split and dense-reference L2 diagnostic show the fitted step is not a uniformly better integrator. Meanwhile, the theoretical motivation rests on the commuting Gaussian model, and the learned-checkpoint diagnostics are described as 'consistent with' the model ordering without directly verifying that the guided terminal exponent on the discriminative subspace of a learned checkpoint approaches 1+w as sigma → 0 (Proposition 1's prediction). The paper would be substantially strengthened by a direct measurement attempting to estimate the effective terminal exponent on the discriminative subspace of the CIFAR-10 EDM denoiser.

    Authors: We agree completely. The gap between 'consistent with the model ordering' and a direct exponent measurement is the weakest link in the current manuscript, and the referee is right that closing it—even partially—would substantially strengthen the paper. We will add a new diagnostic in Section 4 that attempts to estimate the effective terminal exponent on the discriminative subspace of the EDM CIFAR-10 checkpoints. Concretely: along sampling trajectories at high guidance (w=6.5, 8), we will compute the local contraction ratio of the guidance direction (D_u - D_c) across consecutive steps in the small-sigma regime and fit a power-law exponent in sigma. The commuting model predicts this exponent should approach 1+w on discriminative directions; we will report whatever the data shows, including if the exponent does not cleanly approach 1+w. We are honest that this is an indirect measurement: the learned denoiser's Jacobian does not diagonalize into clean eigen-coordinates, so we are measuring an effective exponent on a projected subspace, not verifying Proposition 1 literally. We will state this limitation explicitly. If the measurement is inconclusive or contradicts the prediction, we will say so rather than cherry-pick. We believe this is feasible with the existing checkpoints and sampling infrastructure, and we commit to including it in the revision. revision: yes

  2. Referee: The first-order accuracy result (Theorem 2) is proven only for a single discriminative coordinate (a=0 < b) on the commuting model. The gap to the multi-coordinate learned setting is not bridged even at the level of a multi-coordinate Gaussian model where the spectrum-aware factor (6) could be evaluated. A numerical experiment on a multi-coordinate commuting Gaussian (e.g., 2-3 coordinates with different b_i values) showing that the single-coordinate accuracy result does or does not extend would help clarify the scope of the theory.

    Authors: This is a fair and concrete suggestion. On the commuting model, coordinates decouple, so the single-coordinate accuracy result (Theorem 2) extends component-wise to any number of coordinates with different b_i values—the log-defect bound applies independently per coordinate and the W_2 bound sums. However, the referee's point is deeper: the practical question is whether the spectrum-free coefficient r^(1+w) - r, which is exact only on the pure discriminative terminal layer (A=1, B=0), remains accurate when different coordinates are at different points in their crossovers (different b_i values, hence different B_i at a given sigma). The honest answer is that it does not remain uniformly accurate in that regime—the spectrum-aware factor (6) would be needed for exactness, and the fitted step pays a finite O(1+w) crossover tax per coordinate (Proposition 4). We will add a numerical experiment on a 3-coordinate commuting Gaussian with distinct b_i values, showing the per-coordinate accuracy and the aggregate W_2 error as sigma_min → 0, on both uniform and non-uniform meshes. This will make the scope of Theorem 2 precise: the single-coordinate bound extends by independence, but the practical accuracy depends on how many coordinates are mid-crossover at the terminal steps, which is a mesh- and spectrum-dependent question the theory does not fully resolve. revision: yes

  3. Referee: The class-subset hypothesis (a_i ≤ b_i) is a load-bearing assumption, and the reverse ordering (b=0 < a) creates a continuous-flow instability. The paper acknowledges that 'real learned checkpoints need not obey it' but does not provide any diagnostic or argument for why the EDM CIFAR-10 checkpoints used in §4 might satisfy it, or how common the reverse ordering is in practice. Even an indirect argument (e.g., relating the class-subset property to the training setup where the unconditional model is a mixture of class-conditional models) would help readers assess the applicability of the theory.

    Authors: The referee is right that we left this hanging. We will add an explicit argument and a diagnostic. The argument: in the EDM CIFAR-10 setup, the unconditional model is trained on the full data distribution (a mixture over all classes), while each class-conditional model is trained on a single class. Under a Gaussian approximation, the marginal covariance C_u is the mixture covariance E[C_c | class] + Cov(mean_c), which is at least as large as any single class covariance C_c in the PSD order—this is the class-subset property a_i ≤ b_i at the population level. The reverse ordering (class wider than marginal in some direction) would require a class whose covariance exceeds the mixture covariance in some eigen-direction, which is possible for atypical classes but structurally disfavored by the mixture construction. We will state this argument in Section 2 and add a diagnostic in Section 4: we will estimate the effective class and marginal variances along the guidance direction (D_u - D_c) at small sigma for the EDM checkpoints and check whether the ordering holds in practice. If some classes show reverse ordering, we will report that honestly and note that the theory's applicability is partial. We cannot fully verify the hypothesis on learned checkpoints (the Jacobian structure is not Gaussian), but the indirect argument plus the diagnostic gives readers a concrete basis to assess applicability. revision: yes

Circularity Check

1 steps flagged

Central derivation is parameter-free and self-contained; one load-bearing self-citation establishes the unguided framework but is not itself the target result.

specific steps
  1. self citation load bearing [Section 2, paragraph 'The unguided flow'; Section 1, paragraph 'Guidance re-stiffens the discriminative subspace']
    "A fitted-operator analysis of the unguided terminal layer [12] shows that a frozen-field Euler step is layer-exact if and only if its clock is affine in σ, making the σ-clock step—algebraically the deterministic DDIM update [11]—the unique fitted operator up to affine reparameterization... The present paper asks what survives of that protection under guidance."

    Reference [12] (same author, arXiv:2607.04113) establishes the unguided fitted-operator framework: DDIM is the unique layer-exact frozen-field Euler step on the terminal layer. The present paper builds on this to analyze guidance. However, the self-citation is not circular in the target-result sense: [12] proves a result about the UNGUIDED flow (exponent 1, DDIM exactness), while the present paper's central contribution is about the GUIDED flow (exponent 1+w, the coefficient r^(1+w)-r). The guided exponent (Proposition 1, Eq. 5) and the fitted coefficient (Lemma 2) are derived from scratch within this paper from the guided flow equation (2) and the commuting Gaussian model (Hypothesis 1), not imported from [12]. The self-citation provides scaffolding (the fitted-operator concept, the σ-aff

full rationale

The paper's central derivation is parameter-free and self-contained. The coefficient r^(1+w)-r is derived in Lemma 2 by requiring terminal exactness on the pure discriminative layer (A=1, B=0): the exact factor is r^(1+w) (Lemma 1, Eq. 7), the scalar update gives factor r+α, and exactness forces α = r^(1+w)-r. This is a direct algebraic derivation from the model, not a fit to data. The barrier thresholds (Theorem 1) follow analytically from the elementary function G_w(h) = (1+w)e^{-h} - w. Proposition 4's bounded amplification and Theorem 2's first-order accuracy are proven from the exact factor (Lemma 1) and the fitted factor (Proposition 3) without fitting. The one load-bearing self-citation is to [12] (same author), which establishes the unguided fitted-operator framework. This is a dependency on prior work for scaffolding, but [12]'s result concerns the unguided flow (exponent 1, DDIM exactness), while the present paper's target result concerns the guided flow (exponent 1+w, the new coefficient). The guided exponent and the fitted coefficient are derived within this paper from the guided flow equation and the commuting model, not imported from [12]. The self-citation is real but not circular: it provides the conceptual framework, not the target result. Score 2: one minor self-citation that is not load-bearing for the paper's central claim.

Axiom & Free-Parameter Ledger

1 free parameters · 4 axioms · 0 invented entities

The paper introduces no new physical entities, particles, forces, or dimensions. The 'guided fitted step' is a mathematical construction derived from existing objects. The only free parameter is the practitioner-set guidance weight w. The key axioms are the commuting Gaussian model (Hypothesis 1) and the VE parameterization, both domain assumptions. The dependency on [12] for the unguided fitted-operator framework is an ad-hoc-to-paper axiom, though that result is parameter-free.

free parameters (1)
  • w (guidance weight) = set by practitioner (e.g., 6.5, 8.0)
    The guidance scale w is an input from the practitioner, not fitted by the paper. The repair coefficient r^(1+w)-r is a function of w and r, with no new free parameters introduced.
axioms (4)
  • domain assumption Hypothesis 1: Commuting class-marginal pair with a_i ≤ b_i (class-subset hypothesis)
    Section 2. The entire analytical framework is built on this Gaussian model. Real learned checkpoints need not satisfy it; the reverse case is a genuine model boundary (Remark 4).
  • domain assumption Variance-exploding (VE) family: p_sigma = p_data * N(0, sigma^2 I)
    Section 2. The analysis is specific to the VE parameterization. Extension to variance-preserving or other SDE types is not addressed.
  • ad hoc to paper The unguided DDIM step is the unique fitted operator for the terminal layer (from [12])
    Section 1 and Section 2. The paper's central framework depends on this result from reference [12] (same author). That result is a parameter-free derivation, but its applicability to guided flows is the paper's extension.
  • standard math D_w is not a VE posterior mean with PSD Tweedie covariance
    Section 2. This follows from the linear algebra of the extrapolation D_w = (1+w)D_c - wD_u and the Tweedie formula (Efron [4]). It is the reason the unguided rigidity theorem does not apply to the guided field.

pith-pipeline@v1.1.0-glm · 19286 in / 3190 out tokens · 308598 ms · 2026-07-09T03:02:17.071453+00:00 · methodology

0 comments
read the original abstract

Classifier-free guidance (CFG) is the standard way to strengthen class-conditioning in diffusion and flow-matching samplers, yet at large guidance it oversaturates and destabilizes, symptoms practitioners suppress with more steps or limited-interval schedules. We analyze CFG through an asymptotic-preserving, numerical-analysis lens. Building on a recent result that the deterministic DDIM step is the unique fitted operator for the unguided terminal layer, exact on the final small-sigma stretch of sampling, we show that guidance re-stiffens exactly the discriminative subspace to an anomalous exponent 1+w. DDIM is therefore no longer fitted there, and on coarse meshes its guided residual diverges as sigma_min goes to zero. We prove a guided clock barrier with three ordered step-size thresholds, and read one-step oversaturation as its endpoint: a solver artifact on the calibration model rather than the continuous guided law. The same analysis yields a one-coefficient, zero-extra-NFE repair: replace CFG's w(r-1) by r^(1+w)-r on the guidance direction. On the calibration model's discriminative crossover, this removes CFG's sigma_min-divergent blow-up and is first-order accurate against the exact guided flow as sigma_min goes to zero. On learned CIFAR-10 checkpoints, and as a cross-domain smoke test on Stable Diffusion 1.5 DDIM, it acts as a high-guidance stabilizer at no extra cost rather than a universal quality knob: it cuts residual amplification and saturation, gives 9/9 point-FID wins over CFG on the tested grid, and preserves classifier-proxy target accuracy in the hard-cell blocks. We report the limits alongside: it is not a universal image-quality win, and against a dense vanilla-CFG reference it is not a uniformly better integrator of that field.

Figures

Figures reproduced from arXiv: 2607.07665 by Shiheng Zhang.

Figure 1
Figure 1. Figure 1: The guided terminal exponent µw(σ) by layer type on the commuting model (schematic; noise decreases to the right). Guidance leaves the shared-normal exponent at one and freezes tan￾gential directions, but re-stiffens the discriminative directions to 1 + w—the subspace on which DDIM+CFG exits the class of fitted operators. Here a, b are the class and marginal variances along the direction (Section 2). 1 + w… view at source ↗
Figure 2
Figure 2. Figure 2: Synthetic crossover at w = 8: residual amplification as σmin → 0. CFG develops the σmin-divergent branch of Theorem 1(b) on coarse meshes (N ∈ {8, 16, 32}), stabilizing only once h < h♯—the flat gray curve (N=64) coincides with the exact flow. The fitted step (dashed) stays below the 1 + w tax on every mesh. Hence for every prefix k, |ηw(ξk, σk)| |ηw(ξ0, σ0)| ≤ µk µ0 ≤ 1 + w. (19) Thus the fitted step remo… view at source ↗
Figure 3
Figure 3. Figure 3: Uncurated hard-cell samples (w=8, N=8; first ten CIFAR-10 classes, matched seeds). Rows: CFG, fitted CFG, terminal-interval. CFG’s row is visibly more saturated and higher-contrast; the repair and interval are lower-contrast at zero extra NFE. Qualitative only—KID still favors CFG. Finally, the target-accuracy column of [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

12 extracted references · 12 canonical work pages · 9 internal anchors

  1. [1]

    Classifier-Free Guidance is a Predictor-Corrector

    Arwen Bradley and Preetum Nakkiran. Classifier-free guidance is a predictor-corrector.Trans- actions on Machine Learning Research (TMLR), 2025. arXiv:2408.09000

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  2. [2]

    What does guidance do? A fine-grained analysis in a simple setting

    Muthu Chidambaram, Khashayar Gatmiry, Sitan Chen, Holden Lee, and Jianfeng Lu. What does guidance do? a fine-grained analysis in a simple setting. InAdvances in Neural Informa- tion Processing Systems (NeurIPS), 2024. arXiv:2409.13074

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  3. [3]

    CFG++: Manifold-constrained Classifier Free Guidance for Diffusion Models

    Hyungjin Chung, Jeongsol Kim, Geon Yeong Park, Hyelin Nam, and Jong Chul Ye. CFG++: Manifold-constrained classifier free guidance for diffusion models. InInternational Confer- ence on Learning Representations (ICLR), 2025. arXiv:2406.08070

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  4. [4]

    Tweedie’s formula and selection bias.Journal of the American Statistical Association, 106(496):1602–1614, 2011

    Bradley Efron. Tweedie’s formula and selection bias.Journal of the American Statistical Association, 106(496):1602–1614, 2011

    work page 2011

  5. [5]

    CFG-Zero*: Improved Classifier-Free Guidance for Flow Matching Models

    Weichen Fan, Amber Yijia Zheng, Raymond A. Yeh, and Ziwei Liu. CFG-Zero*: Improved classifier-free guidance for flow matching models.arXiv preprint arXiv:2503.18886, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  6. [6]

    C$^2$FG: Control Classifier-Free Guidance via Score Discrepancy Analysis

    Jiayang Gao, Tianyi Zheng, Jiayang Zou, Fengxiang Yang, Shice Liu, Luyao Fan, Zheyu Zhang, Hao Zhang, Jinwei Chen, Peng-Tao Jiang, Bo Li, and Jia Wang. C 2FG: Control classifier-free guidance via score discrepancy analysis. InProceedings of the IEEE/CVF 10 Conference on Computer Vision and Pattern Recognition (CVPR), pp. 34398–34407, 2026. arXiv:2603.08155

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  7. [7]

    Classifier-Free Diffusion Guidance

    Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance.arXiv preprint arXiv:2207.12598, 2022

    work page internal anchor Pith review Pith/arXiv arXiv 2022

  8. [8]

    Elucidating the design space of diffusion-based generative models

    Tero Karras, Miika Aittala, Timo Aila, and Samuli Laine. Elucidating the design space of diffusion-based generative models. InAdvances in Neural Information Processing Systems (NeurIPS), 2022

    work page 2022

  9. [9]

    Applying Guidance in a Limited Interval Improves Sample and Distribution Quality in Diffusion Models

    Tuomas Kynk ¨a¨anniemi, Miika Aittala, Tero Karras, Samuli Laine, Timo Aila, and Jaakko Lehtinen. Applying guidance in a limited interval improves sample and distribution quality in diffusion models. InAdvances in Neural Information Processing Systems (NeurIPS), 2024. arXiv:2404.07724

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  10. [10]

    Seyedmorteza Sadat, Otmar Hilliges, and Romann M. Weber. Eliminating oversaturation and artifacts of high guidance scales in diffusion models. InInternational Conference on Learning Representations (ICLR), 2025. arXiv:2410.02416

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  11. [11]

    Denoising diffusion implicit models

    Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In International Conference on Learning Representations (ICLR), 2021

    work page 2021

  12. [12]

    Asymptotic-Preserving A Posteriori Analysis of Diffusion and Flow-Matching Samplers

    Shiheng Zhang. Asymptotic-preserving a posteriori analysis of diffusion and flow-matching samplers.arXiv preprint arXiv:2607.04113, 2026. A DEFERRED PROOFS Proof of Proposition 1.Fora= 0one hasA≡1, soµ w = (1 +w)−wB= 1 +w(1−B) = 1+wb/(b+σ 2); the limits and monotonicity are immediate. Cases (i) and (ii) follow fromA, B→1 andA, B→0respectively. Proof of Le...

    work page internal anchor Pith review Pith/arXiv arXiv 2026