Pith. sign in

REVIEW 4 major objections 4 minor 2 cited by

Conditional Diffusion Models with Classifier-Free Gibbs-like Guidance

T0 review · 4 major / 4 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read Classifier-free guidance is not a valid diffusion sampler: the CFG score misses the gradient of a Rényi divergence, a repulsive term that vanishes only at low noise, so the paper corrects it with a Gibbs-like refinement.

desk verdict Correct and useful score decomposition, but the implemented CFG IG sampler's target distribution is not proven; the paper warrants peer review with a request to close or caveat that gap. read the letter →

arxiv 2505.21101 v1 pith:BLPD7SZH submitted 2025-05-27 cs.LG stat.ME

Badr Moufad , Yazid Janati , Alain Durmus , Ahmed Ghorbel , Eric Moulines , Jimmy Olsson This is my paper
classification cs.LGstat.ME
keywords classifier-freeguidancediffusionmodelsRényidivergencesamplediversitytilteddistributionGibbssamplingconditionalgenerationtext-to-audio
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

Classifier-free guidance (CFG) is widely used to sharpen conditional diffusion outputs, but this paper argues that the CFG denoiser is generally not the denoiser of any denoising diffusion model. The distribution CFG appears to aim at—the data distribution tilted by a power w>1 of the conditional density—has a score with an extra term: the gradient of a Rényi divergence between the conditional and unconditional posteriors. This term acts as a repulsive force that preserves diversity, and it disappears only as the noise level approaches zero. The paper proposes CFG IG, a Gibbs-like chain that alternately noises a sample up to a small level sigma* and denoises it with CFG, and shows on ImageNet and AudioCaps that this recovers diversity while matching or improving quality metrics.

What carries the argument

The load-bearing object is the score decomposition in Proposition 1, ∇log pi_sigma(x|c;w) = (w−1)∇R_sigma(x,c;w) + ∇log pi^cfg_sigma(x|c;w), where R_sigma(x,c;w) = R_w(pi_{0|$\sigma$}(·|x,c) ∥ p_{0|$\sigma$}(·|x)) is the Rényi divergence of order w between the conditional and unconditional posterior distributions of the clean data at noise level $\sigma$. This turns the vague complaint that CFG breaks the diffusion process into a specific missing force: the gradient of the divergence, which pushes samples away from collapsed regions. Around it sit the vanishing result ∇R_sigma = O(sigma²) as $\sigma$ → 0, which justifies dropping the term at low noise, and the Gibbs-like update using the PF-ODE flow map as the transition kernel.

What would settle it

Run CFG IG on a one-dimensional mixture of two well-separated Gaussians with known conditional densities and an exact ODE solver, and measure the stationary distribution of the chain for a fixed small sigma*. If samples remain over-concentrated on one mode for many Gibbs rounds while the true tilted distribution has substantial mass on the other mode, the claim that the CFG approximation preserves diversity at small sigma* would be refuted; the paper's Gaussian analysis predicts the opposite, with bias of order sigma*².

Watch

Extended reading notes

Core claim

On the paper's own terms, the central discovery is that the score of the tilted target pi_0(·|c;w) ∝ g_0(c|·)^w p_0 decomposes as ∇log pi_sigma(x|c;w) = (w−1)∇R_sigma(x,c;w) + ∇log pi^cfg_sigma(x|c;w), where R_sigma is the order-w Rényi divergence between conditional and unconditional posteriors at noise level $\sigma$. The CFG denoiser keeps only the second term, so it is not a valid denoiser for the tilted distribution except in the low-noise limit, where the Rényi gradient is O(sigma²). The constructed Gibbs-like chain $X_0^{{(r+1)}}$ = F_{0|$\sigma$*}($X_0^{{(r)}}$ + $\sigma$* $Z^{{(r+1)}}$; w), with the ideal denoiser, has the tilted distribution as its unique stationary distribution, and replacing the ideal denoiser by the CFG denoiser at a small fixed $\sigma$* yields the practical CFG IG algorithm, which is shown in a Gaussian example to converge to the target up to an O($\sigma$*²) bias.

Load-bearing premise

The practical claim rests on assuming that the error from dropping the Rényi repulsive term stays small all along the ODE integration path, and that a small fixed noise level sigma* keeps the noising-denoising chain close to the ideal Gibbs chain; the paper proves the smallness pointwise under smoothness assumptions and proves the chain result only for a scalar Gaussian example.

Editorial extensions

If this is right

  • The diversity loss reported for CFG is not a side effect of high guidance but the signature of a missing repulsive term that is active at medium to high noise levels.
  • CFG IG turns guidance into a controllable noising-denoising refinement: initial diversity comes from a weakly guided sample, and quality is sharpened by a few Gibbs rounds, so sigma* and the number of repetitions are explicit knobs for the trade-off.
  • The two-noise-level score expression in Proposition 4 provides a theoretical grounding for guidance schemes that combine conditional denoisers evaluated at different noise levels without using an unconditional model.
  • In the ideal-denoiser case the chain has the tilted distribution as its unique stationary distribution, and the Gaussian example quantifies the O(sigma*²) bias introduced by the CFG approximation.
  • The same decomposition suggests training objectives that explicitly include the Rényi term, which the paper identifies as a route to guidance that stays consistent over the full noise range.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • A direct stress test would be to run CFG IG on a multimodal ground-truth model where the tilted target can be computed exactly and then compare the empirical stationary distribution with that target; the paper's Gaussian analysis predicts only an O(sigma*²) bias, while a failure to escape the initial mode would refute the diversity claim.
  • The noising-denoising refinement is structurally the SDEdit procedure, suggesting the same Gibbs correction could be applied to image editing and inverse problems, with sigma* controlling how much of the original structure is preserved.
  • Because the correction acts repulsively where conditional and unconditional posteriors disagree, one could estimate it with a learned density-ratio or posterior classifier and use it to correct CFG at all noise levels rather than only at low noise.
  • The two-noise-level identity may also yield a new score-estimation identity for unconditional sampling, potentially useful when only a conditional model is available.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 4 minor

Summary. The paper studies classifier-free guidance (CFG) in conditional diffusion models. It first argues, via Example 1, that the CFG denoiser does not generally define a valid denoising diffusion model targeting the tilted distribution π0(·|c;w) ∝ π0(·|c)^w p0(·)^{1−w}; Proposition 1 identifies the missing term as the gradient of a Rényi divergence, and Proposition 2 shows this term is O(σ²) pointwise in the low-noise limit. The paper then proposes CFG IG, an iterative noising–denoising procedure that alternates adding noise σ* and running the PF-ODE with the CFG denoiser, and analyzes it exactly in a scalar Gaussian case (Proposition 3). Experiments on ImageNet-512 and AudioCaps report improved FID/FDDINOv2/FAD and better diversity metrics relative to CFG, LI-CFG, and CFG++.

Significance. Proposition 1 and Example 1 are clean, self-contained results that clarify why CFG is not a valid DDM: the derivation of (3.2) follows directly from the Radon–Nikodym derivative of the tilted posterior, and Example 1 gives a rigorous counterexample. Proposition 4 is a useful two-noise-level score identity that connects to recent guidance methods. The paper also ships code, reports ablations over w, σ*, and R, and follows standard evaluation protocols. The key weakness is that the central claim about CFG IG is not backed by a theorem: the O(σ²) Rényi-term estimate is pointwise, the only exact analysis is a scalar Gaussian with c=0, and no perturbation result links the ideal Gibbs chain to the implemented CFG-denoiser chain. If those gaps are filled, or the claims appropriately softened, the paper would be a valuable contribution; in its current form, the theoretical and practical claims are not fully aligned.

major comments (4)
  1. [Section 3, Algorithm 1 and Proposition 2] The central practical claim that CFG IG 'generates approximate samples from π0(·|c;w)' is not established by the stated results. Algorithm 1 replaces the ideal denoiser Dσ(·|c;w) with Dcfg_σ(·|c;w) in every Gibbs iteration, but Proposition 2 only shows, pointwise in x and c, that ∇Rσ(x,c;w)=O(σ²) as σ→0 under Q3 smoothness. This pointwise estimate does not imply a bound on the difference between the corresponding PF-ODE flow maps over [0,σ*], nor a perturbation bound on the stationary distribution of the implemented CFG-denoiser chain. Without a uniform-in-x error bound and an invariant-measure perturbation argument, the claim that the implemented chain approximates the tilted target remains a heuristic; the text currently states it as a conclusion.
  2. [Proposition 3 and Appendix A.3] The only quantitative analysis of the CFG-denoiser chain is restricted to the scalar Gaussian model of Example 1 with c=0, where the PF-ODE is exactly solvable. Proposition 3 shows V∞(w)=V(w)+O(σ*²) in that special case; it says nothing about the general high-dimensional setting with c≠0, where the ODE cannot be solved in elementary form. The general claim therefore rests on the heuristic 'σ* small enough' statement in Section 3, not on a theorem. The authors should either extend the perturbation analysis to a broader class or explicitly present the Gaussian case as an illustration rather than support for the general claim.
  3. [Section 5, Tables 1 and 2] The empirical metrics (FID, FDDINOv2, FAD, Precision/Recall, Density/Coverage) measure closeness to the data distribution, not closeness to the tilted distribution π0(·|c;w). A sampler that targets an over-concentrated distribution (e.g., plain CFG with a tuned scale) can score well on these metrics, so the reported gains do not test whether CFG IG samples from the intended target. The paper should either add a diagnostic that compares empirical samples against the tilted target (e.g., in a setting where the target is known) or soften the claim that the method 'generates approximate samples from' the tilted distribution.
  4. [Appendix A.3, Proposition 5 and Assumption A2] The uniqueness of the stationary distribution is established only for the ideal Gibbs chain under Assumption A2, which requires the ideal flow map F0|σ* to be a continuously differentiable diffeomorphism. This assumption is unlikely to hold for practical neural-network denoisers, and no counterpart is given for the implemented CFG-denoiser chain. This does not invalidate the method, but the phrase 'we can verify that the associated Markov chain admits π0(·|c;w) as its unique stationary distribution' in Section 3 overstates what is proved.
minor comments (4)
  1. [Section 3] The word 'refered' should be 'referred'.
  2. [Section 3, Proposition 2] The statement 'Under suitable assumptions on p0 and g0' is vague; the Q3 assumptions used in the proof should be stated in the main text.
  3. [Appendix B] The phrase 'An simple derivation' should be 'A simple derivation'.
  4. [Throughout] The accent in 'Rényi' is inconsistently rendered (the main text uses 'R' with an accent in some places and plain 'Renyi' in others, e.g., in the Proposition 1 statement).

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the theoretical identities are algebraic derivations, the counterexample is self-contained, and the empirical claims are benchmarked rather than predicted from fitted constants.

full rationale

The paper's derivation chain is self-contained. Proposition 1 is a direct Radon–Nikodym computation expressing the tilted score as the CFG score plus the gradient of a Rényi divergence; the target tilted distribution is defined before any approximation, and the decomposition is not used to redefine the target. Example 1 is a standalone Gaussian counterexample showing that no valid denoising diffusion model has the CFG marginals for all noise levels. Proposition 2 follows from Taylor expansions under the stated Q3 smoothness assumptions, and Proposition 3 is an explicit scalar Gaussian calculation with c=0, so neither reduces to the claim it supports. Proposition 4 is another algebraic score identity. Algorithm 1's practical claim that the implemented CFG-based chain approximates the tilted distribution is not fully established for neural denoisers—Proposition 2 is only pointwise in x and c, and Proposition 3 covers only a scalar Gaussian—but this is a rigor gap about uniformity and perturbation of invariant measures, not a circularity: no target distribution, denoiser, or metric is defined in terms of the algorithm's outputs, and no fitted hyperparameter is relabeled as a prediction. The paper also does not rely on load-bearing self-citations; the uniqueness argument uses the external Meyn and Tweedie reference, and the SDEdit connection is an acknowledged precursor rather than a smuggled ansatz. Overall, the central contributions have independent mathematical content, and the empirical evaluations are external benchmarks. Thus no specific circular step can be exhibited, and the appropriate score is 0.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

The main mathematical claims rest on smoothness and absolute-continuity assumptions (A1 and Q3) that exclude realistic low-dimensional image and audio distributions, and on a diffeomorphism assumption (A2) used only for the theoretical uniqueness proof. The empirical headline numbers additionally depend on five hand-tuned hyperparameters.

free parameters (5)
  • initial guidance scale w0 = 1 (EDM2-S), 1 (EDM2-XXL), 1.5 (AudioLDM2)
    Chosen by grid search; controls diversity of the initial sample and affects all reported image and audio metrics.
  • guidance scale w = 2.3 (EDM2-S), 2 (EDM2-XXL), 5 (AudioLDM2)
    Main CFG guidance scale, grid-searched to optimize FID for images and FAD for audio.
  • Gibbs noise level sigma* = 2 (EDM2-S), 1 (EDM2-XXL), 5 (AudioLDM2)
    Fixed noise level used in the noising and denoising loop; the method's bias-variance tradeoff hinges on this value.
  • number of Gibbs repetitions R = 2 (ImageNet and audio main results), 1 in an audio ablation
    Grid-searched; increasing repetitions beyond 2 worsens scores at fixed NFE, and the optimal value depends on the application.
  • initial ODE steps T0 = 12 (ImageNet), 100 (AudioLDM2)
    Allocation of the function-evaluation budget to generating the initial sample; chosen by grid search.
assumptions (4)
  • domain assumption A1: g0(c|x0) >= 0 and 0 < Z(c) < infinity for all c and x0
    Ensures the tilted distribution pi0(·|c;w) is a well-defined probability measure; introduced in Appendix A.1 before Proposition 1.
  • domain assumption Absolute continuity and Q3 smoothness of p0, g0(c|·), and g0(c|·)^w
    The O(sigma^2) vanishing of the Rényi term is proven only under strict positivity, three-times differentiability, and polynomial growth, which real image and audio distributions are not known to satisfy.
  • ad hoc to paper A2: the ideal flow map F0|sigma*(z|c;w) is a continuously differentiable diffeomorphism
    Used to establish uniqueness of the stationary distribution of the ideal Gibbs chain; stated as a simplifying assumption and unlikely to hold for practical neural-network denoisers.
  • domain assumption The desired output is the tilted distribution pi0(·|c;w) proportional to g0(c|·)^w p0
    This target is a modeling choice inherited from CFG intuition; the paper does not derive why this particular tilted distribution is the right objective.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Conditional Diffusion Models with Classifier-Free Gibbs-like Guidance." pith.science (2026). https://pith.science/paper/BLPD7SZH

@misc{pith2026250521101,
  author       = {Pith},
  title        = {Pith review of: Conditional Diffusion Models with Classifier-Free Gibbs-like Guidance},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/BLPD7SZH}},
  note         = {Machine review of arXiv:2505.21101}
}
abstract

Classifier-Free Guidance (CFG) is a widely used technique for improving conditional diffusion models by linearly combining the outputs of conditional and unconditional denoisers. While CFG enhances visual quality and improves alignment with prompts, it often reduces sample diversity, leading to a challenging trade-off between quality and diversity. To address this issue, we make two key contributions. First, CFG generally does not correspond to a well-defined denoising diffusion model (DDM). In particular, contrary to common intuition, CFG does not yield samples from the target distribution associated with the limiting CFG score as the noise level approaches zero -- where the data distribution is tilted by a power $w \gt 1$ of the conditional distribution. We identify the missing component: a R\'enyi divergence term that acts as a repulsive force and is required to correct CFG and render it consistent with a proper DDM. Our analysis shows that this correction term vanishes in the low-noise limit. Second, motivated by this insight, we propose a Gibbs-like sampling procedure to draw samples from the desired tilted distribution. This method starts with an initial sample from the conditional diffusion model without CFG and iteratively refines it, preserving diversity while progressively enhancing sample quality. We evaluate our approach on both image and text-to-audio generation tasks, demonstrating substantial improvements over CFG across all considered metrics. The code is available at https://github.com/yazidjanati/cfgig

Figures

Figures reproduced from arXiv: 2505.21101 by the authors.

Figure 1
Figure 1. Illustration of sample refinement across Gibbs iterations. Left, samples generated using EDM-XXL [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Left: DDIM sampling with CFG denoiser (2.3); Right: DDIM sampling with the ideal denoiser (3.2). The trajectories of 1000 particles are represented with thin red lines and 5 selected trajectories are displayed in black thick line along which scores are being depicted with arrows. The histogram of the simulated particles is represented in light gray. We also plot the ideal score (3.2) (arrow in black) with the contri… view at source ↗
Figure 3
Figure 3. Impact of the hyperparameters for the EDM2- [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Impact of the hyperparameters on the FAD [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]
Figure 5
Figure 5. Figure 5: Precision and recall as a function of σ∗ for EDM2-S (top) and EDM2-XXL (bottom). Ablations. In [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]
Figure 6
Figure 6. Figure 6: Comparison of sample diversity between CFGIG and CFG across a batch of 6 samples generated using EDM-XXL model for two ImageNet classes: 84 (top) 957 (bottom). Both algorithms are run with a total of 32 Heun steps and the same seed. CFG is run with a guidance scale w =…
Figure 7
Figure 7. Figure 7: The difference |VR(w) − V (w)| as a function of σ∗ for various values of R. Analysis as σ∗ → 0 and R < ∞. To investigate the behavior at finite R, we initialize the recursion (A.11) with X (0) 0 ∼ π0 (·|c) = N (0, V0), where V0 := γ 2/(γ 2 + 1). For any R > 0, the fina…

Discussion (0). Sign in to comment.

Forward citations

Cited by 2 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Signed Rectified Flow: Negativity-Controlled Generation

    cs.LG 2026-07 conditional novelty 6.0 of 10

    Signed Rectified Flow adds a negative branch to flow-based generation by targeting the signed measure (1+α)π+ − απ−, provably avoiding negative regions while preserving the positive density on a reachable subset.

  2. Provable Diffusion Posterior Sampling for Bayesian Inversion

    stat.ML 2025-12 conditional novelty 6.0 of 10

    A diffusion posterior sampler using Monte Carlo Langevin score estimation and warm start is proven to converge in Wasserstein-2 distance under semi-log-concavity and sub-Gaussian assumptions, and outperforms DPS/TV on...

Reference graph

Works this paper leans on

44 extracted references · 19 canonical work pages · cited by 2 Pith papers

  1. [1]

    Align your latents: High-resolution video synthesis with latent diffusion models

    Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your latents: High-resolution video synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 22563–22575, 2023

    2023

  2. [2]

    Classifier-free guidance is a predictor-corrector

    Arwen Bradley and Preetum Nakkiran. Classifier-free guidance is a predictor-corrector. arXiv preprint arXiv:2408.09000, 2024

    arXiv 2024

  3. [3]

    Muse: Text-to-image generation via masked generative transformers

    Huiwen Chang, Han Zhang, Jarred Barber, Aaron Maschinot, Jose Lezama, Lu Jiang, Ming- Hsuan Yang, Kevin Patrick Murphy, William T Freeman, Michael Rubinstein, et al. Muse: Text-to-image generation via masked generative transformers. In International Conference on Machine Learning, pages 4055–4075. PMLR, 2023

    work page 2023

  4. [4]

    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. In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024

    work page 2024

  5. [5]

    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. In The Thirteenth Interna- tional Conference on Learning Representations, 2025

    work page 2025

  6. [6]

    Guidance: a cheat code for diffusion models, 2022

    Sander Dieleman. Guidance: a cheat code for diffusion models, 2022

    work page 2022

  7. [7]

    Sequential Monte Carlo methods in practice, volume 1

    Arnaud Doucet, Nando De Freitas, Neil James Gordon, et al. Sequential Monte Carlo methods in practice, volume 1. Springer, 2001

    work page 2001

  8. [8]

    Gibbs sampling

    Alan E Gelfand. Gibbs sampling. Journal of the American statistical Association, 95(452):1300– 1304, 2000

    work page 2000

Show all 44 references
  1. [9]

    Cnn architectures for large-scale audio classification

    Shawn Hershey, Sourish Chaudhuri, Daniel PW Ellis, Jort F Gemmeke, Aren Jansen, R Chan- ning Moore, Manoj Plakal, Devin Platt, Rif A Saurous, Bryan Seybold, et al. Cnn architectures for large-scale audio classification. In 2017 ieee international conference on acoustics, speec...

    2017

  2. [10]

    Gans trained by a two time-scale update rule converge to a local nash equilibrium

    Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 10

    2017

  3. [11]

    Denoising diffusion probabilistic models

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems, 33:6840–6851, 2020

    2020

  4. [12]

    Classifier-free diffusion guidance

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

    2022 arXiv

  5. [13]

    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. Advances in Neural Information Processing Systems, 35:26565–26577, 2022

    2022

  6. [14]

    Guiding a diffusion model with a bad version of itself

    Tero Karras, Miika Aittala, Tuomas Kynkäänniemi, Jaakko Lehtinen, Timo Aila, and Samuli Laine. Guiding a diffusion model with a bad version of itself. Advances in Neural Information Processing Systems, 37:52996–53021, 2024

    2024

  7. [15]

    Analyzing and improving the training dynamics of diffusion models

    Tero Karras, Miika Aittala, Jaakko Lehtinen, Janne Hellsten, Timo Aila, and Samuli Laine. Analyzing and improving the training dynamics of diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24174–24184, 2024

    2024

  8. [16]

    Audiocaps: Generat- ing captions for audios in the wild

    Chris Dongjoo Kim, Byeongchang Kim, Hyunmin Lee, and Gunhee Kim. Audiocaps: Generat- ing captions for audios in the wild. InProceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 ...

    2019

  9. [17]

    Panns: Large-scale pretrained audio neural networks for audio pattern recognition

    Qiuqiang Kong, Yin Cao, Turab Iqbal, Yuxuan Wang, Wenwu Wang, and Mark D Plumbley. Panns: Large-scale pretrained audio neural networks for audio pattern recognition. IEEE/ACM Transactions on Audio, Speech, and Language Processing, 28:2880–2894, 2020

    2020

  10. [18]

    Diffwave: A versatile diffusion model for audio synthesis

    Zhifeng Kong, Wei Ping, Jiaji Huang, Kexin Zhao, and Bryan Catanzaro. Diffwave: A versatile diffusion model for audio synthesis. arXiv preprint arXiv:2009.09761, 2020

    2009 arXiv

  11. [19]

    Applying guidance in a limited interval improves sample and distribution quality in diffusion models

    Tuomas Kynkäänniemi, 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. In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024

    2024

  12. [20]

    Debiasing guidance for discrete diffusion with sequential Monte Carlo

    Cheuk Kit Lee, Paul Jeha, Jes Frellsen, Pietro Lio, Michael Samuel Albergo, and Francisco Vargas. Debiasing guidance for discrete diffusion with sequential Monte Carlo. arXiv preprint arXiv:2502.06079, 2025

    2025 arXiv

  13. [21]

    Self-guidance: Boosting flow and diffusion generation on their own

    Tiancheng Li, Weijian Luo, Zhiyang Chen, Liyuan Ma, and Guo-Jun Qi. Self-guidance: Boosting flow and diffusion generation on their own. arXiv preprint arXiv:2412.05827, 2024

    2024

  14. [22]

    Audioldm 2: Learning holistic audio generation with self-supervised pretraining

    Haohe Liu, Yi Yuan, Xubo Liu, Xinhao Mei, Qiuqiang Kong, Qiao Tian, Yuping Wang, Wenwu Wang, Yuxuan Wang, and Mark D Plumbley. Audioldm 2: Learning holistic audio generation with self-supervised pretraining. IEEE/ACM Transactions on Audio, Speech, and Language Processing, 2024

    2024

  15. [23]

    Sdedit: Image synthesis and editing with stochastic differential equations

    Chenlin Meng, Yang Song, Jiaming Song, Jiajun Wu, Jun-Yan Zhu, and Stefano Ermon. Sdedit: Image synthesis and editing with stochastic differential equations. arXiv preprint arXiv:2108.01073, 2021

    2021 arXiv

  16. [24]

    Markov chains and stochastic stability

    Sean P Meyn and Richard L Tweedie. Markov chains and stochastic stability. Springer Science & Business Media, 2012

    2012

  17. [25]

    Reliable fidelity and diversity metrics for generative models

    Muhammad Ferjad Naeem, Seong Joon Oh, Youngjung Uh, Yunjey Choi, and Jaejun Yoo. Reliable fidelity and diversity metrics for generative models. In International conference on machine learning, pages 7176–7185. PMLR, 2020

    2020

  18. [26]

    Dinov2: Learning robust visual features without supervision

    Maxime Oquab, Timothée Darcet, Théo Moutakanni, Huy V o, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023

    2023 arXiv

  19. [27]

    Understanding classifier-free guidance: High-dimensional theory and non-linear generalizations

    Krunoslav Lehman Pavasovic, Jakob Verbeek, Giulio Biroli, and Marc Mezard. Understanding classifier-free guidance: High-dimensional theory and non-linear generalizations. arXiv preprint arXiv:2502.07849, 2025. 11

    2025 arXiv

  20. [28]

    Sdxl: Improving latent diffusion models for high-resolution image synthesis

    Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023

    2023 arXiv

  21. [29]

    Language models are unsupervised multitask learners

    Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. Language models are unsupervised multitask learners. OpenAI blog, 1(8):9, 2019

    2019

  22. [30]

    Herbert E. Robbins. An empirical bayes approach to statistics. In Proceedings of the Third Berkeley Symposium on Mathematical Statistics and Probability, Volume 1: Contributions to the Theory of Statistics, 1956

    1956

  23. [31]

    High- resolution image synthesis with latent diffusion models

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High- resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10684–10695, 2022

    2022

  24. [32]

    Seyedmorteza Sadat, Jakob Buhmann, Derek Bradley, Otmar Hilliges, and Romann M. Weber. CADS: Unleashing the diversity of diffusion models through condition-annealed sampling. In The Twelfth International Conference on Learning Representations, 2024

    2024

  25. [33]

    Seyedmorteza Sadat, Manuel Kansy, Otmar Hilliges, and Romann M. Weber. No training, no problem: Rethinking classifier-free guidance for diffusion models. In The Thirteenth International Conference on Learning Representations, 2025

    2025

  26. [34]

    Optimal transport for applied mathematicians, volume 87

    Filippo Santambrogio. Optimal transport for applied mathematicians, volume 87. Springer, 2015

    2015

  27. [35]

    Feynman- kac correctors in diffusion: Annealing, guidance, and product of experts

    Marta Skreta, Tara Akhound-Sadegh, Viktor Ohanesian, Roberto Bondesan, Alán Aspuru- Guzik, Arnaud Doucet, Rob Brekelmans, Alexander Tong, and Kirill Neklyudov. Feynman- kac correctors in diffusion: Annealing, guidance, and product of experts. arXiv preprint arXiv:2503.02819, 2025

    2025 arXiv

  28. [36]

    Deep unsuper- vised learning using nonequilibrium thermodynamics

    Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsuper- vised learning using nonequilibrium thermodynamics. In International Conference on Machine Learning, pages 2256–2265. PMLR, 2015

    2015

  29. [37]

    Denoising diffusion implicit models

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

    2021

  30. [38]

    Generative modeling by estimating gradients of the data distribution

    Yang Song and Stefano Ermon. Generative modeling by estimating gradients of the data distribution. Advances in neural information processing systems, 32, 2019

    2019

  31. [39]

    Score-based generative modeling through stochastic differential equations

    Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. In International Conference on Learning Representations, 2021

    2021

  32. [40]

    Exposing flaws of generative model evaluation metrics and their unfair treatment of diffusion models

    George Stein, Jesse Cresswell, Rasa Hosseinzadeh, Yi Sui, Brendan Ross, Valentin Villecroze, Zhaoyan Liu, Anthony L Caterini, Eric Taylor, and Gabriel Loaiza-Ganem. Exposing flaws of generative model evaluation metrics and their unfair treatment of diffusion models. Advances i...

    2023

  33. [41]

    Re- thinking the inception architecture for computer vision

    Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jon Shlens, and Zbigniew Wojna. Re- thinking the inception architecture for computer vision. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2818–2826, 2016

    2016

  34. [42]

    Rényi divergence and kullback-leibler divergence

    Tim Van Erven and Peter Harremos. Rényi divergence and kullback-leibler divergence. IEEE Transactions on Information Theory, 60(7):3797–3820, 2014

    2014

  35. [43]

    Analysis of classifier-free guidance weight schedulers

    W ANG Xi, Nicolas Dufour, Nefeli Andreou, CANI Marie-Paule, Victoria Fernandez Abrevaya, David Picard, and Vicky Kalogeiton. Analysis of classifier-free guidance weight schedulers. Transactions on Machine Learning Research

  36. [44]

    number of

    Hu Xu, Juncheng Li, Alexei Baevski, Michael Auli, Wojciech Galuba, Florian Metze, Christoph Feichtenhofer, et al. Masked autoencoders that listen. arXiv e-prints, pages arXiv–2207, 2022. 12 A Proofs A.1 Score expressions In this section, we adopt measure-theoretic notations. D...

    2022

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.