pith. sign in

arxiv: 2507.05482 · v3 · pith:NJEK4LAVnew · submitted 2025-07-07 · 💻 cs.LG · stat.ML

Stein Diffusion Guidance: Training-Free Posterior Correction for Sampling Beyond High-Density Regions

Van Khoa Nguyen , Lionel Blond\'e , Alexandros Kalousis This is my paper

Pith reviewed 2026-05-22 00:09 UTC · model grok-4.3

classification 💻 cs.LG stat.ML
keywords diffusion guidancetraining-free methodsStein variational inferenceposterior correctionlow-density samplingstochastic optimal controlimage generationmolecular docking
0
0 comments X

The pith

Stein Diffusion Guidance corrects approximate posteriors via Stein variational inference to enable reliable sampling in low-density regions without retraining.

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

Standard training-free diffusion guidance relies on Tweedie's formula for posterior approximation but becomes unreliable away from high-density data regions. This paper introduces Stein Diffusion Guidance as a surrogate to stochastic optimal control that uses a new bound on the value function to justify explicit correction. It applies Stein variational inference to find the direction that minimizes Kullback-Leibler divergence between the approximate and true posteriors while employing a novel running cost. Experiments on image guidance and small-ligand protein docking indicate that the resulting method outperforms prior training-free baselines in those low-density settings.

Core claim

The paper establishes a theoretical bound on the stochastic optimal control value function that demonstrates the necessity of correcting approximate posteriors to match true diffusion dynamics, then shows that Stein variational inference supplies the steepest descent direction for minimizing the Kullback-Leibler divergence to the true posterior; combining this Stein correction with a novel running cost functional produces effective training-free guidance beyond high-density regimes.

What carries the argument

Stein correction mechanism that computes the steepest descent direction minimizing KL divergence between approximate and true posteriors, grounded in a surrogate SOC objective and a new bound on the SOC value function.

If this is right

  • SDG enables effective guidance in low-density regions where Tweedie-based approximations fail.
  • The method consistently outperforms standard training-free guidance on image-guidance tasks.
  • It produces better results on small-ligand sampling for protein docking.
  • The framework extends in principle to other posterior sampling problems outside high-density regimes.

Where Pith is reading between the lines

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

  • The correction idea could be tested in conditional generation settings where rare classes or unusual prompts are involved.
  • If the running cost functional proves robust, similar Stein-based adjustments might apply to non-diffusion generative models.
  • The approach may lower the cost of adapting guidance to new domains without classifier retraining.

Load-bearing premise

The new bound on the SOC value function holds and the Stein correction step actually achieves the claimed reduction in KL divergence to the true posterior in practice.

What would settle it

An experiment measuring whether removing the Stein correction step causes measurable degradation in sample quality or posterior alignment specifically in low-density regions, compared against the full SDG method on the same tasks.

Figures

Figures reproduced from arXiv: 2507.05482 by Alexandros Kalousis, Lionel Blond\'e, Van Khoa Nguyen.

Figure 1
Figure 1. Figure 1: SDG provides a computation￾ally efficient alternative to SOC-based diffusion guidance for molecular sam￾pling in low-density regions. In many scientific domains, key discoveries often depend on identifying rare samples buried within large data dis￾tributions. For instance, while billions of molecules ex￾ist in chemistry (Polishchuk et al., 2013), only a minute fraction possesses properties relevant to drug… view at source ↗
Figure 2
Figure 2. Figure 2: Back-and-forth Stein correc￾tion: Particles are mapped backward to MT to obtain posterior samples, which are corrected via Stein correction, and then mapped forward to Mt for reward￾based guidance. Dashed arrows indicate the standard training-free method, while solid arrows denote SDG. We introduce a novel training-free diffusion guidance framework derived from a surrogate stochastic optimal control (SOC) … view at source ↗
Figure 3
Figure 3. Figure 3: Example docking pose of a sampled ligand bound to the jak2 protein receptor. −10 −5 0 DS (jak2) 0.0 0.2 0.4 Density −15 −10 −5 0 DS (fa7) 0.0 0.2 0.4 Density −20 −10 0 DS (5ht1b) 0.0 0.2 0.4 Density data SDG w/o Stein correction SDG [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Distribution of docking scores (lower is better) for generated molecules of SDG with and [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Temporal sampling dynamics of SDG for the jak2 protein. (a) Percentage of molecules [PITH_FULL_IMAGE:figures/full_fig_p008_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Multiple sampling objec￾tives on jak2; SA denotes normalized ˆ synthetic accessibility (SA) scores. In many applications, true (genuine) rewards are computed by non-differentiable oracle functions, which cannot be directly used in training-free diffusion guidance methods. Reward models and classifiers are trained to learn these genuine rewards and produce approximate (nominal) re￾wards, serving as differen… view at source ↗
Figure 7
Figure 7. Figure 7: Ablation results under different low-density levels ( [PITH_FULL_IMAGE:figures/full_fig_p009_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: SDG performance on radar plots: fa7 (left), jak2 (middle), 5ht1b (right). [PITH_FULL_IMAGE:figures/full_fig_p020_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Temporal sampling dynamics of SDG for the fa7 protein. (a) Percentage of molecules [PITH_FULL_IMAGE:figures/full_fig_p021_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Temporal sampling dynamics of SDG for the jak2 protein. (a) Percentage of molecules [PITH_FULL_IMAGE:figures/full_fig_p021_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Temporal sampling dynamics of SDG for the 5ht1b protein. (a) Percentage of molecules [PITH_FULL_IMAGE:figures/full_fig_p021_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Ablation results under different low-density levels ( [PITH_FULL_IMAGE:figures/full_fig_p023_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: Ablation results under different low-density levels ( [PITH_FULL_IMAGE:figures/full_fig_p023_13.png] view at source ↗
Figure 14
Figure 14. Figure 14: Ablation results under different low-density levels ( [PITH_FULL_IMAGE:figures/full_fig_p023_14.png] view at source ↗
Figure 15
Figure 15. Figure 15: Visualization of image deblurring results: SDG without Stein correction (Left) vs. SDG [PITH_FULL_IMAGE:figures/full_fig_p025_15.png] view at source ↗
Figure 16
Figure 16. Figure 16: Visualization of image super-resolution results: SDG without Stein correction (Left) vs. [PITH_FULL_IMAGE:figures/full_fig_p025_16.png] view at source ↗
Figure 17
Figure 17. Figure 17: Visualization of docking poses for multiple generated ligands bound to the jak2 protein. [PITH_FULL_IMAGE:figures/full_fig_p026_17.png] view at source ↗
Figure 18
Figure 18. Figure 18: Visualization of docking poses for multiple generated ligands bound to the fa7 protein. [PITH_FULL_IMAGE:figures/full_fig_p027_18.png] view at source ↗
Figure 19
Figure 19. Figure 19: Visualization of docking poses for multiple generated ligands bound to the 5ht1b protein. [PITH_FULL_IMAGE:figures/full_fig_p028_19.png] view at source ↗
read the original abstract

Training-free diffusion guidance offers a flexible framework for leveraging off-the-shelf classifiers without additional training. Yet, current approaches hinge on posterior approximations via Tweedie's formula, which often yield unreliable guidance, particularly in low-density regions. Stochastic optimal control (SOC), in contrast, enables principled posterior sampling but remains computationally prohibitive for efficient inference. In this work, we reconcile the strengths of these paradigms by introducing Stein Diffusion Guidance (SDG), a novel training-free framework grounded in a surrogate SOC objective. We establish a new theoretical bound on the SOC value function, revealing the necessity of correcting approximate posteriors to reflect true diffusion dynamics. Building on Stein variational inference, SDG computes the steepest descent direction that minimizes the Kullback-Leibler divergence between approximate and true posteriors. By integrating a principled Stein correction mechanism along with a novel running cost functional, SDG enables effective guidance in low-density regions. Our experiments on diverse image-guidance tasks and on challenging small-ligand sampling for protein docking suggest that SDG consistently outperforms standard training-free guidance methods and highlights its potential for broader posterior sampling problems beyond high-density regimes.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

1 major / 2 minor

Summary. The manuscript introduces Stein Diffusion Guidance (SDG), a training-free framework for diffusion posterior sampling that reconciles Tweedie-style approximations with stochastic optimal control (SOC) via a surrogate SOC objective and Stein variational inference. It claims a new theoretical bound on the SOC value function that necessitates Stein correction of approximate posteriors to match true diffusion dynamics, introduces a novel running cost functional, and reports consistent outperformance on image-guidance tasks and small-ligand protein docking, particularly in low-density regions.

Significance. If the bound and the resulting Stein descent direction are valid, the work offers a principled way to extend training-free guidance beyond high-density regimes without retraining, which is valuable for applications such as molecular docking. Credit is due for the explicit connection between SOC and Stein VI, the novel running cost, and the empirical evaluation on a challenging docking benchmark.

major comments (1)
  1. [§3.2, Theorem 1] §3.2, Theorem 1 (the bound on the SOC value function): the derivation assumes the surrogate objective and the novel running cost satisfy conditions that allow the inequality to hold even when the Tweedie posterior is inaccurate in low-density regions; however, the proof sketch does not explicitly verify that the bound remains non-vacuous under the same conditions used in the experiments, which is load-bearing for the claim that Stein correction is required to match true diffusion dynamics.
minor comments (2)
  1. [§4.1] §4.1: the definition of the novel running cost functional could be stated more explicitly with its dependence on the diffusion time t and the guidance signal.
  2. [Figure 3] Figure 3: the low-density region masks used for quantitative evaluation are not described in sufficient detail to allow exact reproduction.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for their constructive and positive review. The feedback identifies a valuable opportunity to strengthen the presentation of the theoretical bound. We address the single major comment below and will revise the manuscript to improve clarity while preserving the core claims.

read point-by-point responses

  1. Referee: [§3.2, Theorem 1] §3.2, Theorem 1 (the bound on the SOC value function): the derivation assumes the surrogate objective and the novel running cost satisfy conditions that allow the inequality to hold even when the Tweedie posterior is inaccurate in low-density regions; however, the proof sketch does not explicitly verify that the bound remains non-vacuous under the same conditions used in the experiments, which is load-bearing for the claim that Stein correction is required to match true diffusion dynamics.

    Authors: We appreciate the referee's close examination of the proof. Theorem 1 establishes the bound under the stated assumptions on the surrogate SOC objective and the novel running cost functional; these assumptions are formulated precisely so that the inequality remains valid even when the Tweedie approximation is inaccurate in low-density regions. The running cost is constructed to ensure the bound stays informative rather than vacuous. That said, we agree that the current proof sketch would benefit from an explicit verification step confirming non-vacuousness under the precise conditions of the experiments. In the revised version we will expand §3.2 with a short remark (or corollary) that directly checks the relevant conditions for the low-density regime, thereby reinforcing why the Stein correction is necessary to recover the true diffusion dynamics. revision: yes

Circularity Check

0 steps flagged

No significant circularity detected in derivation chain.

full rationale

The paper introduces a new theoretical bound on the SOC value function and a Stein correction derived from variational inference principles. These steps are presented as independent mathematical results that justify the surrogate objective and guidance direction, without reducing by construction to fitted parameters, self-definitions, or prior self-citations that carry the central claim. The derivation chain relies on external SOC and Stein VI foundations rather than renaming or smuggling ansatzes from the authors' own prior unverified work. The method's performance claims rest on this bound and the running cost functional, which do not collapse to tautological inputs in the provided text.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 1 invented entities

The central claim depends on the existence of a theoretical bound linking SOC value functions to the need for posterior correction and on the assumption that Stein variational inference can be applied along the diffusion trajectory without introducing new fitted parameters or violating the diffusion dynamics.

axioms (1)
  • domain assumption Approximate posteriors obtained via Tweedie's formula require explicit correction to reflect true diffusion dynamics in low-density regions
    This premise is invoked to justify the introduction of the Stein correction and is presented as revealed by the new theoretical bound on the SOC value function.
invented entities (1)
  • Stein Diffusion Guidance (SDG) framework no independent evidence
    purpose: Training-free posterior correction mechanism for diffusion sampling
    New method introduced to reconcile training-free guidance with SOC principles; no independent evidence outside the paper is provided.

pith-pipeline@v0.9.0 · 5727 in / 1418 out tokens · 49891 ms · 2026-05-22T00:09:46.293232+00:00 · methodology

discussion (0)

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

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

49 extracted references · 49 canonical work pages · 2 internal anchors

  1. [1]

    Fast, accurate, and reliable molecular docking with quickvina 2

    Amr Alhossary, Stephanus Daniel Handoko, Yuguang Mu, and Chee-Keong Kwoh. Fast, accurate, and reliable molecular docking with quickvina 2. Bioinformatics, 31 0 (13): 0 2214--2216, 2015

    work page 2015

  2. [2]

    An optimal control perspective on diffusion-based generative modeling

    Julius Berner, Lorenz Richter, and Karen Ullrich. An optimal control perspective on diffusion-based generative modeling. Transactions on Machine Learning Research, 2024. ISSN 2835-8856. URL https://openreview.net/forum?id=oYIjw37pTP

    work page 2024

  3. [3]

    Diffusion posterior sampling for general noisy inverse problems

    Hyungjin Chung, Jeongsol Kim, Michael Thompson Mccann, Marc Louis Klasky, and Jong Chul Ye. Diffusion posterior sampling for general noisy inverse problems. In The Eleventh International Conference on Learning Representations, 2023. URL https://openreview.net/forum?id=OnD9zGAGT0k

    work page 2023

  4. [4]

    Mean field simulation for monte carlo integration

    Pierre Del Moral. Mean field simulation for monte carlo integration. Monographs on Statistics and Applied Probability, 126 0 (26): 0 6, 2013

    work page 2013

  5. [5]

    Diffusion models beat gans on image synthesis

    Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34: 0 8780--8794, 2021

    work page 2021

  6. [6]

    Representative guidance: Diffusion model sampling with coherence

    Anh-Dung Dinh, Daochang Liu, and Chang Xu. Representative guidance: Diffusion model sampling with coherence. In The Thirteenth International Conference on Learning Representations, 2025. URL https://openreview.net/forum?id=gWgaypDBs8

    work page 2025

  7. [7]

    Carles Domingo-Enrich, Jiequn Han, Brandon Amos, Joan Bruna, and Ricky T. Q. Chen. Stochastic optimal control matching. In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024. URL https://openreview.net/forum?id=wfU2CdgmWt

    work page 2024

  8. [8]

    Carles Domingo-Enrich, Michal Drozdzal, Brian Karrer, and Ricky T. Q. Chen. Adjoint matching: Fine-tuning flow and diffusion generative models with memoryless stochastic optimal control. In The Thirteenth International Conference on Learning Representations, 2025. URL https://openreview.net/forum?id=xQBRrtQM8u

    work page 2025

  9. [9]

    Controlled Markov processes and viscosity solutions, volume 25

    Wendell H Fleming and Halil Mete Soner. Controlled Markov processes and viscosity solutions, volume 25. Springer Science & Business Media, 2006

    work page 2006

  10. [10]

    Gradient guidance for diffusion models: An optimization perspective

    Yingqing Guo, Hui Yuan, Yukang Yang, Minshuo Chen, and Mengdi Wang. Gradient guidance for diffusion models: An optimization perspective. In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024. URL https://openreview.net/forum?id=X1QeUYBXke

    work page 2024

  11. [11]

    Efficient rare event simulation by optimal nonequilibrium forcing

    Carsten Hartmann and Christof Sch \"u tte. Efficient rare event simulation by optimal nonequilibrium forcing. Journal of Statistical Mechanics: Theory and Experiment, 2012 0 (11): 0 P11004, 2012

    work page 2012

  12. [12]

    Hazan and S

    Elad Hazan and Sham Kakade. Revisiting the polyak step size. arXiv preprint arXiv:1905.00313, 2019

    work page arXiv 1905

  13. [13]

    Denoising diffusion probabilistic models

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33: 0 6840--6851, 2020

    work page 2020

  14. [14]

    A simulation-free deep learning approach to stochastic optimal control

    Mengjian Hua, Matthieu Lauri \`e re, and Eric Vanden-Eijnden. A simulation-free deep learning approach to stochastic optimal control. arXiv preprint arXiv:2410.05163, 2024

    work page arXiv 2024

  15. [15]

    Estimation of non-normalized statistical models by score matching

    Aapo Hyv \"a rinen and Peter Dayan. Estimation of non-normalized statistical models by score matching. Journal of Machine Learning Research, 6 0 (4), 2005

    work page 2005

  16. [16]

    Zinc: a free tool to discover chemistry for biology

    John J Irwin, Teague Sterling, Michael M Mysinger, Erin S Bolstad, and Ryan G Coleman. Zinc: a free tool to discover chemistry for biology. Journal of chemical information and modeling, 52 0 (7): 0 1757--1768, 2012

    work page 2012

  17. [17]

    A mixture-based framework for guiding diffusion models

    Yazid Janati, Badr Moufad, Mehdi Abou El Qassime, Alain Durmus, Eric Moulines, and Jimmy Olsson. A mixture-based framework for guiding diffusion models. arXiv preprint arXiv:2502.03332, 2025

    work page arXiv 2025

  18. [18]

    Autonomous molecule generation using reinforcement learning and docking to develop potential novel inhibitors

    Woosung Jeon and Dongsup Kim. Autonomous molecule generation using reinforcement learning and docking to develop potential novel inhibitors. Scientific reports, 10 0 (1): 0 22104, 2020

    work page 2020

  19. [19]

    Hierarchical generation of molecular graphs using structural motifs

    Wengong Jin, Regina Barzilay, and Tommi Jaakkola. Hierarchical generation of molecular graphs using structural motifs. In International conference on machine learning, pages 4839--4848. PMLR, 2020

    work page 2020

  20. [20]

    Score-based generative modeling of graphs via the system of stochastic differential equations

    Jaehyeong Jo, Seul Lee, and Sung Ju Hwang. Score-based generative modeling of graphs via the system of stochastic differential equations. In International conference on machine learning, pages 10362--10383. PMLR, 2022

    work page 2022

  21. [21]

    Optimal control as a graphical model inference problem

    Hilbert J Kappen, Vicen c G \'o mez, and Manfred Opper. Optimal control as a graphical model inference problem. Machine learning, 87: 0 159--182, 2012

    work page 2012

  22. [22]

    Exploring chemical space with score-based out-of-distribution generation

    Seul Lee, Jaehyeong Jo, and Sung Ju Hwang. Exploring chemical space with score-based out-of-distribution generation. In International Conference on Machine Learning, pages 18872--18892. PMLR, 2023

    work page 2023

  23. [23]

    Stein variational gradient descent as gradient flow

    Qiang Liu. Stein variational gradient descent as gradient flow. Advances in neural information processing systems, 30, 2017

    work page 2017

  24. [24]

    Stein variational gradient descent: A general purpose bayesian inference algorithm

    Qiang Liu and Dilin Wang. Stein variational gradient descent: A general purpose bayesian inference algorithm. Advances in neural information processing systems, 29, 2016

    work page 2016

  25. [25]

    A kernelized stein discrepancy for goodness-of-fit tests

    Qiang Liu, Jason Lee, and Michael Jordan. A kernelized stein discrepancy for goodness-of-fit tests. In International conference on machine learning, pages 276--284. PMLR, 2016

    work page 2016

  26. [26]

    Decoupled weight decay regularization

    Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In International Conference on Learning Representations, 2019. URL https://openreview.net/forum?id=Bkg6RiCqY7

    work page 2019

  27. [27]

    Elucidating the design space of classifier-guided diffusion generation

    Jiajun Ma, Tianyang Hu, Wenjia Wang, and Jiacheng Sun. Elucidating the design space of classifier-guided diffusion generation. In The Twelfth International Conference on Learning Representations, 2024. URL https://openreview.net/forum?id=9DXXMXnIGm

    work page 2024

  28. [28]

    Variational diffusion posterior sampling with midpoint guidance

    Badr MOUFAD, Yazid Janati, Lisa Bedin, Alain Oliviero Durmus, randal douc, Eric Moulines, and Jimmy Olsson. Variational diffusion posterior sampling with midpoint guidance. In The Thirteenth International Conference on Learning Representations, 2025. URL https://openreview.net/forum?id=6EUtjXAvmj

    work page 2025

  29. [29]

    Solving high-dimensional hamilton--jacobi--bellman pdes using neural networks: perspectives from the theory of controlled diffusions and measures on path space

    Nikolas N \"u sken and Lorenz Richter. Solving high-dimensional hamilton--jacobi--bellman pdes using neural networks: perspectives from the theory of controlled diffusions and measures on path space. Partial differential equations and applications, 2 0 (4): 0 48, 2021

    work page 2021

  30. [30]

    Stochastic processes and applications

    Grigorios A Pavliotis. Stochastic processes and applications. Texts in applied mathematics, 60, 2014

    work page 2014

  31. [31]

    Continuous-time stochastic control and optimization with financial applications, volume 61

    Huy \^e n Pham. Continuous-time stochastic control and optimization with financial applications, volume 61. Springer Science & Business Media, 2009

    work page 2009

  32. [32]

    Estimation of the size of drug-like chemical space based on gdb-17 data

    Pavel G Polishchuk, Timur I Madzhidov, and Alexandre Varnek. Estimation of the size of drug-like chemical space based on gdb-17 data. Journal of computer-aided molecular design, 27: 0 675--679, 2013

    work page 2013

  33. [33]

    From reinforcement learning to optimal control: A unified framework for sequential decisions

    Warren B Powell. From reinforcement learning to optimal control: A unified framework for sequential decisions. In Handbook of Reinforcement Learning and Control, pages 29--74. Springer, 2021

    work page 2021

  34. [34]

    An empirical bayes approach to statistics

    Herbert E Robbins. An empirical bayes approach to statistics. In Breakthroughs in Statistics: Foundations and basic theory, pages 388--394. Springer, 1992

    work page 1992

  35. [35]

    RB -modulation: Training-free stylization using reference-based modulation

    Litu Rout, Yujia Chen, Nataniel Ruiz, Abhishek Kumar, Constantine Caramanis, Sanjay Shakkottai, and Wen-Sheng Chu. RB -modulation: Training-free stylization using reference-based modulation. In The Thirteenth International Conference on Learning Representations, 2025. URL https://openreview.net/forum?id=bnINPG5A32

    work page 2025

  36. [36]

    Generating high fidelity data from low-density regions using diffusion models

    Vikash Sehwag, Caner Hazirbas, Albert Gordo, Firat Ozgenel, and Cristian Canton. Generating high fidelity data from low-density regions using diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11492--11501, 2022

    work page 2022

  37. [37]

    Understanding and improving training-free loss-based diffusion guidance

    Yifei Shen, Xinyang Jiang, Yifan Yang, Yezhen Wang, Dongqi Han, and Dongsheng Li. Understanding and improving training-free loss-based diffusion guidance. Advances in Neural Information Processing Systems, 37: 0 108974--109002, 2024

    work page 2024

  38. [38]

    Deep unsupervised learning using nonequilibrium thermodynamics

    Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International conference on machine learning, pages 2256--2265. pmlr, 2015

    work page 2015

  39. [39]

    Denoising Diffusion Implicit Models

    Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020 a

    work page internal anchor Pith review Pith/arXiv arXiv 2010

  40. [40]

    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

    work page 2019

  41. [41]

    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. arXiv preprint arXiv:2011.13456, 2020 b

    work page internal anchor Pith review Pith/arXiv arXiv 2011

  42. [42]

    Maximum likelihood training of score-based diffusion models

    Yang Song, Conor Durkan, Iain Murray, and Stefano Ermon. Maximum likelihood training of score-based diffusion models. Advances in neural information processing systems, 34: 0 1415--1428, 2021

    work page 2021

  43. [43]

    Neural Stochastic Differ- ential Equations: Deep Latent Gaussian Models in the Diffu- sion Limit, 2019

    Belinda Tzen and Maxim Raginsky. Neural stochastic differential equations: Deep latent gaussian models in the diffusion limit. arXiv preprint arXiv:1905.09883, 2019

    work page arXiv 1905

  44. [44]

    Fine- tuning of continuous-time diffusion models as entropy-regularized control.arXiv preprint arXiv:2402.15194, 2024

    Masatoshi Uehara, Yulai Zhao, Kevin Black, Ehsan Hajiramezanali, Gabriele Scalia, Nathaniel Lee Diamant, Alex M Tseng, Tommaso Biancalani, and Sergey Levine. Fine-tuning of continuous-time diffusion models as entropy-regularized control. arXiv preprint arXiv:2402.15194, 2024

    work page arXiv 2024

  45. [45]

    Don't play favorites: Minority guidance for diffusion models

    Soobin Um, Suhyeon Lee, and Jong Chul Ye. Don't play favorites: Minority guidance for diffusion models. In The Twelfth International Conference on Learning Representations, 2024. URL https://openreview.net/forum?id=3NmO9lY4Jn

    work page 2024

  46. [46]

    Stochastic voyages into uncharted chemical space produce a representative library of all possible drug-like compounds

    Aaron M Virshup, Julia Contreras-Garc \' a, Peter Wipf, Weitao Yang, and David N Beratan. Stochastic voyages into uncharted chemical space produce a representative library of all possible drug-like compounds. Journal of the American Chemical Society, 135 0 (19): 0 7296--7303, 2013

    work page 2013

  47. [47]

    Fine-tuning discrete diffusion models via reward optimization with applications to DNA and protein design

    Chenyu Wang, Masatoshi Uehara, Yichun He, Amy Wang, Avantika Lal, Tommi Jaakkola, Sergey Levine, Aviv Regev, Hanchen, and Tommaso Biancalani. Fine-tuning discrete diffusion models via reward optimization with applications to DNA and protein design. In The Thirteenth International Conference on Learning Representations, 2025. URL https://openreview.net/for...

    work page 2025

  48. [48]

    Hit and lead discovery with explorative rl and fragment-based molecule generation

    Soojung Yang, Doyeong Hwang, Seul Lee, Seongok Ryu, and Sung Ju Hwang. Hit and lead discovery with explorative rl and fragment-based molecule generation. Advances in Neural Information Processing Systems, 34: 0 7924--7936, 2021

    work page 2021

  49. [49]

    Freedom: Training-free energy-guided conditional diffusion model

    Jiwen Yu, Yinhuai Wang, Chen Zhao, Bernard Ghanem, and Jian Zhang. Freedom: Training-free energy-guided conditional diffusion model. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 23174--23184, 2023

    work page 2023