pith. sign in

arxiv: 2607.00947 · v1 · pith:6HCM4Y5Rnew · submitted 2026-07-01 · 💻 cs.LG

Diffeomorphic Optimization

Ludwig Winkler , Andrew Leaver-Fay , Joseph Kleinhenz , Pan Kessel This is my paper

Pith reviewed 2026-07-02 15:36 UTC · model grok-4.3

classification 💻 cs.LG
keywords diffeomorphic optimizationRiemannian gradient descentdata manifolddiffusion modelsflow modelsprotein designLie groupsSO(3)
0
0 comments X

The pith

Diffeomorphic optimization performs gradient descent in the base space of flow models, equivalent to Riemannian gradient descent on the data manifold up to O(λ²) corrections.

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

The paper establishes that diffusion and flow models supply a map from a data manifold to a simpler base space where ordinary gradient descent can be run. Differential geometry shows this procedure matches Riemannian gradient descent on the manifold itself, with the added property that trajectories stay on the manifold automatically and the loss surface becomes smoother. The authors extend the approach to the Lie groups SO(3) and SE(3) for protein design by supplying an autograd-compatible SO(3) gradient and a generalized adjoint method for Lie-group ODE solvers. Concrete gains appear on secondary-structure targeting, peptide binding affinity, and Rosetta energy reduction across PDB structures.

Core claim

Diffeomorphic optimization starts from the observation that diffusion and flow models provide a map from the data manifold to a much simpler base space in which we perform gradient descent. Using differential geometry, we show this is equivalent to Riemannian gradient descent on the data manifold up to O(λ²) corrections, keeping trajectories on-manifold by construction and yielding a smoother optimization surface. For protein design, we extend diffeomorphic optimization to the matrix Lie groups SO(3) and SE(3), deriving an autograd-compatible SO(3) gradient and a generalized adjoint-state method for backpropagation through Lie-group ODE solvers.

What carries the argument

The diffeomorphic map supplied by a diffusion or flow model that sends the data manifold to a simpler base space for gradient descent.

If this is right

  • Optimization trajectories remain on the data manifold by construction rather than drifting off it.
  • The effective loss surface is smoother than the ambient high-dimensional landscape.
  • The method yields 91.3 percent of residues in the target Ramachandran region versus 63.3 percent for tuned guidance on FrameFlow.
  • It outperforms OC-Flow on peptide binding affinity while running at twice the speed.
  • Rosetta energies drop by thousands of units on PDB test structures containing hundreds of residues.

Where Pith is reading between the lines

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

  • The same base-space trick could be applied to other classes of generative models that learn manifold-to-base maps.
  • The O(λ²) error term could be bounded or corrected explicitly to enlarge the regime where the equivalence holds.
  • The Lie-group extensions might transfer to other robotics or molecular tasks that already use SO(3) or SE(3) representations.
  • Because the method is autograd-compatible, it could be dropped into existing differentiable simulators without custom manifold layers.

Load-bearing premise

The diffusion and flow models supply an accurate enough map from the data manifold to the base space that the O(λ²) corrections stay negligible for the protein design tasks considered.

What would settle it

Run the same optimization objective once with the diffeomorphic procedure and once with explicit Riemannian gradient descent on the manifold; the trajectories should agree up to O(λ²) differences.

Figures

Figures reproduced from arXiv: 2607.00947 by Andrew Leaver-Fay, Joseph Kleinhenz, Ludwig Winkler, Pan Kessel.

Figure 1
Figure 1. Figure 1: Optimization of protein with undesired properties (red) to a protein of desired properties [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: SE(3) roto￾translation of the idealized backbone positions. where Ti = (Ri , ti) is an element of the three-dimensional special Euclidean group SE(3) = SO(3) ⋉ R3 . The positions of the remaining heavy atoms (backbone oxygen and sidechain carbons) can be fixed by dihedral angles. Therefore, the vector field vθ in the ODE takes value in the tangent space of the corresponding Lie groups SE(3) and SO(2). Inte… view at source ↗
Figure 3
Figure 3. Figure 3: Diffeomorphic Optimization on SO(3): left hand side visualizes the gradient descent trajectory in the base space Z. Right hand side visualizes the same trajectory when mapped to the target space X . Diffeomorphic optimization clearly stays on manifold. The green plane is spanned by the left-singular vectors of the Jacobian ∇Zg(Z) scaled by the corresponding singular vectors. This shows that the diffeomorph… view at source ↗
Figure 4
Figure 4. Figure 4: We show various snapshots of the trajectory of diffeomorphic maximization of the distance [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Diffeomorphic optimization is used to change the secondary structure of the protein. For [PITH_FULL_IMAGE:figures/full_fig_p008_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Diffeomorphic optimiza￾tion applied to AlphaFlow exhibits strong improvement as measured in the Rosetta energy score. For this, we minimize an energy function that is based on the Ramachandran-PAAPP-combined term in the Rosetta energy Alford et al. (2017); Leaver-Fay et al. (2025) which interpo￾lates energies derived from statistics of backbone dihedral pref￾erences in the PDB on a toroidal grid. The grid … view at source ↗
Figure 7
Figure 7. Figure 7: Diffeomorphic optimiza￾tion leads to better/lower vina score than iid sampling with the same budget. Minimization of Rosetta Energy with AlphaFlow: we con￾sider the Rosetta energy function Alford et al. (2017) which is widely used in the protein community using the tmol pytorch implementation Leaver-Fay et al. (2025) of beta_nov2016_cart. We select the same pdb test set as in the AlphaFlow publication Jing… view at source ↗
Figure 8
Figure 8. Figure 8: Gradient Error as a function of dimensionality and the number of steps [PITH_FULL_IMAGE:figures/full_fig_p024_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Gradient Error as a function of dimensionality and the number of steps [PITH_FULL_IMAGE:figures/full_fig_p025_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Gradient Error as a function of dimensionality and the number of steps [PITH_FULL_IMAGE:figures/full_fig_p025_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Grid used for Ramachandran guidance [PITH_FULL_IMAGE:figures/full_fig_p027_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Self-consistent RMSD and the optimized ABEGO percentage over the course of diffeo [PITH_FULL_IMAGE:figures/full_fig_p027_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: The performance of diffeomorphic optimization (higher is better) as a function of the [PITH_FULL_IMAGE:figures/full_fig_p029_13.png] view at source ↗
Figure 16
Figure 16. Figure 16: In addition to the experiments reported in the main part of the paper, we also compared the per￾formance of our diffeomorphic Rosetta Relax to a sampling-based baseline. For this baseline, we simply sample from AlphaFlow with the equivalent computational budget used by diffeomorphic 30 [PITH_FULL_IMAGE:figures/full_fig_p030_16.png] view at source ↗
Figure 14
Figure 14. Figure 14: Diffeomorphic optimization of the Rosetta energy function: each point denotes a prediction [PITH_FULL_IMAGE:figures/full_fig_p031_14.png] view at source ↗
Figure 15
Figure 15. Figure 15: Diffeomorphic optimization in combination with Rosetta Relax yields improved (lower) [PITH_FULL_IMAGE:figures/full_fig_p031_15.png] view at source ↗
Figure 16
Figure 16. Figure 16: Diffeomorphic Optimization in combination with Rosetta Relax can achieve substantial [PITH_FULL_IMAGE:figures/full_fig_p032_16.png] view at source ↗
read the original abstract

Generative models learn data distributions that reside on a low-dimensional manifold within a higher-dimensional ambient space. Optimizing differentiable objectives on this manifold is challenging: the ambient loss landscape is high-dimensional, rugged, and non-convex. Direct gradient descent, blind to the manifold's geometry, quickly drifts off it. Diffeomorphic optimization starts from the observation that diffusion and flow models provide a map from the data manifold to a much simpler base space in which we perform gradient descent. Using differential geometry, we show this is equivalent to Riemannian gradient descent on the data manifold up to $\mathcal{O}(\lambda^2)$ corrections, keeping trajectories on-manifold by construction and yielding a smoother optimization surface. For protein design, we extend diffeomorphic optimization to the matrix Lie groups $\mathrm{SO}(3)$ and $\mathrm{SE}(3)$, deriving an autograd-compatible $\mathrm{SO}(3)$ gradient and a generalized adjoint-state method for backpropagation through Lie-group ODE solvers. Diffeomorphic optimization improves over tuned guidance on secondary-structure targeting with FrameFlow ($91.3\%$ vs. $63.3\%$ of residues in the Ramachandran target), outperforms OC-Flow on peptide binding affinity at $2\times$ the speed, and reduces Rosetta energies by thousands of units across the PDB test set for structures with hundreds of residues.

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

2 major / 1 minor

Summary. The manuscript introduces diffeomorphic optimization, which uses diffusion and flow models to map data-manifold optimization problems into a simpler base space where gradient descent is performed. Using differential geometry, it claims this procedure is equivalent to Riemannian gradient descent on the data manifold up to O(λ²) corrections, thereby keeping trajectories on-manifold by construction and producing a smoother loss landscape. The method is extended to the Lie groups SO(3) and SE(3) for protein design, with derivations of an autograd-compatible SO(3) gradient and a generalized adjoint-state method for backpropagation through Lie-group ODE solvers. Empirical results are reported for FrameFlow on secondary-structure targeting (91.3% vs. 63.3% Ramachandran compliance), OC-Flow on peptide binding affinity (outperformance at 2× speed), and Rosetta energy minimization across a PDB test set.

Significance. If the O(λ²) approximation is shown to be negligible in the operating regimes of the cited generative models, the framework supplies a geometrically grounded alternative to guidance-based or projection-based manifold optimization that could improve stability and sample quality in generative modeling for structural biology. The explicit Lie-group extensions and autograd-compatible adjoint methods constitute concrete implementation contributions that lower the barrier to adoption.

major comments (2)
  1. [Abstract and theoretical derivation section] Abstract and the differential-geometry derivation (the section establishing the Riemannian equivalence): the central claim that trajectories remain on-manifold by construction rests on the O(λ²) corrections being negligible, yet no bounds on λ, no numerical estimates of the higher-order terms, and no sensitivity analysis are supplied for the FrameFlow or OC-Flow models used in the protein-design experiments. This directly affects whether the reported gains (e.g., Ramachandran compliance and energy reductions) can be attributed to the claimed geometric property.
  2. [Protein-design experimental section] Protein-design experimental section (the paragraphs reporting 91.3% Ramachandran compliance and energy reductions): the comparisons to tuned guidance and OC-Flow are presented without any verification that the diffusion/flow map remains sufficiently accurate for the O(λ²) remainder to stay small across the tested λ schedules or residue lengths; absent such checks, the on-manifold-by-construction advantage is not yet demonstrated in the reported regimes.
minor comments (1)
  1. [Notation and setup] The notation for the base-space map and the precise definition of λ should be introduced with an explicit equation reference before the equivalence statement to aid readers who are not already familiar with the underlying differential-geometry construction.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive comments. We agree that additional validation of the O(λ²) approximation is needed to fully support the claims and will incorporate the suggested analyses in the revised manuscript.

read point-by-point responses

  1. Referee: [Abstract and theoretical derivation section] Abstract and the differential-geometry derivation (the section establishing the Riemannian equivalence): the central claim that trajectories remain on-manifold by construction rests on the O(λ²) corrections being negligible, yet no bounds on λ, no numerical estimates of the higher-order terms, and no sensitivity analysis are supplied for the FrameFlow or OC-Flow models used in the protein-design experiments. This directly affects whether the reported gains (e.g., Ramachandran compliance and energy reductions) can be attributed to the claimed geometric property.

    Authors: We concur that the manuscript lacks explicit bounds, numerical estimates, and sensitivity analysis for the O(λ²) terms in the cited models. While the theoretical derivation establishes the equivalence, empirical validation of the approximation's accuracy is indeed absent. In the revision, we will add numerical estimates of the higher-order terms and a sensitivity analysis for the λ schedules and residue lengths in the FrameFlow and OC-Flow experiments. This will help attribute the reported gains to the geometric properties. revision: yes

  2. Referee: [Protein-design experimental section] Protein-design experimental section (the paragraphs reporting 91.3% Ramachandran compliance and energy reductions): the comparisons to tuned guidance and OC-Flow are presented without any verification that the diffusion/flow map remains sufficiently accurate for the O(λ²) remainder to stay small across the tested λ schedules or residue lengths; absent such checks, the on-manifold-by-construction advantage is not yet demonstrated in the reported regimes.

    Authors: We acknowledge that the experimental section does not include explicit verification of the flow map accuracy or the size of the O(λ²) remainder for the tested conditions. The reported improvements are observed, but to demonstrate the on-manifold advantage specifically, we will include in the revision the requested checks on the approximation across the λ schedules and residue lengths used. revision: yes

Circularity Check

0 steps flagged

No significant circularity; central equivalence is a differential-geometry derivation independent of fitting

full rationale

The paper states it uses differential geometry to establish equivalence to Riemannian gradient descent up to O(λ²) corrections. This is presented as a mathematical result, not a fit or self-citation reduction. No equations, parameters, or load-bearing steps in the abstract or described claims reduce by construction to inputs, fitted quantities, or prior self-citations. The Lie-group extensions (autograd-compatible SO(3) gradient, adjoint method) are described as derivations. Empirical results are reported separately and do not serve as the justification for the equivalence. The derivation chain is therefore self-contained against external mathematical benchmarks.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

No information available from abstract to populate free parameters, axioms, or invented entities.

pith-pipeline@v0.9.1-grok · 5765 in / 1320 out tokens · 35344 ms · 2026-07-02T15:36:09.936580+00:00 · methodology

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

63 extracted references · 19 canonical work pages · 7 internal anchors

  1. [1]

    Proceedings of the National Academy of Sciences , volume=

    Explaining neural scaling laws , author=. Proceedings of the National Academy of Sciences , volume=. 2024 , publisher=

    2024

  2. [2]

    Journal of the American Mathematical Society , volume=

    Testing the manifold hypothesis , author=. Journal of the American Mathematical Society , volume=

  3. [3]

    arXiv preprint arXiv:2207.02862 , year=

    Verifying the union of manifolds hypothesis for image data , author=. arXiv preprint arXiv:2207.02862 , year=

    work page arXiv

  4. [4]

    Advances in Neural Information Processing Systems , volume=

    Hardness of Learning Neural Networks Under the Manifold Hypothesis , author=. Advances in Neural Information Processing Systems , volume=

  5. [5]

    arXiv preprint arXiv:2310.05297 , year=

    Fast protein backbone generation with se (3) flow matching , author=. arXiv preprint arXiv:2310.05297 , year=

    work page arXiv

  6. [6]

    Advances in Neural Information Processing Systems , volume=

    Abdiffuser: full-atom generation of in-vitro functioning antibodies , author=. Advances in Neural Information Processing Systems , volume=

  7. [7]

    arXiv preprint arXiv:2302.02277 , year=

    SE (3) diffusion model with application to protein backbone generation , author=. arXiv preprint arXiv:2302.02277 , year=

    work page arXiv

  8. [8]

    2024 , booktitle=

    AlphaFold Meets Flow Matching for Generating Protein Ensembles , author=. 2024 , booktitle=

    2024

  9. [9]

    International Conference on Learning Representations (ICLR) , year=

    DiffDock: Diffusion Steps, Twists, and Turns for Molecular Docking , author =. International Conference on Learning Representations (ICLR) , year=

  10. [10]

    Journal of Computational Chemistry , volume=

    AutoDock Vina: Improving the speed and accuracy of docking with a new scoring function, efficient optimization, and multithreading , author=. Journal of Computational Chemistry , volume=. 2010 , publisher=

    2010

  11. [11]

    Building Normalizing Flows with Stochastic Interpolants

    Building Normalizing Flows with Stochastic Interpolants , author=. arXiv preprint arXiv:2209.15571 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  12. [12]

    Flow Matching for Generative Modeling

    Flow Matching for Generative Modeling , author=. arXiv preprint arXiv:2210.02747 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  13. [13]

    Advances in neural information processing systems , volume=

    Denoising diffusion probabilistic models , author=. Advances in neural information processing systems , volume=

  14. [14]

    Score-Based Generative Modeling through Stochastic Differential Equations

    Score-based generative modeling through stochastic differential equations , author=. arXiv preprint arXiv:2011.13456 , year=

    work page internal anchor Pith review Pith/arXiv arXiv 2011

  15. [15]

    Journal of Machine Learning Research , volume=

    Normalizing flows for probabilistic modeling and inference , author=. Journal of Machine Learning Research , volume=

  16. [16]

    nature , volume=

    Highly accurate protein structure prediction with AlphaFold , author=. nature , volume=. 2021 , publisher=

    2021

  17. [17]

    2025 , url =

    Peter Holderrieth and Ezra Erives , title =. 2025 , url =

    2025

  18. [18]

    2025 , url =

    Leaver-Fay, Andrew and Flatten, Jeff and Ford, Alex and Kleinhenz, Joseph and Solberg, Henry amd Baker, David and Watkins, Andrew M and Kuhlman, Brian and DiMaio, Frank , title =. 2025 , url =

    2025

  19. [19]

    2018 , publisher=

    Introduction to Riemannian manifolds , author=. 2018 , publisher=

    2018

  20. [20]

    Advances in neural information processing systems , volume=

    Neural ordinary differential equations , author=. Advances in neural information processing systems , volume=

  21. [21]

    Journal of chemical theory and computation , volume=

    The Rosetta all-atom energy function for macromolecular modeling and design , author=. Journal of chemical theory and computation , volume=. 2017 , publisher=

    2017

  22. [22]

    Journal of molecular biology , volume=

    Automatic classification and analysis of -turn motifs in proteins , author=. Journal of molecular biology , volume=. 1996 , publisher=

    1996

  23. [23]

    Journal of medicinal chemistry , volume=

    The PDBbind database: methodologies and updates , author=. Journal of medicinal chemistry , volume=. 2005 , publisher=

    2005

  24. [24]

    Bioinformatics , pages =

    Hu, Qiuyue and Wang, Zechen and Meng, Jintao and Li, Weifeng and Guo, Jingjing and Mu, Yuguang and Wang, Sheng and Liangzhen, Zheng and Wei, Yanjie , title = ". Bioinformatics , pages =. 2024 , month =. doi:10.1093/bioinformatics/btae628 , url =

    work page doi:10.1093/bioinformatics/btae628 2024

  25. [25]

    Advances in Neural Information Processing Systems , volume=

    Neural manifold ordinary differential equations , author=. Advances in Neural Information Processing Systems , volume=

  26. [26]

    Advances in Neural Information Processing Systems , volume=

    Riemannian continuous normalizing flows , author=. Advances in Neural Information Processing Systems , volume=

  27. [27]

    Physical Review D , volume=

    Learning trivializing gradient flows for lattice gauge theories , author=. Physical Review D , volume=. 2023 , publisher=

    2023

  28. [28]

    arXiv preprint arXiv:2302.03660 , year=

    Flow matching on general geometries , author=. arXiv preprint arXiv:2302.03660 , year=

    work page arXiv

  29. [29]

    arXiv preprint arXiv:2006.06663 , year=

    Neural ordinary differential equations on manifolds , author=. arXiv preprint arXiv:2006.06663 , year=

    work page arXiv 2006

  30. [30]

    International Conference on Machine Learning , pages=

    Normalizing Flows on Tori and Spheres , author=. International Conference on Machine Learning , pages=

  31. [31]

    arXiv preprint arXiv:2305.02402 , year=

    Normalizing flows for lattice gauge theory in arbitrary space-time dimension , author=. arXiv preprint arXiv:2305.02402 , year=

    work page arXiv

  32. [32]

    arXiv preprint arXiv:2402.14017 , year=

    D-flow: Differentiating through flows for controlled generation , author=. arXiv preprint arXiv:2402.14017 , year=

    work page arXiv

  33. [33]

    ICML Workshop on Invertible Neural Networks, Normalizing Flows, and Explicit Likelihood Models , year=

    Diffeomorphic explanations with normalizing flows , author=. ICML Workshop on Invertible Neural Networks, Normalizing Flows, and Explicit Likelihood Models , year=

  34. [34]

    IEEE Transactions on Pattern Analysis and Machine Intelligence , volume=

    Diffeomorphic counterfactuals with generative models , author=. IEEE Transactions on Pattern Analysis and Machine Intelligence , volume=. 2023 , publisher=

    2023

  35. [35]

    Nature , volume=

    Accurate structure prediction of biomolecular interactions with AlphaFold 3 , author=. Nature , volume=. 2024 , publisher=

    2024

  36. [36]

    Journal of molecular biology , volume=

    Sampling bottlenecks in de novo protein structure prediction , author=. Journal of molecular biology , volume=. 2009 , publisher=

    2009

  37. [37]

    biorxiv , year=

    Atomically accurate de novo design of single-domain antibodies , author=. biorxiv , year=

  38. [38]

    Nature , volume=

    De novo design of protein structure and function with RFdiffusion , author=. Nature , volume=. 2023 , publisher=

    2023

  39. [39]

    bioRxiv , pages=

    Lab-in-the-loop therapeutic antibody design with deep learning , author=. bioRxiv , pages=. 2025 , publisher=

    2025

  40. [40]

    Towards Realistic Individual Recourse and Actionable Explanations in Black-Box Decision Making Systems

    Towards realistic individual recourse and actionable explanations in black-box decision making systems , author=. arXiv preprint arXiv:1907.09615 , year=

    work page internal anchor Pith review Pith/arXiv arXiv 1907

  41. [41]

    Advances in neural information processing systems , volume=

    Explanations based on the missing: Towards contrastive explanations with pertinent negatives , author=. Advances in neural information processing systems , volume=

  42. [42]

    bioRxiv , pages=

    BindCraft: one-shot design of functional protein binders , author=. bioRxiv , pages=. 2024 , publisher=

    2024

  43. [43]

    bioRxiv , pages=

    BoltzDesign1: Inverting All-Atom Structure Prediction Model for Generalized Biomolecular Binder Design , author=. bioRxiv , pages=. 2025 , publisher=

    2025

  44. [44]

    Nature , volume=

    De novo protein design by deep network hallucination , author=. Nature , volume=. 2021 , publisher=

    2021

  45. [45]

    Protein Science , volume=

    De novo protein design by inversion of the AlphaFold structure prediction network , author=. Protein Science , volume=. 2023 , publisher=

    2023

  46. [46]

    Biomedicines , volume=

    Solubility-aware protein binding peptide design using AlphaFold , author=. Biomedicines , volume=. 2022 , publisher=

    2022

  47. [47]

    Advances in neural information processing systems , volume=

    Diffusion models beat gans on image synthesis , author=. Advances in neural information processing systems , volume=

  48. [48]

    Classifier-Free Diffusion Guidance

    Classifier-free diffusion guidance , author=. arXiv preprint arXiv:2207.12598 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  49. [49]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    Universal guidance for diffusion models , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

  50. [50]

    Advances in Neural Information Processing Systems , volume=

    Practical and asymptotically exact conditional sampling in diffusion models , author=. Advances in Neural Information Processing Systems , volume=

  51. [51]

    arXiv preprint arXiv:2206.04119 , year=

    Diffusion probabilistic modeling of protein backbones in 3d for the motif-scaffolding problem , author=. arXiv preprint arXiv:2206.04119 , year=

    work page arXiv

  52. [52]

    arXiv preprint arXiv:2308.07983 , year=

    Monte Carlo guided diffusion for Bayesian linear inverse problems , author=. arXiv preprint arXiv:2308.07983 , year=

    work page arXiv

  53. [53]

    The Twelfth International Conference on Learning Representations , year=

    Diffusion posterior sampling for linear inverse problem solving: A filtering perspective , author=. The Twelfth International Conference on Learning Representations , year=

  54. [54]

    arXiv preprint arXiv:2501.06848 , year=

    A general framework for inference-time scaling and steering of diffusion models , author=. arXiv preprint arXiv:2501.06848 , year=

    work page arXiv

  55. [55]

    Entropy , volume=

    Stochastic control for bayesian neural network training , author=. Entropy , volume=. 2022 , publisher=

    2022

  56. [56]

    Flow Matching Guide and Code

    Flow matching guide and code , author=. arXiv preprint arXiv:2412.06264 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  57. [57]

    Advances in Neural Information Processing Systems , volume=

    Efficient and accurate gradients for neural sdes , author=. Advances in Neural Information Processing Systems , volume=

  58. [58]

    International Conference on Machine Learning , pages=

    Bridging discrete and continuous state spaces: Exploring the Ehrenfest process in time-continuous diffusion models , author=. International Conference on Machine Learning , pages=. 2024 , organization=

    2024

  59. [59]

    Symposium on Advances in Approximate Bayesian Inference , pages=

    Scalable gradients and variational inference for stochastic differential equations , author=. Symposium on Advances in Approximate Bayesian Inference , pages=. 2020 , organization=

    2020

  60. [60]

    arXiv preprint arXiv:2410.18070 , year=

    Training free guided flow matching with optimal control , author=. arXiv preprint arXiv:2410.18070 , year=

    work page arXiv

  61. [61]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    Flowgrad: Controlling the output of generative odes with gradients , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

  62. [62]

    ICML 2026 Workshop on Structured Probabilistic Inference \ & \ Generative Modeling , year=

    Enhanced Diffusion Sampling: Efficient Rare Event Sampling and Free Energy Calculation with Diffusion Models , author=. ICML 2026 Workshop on Structured Probabilistic Inference \ & \ Generative Modeling , year=

    2026

  63. [63]

    Directly Fine-Tuning Diffusion Models on Differentiable Rewards

    Directly fine-tuning diffusion models on differentiable rewards , author=. arXiv preprint arXiv:2309.17400 , year=

    work page internal anchor Pith review Pith/arXiv arXiv