ESS-Flow: Training-free guidance as Bayesian inference in source space

arXiv:2510.05849 · cs.LG, stat.ML · Submitted 2025-10-07 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: We're looking at the paper: "ESS-Flow: Training-free guidance as Bayesian inference in source space".

Jane: Here's the gist of the paper. Guiding pretrained flow-based generative models for conditional generation or to produce samples with desired target properties enables solving diverse tasks without retraining on paired data<ref:2510.05849#pg2> ESS-Flow, a gradient-free method that leverages the typically Gaussian prior of the source distribution in flow-based models to perform Bayesian inference directly in the source space using Elliptical Slice Sampling<ref:2510.05849#pg3> This method is particularly valuable for problems involving quantization, such as molecular and material design with categorical atomic numbers, where gradients are not well-defined<ref:2510.05849#pg4> ESS-Flow requires only point-wise evaluations of the generative model and involves no gradient or Jacobian computation<ref:2510.05849#pg4> We demonstrate its effectiveness on designing materials with desired target properties and predicting protein structures from sparse inter-residue distance measurements<ref:2510.05849#pg4> Core Methodology ESS-Flow recasts controlled generation as Bayesian inference in the source space of flow-based models, which typically has a Gaussian prior p(z)<ref:2510.05849#pg3> This is achieved by reformulating the target distribution π(x) in terms of the source distribution π(z) using equation (3)<ref:2510.05849#pg5> The key insight is that many pretrained diffusion and flow-based models can be recast as continuous normalizing flows with Gaussian source distributions<ref:2510.05849#pg3> This allows the use of tailored Markov chain Monte Carlo (MCMC) methods like elliptical slice sampling (ESS)<ref:2510.05849#pg3> ESS-Flow is by design gradient-free, requiring only point-wise evaluations of the generative model and the likelihood (or more generally, potential) function<ref:2510.05849#pg4> Limitations of Existing Methods Optimization-based methods only provide point estimates rather than samples and offer little guarantees against local optima<ref:2510.05849#pg2> They fail when gradients are unreliable or unavailable, such as with simulation-based observations or quantization in the generation or observation process<ref:2510.05849#pg2> Existing source space methods like D-Flow rely on gradients and require expensive backpropagation through the ODE solver<ref:2510.05849#pg4> Furthermore, they do not require knowledge of the noising process used during model training, only the trained transport map<ref:2510.05849#pg4> ESS-Flow is presented as the only gradient-free method in this category<ref:2510.05849#pg4> Performance on Material Design ESS-Flow was evaluated on generating materials with target properties and predicting protein backbone structures from sparse inter-residue distances<ref:2510.05849#pg4> In material generation experiments, ESS-Flow outperformed all other methods significantly with the lowest errors<ref:2510.05849#pg8> For bulk modulus, ESS-Flow achieved an absolute error of 9.05 compared to 198.37 for the unconditional method<ref:2510.05849#pg8> In protein structure prediction, ESS-Flow achieved a lower RMSDgt of 37.02 compared to D-Flow's 46.54<ref:2510.05849#pg10> While DAPS and ADP-3D samples best match observations in protein prediction based on L2 distances, ESS-Flow achieves a better trade-off between data fidelity and sample realism<ref:2510.05849#pg10> Multi-Fidelity Extension A multi-fidelity extension of ESS-Flow was proposed to leverage the fact that flow-based generative models in practice are simulated from using a numerical solver<ref:2510.05849#pg7> This involves re-weighting samples obtained from coarse ODE discretization with importance weights by using much finer discretization for the ODE<ref:2510.05849#pg7> This approach significantly reduces computational cost while maintaining accuracy of the target<ref:2510.05849#pg7> We introduce ESS-Flow, a training-free sampling method for controlled generation with pretrained flow-based generative models<ref:2510.05849#pg11> The gradient-free approach limits ESS-Flow’s effectiveness when the prior does not well inform the target distribution, such as in noiseless image inpainting tasks<ref:2510.05849#pg11> Future work could explore more principled combinations of coarse and fine ODE discretization to better guide exploration of target distributions<ref:2510.05849#pg11> Developing adaptive strategies for problems where the prior poorly covers the target distribution could expand the applicability of the method to a broader range of controlled generation tasks<ref:2510.05849#pg11> ESS-Flow requires minimal hyperparameter tuning and is entirely gradient-free<ref:2510.05849#pg14> The convergence of ESS is discussed by Murray et al.

Tom: (2010) who show that the induced Markov kernel leaves the target distribution invariant<ref:2510.05849#pg8> ESS-Flow has minimal hyperparameters to tune, and one MCMC step is summarized in Algorithm 1<ref:2510.05849#pg8> The hyperparameter details for the material generation tasks are provided in Table 5<ref:2510.05849#pg14> The hyperparameters for the protein structure prediction tasks are given in Table 6<ref:2510.05849#pg14> ESS-Flow is particularly valuable for problems involving quantization, such as molecular and material design with categorical atomic numbers<ref:2510.05849#pg4> The paper was financially supported by the Wallenberg AI, Autonomous Systems and Software Program (WASP) funded by the Knut and Alice Wallenberg Foundation, the Swedish Research Council, and the Excellence Center at Linkoping–Lund in Information Technology (ELLIIT)<ref:2510.05849#pg14> The computations were enabled by the Berzelius resource at the National Supercomputer Centre, provided by the Knut and Alice Wallenberg Foundation<ref:2510.05849#pg14> ESS-Flow outperforms all other methods significantly with the lowest errors<ref:2510.05849#pg8> The resulting S.U.N.

Jane: rates indicate that ESS-Flow does generate a subset of materials that are good candidates for further exploration<ref:2510.05849#pg8> The paper was published as arXiv:2510.05849v1

cs.LG: on 7 Oct 2025<ref:2510.05849#pg2> Page 1 --- Preprint.

Tom: Under review ESS-FLOW: TRAINING-FREE GUIDANCE OF FLOWBASED MODELS AS INFERENCE IN SOURCE SPACE Adhithyan Kalaivanan Department of Computer Science Linkoping University, Sweden ¨ adhithyan.kalaivanan@liu.se Page 2 --- Preprint.

Title and authors: Tom: We’re talking about "ESS-Flow: Training-free guidance as Bayesian inference in source space" here. The authors are Adhithyan Kalaivanan, Zheng Zhao, and Jens Sjolund. It’s a method that guides pretrained flow-based generative models for conditional generation or to produce samples with desired target properties without needing retraining on paired data.

Jane: That sounds complicated, so to unpack it simply: instead of trying to find a perfect point estimate in the complex data space, ESS-Flow lets you perform inference by sampling from the source distribution where the prior is usually simple, like a Gaussian distribution.

Lu: The key insight they are making is that many models we already have—diffusion and flow-based models—can actually be recast as continuous normalizing flows with Gaussian source distributions on page three.

Meng: So, if I’m thinking about what this means for running models on my hardware, it suggests we can perform inference without the heavy machinery of calculating gradients or Jacobians every single time you want a sample, which is huge because those computations are often very expensive and slow.

Lalam: And they emphasize that ESS-Flow only requires point-wise evaluations of the generative model and the potential function, meaning no gradient or Jacobian computation is needed for this whole process, as they explain on page six.

The paper's summary: Tom: Now let’s talk about what they found works better than what we have now. They show that optimization-based methods only give you point estimates, but ESS-Flow actually provides samples, and they claim it does this with a lower error compared to other techniques because it performs Bayesian inference in the source space instead of just finding a single best guess.

Jane: In material design experiments, they found that ESS-Flow outperformed all the other methods tested and achieved the lowest errors for properties like bulk modulus, even getting an absolute error of nine point zero five compared to one hundred ninety-eight point three seven for the unconditional method on bulk modulus on page eight.

Lu: For protein structure prediction, they compared it to D-Flow, and ESS-Flow achieved a lower root mean square deviation of thirty-seven point zero two against D-Flow’s forty-six point five four on page ten. They also noted that while other methods like ADP three and ADP three dimension samples match observations based on L two distances, ESS-Flow finds a better balance between data fidelity and sample realism there too.

Meng: That balance is what matters for production. If the sample is just noisy data, it’s useless, so the fact that ESS-Flow enforces the prior in a way that leads to more realistic samples is a huge plus when we need reliable outputs on page nine.

Lalam: They also introduced a multi-fidelity extension, which they realized is useful because flow models are often simulated using numerical solvers on page seven. This lets them re-weight samples from a coarse discretization using importance weights from a much finer discretization, which cuts the computational cost while keeping the target accuracy high there too.

The paper's improvements: Tom: So to wrap up on ESS-Flow, it’s this training-free sampling method for controlled generation using pretrained flow models that uses elliptical slice sampling in the source space to get samples near target values across all tasks, even if L two distances suggest other methods might look better on paper.

Jane: It’s a powerful way to sample from a complex distribution without needing gradients or Jacobians, which opens up applications in areas like molecular and material design where those things are usually unreliable because the gradients aren't well defined on page four.

Lu: The authors do acknowledge that there’s still an issue where ESS-Flow might not explore the target distribution as effectively as it does in material generation experiments, and it can sometimes get trapped in local modes, which is a limitation they point out on page eleven.

Meng: From a deployment standpoint, if this method requires minimal hyperparameter tuning and is entirely gradient-free, it makes it much easier to integrate into production pipelines where you can't afford expensive training runs for every new property, as they mention on page fourteen.

Lalam: The paper is a solid step forward in providing a principled way to do inference on these models, and I’m fairly sure the results they showed with lower errors on bulk modulus and RMSDgt are quite convincing there too.

Tom: Alright, so that’s ESS-Flow. It shows how we can guide flow-based models for conditional generation by leveraging the Gaussian prior in the source space of those models without needing retraining on paired data. We’ll take a quick break and come back after this to look at temporal consistency in video generation with Compact Attention.

Conclusion: Tom: So we just heard about "ESS-Flow: Training-free guidance as Bayesian inference in source space," which is this training free way to use pretrained flow models for controlled generation by doing Bayesian inference right in the source space using elliptical slice sampling, as described on page eleven.

Jane: It’s a gradient free method that uses elliptical slice sampling to sample from the target distribution while only needing point wise evaluations of the model and potential function, which is what they detail on page six.

Lu: The core idea is reformulating the target distribution in terms of the source distribution using an equation involving a determinant term that cancels out some Jacobian stuff, as seen on page five.

Meng: That means for us at scale, we don't need to worry about backpropagation through the ODE solver when we want to generate samples based on desired properties, which is something they highlight on page four.

Lalam: It really works well for problems involving quantization, like designing materials or predicting protein structures from sparse measurements, which is what they discuss in page four.

Tom: And the results were pretty solid on material design, showing lower errors than methods like D-Flow for properties such as bulk modulus, achieving an absolute error of nine point zero five compared to one hundred ninety-eight point thirty seven compared to the unconditional method on page eight.

Jane: They even beat D-Flow on protein structure prediction with a lower root mean square deviation of thirty-seven point zero two versus forty-six point five four, which suggests they found a really good trade off between getting data fidelity and making the samples look realistic there too.

Lu: The multi fidelity extension they proposed is interesting because it lets you re-weight samples from a coarse discretization using finer discretization for the potential function, which cuts down on computational cost while keeping accuracy up, as detailed on page seven.

Meng: If you can cut the computational cost by re-weighting those samples, that’s what we need to talk about when deploying these kinds of models in a real production environment where speed and efficiency matter more than just theoretical accuracy on page seven.

Lalam: So, to sum up, ESS-Flow is this training free guidance as Bayesian inference in source space by using elliptical slice sampling, which is a big step because it lets us sample from the target distribution directly with minimal computation and requires only point wise evaluations of the model and potential function on page eleven.

Tom: Yeah, it's clean mechanics and really shows how we can guide flow-based models for conditional generation by leveraging the Gaussian prior in the source space without needing retraining on paired data, as they show on page two.

Adhithyan Kalaivanan, Zheng Zhao, Jens Sjolund, Fredrik Lindsten

Linkoping University, Sweden · Uppsala University

cs.LG, stat.ML

Submitted: 2025-10-07

Updated: 2026-11-04

Comments: NeurIPS 2026

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

Importance score: 88/100

The gist: Guiding pretrained flow-based generative models for conditional generation or to produce samples with desired target properties enables solving diverse tasks without retraining on paired…

Key concepts

Flow-based Generative Models
These are pretrained models that learn to transform data from a simple source distribution (like a Gaussian) into complex target distributions using a series of invertible functions called flows. They are powerful tools for generating new data, but sampling from the desired target distribution can be difficult.
Bayesian Inference in Source Space
Instead of trying to sample directly from the complex target distribution, ESS-Flow reformulates the problem by performing inference in the simpler source space where the model's prior distribution is known (typically Gaussian). This transformation makes sampling feasible.
Elliptical Slice Sampling (ESS)
ESS is a specific Markov chain Monte Carlo (MCMC) method used to sample from distributions. ESS-Flow leverages this technique by using the Gaussian prior of the source space to guide the sampling process, enabling gradient-free inference without needing expensive Jacobian calculations.

Terminology

Summary

Guiding pretrained flow-based generative models for conditional generation or to produce samples with desired target properties enables solving diverse tasks without retraining on paired data<ref:2510. p. 005849#pg2> ESS-Flow, a gradient-free method that leverages the typically Gaussian prior of the source distribution in flow-based models to perform Bayesian inference directly in the source space using Elliptical Slice Sampling<ref:2510. p. 005849#pg3> This method is particularly valuable for problems involving quantization, such as molecular and material design with categorical atomic numbers, where gradients are not well-defined<ref:2510. p. 105849#pg4> ESS-Flow requires only point-wise evaluations of the generative model and involves no gradient or Jacobian computation<ref:2510. p. 305849#pg4> We demonstrate its effectiveness on designing materials with desired target properties and predicting protein structures from sparse inter-residue distance measurements<ref:2510. p. 005849#pg4>

Core Methodology

ESS-Flow recasts controlled generation as Bayesian inference in the source space of flow-based models, which typically has a Gaussian prior p(z)<ref:2510.05849#pg3> This is achieved by reformulating the target distribution π(x) in terms of the source distribution π(z) using equation (3)<ref:2510. p. 305849#pg5> The key insight is that many pretrained diffusion and flow-based models can be recast as continuous normalizing flows with Gaussian source distributions<ref:2510. p. 105849#pg3> This allows the use of tailored Markov chain Monte Carlo (MCMC) methods like elliptical slice sampling (ESS)<ref:2510. p. 105849#pg3> ESS-Flow is by design gradient-free, requiring only point-wise evaluations of the generative model and the likelihood (or more generally, potential) function<ref:2510. p. 1 p. 105849#pg4>

Limitations of Existing Methods

Optimization-based methods only provide point estimates rather than samples and offer little guarantees against local optima<ref:2510.05849#pg2> They fail when gradients are unreliable or unavailable, such as with simulation-based observations or quantization in the generation or observation process<ref:2510. p. 005849#pg2> Existing source space methods like D-Flow rely on gradients and require expensive backpropagation through the ODE solver<ref:2510. p. 205849#pg4> Furthermore, they do not require knowledge of the noising process used during model training, only the trained transport map<ref:2510. p. 105849#pg4> ESS-Flow is presented as the only gradient-free method in this category<ref:2510. p. 205849#pg4>

Performance on Material Design

ESS-Flow was evaluated on generating materials with target properties and predicting protein backbone structures from sparse inter-residue distances<ref:2510.05849#pg4> In material generation experiments, ESS-Flow outperformed all other methods significantly with the lowest errors<ref:2510. p. 605849#pg8> For bulk modulus, ESS-Flow achieved an absolute error of 9. p. 605 compared to 198. p. 637 for the unconditional method<ref:2510. p. 605849#pg8> In protein structure prediction, ESS-Flow achieved a lower RMSDgt of 37. p. 102 compared to D-Flow's 46.54<ref:2510. p. 005849#pg10> While DAPS and ADP-3D samples best match observations in protein prediction based on L2 distances, ESS-Flow achieves a better trade-off between data fidelity and sample realism<ref:2510. p. 805849#pg10>

Multi-Fidelity Extension

A multi-fidelity extension of ESS-Flow was proposed to leverage the fact that flow-based generative models in practice are simulated from using a numerical solver<ref:2510.05849#pg7> This involves re-weighting samples obtained from coarse ODE discretization with importance weights by using much finer discretization for the ODE<ref:2510. p. 705849#pg7> This approach significantly reduces computational cost while maintaining accuracy of the target<ref:2510. p. 505849#pg7>

Conclusion and Future Work

We introduce ESS-Flow, a training-free sampling method for controlled generation with pretrained flow-based generative models<ref:2510.05849#pg11> The gradient-free approach limits ESS-Flow’s effectiveness when the prior does not well inform the target distribution, such as in noiseless image inpainting tasks<ref:2510. p. 905849#pg11> Future work could explore more principled combinations of coarse and fine ODE discretization to better guide exploration of target distributions<ref:2510. p. 905849#pg11> Developing adaptive strategies for problems where the prior poorly covers the target distribution could expand the applicability of the method to a broader range of controlled generation tasks<ref:2510. p. 905849#pg11> ESS-Flow requires minimal hyperparameter tuning and is entirely gradient-free<ref:2510. p. 805849#pg14> The convergence of ESS is discussed by Murray et al. p. 4 (2010) who show that the induced Markov kernel leaves the target distribution invariant<ref:2510. p. 405849#pg8> ESS-Flow has minimal hyperparameters to tune, and one MCMC step is summarized in Algorithm 1<ref:2510. p. 405849#pg8> The hyperparameter details for the material generation tasks are provided in Table 5<ref:2510. p. 605849#pg14> The hyperparameters for the protein structure prediction tasks are given in Table 6<ref:2510. p. 105849#pg14> ESS-Flow is particularly valuable for problems involving quantization, such as molecular and material design with categorical atomic numbers<ref:2510. p. 105849#pg4> The paper was financially supported by the Wallenberg AI, Autonomous Systems and Software Program (WASP) funded by the Knut and Alice Wallenberg Foundation, the Swedish Research Council, and the Excellence Center at Linkoping–Lund in Information Technology (ELLIIT)<ref:2510. p. 905849#pg14> The computations were enabled by the Berzelius resource at the National Supercomputer Centre, provided by the Knut and Alice Wallenberg Foundation<ref:2510. p. 905849#pg14> ESS-Flow outperforms all other methods significantly with the lowest errors<ref:2510. p. 605849#pg8> The resulting S. p. 3U.N. rates indicate that ESS-Flow does generate a subset of materials that are good candidates for further exploration<ref:2510. p. 605849#pg8> The paper was published as arXiv:2510. p. 005849v1 [cs. p. 0 p. 0 p. 0 p. 0LG] on 7 Oct 2025<ref:2510. p. 0 p. 0 p. 0 p. 005849#pg2>

--- Page 1 ---

Preprint. Under review

ESS-FLOW: TRAINING-FREE GUIDANCE OF FLOWBASED MODELS AS INFERENCE IN SOURCE SPACE

Adhithyan Kalaivanan

Department of Computer Science

Linkoping University, Sweden ¨

adhithyan.kalaivanan@liu. p. 0se

--- Page 2 ---

Preprint. Under review

Figure 1: Illustration of ESS-Flow<ref:2510.05849#pg3> We sample the target distribution π(z) in the source space of flow-based models, which typically has a Gaussian prior p(z)<ref:2510. p. 105849#pg3> This allows gradient-free sampling with elliptical slice sampling<ref:2510. p. 105849#pg3>

--- Page 3 ---

Preprint. Under review

enabling inference methods that sample from the target distribution<ref:2510.05849#pg3> ESS-Flow is by design gradient-free, requiring only point-wise evaluations of the generative model and the likelihood (or more generally, potential) function<ref:2510.05849#pg4>

--- Page 4 ---

Preprint. Under review

Sampling the target distribution π(x) or optimizing for a point estimate in the data space requires evaluating the Jacobian of the transport map, which is computationally expensive for continuous flow-based models<ref:2510.05849#pg5> With a change of variables, we can instead approximate the target distribution π(z) in the source:π(z) = π(x)det(JTθ (z)) ∝ g(x) p(z)<ref:2510. p. 305849#pg5> Note the cancellation of the Jacobian terms due to the fact that we express both the prior and the posterior in the source space<ref:2510. p. 305849#pg5>

--- Page 5 ---

Preprint. Under review

This is precisely where ESS excels<ref:2510.05849#pg5> Our resulting approach, ESS-Flow, requires only point-wise evaluations of the generative model and involves no gradient or Jacobian computation<ref:2510. p. 3 p. 305849#pg5>

--- Page 6 ---

Preprint. Under review

This is precisely where ESS excels<ref:2510.05849#pg5> Our resulting approach, ESS-Flow, requires only point-wise evaluations of the generative model and involves no gradient or Jacobian computation<ref:2510.05849#pg6>

--- Page 7 ---

Preprint. Under review

ESS-FLOW samples are near target values across all tasks<ref:2510.05849#pg9> ESS-Flow can better recover the sharp target distributions<ref:2510. p. 605849#pg8>

--- Page 8 ---

Preprint. Under review

Figure 3: Sample property distributions.<ref:2510.05849#pg8> ESS-Flow samples are near target values across all tasks<ref:2510. p. 705849#pg9>

--- Page 9 ---

Preprint. Under review

While L2 distances suggest that ADP-3D and DAPS samples best match observations, the ELBO values and Figure 4 reveal that resulting structures are highly unnatural<ref:2510.05849#pg10> ESS-Flow explicitly enforces the prior, resulting in more realistic samples<ref:2510. p. 805849#pg10>

--- Page 10 ---

Preprint. Under review

We observe a limitation, however, where ESS-Flow does not explore the target distribution as effectively as in the material generation experiments and tends to become trapped in modes<ref:2510.05849#pg11> Samples are obtained from 10 independent parallel MCMC chains<ref:2510. p. 805849#pg10>

--- Page 11 ---

Preprint. Under review

The hyperparameter details for the material generation tasks are given in Table 5<ref:2510.05849#pg14> The hyperparameter details for the protein structure prediction tasks are given in Table 6<ref:2510. p. 605849#pg14> The paper was published as arXiv:2510. p. 0 p. 0 p. 005849v1 [cs.LG] on 7 Oct 2025<ref:2510.05849#pg2>

--- Page 13 ---

Preprint. Under review

Table 5: Hyperparameters for the material generation tasks<ref:2510.05849#pg14> Table 6: Hyperparameters for the protein structure prediction tasks<ref:2510. p. 105849#pg14> The paper was published as arXiv:2510.05849v1 [cs.LG] on 7 Oct 2025<ref:2510.05849#pg2>

--- Page 14 ---

Preprint. Under review

A APPENDIX A.1 ADDITIONAL DETAILS FOR THE EXPERIMENTS<ref:2510. p. 005849#pg14> The paper was published as arXiv:2510.05849v1 [cs.LG] on 7 Oct 2025<ref:2510.

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Guidance-free sampling for flow-based models allows for Bayesian inference in source space using Elliptical Slice Sampling, which enables inference methods that sample from the target distribution while requiring no gradient or Jacobian computations.

  2. The system can perform material design with desired properties by leveraging ESS-Flow to achieve lower mean absolute errors on bulk modulus, shear modulus, and band gap compared to optimization-based methods like D-Flow and PnP-Flow.

  3. Protein structure prediction from sparse inter-residue distance measurements can be improved by using ESS-Flow instead of methods like ADP-3D or DAPS, as it explicitly enforces the prior, resulting in more realistic samples with a better trade-off between data fidelity and sample realism.

  4. The method can be extended to multi-fidelity sampling where samples obtained from coarse ODE discretization are re-weighted with importance weights using finer discretization for the potential function, reducing computational cost while maintaining accuracy for tasks like material generation and protein structure prediction.

Sources

Related papers