ESS-Flow: Training-free guidance of flow-based models as inference in source space

2025-10-07

AI Summary

This work addresses the challenge of enabling attribute-guided generation from pretrained flow-based generative models without fine-tuning or gradient computation. To this end, the authors perform Bayesian inference directly in the latent (source) space, leveraging the inherent Gaussian prior structure of flow models to achieve observation modeling and sampling via forward passes only. A key methodological contribution is the introduction of a gradient-free elliptical slice sampling algorithm, which overcomes optimization bottlenecks arising from unreliable or absent gradients. The framework is empirically validated on materials design and protein structure prediction tasks: it accurately reconstructs 3D protein structures from sparse residue distance constraints and efficiently generates material samples satisfying target physical or chemical properties—demonstrating strong performance without backpropagation or model adaptation.

Technology Category

• Machine Learning: Deep Generative Models & Autoencoders
• Search and Optimization: Sampling/Simulation-based Search
• Computer Vision: Generative Adversarial Networks (GANs) for Vision

Application Category

• Graph Algorithms and Modeling for the Web: Foundation models and LLMs for Web-related graphs
• Economics, Online Markets and Human Computation: Economic ramifications for generative AI infrastructure and applications
• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking

Abstract

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. We present 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. ESS-Flow only requires forward passes through the generative model and observation process, no gradient or Jacobian computations, and is applicable even when gradients are unreliable or unavailable, such as with simulation-based observations or quantization in the generation or observation process. We demonstrate its effectiveness on designing materials with desired target properties and predicting protein structures from sparse inter-residue distance measurements.

Problem

• Guiding flow-based models for conditional generation without retraining
• Performing Bayesian inference in source space using Gaussian prior
• Enabling gradient-free sampling with simulation-based observations

Innovation

• Bayesian inference in source space via Elliptical Slice Sampling
• Gradient-free guidance using only forward model passes
• Applicable to simulation-based and quantization observation processes