Train once. Infer on repeat.
Turn your simulator into a fast inference engine. Train on simulations, get posteriors for new datasets without refitting, and check your results with built-in diagnostics.
Research papers
Supporting methodological research and applications across scientific disciplines.
Inference pipelines
Combine generative heads, embeddings, and data types for any Bayesian task. [4]
Diagnostics utilities
Built-in plots and metrics for everything you need to trust your inference. [1]
Bring your simulator.Train your estimator.
If you can simulate from your model, BayesFlow can learn to estimate its parameters, even when the likelihood is intractable.
Lotka–Volterra predator–prey model
Prey growth
α = 1.00
Hunting pressure
β = 0.90
Predator decline
γ = 0.80
Predator growth
δ = 0.90
Dots are noisy observations; the network sees only these.
One BayesFlow.Three backends.
BayesFlow is part of the 
eras ecosystem. Write your workflow, then train it with your choice of PyTorch, JAX, or TensorFlow.
Get the posterior.Then check it.
An amortized posterior is only as good as its training. BayesFlow provides built-in diagnostics so that you can inspect the accuracy and calibration of your trained estimator.
32,768 training simulations · 512 posterior draws per validation dataset
Documentation & evidence
- BayesFlow documentation
Workflows, networks, diagnostics, and backends.
- Diagnostics guide
Recovery, simulation-based calibration, and contraction.
- Lotka-Volterra inference example
The tutorial behind the demo on this page, which uses its own trained network.
- BayesFlow 2 paper
Framework design, inference workflows, multi-backend support, and new capabilities.