bayesflow

Why we included this project

If your research or product depends on simulators that don't have a closed-form likelihood, BayesFlow is worth a look. It trains neural networks to do Bayesian inference: you train once on simulated data, then get fast posterior estimates, model comparisons, and parameter point estimates on demand, instead of running MCMC on every new dataset. The library includes a range of generative models, from diffusion to consistency models, and runs on Keras 3, so you can use PyTorch, TensorFlow, or JAX as the backend. The workflow API is straightforward: define a simulator, pick an inference network, and call fit_online to get diagnostics. Researchers working with epidemiological or cognitive models, where likelihoods are often intractable, will find the tutorials and worked examples a practical starting point.