The Future of Tractable Deep Generative Models

PyJuice is a new system for training and inference of probabilistic circuits that outperforms prior baselines in speed, memory efficiency, and reproducibility. Benchmarking on PC and HCLT structures across datasets like ImageNet32, PyJuice consistently delivers stronger results—achieving 4.33 bpd vs. 4.82 bpd reported in past work. While not aiming for record-breaking scores, the framework establishes reproducible baselines and opens the door for scalable research on tractable deep generative models.