<ul><li>Normalizing flows are powerful generative models due to their theoretical rigor and analytical log-likelihood computation.</li><li>Autoregressive modeling has improved the expressive power of normalizing flows but has limitations in parallel computation during inference, resulting in slow generation.</li><li>A new method called Selective Jacobi Decoding (SeJD) is proposed to accelerate autoregressive inference through parallel iterative optimization.</li><li>Empirical evaluations show that the SeJD strategy can achieve up to 4.7 times faster inference while maintaining generation quality and fidelity.</li></ul>

Inference Acceleration of Autoregressive Normalizing Flows by Selective Jacobi Decoding

Discover more