PixelCNN++ enhancements add discretized logistic mixture likelihood and related modifications

AI Impact Summary

PixelCNN++ is being updated to use a discretized logistic mixture likelihood with additional modifications, aiming to better capture the distribution of pixel values and improve density estimates. This can translate into higher-quality samples and more accurate likelihoods for image datasets, benefiting applications in image synthesis and compression that rely on autoregressive models. Teams should anticipate changes to the training objective and hyperparameters (e.g., number of mixture components, discretization handling) and potential impacts on training time and memory usage; existing code may require updates to accommodate the new likelihood formulation and sampling pipeline.