We describe an end-to-end trainable model for image compression based onvariational autoencoders. The model incorporates a hyperprior to effectivelycapture spatial dependencies in the latent representation. This hyperpriorrelates to side information, a concept universal to virtually all modern imagecodecs, but largely unexplored in image compression using artificial neuralnetworks (ANNs). Unlike existing autoencoder compression methods, our modeltrains a complex prior jointly with the underlying autoencoder. We demonstratethat this model leads to state-of-the-art image compression when measuringvisual quality using the popular MS-SSIM index, and yields rate-distortionperformance surpassing published ANN-based methods when evaluated using a moretraditional metric based on squared error (PSNR). Furthermore, we provide aqualitative comparison of models trained for different distortion metrics.

Variational image compression with a scale hyperprior

Scan to connect with one of our mobile apps

Coinbase Wallet app

Coinbase app

Or try the Coinbase Wallet browser extension