SDXL is a latent diffusion model, where the diffusion operates in a pretrained, learned (and fixed) latent space of an autoencoder. While the bulk of the semantic composition is done by the latent diffusion model, we can improve local, high-frequency details in generated images by improving the quality of the autoencoder. To this end, we train the same autoencoder architecture used for the original Stable Diffusion at a larger batch-size (256 vs 9) and additionally track the weights with an exponential moving average (EMA). The resulting autoencoder outperforms the original model in all evaluated reconstruction metrics, see the table below.
baidu