LayerDiffusion is a method that enables large-scale pre-trained latent diffusion models to generate transparent images. This method allows for the generation of single transparent images or multiple transparent layers. It learns a 'latent transparency' by encoding the Alpha channel transparency into the latent space of a pre-trained latent diffusion model. By adjusting the added transparency as a latent offset, the method minimally alters the original latent distribution of the pre-trained model, preserving the production-ready quality of large diffusion models. By fine-tuning the latent space, any latent diffusion model can be converted into a transparent image generator. We trained the model on a dataset of 1 million transparent image layers collected through human-in-the-loop data gathering. We demonstrate that latent transparency can be applied to different open-source image generators or adapted to various conditioning systems, enabling applications such as foreground/background conditioned layer generation, joint layer generation, and content structure control of layers. User studies revealed that in most cases (97%), users preferred our locally generated transparent content over previous makeshift solutions like generating and then removing the background. Users also reported that the quality of our generated transparent images is comparable to real commercial transparent assets from sources like Adobe Stock.