Unified Latents trains latent representations together with a diffusion prior and diffusion decoder. The method links reconstruction quality, bitrate, and generative quality through a single diffusion-based latent framework.