Yes, it's mostly a new training technique (that is impressive: "PIXART-α only takes 10.8% of Stable Diffusion v1.5's training time"). I'm not really sure if it really improves the image quality over SDXL, but it may a bit: https://pixart-alpha.github.io