No, there are publicly available models that are about this good at composition with just text prompts. Having an image prompt makes things significantly easier. Their model might be doing some depth masking, but it doesn't look like it from the examples I've seen.