You example is perhaps misleading. I tried to reproduce it using stable diffusion 1.5 and it’s possible using negative weights and some creative keywords but I doubt the title of the post is the actual prompt.
I can get some thing like this by adding context distractions. For example, "bear eating salmon in river" shows whole fishes, but "bear wearing tophat eating salmon in river" gives a bear (without a tophat usually) eating supermarket sliced salmon in a river.