It's largely an artifact of classifier-free guidance used in diffusion models. It makes the image generation more closely follow the prompt but also makes everything look more saturated and extreme.