The correct answer is Diffusion models synthesize images by starting from pure noise and repeatedly denoising with a learned reverse process.

Diffusion models learn a forward process that gradually corrupts training images with noise and a learned reverse process that removes noise step by step so new images can be sampled by starting from random noise and iteratively denoising. The reverse model is trained to predict either the less noisy image or the noise itself and that training objective lets the model generate diverse and high fidelity concept art from scratch.

Diffusion models search the training corpus for the closest visual matches using vector similarity and return them is incorrect because that describes retrieval where examples are fetched instead of being synthesized from a learned probability distribution.

Diffusion models refine pictures by applying supervised filters that accentuate boundaries between pixels is incorrect because that describes enhancement of existing images rather than the generative denoising process that creates new images from noise.

Diffusion models build images primarily through an encoder-decoder with attention layers is misleading because some diffusion architectures incorporate encoder-decoder blocks and attention but the defining mechanism is the iterative denoising reverse process rather than any single encoder-decoder design.

Full AWS Practitioner Certification Question