GCP Generative AI Leader Practice Question 22: Diffusion Model Image Generation
Google Cloud Generative AI Leader · Test 3 · question 2 of 10
A product manager is reading a vendor briefing on how text-to-image models such as Imagen produce a picture from a written prompt. The briefing describes the model as being trained by gradually corrupting real images with random noise over many steps, then learning to reverse that process.
What is this training approach called, and how does the trained model generate a new image?
Show answer and explanation
Answer:
This is a diffusion model. During training, the model learns to reverse a process where noise is added to real images in small increments until only noise remains. At generation time, the model runs that learned process backward: it starts from pure random noise and, guided by the text prompt, removes a little noise at each step until a realistic image emerges. This is the technique behind Google's Imagen and other modern text-to-image systems.
Option A describes reinforcement learning, a technique used elsewhere in gen AI (for example to align chat models with human preferences), but it is not how diffusion image models generate pixels.
Option B describes a form of supervised super-resolution training, not the noise-to-image process that defines diffusion models.
Option C describes retrieval, which returns an existing item rather than generating a new one. Diffusion models synthesize a novel image; they do not look one up in a library.
Reference: Introduction to Image Generation