Main Concept

Diffusion models are a type of generative AI model that produces images by learning to reverse a noise-adding process. They are trained to understand how to reconstruct a clean image from pure noise, guided by a text prompt. Well-known examples: DALL-E, Stable Diffusion, and Amazon Nova Canvas.

Two Key Processes

1. Training (Forward Diffusion)

  • Start with a real image (e.g., a photo of a cat)
  • Gradually add random noise in small steps
  • The image becomes progressively more distorted
  • End result: pure noise
  • What the model learns: what noise was added at each step

2. Generation (Reverse Diffusion)

  • Start with pure random noise
  • Provide a text prompt (“a cat with a computer”)
  • The model removes noise step-by-step, guided by the prompt
  • End result: a new image that matches the description

Simple Analogy

Training: Learn how to destroy a sculpture step-by-step. Generation: Recreate a sculpture by reversing the destruction process.


AIF-C01 Exam Relevance

Diffusion models are explicitly listed in the exam guide (Domain 2) as a generative AI model type to know. Key points:

  • They are used primarily for image generation (not text)
  • They are a separate architecture from Transformer-based LLMs
  • AWS offers diffusion-based image generation through Amazon Nova Canvas (via Amazon Bedrock) and Amazon Titan Image Generator

Exam tip: If a question involves generating images from text prompts, the underlying model type is a diffusion model — not a transformer. The transformer handles text; diffusion handles images.