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.