During training, the model sees real images that have been gradually covered in static, and it learns to predict how to undo a little of that noise at each step. To generate something new, it starts from pure static and repeats the clean-up many times. A text prompt such as “a red bicycle on a beach at sunset” steers each step, which is how many text-to-image tools work.
Diffusion models are a different design from the large language models behind chatbots, which write text one token at a time. Both are kinds of generative AI.
Why it matters
Diffusion models made high-quality AI image generation practical and now appear in design, marketing and video tools. The many clean-up steps also make them comparatively slow, so speeding them up is an active area of work.
No published Chain of Thought episode discusses diffusion models directly yet, so this entry relies on the research papers below.
Related terms
A generative adversarial network is an earlier approach to image generation. Generative AI is the broader category.
Sources
- Ho, Jain and Abbeel: Denoising Diffusion Probabilistic Models — The 2020 paper that showed high-quality image synthesis by training a model to reverse a gradual noising process; checked October 24, 2026.
- Rombach et al.: High-Resolution Image Synthesis with Latent Diffusion Models — Introduces latent diffusion, the approach behind Stable Diffusion, which runs the denoising process in a compressed space and conditions it on inputs such as text; checked October 24, 2026.