Diffusion Model
An image generator that learns to reverse a process of adding noise.
A generative model trained by progressively adding noise to data and learning to undo each step. To generate, it starts from pure noise and denoises, guided by a text prompt.
This is a different family from the transformer-based language models. Image and video generators are usually diffusion; language models are autoregressive.
The mental model worth keeping is "learned reversal of a corruption process". Training shows the model many levels of noise and asks it to undo one step; generation runs that in reverse from pure noise. Text-to-image is that same loop with the prompt biasing each denoising step.
Where images come from in an agent pipeline
Generating diagrams for a document that an agent is writing.
The agent produces the spec, the diffusion model produces the pixels. Keep them separate in your architecture: image models have no notion of your facts, so the text has to carry the meaning and the image is decoration unless a human checks it.
- paperGemini: a family of highly capable multimodal models ↗
A natively multimodal model family — the technical counterpart to "multimodal model" as a product category.
- guideIntroduction to large language models ↗
A plain-language official introduction to models, tokens and context. Teaching material rather than origin research — good for the definition, not for a claim about the frontier.
Multimodal
A model that handles more than text — images, audio, video, screen pixels.
Foundation Model · Base model
A large model trained broadly, meant to be adapted to many downstream tasks.
Non-determinism
The same input can produce different output, and that is by design.