Diffusion model

Diffusion models learn by watching images being destroyed. During training, noise is added to millions of images in small steps, and the model learns to predict and remove that noise. Once trained, it can run the process in reverse: start from pure noise and, step by step, remove it until a coherent image appears. A text prompt steers each step toward an image that matches the description, and a reference image, such as a sketch or model view, can constrain the composition.

How it shows up in a project

• It explains why results vary. Each run starts from different noise, set by the seed, so the same prompt gives different images.
• It explains why reference images matter. Without one, the model decides the geometry; with one, it follows your lines and volumes.
• The model reproduces patterns it has seen, so common building types and materials come out more reliably than unusual ones.
• Fine details such as text, exact counts of windows or precise dimensions are weaker points; check them against the drawings.
• Outputs are plausible images, not calculations, so they are not a substitute for daylight or lighting analysis.

In Rendair

You do not need to manage the model directly. Tools such as Sketch to Render and Text to Render wrap the process in a simple input, prompt and style. See AI rendering for how this approach compares with traditional ray-traced rendering.