Module 3: Image, audio and video models · 10 min
How image models work: diffusion
Most image generators use diffusion. The idea has two halves.
- Training: take real pictures and add a little random noise, again and again, until only noise is left. Teach a network to undo one step of noise at a time.
- Generating: start from pure random noise and let the network remove noise step by step, guided by your text prompt, until a clear picture appears.
The text prompt is turned into embeddings, and these steer the denoising so that the picture matches your words.
Because the process starts from random noise, the same prompt gives a different picture every time unless you fix the random seed.
← Pre-training, fine-tuning and alignment | Multimodal, audio and video AI →