Diffusion Model
A diffusion model creates an image by gradually removing noise: it starts from random pixels and over dozens of steps turns them into a picture matching the request. Most image generators work this way.
How it works
In training, noise is added to real images step by step until only randomness remains, and the model learns to reverse the process, predicting what the frame looked like one step earlier. Having learned that, it can walk from pure noise to an image.
A text prompt steers the path. The description becomes a vector, and at every denoising step the model nudges the result toward matching it. Which is why the same wording with different starting noise yields different pictures.
The number of steps sets the trade-off between quality and speed: fewer steps are faster and rougher, more are slower and finer.
What to keep in mind
- It does not copy picturesIt absorbed patterns rather than storing source images, though overfitting can reproduce individual fragments almost verbatim.
- Text inside images is hardLetters are just another texture to such a model, which is why captions often come out garbled.
- Rights to the output are unsettledThe legal status of generated images and of training data differs by jurisdiction and keeps changing.
- The same prompt, different resultsReproducibility comes from fixing the starting noise, not from the wording of the prompt.