This section demonstrates how training data is prepared. The process starts with a clean image that is progressively corrupted with noise. To see this in action, move the "Diffusion Step" slider.
Original Image (x0)
Noise Added (ε)
Corrupted Image (xt)
The model learns to reverse the noise patterns from the forward process. This section simulates that trained model at work. Click "Generate New Images" to watch it start with pure noise and iteratively create new images.
An "off-target" generation—a common challenge in AI.