Forward diffusion explorer

Forward diffusion adds noise to a picture in small steps until nothing is left. A real model learns to run this backwards; here you can watch the forward half and see why the schedule matters. Everything runs in your browser: the bundled photo and any picture you load stay on your device and are never uploaded.

Image

Non-square pictures are centre-cropped to a square. Or use the bundled sample image below. Loading a new image resets the step to 0 and generates a fresh noise field.

Diffusion

The denoise toggle applies a small box blur to the noised image as a rough illustration of "removing noise". It is not a trained model and does not reconstruct the original picture.

Step 0. alpha-bar = 1.000. Signal-to-noise ratio: +infinity dB.

Image at step t

alpha-bar(t) for the chosen schedule

What to notice: at step 0 the image is unchanged; by step 1000 it is indistinguishable from static. The cosine schedule keeps more signal in the middle steps than the linear schedule, so a cosine-trained model spends more of its steps on fine detail.