Denoising is the reverse-diffusion process: a model removes noise from an image estimate step by step until random noise becomes a clean image. Multiple inference steps make reconstruction more controlled, stable, and detailed than a single step. section 15.1.2 section 15.6
A denoising U-Net receives a noisy image and its time step, then predicts the noise present in that image. Because the noisy image combines clean content with noise, the prediction supports reconstructing the original. Its encoder, bottleneck, decoder, and skip connections capture features while preserving edges and textures. section 15.1.2 section 15.6
A practical generation loop proceeds as follows:
- Start with pure random noise at t = 1000. section 15.4.2
- Set a sequence of inference steps, such as 50 steps from t = 980 down to t = 0. section 15.4.2
- Store intermediate images to inspect how structure emerges before displaying the final clean result. section 15.4.2
Generating and inspecting flower images
A trained denoising U-Net can generate a batch of 10 flower images with a fixed random seed and 50 inference steps. Plotting selected states at t = 800, 600, 400, 200, and 0 shows the gradual transition from noise to recognizable flowers. Changing the seed produces a different batch. section 15.4.2
