Diffusion inference iteratively removes predicted noise from a latent tensor until an image emerges

At inference the diffusion process starts from a random noise tensor (latents). A trained noise predictor (UNet) receives the noisy tensor plus a step number and predicts a slice of noise; subtracting it yields a slightly cleaner tensor whose latents ‘better resemble the input text’. This repeats over many steps — the ‘steps’ parameter in SD interfaces, often defaulting to 50 or 100. Each step adds more relevant information; in the article’s visualization the outline emerges from noise around steps 2–4. The scheduling algorithm controls how much is removed per step. After the specified steps, the autoencoder decoder (which runs only once, at the end) converts the final latents to pixels.

Examples

steps=20: fast, less detail. steps=50: standard. steps=100: slower. In Alammar’s denoising visualization, ‘the outline emerges from the noise’ between steps 2 and 4.

Assessment

Explain why reducing the steps parameter tends to degrade image quality, and describe what each denoising step does to the latents array. Use a 50-step run as the example.