An image pipeline starts from noise and repeatedly updates it towards an image described by a prompt.
In a basic diffusion setup, training uses examples with added noise. The model learns to predict a denoising update. During generation, a sampler repeatedly applies updates, guided by the model and conditioning inputs. Hugging Face introduces this process.
Practical pipelines can work in a compressed representation, called a latent, rather than directly editing ordinary pixels at every step. The model, sampler and associated encoders contribute to the pipeline, so “diffusion” does not name one app or image style.
Read the chosen model's step-count instructions. More steps are not a universal improvement: different models and sampling methods expect different configurations. When comparing results, preserve the model, dimensions, prompt and other settings so you can identify what changed. Inspect the image rather than assuming that a completed progress bar means a usable result.
Sources
Primary references checked 11 October 2026. This explanation is not a benchmark of a particular model or computer.