Diffusion models (Stable Diffusion, DALL·E 3, Sora, Suno) generate content by reversing a noising process: starting from pure noise, a network predicts and removes a little noise at each step until a coherent sample emerges.
x(t) = (1 - α_t)·noise + α_t·target
α_t = (step / steps)^(1 / guidance)
- Diffusion steps — how many denoising iterations the sampler runs before the output is "final".
- Noise level — how far particles start scattered from the target shape (the latent's initial variance).
- Guidance scale (CFG) — classifier-free guidance strength; higher values snap particles to the target faster and more sharply, at the cost of variety.
- Denoising speed — how many steps the sampler advances per second.
Each particle is one point of the target's latent representation; α_t plays the role of the noise schedule that every real diffusion sampler (DDPM, DDIM, etc.) follows from pure noise to a clean sample.