The Denoising Loop

Reverse denoising runs the noising process backwards. At each step the denoiser — typically a U-Net or a diffusion transformer, a DiT — receives three inputs: the current noisy latent, the timestep, and the prompt conditioning embeddings from the text encoder. It outputs a prediction of the noise that was added. That prediction is subtracted from the noisy latent to produce a slightly cleaner one, and the loop repeats. The timestep matters because the amount of noise present changes how aggressively the model should correct. The prompt conditioning enters through cross-attention at every step, so the same prompt steers the latent from pure noise all the way to a coherent video.