NCA-GENM - Software Development - Section 4.1

Generate images from pure noise.

Explain how diffusion models iteratively denoise a Gaussian noise tensor through a learned reverse process to synthesise a coherent image. Recognise the role of the noise schedule and the denoising network in converting pure noise into a structured output.

Diffusion models

Practice question for this objective

Free sampleSoftware Developmentmedium

In a denoising diffusion probabilistic model, the forward process and the reverse process play distinct roles. Which statement correctly characterises the difference between the two processes?

  • AThe forward process is a learned encoder that compresses data into a latent vector, while the reverse process is a fixed decoder that reconstructs the original sample deterministically.
  • BBoth the forward and reverse processes are learned jointly, with the forward network predicting noise and the reverse network predicting the clean image at each step.
  • CThe forward process is a fixed, non-learned procedure that gradually adds Gaussian noise to data, while the reverse process is a learned model that progressively removes noise to recover a sample. Correct
  • DThe forward process removes noise using a fixed schedule, while the reverse process is a learned model that adds structured noise to increase sample diversity.
Understand that diffusion's forward noising process is fixed while only the reverse denoising process is learned. A DDPM defines the forward process as a fixed Markov chain that adds Gaussian noise according to a preset variance schedule, so it has no parameters. The network is trained only on the reverse process, learning to predict and remove the noise at each timestep, which lets it generate images starting from pure noise.

Why A is wrong: Tempting because it echoes the encoder-decoder framing of an autoencoder, but it is wrong: diffusion's forward step is fixed and adds noise rather than being a learned compressor, and the reverse step is the learned component.

Why B is wrong: Tempting because diffusion does involve noise prediction, but it is wrong: only the reverse process is learned, and the forward process carries no trainable parameters at all.

Why C is correct: Correct: the forward (diffusion) process applies a predefined noise schedule with no trainable parameters, and the reverse process is the neural network trained to undo each noising step, turning pure noise into a coherent image.

Why D is wrong: Tempting because it names the right components, but it inverts their roles: the forward process adds noise and the reverse process removes it, not the other way around.

See more NCA-GENM practice questions, answers explained.

More in this domain

Back to all Software Development objectives, or the NCA-GENM cert hub.

Examworthy is not affiliated with or endorsed by NVIDIA. Original, blueprint-aligned practice material only.