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.
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.