During the forward diffusion process of a DDPM, a practitioner wants the model to spend proportionally more training time on moderately noisy images rather than near-pure-noise states. Which noise schedule property best achieves this?
- AA cosine noise schedule defined by the signal-to-noise ratio following a cosine curve, which keeps images perceptually structured for more timesteps before rapid noise saturation. Correct
- BA linear beta schedule that increments the noise variance by a constant amount at each timestep from a small starting value to a maximum near one.
- CA quadratic beta schedule that increases variance slowly at first and rapidly at the end, compressing most variance growth into the final timesteps.
- DA sigmoid beta schedule that applies maximum variance in the early timesteps and tapers off, ensuring clean images receive the greatest denoising signal.
Why A is correct: The cosine schedule, introduced to address linear schedule deficiencies, defines alpha-bar values along a cosine curve so the signal-to-noise ratio decays slowly through intermediate timesteps and only reaches near-zero close to T, naturally concentrating training effort on the perceptually rich moderate-noise regime.
Why B is wrong: A linear schedule allocates equal variance increments across all timesteps, meaning it devotes as many steps to the near-pure-noise end as to the structured intermediate range, which does not concentrate training on moderately noisy images.
Why C is wrong: A quadratic schedule accelerates variance growth toward the end of the chain, which concentrates steps near the heavily corrupted end rather than in the intermediate regime, worsening the imbalance the practitioner wants to correct.
Why D is wrong: Applying maximum variance early would corrupt clean images aggressively in the first few steps, destroying semantic structure before the model has learned to handle it, and does not correspond to any established noise schedule used in diffusion model literature.