Gaussian noise's role in DDPM training and sampling

What's actually being asked

Pin down specifically why Gaussian noise (not noise in general) is used in DDPM, then walk through the training/sampling mechanics that turn a noise-predicting network into a generator of clean samples.

Answer

Why Gaussian noise, specifically

Gaussian noise is used because it's mathematically tractable to compose: the sum of Gaussian-distributed steps is itself Gaussian. This single property gives DDPM two things it relies on throughout:

1. The forward process can jump directly from x_0 to any x_t with one closed-form expression, x_t = √(ᾱ_t) x_0 + √(1−ᾱ_t) ε, ε ~ N(0,I), where α_t = 1 − β_t, ᾱ_t = Π_{s=1}^{t} α_s, instead of needing to simulate t individual noising steps one by one.
2. The true reverse conditional q(x_{t-1}|x_t, x_0) is itself Gaussian with a known closed-form mean and variance (given the β_t schedule) — exactly the family of distribution the learned reverse model p_θ(x_{t-1}|x_t) is built to match.

The three roles Gaussian noise plays

• Corruption mechanism (forward process): at each forward step, q(x_t|x_{t-1}) = N(x_t; √(1−β_t) x_{t-1}, β_t I) — the previous sample is scaled down slightly and a controlled amount of Gaussian noise is added. Repeated over T steps with an increasing schedule β_t, this reliably drives any x_0 to an isotropic Gaussian x_T ~ N(0,I) — trivial to sample from at the start of generation.
• Training target: rather than asking the network to predict the clean image x_0 directly (a much harder, under-constrained regression), DDPM trains it to predict the exact noise ε that was mixed into x_0 to produce x_t. This is a simpler, better-behaved target, and it's the quantity the closed-form forward equation makes exactly knowable during training (we chose ε ourselves when constructing x_t).
• Sampling stochasticity (reverse process): after estimating ε_θ(x_t, t) and stepping from x_t toward x_{t-1}, small fresh Gaussian noise is injected back in at most steps (except typically the final one) to match the true stochastic reverse distribution — without it, the reverse chain would collapse toward a single deterministic, blurry "average" output rather than producing diverse, sharp samples.

How the model learns to generate clean samples from noisy data

Training: sample a real x_0 from the dataset, sample a random timestep t, sample ε ~ N(0,I), compute x_t via the closed-form equation above, feed (x_t, t) into the network (usually a U-Net with a time embedding), and minimize 𝔼[‖ε − ε_θ(x_t, t)‖²]. Across many (x_0, t, ε) combinations, the network learns to recognize "how much, and what shape of, noise is present" at every corruption level.

Sampling (generation): start from x_T ~ N(0,I), pure noise. At each step, feed (x_t, t) into the trained network to get ε_θ(x_t, t), use it to estimate the noise component and take a small step back toward x_{t-1} (adding fresh Gaussian noise back in except near the end), and repeat down to x_0. The final x_0 is a new, realistic sample, constructed entirely by iteratively undoing noise the network was trained to recognize — not by directly decoding anything.

Exam tip

If asked "why not just predict x_0 directly," the correct answer is that predicting the noise ε is a simpler, more consistent regression target across all timesteps t, and it's exactly the quantity known at training time from the closed-form forward equation — don't just say "it works better" without that reasoning.