Question 1

True or false: Diffusion models generate images by starting from pure noise and iteratively denoising it, similar to reversing a video of a sandcastle being blown away.

Accepted Answer

True. That analogy is spot-on: the forward process gradually erodes structure (like wind on sand), and the model learns to rebuild it step by step.

Question 2

True or false: Unlike GANs, diffusion models always produce blurry or low-resolution images due to their iterative noise-removal process.

Accepted Answer

False. Modern diffusion models, like Stable Diffusion, can generate sharp, high-resolution images—often surpassing GANs in fidelity and diversity.

Question 3

True or false: Training a diffusion model typically requires only a single forward pass through the data, making it faster than training a GAN.

Accepted Answer

False. Training involves thousands of timesteps per image, so it's computationally heavy—far slower than GAN training, though inference can be fast with shortcuts.

Question 4

True or false: The 'noise schedule' in a diffusion model determines how much noise is added at each step, but has no effect on the final image quality.

Accepted Answer

False. The noise schedule critically affects sample quality—poor schedules can cause mode collapse or blurry outputs, so it's carefully tuned.

Question 5

True or false: Diffusion models are inherently deterministic—given the same random seed and input noise, they will always produce identical outputs.

Accepted Answer

False. While sampling can be made deterministic (e.g., with DDIM), the standard stochastic process introduces randomness at each step, yielding varied results.

Question 6

True or false: The first successful diffusion model for image generation was published in 2015, years before the famous DDPM paper.

Accepted Answer

True. Sohl-Dickstein and colleagues introduced the concept in 2015, but it was the 2020 DDPM paper that made it practical and popular.

Question 7

True or false: Diffusion models can be used for tasks beyond image generation, including text-to-speech, protein folding, and time series forecasting.

Accepted Answer

True. The framework is domain-agnostic; applications now span audio (e.g., WaveGrad), molecular design, and even weather prediction.

Question 8

True or false: Diffusion models were originally inspired by non-equilibrium thermodynamics, specifically the physics of particle diffusion.

Accepted Answer

True. The original 2015 paper by Sohl-Dickstein et al. explicitly used concepts from thermodynamics to gradually destroy and then reverse data structure.

Diffusion model Trivia Questions