Diffusion forward and reverse processes
the noise-and-denoise framing.
forward process gradually adds Gaussian noise until data is pure noise; reverse process learns to denoise step by step; the network predicts the noise added at each timestep.
What's really being asked
Whether you can separate the fixed corruption process from the learned generative process and state the concrete training objective.
The full answer
The forward process is a fixed schedule that gradually adds small amounts of Gaussian noise to a real image over many timesteps. It has no learnable parameters; the noise variances follow a predefined schedule. After enough steps the image is statistically indistinguishable from pure Gaussian noise. The reverse process is the generative direction: starting from random noise, the model takes many small denoising steps to walk back toward a clean sample. This reverse direction is what the neural network learns. In the standard formulation the network receives a noisy image and its timestep and is trained to predict the noise that was added, using a simple mean-squared error between predicted and actual noise. Predicting the noise is equivalent to estimating the score, the gradient of the log density, which tells the sampler which direction reduces noise.
The mistakes people make
Saying the forward process is learned; claiming the model predicts the final clean image directly in one step rather than the per-step noise; conflating diffusion with GANs and introducing a discriminator; forgetting that the timestep is an input so a single network handles all noise levels.
What usually comes next
Why train on a randomly sampled timestep each iteration? Why is predicting noise preferred over predicting the clean image? How does the reparameterization let you jump to any noisy timestep in one step? What is the link to score-based models?
A concrete example
Take a photo, pick timestep 400 of 1000, add the corresponding amount of Gaussian noise in closed form, and feed the noisy image plus the number 400 to the network. It outputs an estimate of the exact noise tensor that was added; the loss is the squared error to the true noise. At generation time you start from pure noise and repeatedly subtract the predicted noise, gradually revealing a coherent image.
Interview question
In a standard diffusion model, which component is learned by the neural network?
- a.The forward noise-adding schedule
- b.Neither; both are fixed analytical formulas
- c.Both forward and reverse processes jointly
- d.The reverse denoising processCorrect
Why? this is the answer
The forward process is a fixed, non-learned noise schedule, while the reverse denoising process is what the network learns. Claiming the forward process is learned is the most common misconception.
Just read this? Test yourself on what you have been reading.
Read the original → en.wikipedia.org
- #diffusion
- #generative-models
- #denoising
- #forward-process
- #reverse-process
Put your scrolling time to good use
Learn one idea, try a quiz and save useful cards for revision. Tezvyn makes it easy to learn and stay current in your tech field, a few minutes at a time.
The iPhone app is on the way
We are building it. Until it lands, nothing here is held back from you: every interview card, your saved cards, streaks and the job board all work in Safari, plus hundreds of free practice quizzes of thirty questions each. Sign in and it all carries over to the app the day it arrives.
Want it as an icon? Tap Share at the bottom of Safari, then Add to Home Screen. It opens full screen and the cards you have read stay available offline.
We are hiring for this. Every open role lists the topics its interview covers, so you can prepare for the real thing rather than guessing.
See open roles