All concepts
Diffusion Denoising
Learn to reverse a gradual noising process to create samples.
Deep Learning · Intermediate · ~8 min
In plain English
Teach a model to remove a little noise from a picture. Then hand it pure static and ask it to clean it up, over and over, until an image appears.
Why it's worth your time
It's how essentially every modern image and video generator works, and the sampling loop is the part people get wrong in interviews.
If you remember three things
- Forward process adds noise on a schedule; the model learns to reverse one step
- The model predicts the noise, not the image
- Conditioning (text, image) steers each denoising step
Overview
Generative models that learn to reverse a gradual noising process. A forward process corrupts data with Gaussian noise over many timesteps; a network learns to predict that noise, and at inference you iteratively denoise pure noise into a realistic sample. Powers modern image and audio generation.
How it works
- Start: Clean Data Training starts with real images, audio, or latents.
- Clean Data -> Add Noise A forward process corrupts the sample over timesteps.
- Add Noise -> Predict Noise The model learns to estimate the noise component given the noisy sample and timestep.
- Predict Noise -> Denoise Step At inference, repeatedly subtract predicted noise.
- Denoise Step -> Generated Sample After many steps, pure noise becomes a realistic sample.
In an interview
Diffusion models generate by learning to undo noise. During training you add noise to real data across timesteps and train a network to predict the noise; at inference you start from pure noise and repeatedly subtract the predicted noise to recover a clean sample. The training target εθ(x_t, t) makes the loss a simple noise-regression.
Production defaults
- Steps
- 20–50 with a modern sampler. 1000-step sampling is a training-era artifact
- Guidance
- CFG scale 5–8. Higher means more prompt-obedient and less diverse; too high looks burnt
- Adapt it
- LoRA on the attention layers for style; ControlNet when you need structural control
What breaks
- Outputs look oversaturated and harsh — Guidance scale too high. Drop it, or use a sampler with rescaling.
- Generation is far too slow — Step count. Distilled or consistency models get usable results in 1–4 steps.