Noise schedule

On this page 7

A noise schedule is the function that specifies how much Gaussian noise is present at each timestep of a diffusion process, from clean data at the start to almost pure noise at the end. In Generative AI, it is a core design choice of every diffusion model, because it determines which noise levels the network learns during training and which levels a sampler visits when it turns noise back into an image, audio clip or video.

The schedule is usually written in terms of a per-step variance beta, the cumulative signal factor alpha-bar, or the signal-to-noise ratio (SNR). These are equivalent descriptions of the same curve, and choosing that curve well is one of the cheapest ways to improve quality without changing the network.

How a noise schedule shapes the diffusion process

In the forward process, a clean sample is mixed with Gaussian noise so that a noisy version at timestep t equals the square root of alpha-bar times the data plus the square root of one minus alpha-bar times noise. The noise schedule fixes alpha-bar for every t, so it controls the balance between signal and noise throughout the process.

During training, the network sees inputs drawn at random timesteps and learns to predict the noise, the clean image or a velocity target. During generation, the reverse process walks back through the same noise levels. A schedule that spends too few steps at a critical noise range leaves the model undertrained exactly where detail is formed.

The quantities below are the standard ways to express a schedule. Converting between them is straightforward, but each makes different properties easy to see.

QuantityDefinitionWhat it reveals
Beta (per-step variance)Noise added at one stepLocal step size
Alpha-bar (cumulative)Product of one minus betaRemaining signal fraction
Signal-to-noise ratioAlpha-bar divided by one minus alpha-barDifficulty at each level
Sigma (noise scale)Noise standard deviation in variance-exploding formAbsolute noise magnitude
Log-SNRLogarithm of the signal-to-noise ratioEven spacing for comparison
Quantities used to describe a noise schedule

Common types of noise schedules

Several schedule families are widely used, and they differ mainly in how quickly they destroy information in the early and late parts of the process. The original DDPM work used a linear beta schedule with 1,000 steps, beta rising from 0.0001 to 0.02.

The cosine schedule was introduced because the linear one destroys low-resolution images too fast, wasting many late steps on near-pure noise. Later, the EDM formulation separated training noise levels from sampling noise levels, sampling sigma on a log-normal distribution for training and on a power-law spacing for inference.

ScheduleShapeTypical use
Linear betaBeta grows linearly with tDDPM, early latent diffusion
Scaled linearLinear in square root of betaStable Diffusion 1.x and 2.x
CosineAlpha-bar follows a cosine curvePixel-space models, small images
Karras (EDM) sigmasPower-law spacing of sigmaModern samplers and fine-tuning
Shifted or logit-normalEmphasis on middle noise levelsFlow matching, high resolution
Main noise schedule families

The term noise schedule is often used loosely, and it helps to separate the training schedule from the sampling schedule. The first defines which noise levels the model learns; the second defines which levels a diffusion sampler visits at inference, often far fewer than during training.

TermWhat it controlsRelation to the schedule
Noise scheduleNoise level per timestepDefines the path itself
Diffusion samplerNumerical method for the reverse pathFollows a chosen schedule
Sampling stepsNumber of evaluations at inferenceDiscretizes the schedule
Denoising strengthStarting noise level in image-to-imagePicks an entry point on it
Classifier-free guidanceStrength of prompt conditioningIndependent, but interacts
Noise schedule versus neighboring terms

In flow matching, the same role is played by the interpolation path and the timestep sampling distribution. Rectified flow models therefore still need a schedule-like decision, usually a resolution-dependent shift toward noisier timesteps.

Choosing and tuning a noise schedule

The right schedule depends on resolution, data domain and sampler budget. Larger images contain more redundant pixels, so the same nominal noise level leaves more recognizable structure; schedules must be shifted toward higher noise as resolution grows, or the model never learns to form global composition.

SituationRecommended choiceReason
Training at higher resolutionShift SNR toward more noisePreserves learning of global layout
Few-step inferenceKarras or similar spacingConcentrates steps where detail forms
Very dark or bright images neededZero terminal SNRRemoves residual signal at the last step
Fine-tuning an existing modelKeep the original scheduleAvoids mismatch with learned weights
Latent model with unusual scalingRescale latents firstSchedule assumes unit variance
Decision guide for schedule choices

Tuning is validated by sample quality metrics such as FID together with visual inspection across prompts. Small changes to the endpoints of the schedule often matter more than changes to its overall shape.

Limitations and common pitfalls

A frequent defect is a non-zero terminal signal-to-noise ratio. Many older schedules leave a trace of the original image at the final timestep, so the model never learns to start from pure noise and cannot generate very dark or very bright scenes, drifting toward medium brightness instead.

  • Training and sampling mismatch: a sampler that visits noise levels the model rarely saw produces artifacts.
  • Resolution transfer without a shift: a schedule tuned at 256 pixels oversmooths results at 1,024 pixels.
  • Confusing steps with quality: more sampling steps cannot repair a poorly chosen schedule.
  • Ignoring parameterization: noise, clean-data and velocity prediction weight timesteps differently.

A non-technical view

Image generators that use diffusion learn by watching pictures dissolve into static and then practicing how to reverse it. The noise schedule is the plan that says how fast the picture dissolves at each step.

Think of slowly fogging a window. If the fog thickens too fast, the learner never sees the stage where shapes are still faintly visible. If it thickens too slowly, time is wasted on views that barely change. A good schedule spends the most practice where the picture is hardest to recover.

A concrete example

A research team trains a latent diffusion model for product photography and moves from 512 to 1,024 pixel training images. After the switch, generated scenes look sharp in detail but show broken layouts, with objects floating or duplicated.

The team diagnoses that the schedule tuned for 512 pixels leaves too much structure visible at high noise levels. They shift the timestep distribution toward noisier levels by a factor that roughly matches the fourfold increase in pixel count, and they rescale the schedule to reach zero terminal SNR.

After retraining for the same number of steps, layout errors in a 500-prompt evaluation set drop sharply, and the model can now produce true black backgrounds that the previous version rendered as dark gray.