A noise schedule is a recipe for how much noise enters an image at each diffusion timestep. It controls how the image gradually changes from clear content to near-pure noise, and how the reverse process removes that noise. section 12.28 section 12.9
The schedule is built from related values:
- Beta specifies the noise amount added at an individual timestep, usually increasing from a small starting value to a larger final value. section 12.9
- Alpha equals 1 minus beta and represents the signal retained at that timestep. section 12.9
- Alpha-bar is the cumulative product of the alpha values, representing how much original signal remains after several steps. section 12.9
- Linear schedules are simple, cosine schedules degrade images more gently at first, and sigmoid or custom schedules trade flexibility for tuning difficulty. section 12.28
A three-step intuition
Suppose beta values begin at 0.00010, 0.00012, and 0.00014. Alpha becomes 0.99990, 0.99988, and 0.99986, while alpha-bar multiplies these values together. Forward diffusion then combines the original image, scaled by the square root of alpha-bar, with random noise, scaled by the square root of one minus alpha-bar. During denoising, the same schedule constants determine how much predicted noise to subtract. section 12.9 section 12.25
Table 12.1 A subset of values for the noise schedule
|
Timestep (t) |
Beta |
Alpha calculation (1 - beta) |
Alpha |
Alpha-Bar calculation |
Alpha-Bar |
| 1 |
0.00010 |
1 - 0.00010 |
0.99990 |
0.99990 |
0.99990 |
| 2 |
0.00012 |
1 - 0.00012 |
0.99988 |
0.99990 × 0.99988 |
0.99978 |
| 3 |
0.00014 |
1 - 0.00014 |
0.99986 |
0.99990 × 0.99988 × 0.99986 |
0.99964 |
| … |
|||||
| 1000 |
0.02000 |
1 - 0.02000 |
0.98000 |
0.00004 |
|
Use the schedule table to trace beta, alpha, and alpha-bar across early and late timesteps. section 12.9
