DDPM: The Linear Noise Schedule
Intermediate
The problem, from first principles
DDPM uses a fixed schedule of noise levels. A linear schedule of small variances, accumulated as a cumulative product, gives the fraction of the original signal left at each step.
From theory to code
Implement alpha_bars(T, b0, b1), the cumulative product of 1 - beta_t for a linear schedule from b0 to b1.
Constraints
- The schedule is linear in beta.
Hints
Hint 1
Make T evenly spaced betas, subtract them from one, then take the running product.
Keywords
research-paperscomputer-visiondiffusiongenerative-models
Python
Run your code to see results
No results yet
Run your code to check it against the examples, or submit to run every test.