DDPMScheduler<T>
Models & Types · AiDotNet.Diffusion.Schedulers
DDPM (Denoising Diffusion Probabilistic Models) scheduler implementation.
Inheritance: NoiseSchedulerBase<T> → DDPMScheduler<T>
For Beginners
DDPM is the foundational method for diffusion image generation.
Think of it like restoring a photograph that has been progressively damaged:
- Training: Learn how each level of damage looks
- Generation: Start with pure static and remove damage one tiny step at a time
Key characteristics:
- Stochastic: Each step adds a small amount of random noise
- Many steps needed: Typically 1000 steps for good quality
- Well-studied: Strong theoretical guarantees on output quality
- Variance can be learned or fixed
The step formula: x_{t-1} = (1/sqrt(alpha_t)) * (x_t - (beta_t / sqrt(1-alpha_cumprod_t)) * eps) + sigma_t * z
Where z is random noise and sigma_t controls the stochasticity.
How It Works
DDPM is the original diffusion model scheduler that uses a Markov chain of Gaussian transitions to gradually denoise samples. It requires many steps (typically 1000) but produces high-quality results with well-understood theoretical properties.
Reference: Ho et al., “Denoising Diffusion Probabilistic Models”, NeurIPS 2020
Constructors
| Constructor | Summary |
|---|---|
DDPMScheduler(SchedulerConfig<T> config) |
Initializes a new instance of the DDPM scheduler. |
Methods
| Method | Summary |
|---|---|
Vector<T> Step(Vector<T> modelOutput, int timestep, Vector<T> sample, T eta, Vector<T> noise) |
Performs one DDPM denoising step. |