ConsistencyModelScheduler<T>
Models & Types · AiDotNet.Diffusion.Schedulers
Consistency Model scheduler for single-step or few-step diffusion sampling.
Inheritance: NoiseSchedulerBase<T> → ConsistencyModelScheduler<T>
For Beginners
Consistency models are the fastest type of diffusion model:
Normal diffusion: Start with noise, take 20-50 small steps to get a clean image Consistency model: Start with noise, jump directly to the clean image in 1 step!
Key characteristics:
- Single-step generation possible (fastest diffusion method)
- Multi-step mode (2-4 steps) improves quality
- Maps noisy samples directly to clean data predictions
- Works with both distilled and directly trained consistency models
How multi-step consistency works:
- Start with pure noise at sigma_max
- Apply consistency function → get approximate clean image
- Add noise back at a lower sigma level
- Apply consistency function again → better clean image
- Repeat for desired number of steps
This “denoise-then-add-noise” cycle progressively refines the output.
How It Works
Consistency Models map any point on the probability flow ODE trajectory directly to the trajectory’s origin (the clean data). This allows single-step generation, with optional multi-step refinement for higher quality.
Reference: Song et al., “Consistency Models”, ICML 2023
Constructors
| Constructor | Summary |
|---|---|
ConsistencyModelScheduler(SchedulerConfig<T> config, double sigmaMin, double sigmaMax, int? seed) |
Initializes a new instance of the Consistency Model scheduler. |
Methods
| Method | Summary |
|---|---|
ConsistencyModelScheduler<T> CreateDefault(int? seed) |
Creates a Consistency Model scheduler with default settings. |
Vector<T> Step(Vector<T> modelOutput, int timestep, Vector<T> sample, T eta, Vector<T> noise) |
Performs one consistency model denoising step. |
void SetTimesteps(int inferenceSteps) |