HeunDiscreteScheduler<T>
Models & Types · AiDotNet.Diffusion.Schedulers
Heun discrete scheduler for diffusion model sampling using second-order Heun’s method.
Inheritance: NoiseSchedulerBase<T> → HeunDiscreteScheduler<T>
For Beginners
Heun’s method is like Euler but smarter:
- Euler takes one step and hopes for the best
- Heun takes a trial step, evaluates the derivative there too,
then averages both derivatives for a more accurate step
Key characteristics:
- Second-order accuracy (better than Euler per step)
- Two model evaluations per step (so 20 Heun steps = 40 model calls)
- Good quality with fewer steps than Euler
- Deterministic: same seed always produces the same result
When to use Heun:
- When you want higher quality per step than Euler
- When model evaluation cost is acceptable
- When you want smooth, accurate trajectories
How It Works
The Heun scheduler implements second-order ODE solving for diffusion sampling. It performs two model evaluations per step (predictor + corrector) to achieve higher accuracy than first-order methods like Euler.
Reference: Karras et al., “Elucidating the Design Space of Diffusion-Based Generative Models”, NeurIPS 2022
Constructors
| Constructor | Summary |
|---|---|
HeunDiscreteScheduler(SchedulerConfig<T> config) |
Initializes a new instance of the Heun discrete scheduler. |
Methods
| Method | Summary |
|---|---|
HeunDiscreteScheduler<T> CreateDefault() |
Creates a Heun scheduler with default Stable Diffusion settings. |
Vector<T> Step(Vector<T> modelOutput, int timestep, Vector<T> sample, T eta, Vector<T> noise) |
Performs one Heun denoising step (second-order method). |
void SetTimesteps(int inferenceSteps) |
Sets up the inference timesteps and computes sigma schedule. |