Scheduler Parameters
An algorithm that controls how noise is added and removed across sampling steps. Different schedulers produce different quality characteristics at various step counts.
Schedulers (also called noise schedules) determine the pattern of denoising during generation.
How Schedulers Work
The scheduler defines:
- How much noise at each step
- The “speed” of denoising
- Distribution of work across steps
Common Schedulers
Karras
- Front-loads denoising
- Often produces sharper results
- Works well at lower step counts
- Most recommended scheduler
Linear
- Even noise distribution
- Traditional approach
- May need more steps
Exponential
- Rapid initial denoising
- Slower refinement at end
- Good for some models
SGM Uniform
- Used in some newer models
- Specific noise patterns
- Model-dependent results
Scheduler + Sampler
They work together:
- Sampler: How to denoise
- Scheduler: When and how much
Common combinations:
- DPM++ 2M Karras
- Euler with default scheduler
- DPM++ SDE Karras
Choosing a Scheduler
| Priority | Scheduler |
|---|---|
| Speed + Quality | Karras |
| Traditional | Linear |
| Model default | Check docs |
Tips
- Karras is usually best default
- Some models are tuned for specific schedulers
- Experiment when quality seems off
- Scheduler affects optimal step count