Skip to main content
A Step Scheduler determines how much freedom your model will be able to have when generating outputs. This is because the scheduler mixes two core ingredients:
  • Input Structure: Your original image (Reality).
  • Diffusion Noise: Random static (Creativity).
A diffusion model generate images by adding noise to ‘damage’ an input. The model then tries it’s best to ‘repair’ by changing that ‘damaging noise’ into details one could find in an image.

​
Single Step Schedule

In StreamDiffusionTD, users can choose a single step between 1–50.
Where the step is placed determines the mixture of noise vs. structure:
  • Low steps (1–15):
    More emphasis on Diffusion Noise → more creativity, looser structure, less adherence to the input.
  • High steps (35–50):
    More emphasis on Input Structure → outputs that strongly match, or even nearly replicate, your input.
Step Schedule Ezgif Com Resize Gi single step schedule from moving from 1-49; prompt: blueberries

​
Multi-Step Schedule

StreamDiffusionTD also allows a multi-step schedule, where you can place up to four steps in order from lowest to highest.
It is important to note that a multi-step schedule requires MORE machine compute power
A Multi-Step Schedule allows StreamDiffusion to run several refinement passes within a single frame. Instead of jumping directly from one noise level to the final output, the model moves through multiple steps in sequence (e.g., 1 → 24 → 49). Each step receives the result of the previous one and refines it further. Step Scheduler Pn 1 step schedule vs 3 step schedule; prompt: blueberries Stacking steps creates a smoother progression from “creative” to “accurate,” resulting in outputs that can be more coherent and stable than a single-step schedule.