The Role of a Scheduler in Generative AI
A scheduler helps in generating images by managing the process of refining them step by step.
- Denoising Samples: It takes the model’s output and a specific point in the process (called a timestep) to clean up the image.
- Timesteps: These are points in the process that indicate how far along the image is in its development.
- Forward and Backward Process: The image is created by moving forward through these timesteps, and fine-tuning (inference) happens by moving backward through them.
They offer a wide variety of results. Depending on your use case, select the one that best fits your needs. For a full list of schedulers accessible via the API please see:
https://docs.scenario.com/docs/schedulers
We generally recommend the default scheduler for a majority of use cases. Adjusting the scheduler is for advanced users.
For a deep dive into Schedulers (also called Samplers) follow this link!
Results Comparison
The following grids of images have been generated using:
- the same model
- the same seed
- the same prompt: "a bear alone in dark woods"
You'll notice that the highest number of steps or guidance doesn't necessarily give the best results.
30 steps
40 steps
50 steps
Using LCM Scheduler:
Benefits:
- Significantly increases the speed and efficiency of image generation on Stable Diffusion 1.5 and SDXL.
Guidelines for Using LCM:
- Default Settings: After selecting the LCM Scheduler, you’ll see lower default values for Sampling Steps and Guidance compared to other schedulers.
- Quality Trade-Off: Be aware that using the LCM Scheduler can decrease image quality, leading to some undesirable effects in certain outputs.
Adjusting Settings:
- Sampling Steps: Increase by 1 per iteration to test the response (maximum of 6 for SDXL recommended).
- Guidance: Decrease in increments of 0.5. Effective values for SDXL are usually around 1.0 to 1.5.
Caution:
- Use caution when adjusting settings to find the right balance between speed and quality.