Schedulers

Schedulers are crucial in the denoising process, executing multiple iterative steps to transform a noisy image into a clean one. They manage the diffusion process, introducing random noise and gradually removing it to generate new data samples.

These components define a trade-off between denoising speed and quality. Choosing the right scheduler can be complex, with no clear quantitative measure for optimal performance. It's recommended to experiment and find the scheduler that works best for your specific creative project and system.
‍

Available Schedulers and Time Step Spacing Algorithms Schedulers :

DPM Solver Multistep: Default selection, employing the DPM-Solver++ algorithm.

PNDM (Pseudo-linear multi-step): Utilizes the pseudo linear multi-step (PLMS) method, excluding pseudo Runge-Kutta (PRK) steps.

LCM (Latent Consistency Models): Designed for LCM models and SDXL-Turbo, requiring only 1 to 8 steps compared to the 20-50 steps required by regular models.LCM models typically yield better results with 2 to 8 steps and a low Guidance Scale of 1-2. When you select the LCM scheduler, the Steps and Guidance Controls' value scale is automatically adjusted to allow for a more fine-grained value setup. On macOS, you also get the option to enable "Real-Time Generation".

Time Spacing:

‍Time Spacing determines how time steps are spaced and offers three options: Linear, Leading, and Karras.

‍Note: The Time Spacing setting has no effect when the LCM Scheduler is selected.

‍Tip: When utilizing SDXL models, the default Time Spacing option is Karras. However, when generating animations with the use of a Guide Image, it is advisable to switch to Linear or Leading to prevent image degradation.


You can make changes to your Scheduler within the Advanced Tab in Sogni