ComfyUI Custom Sampler nodes that implement Zheng et al.'s Trajectory Consistency Distillation based on https://mhh0318.github.io/tcd
This custom node repository adds TCDScheduler, SamplerTCD Euler A, and SamplerTCD nodes to the Custom Sampler category.
Just clone it into your custom_nodes folder and you can start using it as soon as you restart ComfyUI.
NOTE: SamplerTCD is a WIP and currently just operates as DDIM with no gamma parameter. Please use SamplerTCD Euler A for the time being.
LCMScheduler has one special parameter:
-
gamma, a parameter used to control the stochasticity in every step. When gamma = 0, it represents deterministic sampling, whereas gamma = 1 indicates full stochastic sampling. In a way it acts as a sort of a crossfade between Karras and Euler.The default value is 0.3, but it is recommend using a higher value when increasing the number of inference steps.
Thanks to @laksjdjf for their help with converting from sigmas to timestep and SamplerTCDEulerA.
BTW If you're curious about learning more about samplers and schedulers, check out this article: https://www.felixsanz.dev/articles/complete-guide-to-samplers-in-stable-diffusion