The PLMS sampler is a sampler in text-to-image diffusion models that is based on the Pseudo-Linear Multi-Step method for solving differential equations. It is a fast and efficient sampler that can often generate good outputs in 20-30 steps.

The PLMS sampler works by sampling noise from the diffusion process and then subtracting it from the image. However, unlike other samplers, the PLMS sampler also adds a weighted sum of the previous noise samples back to the image at each step. This helps to improve the stability of the sampling process and to generate more realistic images.

The PLMS sampler is no longer widely used in the text-to-image diffusion community, as it has been superseded by newer and more advanced samplers, such as the DDIM and PNDM samplers.