Make sure that text encoder is not loaded twice which should save around 800MB GPU RAM.
Make sure that seed is already on GPU (should not make a big difference)
Make sure to use CUDA 11.8

AP123 changed pull request status to merged

Sign up or log in to comment