DPMSolverMultistepScheduler, how to correctly denoise the specific noise step ? #7869
Unanswered
HuangChiEn
asked this question in
Q&A
Replies: 0 comments
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
-
I'm trying to add the small noise (step t=1) on the latent. Although the document doesn't cover how to add the noise to it, but I found a method called add_noise, which seems fit to my requirements.
The issue I faced is that I have no idea how to denoise the noise latent by calling the
scheduler.step(.)
method.The following image shows the way I had tried :
sample_latent
is the ground truth (GT), while I calculate the difference between the noise latent with the GT. Then no surprised, the 49-th step noise is the latent with less noise, and i take it as my noise latent.This is how I call the step method :
apprx_latent = self.scheduler.step(noise, t, noise_latent, **extra_step_kwargs, return_dict=False)[0]
, wheret=timesteps[0]
.Unfortunately, the noise latent seems closer to the GT comparing with the denoised latent..
Any suggestion will be appreciated!!
Beta Was this translation helpful? Give feedback.
All reactions