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
Uh oh!
There was an error while loading. Please reload this page.
Uh oh!
There was an error while loading. Please reload this page.
-
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