Hello everyone, hope you are doing great.
During my experiments, I noticed DiffusionPipeline
takes a timesteps
argument which is set to None
by default.
Given that num_inference_steps
also exists and is used to initialize timesteps
(correct me if Im wrong please) and that the majority of schedulers/samplers seem not to accept timesteps through this argument, is this simply deprecated? or is it me that’s missing something?
I’ve tried LMSDiscreteScheduler
, DDIMScheduler
, and the PNDMScheduler
all to no avail.
my snippet in using a custom timesteps tensor:
timesteps = torch.cat([torch.linspace(0, 0.5, steps=30), torch.linspace(0.5, 1, steps=10)])
text2image.scheduler = dfs.schedulers.scheduling_ddim.DDIMScheduler.from_config(text2image.scheduler.config)
text2image.scheduler.set_timesteps(timesteps)
# ...
Could someone kindly enlighten me here? am I using the wrong schedulers? if so what scheduler can be used in this situation?
Thanks a lot in advance