Sampler change

[BoazKG93]

I'm trying to control the sampler, but can't find a way to do it. How can I change the sampler?

[exo-pla-net]

A sampler is also called a scheduler.

There is an example of changing schedulers on the main readme:

from diffusers import LMSDiscreteScheduler

lms = LMSDiscreteScheduler(
    beta_start=0.00085, 
    beta_end=0.012, 
    beta_schedule="scaled_linear"
)

pipe = StableDiffusionPipeline.from_pretrained(
    "CompVis/stable-diffusion-v1-4", 
    revision="fp16", 
    torch_dtype=torch.float16,
    scheduler=lms,
    use_auth_token=True
)
pipe = pipe.to("cuda")

prompt = "a photo of an astronaut riding a horse on mars"
with autocast("cuda"):
    image = pipe(prompt).images[0]  
    
image.save("astronaut_rides_horse.png")

[patil-suraj]

thanks a lot for the answer @exo-pla-net ! We can always pass different schedulers when loading the pipeline.

Note: it's hacky and not recommneded but one can also set pipe.scheduler on already loaded pipeline.

[patrickvonplaten]

Related to: #336

[patil-suraj]

As discussed offline with @patrickvonplaten , re-opening this issue:

It has become a common practice to use different schedulers with SD but right now there isn't a clean way to change scheduler in a loaded pipeline. Also, it's not recommended to set
pipeline.scheduler directly as it could lead to silent errors cf #1089 (comment)

We will now allow this by making pipeline.scheduler a setter property method, which will do checks to see if the scheduler can be swapped and adopt the internal config accordingly.
Also cc @pcuenca @anton-l