Implement Euler Ancestral Discrete scheduler

[mspronesti]

Hi @LaurentMazare, this PR aims at integrating the Euler Ancestral Discrete Scheduler into this repository, solving the second task mentioned in #23 .

It includes all the features implemented in HF's Python version.

All the considerations made in my last PR, regarding the need to generalize the examples, still apply here.

[sssemil]

Tried this with sd1.5 and this prompt:

    "n_steps": 21,
    "seed": 1573789502,
    "prompt": "A rusty robot holding a fire torch.",
    "guidance_scale": 7.5

I think something is wrong here, or in my code.

[mspronesti]

Hi @sssemil, I tried with HF python diffusers, using your same config (expect the seed) and here are some images with different seeds. To me they seem aligned with the one you obtained with my code.

[sssemil]

Hm, but at the same time, this is what I get with python snippet you wrote in another thread:

And this looks better.

[mspronesti]

Did you change the scheduler and the number of inference steps (some of those snippets have 1000)? I tried several times with that same snippet and the appropriate scheduler. Here are other examples I just produced

I leave the exact snippet hereby, to make sure we're using the same one

import torch
from diffusers import StableDiffusionPipeline, EulerAncestralDiscreteScheduler

pipe = StableDiffusionPipeline.from_pretrained("runwayml/stable-diffusion-v1-5", torch_dtype=torch.float16)

pipe.scheduler = EulerAncestralDiscreteScheduler()

pipe = pipe.to("cuda")
prompt = "A very rusty robot holding a fire torch."
image = pipe(prompt, num_inference_steps=21).images[0]  
image

[sssemil]

I think I ran it with 50 steps before, here's with 21:

Welp, then it's similar results.

[mspronesti]

Happy to help :+1:

[LaurentMazare]

Merged, thanks for all the scheduler hard work!

[mspronesti]

Thanks for merging! (maybe forgot to push after git merge?)

[LaurentMazare]

Oh right sorry about that!