Stable diffusion 2-1

[tralala87]

Is it possible to run this with stabilityai/stable-diffusion-2-1?

[davidrs]

+1 to requesting this guidance.
Here are some more details on current attempts and error.
I've not yet dug in to pipeline_utils to see how it was extracting pipeline args and how they changed between 1.4 and 2

import torch

from stable_diffusion_videos import StableDiffusionWalkPipeline, Interface

from diffusers import StableDiffusionPipeline, EulerDiscreteScheduler

model_id = "stabilityai/stable-diffusion-2"

pipeline = StableDiffusionWalkPipeline.from_pretrained(
    model_id,
    torch_dtype=torch.float16,
    revision="fp16",
).to("cuda")

interface = Interface(pipeline)

Throws error:

---------------------------------------------------------------------------
ValueError                                Traceback (most recent call last)
[<ipython-input-19-3739427ac16b>](https://localhost:8080/#) in <module>
      1 model_id = "stabilityai/stable-diffusion-2"
      2 
----> 3 pipeline = StableDiffusionWalkPipeline.from_pretrained(
      4     model_id,
      5     torch_dtype=torch.float16,

1 frames
[/usr/local/lib/python3.8/dist-packages/diffusers/pipeline_utils.py](https://localhost:8080/#) in from_pretrained(cls, pretrained_model_name_or_path, **kwargs)
    672         elif len(missing_modules) > 0:
    673             passed_modules = set(list(init_kwargs.keys()) + list(passed_class_obj.keys())) - optional_kwargs
--> 674             raise ValueError(
    675                 f"Pipeline {pipeline_class} expected {expected_modules}, but only {passed_modules} were passed."
    676             )

ValueError: Pipeline <class 'stable_diffusion_videos.stable_diffusion_pipeline.StableDiffusionWalkPipeline'> expected {'vae', 'scheduler', 'unet', 'text_encoder', 'feature_extractor', 'tokenizer', 'safety_checker'}, but only {'vae', 'scheduler', 'unet', 'text_encoder', 'tokenizer'} were passed.

[nateraw]

taking a look

[nateraw]

hmm the first error goes away if I pass feature_extractor=CLIPFeatureExtractor() or feature_extractor=None, but now I'm getting black images. I think perhaps its either due to diffusers dep being out of date or the different scheduler they use...investigating.

[nateraw]

btw related issue here for 2.0 #124

[nateraw]

ah some insight here: https://huggingface.co/stabilityai/stable-diffusion-2-1/discussions/9#639b847134967bcf45576059

[nateraw]

Ok so you can successfully initialize it like this if you have the latest diffusers version installed (pip install --upgrade diffusers)

import torch

from stable_diffusion_videos import StableDiffusionWalkPipeline
from diffusers import DPMSolverMultistepScheduler

model_id = "stabilityai/stable-diffusion-2-1"

pipe = StableDiffusionWalkPipeline.from_pretrained(
    model_id,
    feature_extractor=None,
    safety_checker=None,
    revision="fp16",
    torch_dtype=torch.float16,
).to("cuda")
pipe.scheduler = DPMSolverMultistepScheduler.from_config(pipe.scheduler.config)

and generate images like this:

prompt = "a cat"
image = pipe(
    prompt,
    num_inference_steps=5,
    generator=torch.Generator(device='cuda').manual_seed(42),
    height=768,
    width=768
).images[0]

But the walk function of the pipeline currently will generate just black images due to the use of torch.autocast('cuda'). Need to remove it. If you run the same code but with autocast, like this, you'll see what I mean...

prompt = "a cat"

with torch.autocast('cuda'):
    image = pipe(
        prompt,
        num_inference_steps=5,
        generator=torch.Generator(device='cuda').manual_seed(42),
        height=768,
        width=768
    ).images[0]
image

[nateraw]

@davidrs if you're around feel free to try out the pull request above. This should work in colab:

! pip install git+https://github.com/nateraw/stable-diffusion-videos@remove-autocast

Init the pipe

import torch

from stable_diffusion_videos import StableDiffusionWalkPipeline
from diffusers import DPMSolverMultistepScheduler

model_id = "stabilityai/stable-diffusion-2-1"

pipe = StableDiffusionWalkPipeline.from_pretrained(
    model_id,
    feature_extractor=None,
    safety_checker=None,
    revision="fp16",
    torch_dtype=torch.float16,
).to("cuda")
pipe.scheduler = DPMSolverMultistepScheduler.from_config(pipe.scheduler.config)

Make videos

video_path = pipe.walk(
    ['a cat', 'a dog'],
    [12345, 54321],
)

I did all the edits from the browser haha so not sure I want to merge and make a release til its tested a little

[davidrs]

Excellent @nateraw thanks for the speedy reply.
I have one animation running that I'll wait to finish, then I'll restart kernel and try the new branch!

[davidrs]

UPDATE: it is working, must have been stale deps, or autocast. sorry.
Thanks!!!

---

For faster testing loop:

video_path = pipe.walk(
    ['a cat', 'a dog'],
    [12345, 54321],
    num_inference_steps=20,
    num_interpolation_steps=2,
)

[nateraw]

Awesome. merged and made a release. Thanks for checking it so quick!!