Combine `UnCLIPPipeline` and `StableDiffusionImageVariationPipeline`

[budui]

Model/Pipeline/Scheduler description

UnCLIPPipeline("kakaobrain/karlo-v1-alpha") provide a prior model that can generate clip image embedding from text.
StableDiffusionImageVariationPipeline("lambdalabs/sd-image-variations-diffusers") provide a decoder model than can generate images from clip image embedding.

So, I test combineUnCLIPPipeline and StableDiffusionImageVariationPipeline:

    ...
    decoder_pipe = diffusers.StableDiffusionImageVariationPipeline.from_pretrained(
        "lambdalabs/sd-image-variations-diffusers",
        revision="v2.0",
        torch_dtype=torch.float16,
        local_files_only=True,
        safety_checker=None,
    )
    ...
    decoder_pipe.to(device)

    prior_pipe = diffusers.UnCLIPPipeline.from_pretrained(
        "kakaobrain/karlo-v1-alpha",
        torch_dtype=torch.float16,
        local_files_only=True,
    )
    ...
    prior_pipe.to(device)

    #  see function `karlo_prior` and `sd_image_variations_decoder` detail in gist link bellow.
    prior_pipe.text_to_image_embedding = types.MethodType(karlo_prior, prior_pipe)
    decoder_pipe.image_embedding_to_image = types.MethodType(sd_image_variations_decoder, decoder_pipe)

    random_generator = torch.Generator(device=device).manual_seed(1000)

    prompt = "a shiba inu wearing a beret and black turtleneck"
    image_embeddings = prior_pipe.text_to_image_embedding(prompt, generator=random_generator)

    image = decoder_pipe.image_embedding_to_image(image_embeddings, generator=random_generator).images[0]

It works! And I got reasonable results:

this Stable DALLE2-like Diffusion is relatively lightweight: only need 6.5G GPU RAM use default code and 21s on half T4 GPU.

Open source status

• The model implementation is available
• The model weights are available (Only relevant if addition is not a scheduler).

Provide useful links for the implementation

full code above: my gist

[patrickvonplaten]

Very cool! Would you maybe like to add this as a community pipeline as explained here with you being the author? This sounds like a great contribution :-)

[budui]

Hi, I tried it as you suggested (gist here).

Now the pipeline is mostly OK, but I have a minor problems:

# https://gist.github.com/budui/416b82e489d341f2495b155cb9cb1914#file-stable_unclip_pipeline-py-L289-L300
    pipeline = StableUnCLIPPipeline.from_pretrained(
        "kakaobrain/karlo-v1-alpha",
        torch_dtype=torch.float16,
        decoder_pipe_kwargs=dict(
            image_encoder=None,
            torch_dtype=torch.float16,
        ),
    )
    pipeline.to(device)
    pipeline.decoder_pipe.to(device)

How can I make sure that the pipeline (pipeline) I create has the same dtype and device as it's attribute pipeline (pipeline.decoder_pipe) ? Now I have to set it manually.

[budui]

I know how to do it now. I will create a PR soon..., seems that I need to write some documents.

[yabenz]

Can you please show me how can I add a parameter to specify the image resolution? Also, the seed.

[budui]

   # https://gist.github.com/budui/416b82e489d341f2495b155cb9cb1914#file-stable_unclip_pipeline-py-L288-L314
    prompt = "a shiba inu wearing a beret and black turtleneck"
    random_generator = torch.Generator(device=device).manual_seed(1000)
    output = pipeline(prompt=prompt, generator=random_generator, width=512, height=512)

[yabenz]

Thanks a lot! Going to try that!

[yabenz]

Can I do this with Karlo because that's what I'm using. I use the following:

from diffusers import UnCLIPPipeline
import torch

pipe = UnCLIPPipeline.from_pretrained("kakaobrain/karlo-v1-alpha")
pipe = pipe.to('cpu')

prompt = "Cat on a yellow leaf."

image = pipe([prompt]).images[0]

timage.save("./pop.png")