Community Pipeline: Add z-image differential img2img

[r4inm4ker]

What does this PR do?

Adds Differential Diffusion to Z-Image
Mostly copied from pipeline_flux_differential_img2img.py implementation.

Before submitting

• Did you read the contributor guideline?
• Did you read our philosophy doc (important for complex PRs)?

How to test:

import torch
from pipeline_z_image_differential_img2img import ZImageDifferentialImg2ImgPipeline
from diffusers.utils import load_image

pipe = ZImageDifferentialImg2ImgPipeline.from_pretrained("Z-a-o/Z-Image-Turbo", torch_dtype=torch.bfloat16)
pipe.to("cuda")

init_image = load_image("https://github.com/exx8/differential-diffusion/blob/main/assets/input.jpg?raw=true")

mask = load_image("https://github.com/exx8/differential-diffusion/blob/main/assets/map.jpg?raw=true")

prompt = "painting of a mountain landscape with a meadow and a forest, meadow background, anime countryside landscape"

image = pipe(
 prompt,
 image=init_image,
 mask_image=mask,
 strength=0.75,
 num_inference_steps=9,
 guidance_scale=0.0,
 generator=torch.Generator("cuda").manual_seed(0),
).images[0]
image.save("image.png")

Who can review?

@yiyixuxu @asomoza

[asomoza]

@bot /style

[github-actions]

Style fix is beginning .... View the workflow run here.

[asomoza]

@r4inm4ker thanks a lot! the bot can't fix the style, can you please run:

make style
make quality

also if it doesn't fix it, there's some white spaces in some empty lines in the example doc string

[asomoza]

in the meantime, this is the apple pear test with this one:

original	mask	result

this one is more finicky too than SDXL, I had to up the strength to 0.9 and use 12 steps to get that result

[bghira]

need sigma shift

[r4inm4ker]

Thanks for the feedback ! I will work on the updates after coming back from holiday.

[r4inm4ker]

Hi all, Happy New Year !
I have run the make style and make quality and pushed an update.

@bghira Regarding the "need sigma shift" comment : I am not familiar with what it means as I am still very new to ML/diffusion and this is my first attempt to contribute to the community, so any help to elaborate or explain a bit more would be really appreciated. Thanks !

In the meanwhile, I have tried rerunning example inputs from https://differential-diffusion.github.io/ , and this is the result with strength 1.0 and steps 9. prompt : ""tree of life under the sea"

original	mask	result