Add ZImageImg2ImgPipeline

[CalamitousFelicitousness]

What does this PR do?

This PR adds img2img pipeline for Z-Image. The summary of changes are below

• Updated the pipeline structure to include ZImageImg2ImgPipeline alongside ZImagePipeline.
• Implemented the ZImageImg2ImgPipeline class
• Mapped the new ZImageImg2ImgPipeline for image generation tasks.
• Added unit tests for ZImageImg2ImgPipeline
• Updated dummy objects to include ZImageImg2ImgPipeline for testing

Closes issue #12752

Tested using a simple script:

Testing script

#!/usr/bin/env python
"""Test script for ZImage img2img support (without LoRA)."""

import sys
sys.path.insert(0, '/home/ohiom/diffusers/src')

import torch
from PIL import Image
from diffusers import ZImageImg2ImgPipeline

# Paths
MODEL_PATH = "database/models/huggingface/models--Tongyi-MAI--Z-Image-Turbo/snapshots/78771b7e11b922c868dd766476bda1f4fc6bfc96"
INPUT_IMAGE_PATH = "aline_1024.jpg"  # Use existing image as input

print("Loading ZImageImg2ImgPipeline...")
pipe = ZImageImg2ImgPipeline.from_pretrained(
    MODEL_PATH,
    torch_dtype=torch.bfloat16,
    local_files_only=True,
)
pipe.to("cuda")
print("Pipeline loaded.")

# Load input image
print(f"\nLoading input image from {INPUT_IMAGE_PATH}...")
input_image = Image.open(INPUT_IMAGE_PATH).convert("RGB")
print(f"Input image size: {input_image.size}")

# Generate an image
prompt = "a woman sitting under a tree, oil painting style, impressionist, vibrant colors"
strength = 0.6  # 0.0 = no change, 1.0 = full transformation

print(f"\nGenerating image with prompt: {prompt}")
print(f"Strength: {strength}")

image = pipe(
    prompt=prompt,
    image=input_image,
    strength=strength,
    num_inference_steps=8,
    guidance_scale=3.0,
    generator=torch.Generator(device="cuda").manual_seed(42),
).images[0]

output_path = "test_zimage_img2img_output.png"
image.save(output_path)
print(f"\nImage saved to {output_path}")

Prompt: a woman sitting in a dark room, oil painting style, impressionist, vibrant colors

LoRA functionality depends on my other PR #12750, so they will have to be merged sequentially. I did not think there was much point in leaving it out.

Before submitting

• Did you read the contributor guideline?
• Did you read our philosophy doc (important for complex PRs)?
• Was this discussed/approved via a GitHub issue or the forum? Please add a link to it if that's the case.
• Did you make sure to update the documentation with your changes? Here are the
  documentation guidelines, and
  here are tips on formatting docstrings.
• Did you write any new necessary tests?

Who can review?

Anyone in the community is free to review the PR once the tests have passed. Feel free to tag
members/contributors who may be interested in your PR.

@sayakpaul @asomoza

[CalamitousFelicitousness]

For some reason the VAE Tiling couldn't meet the 0.2 diff threshold, my test has upped that to 0.3, whether further investigation is warranted I am not sure.

[asomoza]

thanks a lot again! for this one we should probably wait for the lora one to be merged. I left a few comments

[asomoza]

can this function be also with Copied from

[asomoza]

same as before, this one can be Copied from no?

[asomoza]

since you're adding this pipeline here, can you also add the t2i too

[asomoza]

this model is really finicky with the img2img but it seems to be working ok.

source	img2img

[CalamitousFelicitousness]

@asomoza I just thought, I have inpainting PR lined up, do you think keeping this one img2img only and inpainting after that, separately, is the better approach, to keep the PR review easier? Or is it less work for you guys if I also merge this in this PR?

[asomoza]

I prefer to keep them separated, I'm not really sure the inpainting can be good with this model so I want to test it and maybe we can add something like differential diffusion as a switch for it to be better

[CalamitousFelicitousness]

Alrighty, that's how I felt as well.

Inpainting seems alright