feat(qwenimage): add modular area composition denoise route

[yaoqih]

What does this PR do?

This PR adds modular area composition routing for QwenImage in a way that follows the modular diffusers philosophy: compose new blocks instead of mutating existing default workflows.

Summary of changes

• Adds two new area-specific core denoise workflows:
  • QwenImageAreaCompositionCoreDenoiseStep (text2image + area composition)
  • QwenImageAreaCompositionImg2ImgCoreDenoiseStep (img2img + area composition)
• Extends QwenImageAutoCoreDenoiseStep routing with an explicit area_composition trigger:
  • uses area route only when area_composition is non-empty
  • preserves existing routing for text2image / img2img / inpaint / controlnet when area composition is not enabled
• Keeps the default non-area denoise path intact (backward-compatible behavior).
• Adds/updates area composition plumbing in QwenImage modular components (inputs, encoders, and denoise path) to support regional conditioning and weighted regional merging.
• Exports newly added area core blocks in qwenimage/__init__.py.

Motivation

The goal is to support area composition as an opt-in capability without replacing default denoise behavior globally. This keeps the architecture composable and explicit, and better aligns with modular pipeline design principles.

Backward compatibility

• Existing workflows remain unchanged unless area_composition is provided.
• No behavior change for standard text2image/img2img/inpaint/controlnet usage.

Fixes # (issue)

Before submitting

• Did you read the contributor guideline?
• Did you read our philosophy doc (important for complex PRs)?
• Did you write any new necessary tests?

Who can review?

This PR touches modular pipelines / denoising route selection for QwenImage.

Suggested reviewers:

• @yiyixuxu

test code:

import torch
import diffusers
from diffusers import QwenImageModularPipeline
from diffusers.modular_pipelines.qwenimage import QwenImageAutoBlocks


cfg = {
    "prompt": "(masterpiece) (best quality) beautiful landscape breathtaking amazing view nature photograph forest mountains ocean (sky) national park scenery, full body (flat chest:1.0) (girl:1.0) with (fennec fox:0.9) (ears:1.0) (short blonde:1.0) hair (blue eyes:1.0) school uniform sweater standing long skirt",
    "negative_prompt": "(hands), text, error, cropped, (worst quality:1.2), (low quality:1.2), normal quality, (jpeg artifacts:1.3), signature, watermark, username, blurry, artist name, monochrome, sketch, censorship, censor, (copyright:1.2), extra legs, (forehead mark) (depth of field) (emotionless) (penis) (pumpkin)",
    "area_composition": [
        {
            "prompt": "(best quality) (night:1.3) (darkness) sky (black) (stars:1.2) (galaxy:1.2) (space) (universe)",
            "x": 0.0,
            "y": 0.0,
            "width": 1.0,
            "height": 0.3,
            "strength": 3,
        },
        {
            "prompt": "(best quality) (evening:1.2) (sky:1.2) (clouds) (colorful) (HDR:1.2) (sunset:1.3)",
            "x": 0.0,
            "y": 0.25,
            "width": 1.0,
            "height": 0.3,
            "strength": 3,
        },
        {
            "prompt": "(best quality) (daytime:1.2) sky (blue)",
            "x": 0.0,
            "y": 0.4,
            "width": 1.0,
            "height": 0.3,
            "strength": 3,
        },
        {
            "prompt": "(masterpiece) (best quality) morning sky",
            "x": 0.0,
            "y": 0.55,
            "width": 1.0,
            "height": 0.3,
            "strength": 3,
        },
    ],
    "seed": 853374162509361,
    "num_inference_steps": 13,
    "width": 704,
    "height": 1280,
}

MODEL_PATH = "/gemini/platform/public/aigc/cv_banc/zsw/zhuangcailin/pretrain/Qwen/Qwen-Image-2512/"
DEVICE = "cuda:1"

blocks = QwenImageAutoBlocks()

pipe = QwenImageModularPipeline(blocks, pretrained_model_name_or_path=MODEL_PATH)
pipe.load_components(
    names=[name for name in pipe.pretrained_component_names if name != "controlnet"],
    torch_dtype={"default": torch.bfloat16, "vae": torch.float16},
)
pipe.to(DEVICE)

with_area = pipe(
    prompt=cfg["prompt"],
    negative_prompt=cfg["negative_prompt"],
    width=cfg["width"],
    height=cfg["height"],
    num_inference_steps=cfg["num_inference_steps"],
    area_composition=cfg["area_composition"],
    generator=torch.Generator(device=DEVICE).manual_seed(cfg["seed"]),
    output=["images"],
)["images"][0]

without_area = pipe(
    prompt=cfg["prompt"],
    negative_prompt=cfg["negative_prompt"],
    width=cfg["width"],
    height=cfg["height"],
    num_inference_steps=cfg["num_inference_steps"],
    area_composition=[],
    generator=torch.Generator(device=DEVICE).manual_seed(cfg["seed"]),
    output=["images"],
)["images"][0]

with_area.save("with_area_composition.png")
without_area.save("without_area_composition.png")

with_area_composition.png(strength=3)

with_area_composition.png(strength=1.5)

without_area_composition.png