Feature/qwenimage21 modular

docs/source/en/api/pipelines/qwenimage21.md

@@ -62,6 +62,83 @@ pipe.transformer.set_attn_processor(QwenImage21FlexAttnProcessor())
 pipe.transformer.compile()
 ```
 
+## Modular generation and inpainting
+
+[`QwenImage21ModularPipeline`] selects text-to-image, image-conditioned generation, or inpainting from the inputs.
+It reuses the QwenImage21 transformer, RGBA VAE, and Qwen3-VL processor from the standard checkpoint.
+
+```python
+import torch
+from diffusers import ModularPipeline
+from diffusers.utils import load_image
+
+pipe = ModularPipeline.from_pretrained("Qwen/Qwen-Image-2.1")
+pipe.load_components(dtype=torch.bfloat16)
+pipe.to("cuda")
+
+images = pipe(prompt="A capybara wearing a wizard hat", output="images")
+
+source = load_image("source.png")
+mask = load_image("mask.png").convert("L")
+images = pipe(
+    prompt="Give the capybara a red scarf",
+    image=source,
+    mask_image=mask,
+    strength=1.0,
+    generator=torch.Generator("cpu").manual_seed(0),
+    output="images",
+)
+images[0].save("inpaint.png")
+```
+
+White mask pixels are repainted; black pixels restore the source latents after each denoising step. Preservation
+is in latent space, so decoding can change unmasked pixels through VAE reconstruction. `strength` must be in
+`(0, 1]` and leave at least one denoising step. Lower values start from a less noisy source image.
+
+For multiple-reference inpainting, keep one source in `image` and pass the extra images as `reference_images`:
+
+```python
+images = pipe(
+    prompt="Use the scarf from image 2 and the fabric pattern from image 3 in the masked area of image 1",
+    image=source,
+    mask_image=mask,
+    reference_images=[load_image("scarf.png"), load_image("fabric.png")],
+    output="images",
+)
+```
+
+The source is the first condition image, followed by references in the supplied order. All condition images feed
+both Qwen3-VL and the VAE. References can have different aspect ratios; only the source has a repaint mask. The
+source, mask, and references are shared across a batch of prompts. Nested per-prompt reference lists are unsupported.
+Without a mask, `image` can be a single image or a flat list for image-conditioned generation.
+
+`height` and `width` must be multiples of 32. `output_resolution` defaults to 1024 and controls the area used to
+resize condition images, independently of explicit output dimensions. Inpainting derives omitted dimensions from
+the source; ordinary image-conditioned generation derives them from the last condition image.
+
+The default guider disables classifier-free guidance. Configure guidance through the component rather than a
+pipeline argument:
+
+```python
+from diffusers import ClassifierFreeGuidance
+
+pipe.update_components(guider=ClassifierFreeGuidance(guidance_scale=3.0))
+images = pipe(prompt="A capybara", negative_prompt="blurry", output="images")
+```
+
+`use_kv_cache=True` caches step-independent condition tokens when the transformer has `causal_condition` enabled.
+Each generation and guidance branch has its own cache. As with the standard pipeline, reduced-precision cached
+and uncached outputs can differ because the attention layouts differ.
+
+For standalone encoding or denoising, obtain a workflow with
+`pipe.blocks.get_workflow("text2image")`, `"image_conditioned"`, or `"inpainting"`, and initialize a pipeline from
+its individual blocks using `init_pipeline()`. Encoders return unexpanded embeddings; the denoise input block
+applies `num_images_per_prompt`, so encoded features can be reused for different output counts.
+
+## QwenImage21ModularPipeline
+
+[[autodoc]] QwenImage21ModularPipeline
+
 ## QwenImage21Pipeline
 
 [[autodoc]] QwenImage21Pipeline

src/diffusers/__init__.py

@@ -557,6 +557,8 @@
             "MiniMaxH3ModularPipeline",
             "MiniMaxMusic3Blocks",
             "MiniMaxMusic3ModularPipeline",
+            "QwenImage21AutoBlocks",
+            "QwenImage21ModularPipeline",
             "QwenImageAutoBlocks",
             "QwenImageEditAutoBlocks",
             "QwenImageEditModularPipeline",
@@ -1418,6 +1420,8 @@
             MiniMaxH3ModularPipeline,
             MiniMaxMusic3Blocks,
             MiniMaxMusic3ModularPipeline,
+            QwenImage21AutoBlocks,
+            QwenImage21ModularPipeline,
             QwenImageAutoBlocks,
             QwenImageEditAutoBlocks,
             QwenImageEditModularPipeline,

src/diffusers/modular_pipelines/__init__.py

@@ -91,6 +91,7 @@
         "Krea2TurboAutoBlocks",
         "Krea2TurboModularPipeline",
     ]
+    _import_structure["qwenimage21"] = ["QwenImage21AutoBlocks", "QwenImage21ModularPipeline"]
     _import_structure["qwenimage"] = [
         "QwenImageAutoBlocks",
         "QwenImageModularPipeline",
@@ -221,6 +222,7 @@
             QwenImageLayeredModularPipeline,
             QwenImageModularPipeline,
         )
+        from .qwenimage21 import QwenImage21AutoBlocks, QwenImage21ModularPipeline
         from .stable_diffusion_3 import StableDiffusion3AutoBlocks, StableDiffusion3ModularPipeline
         from .stable_diffusion_xl import StableDiffusionXLAutoBlocks, StableDiffusionXLModularPipeline
         from .wan import (

src/diffusers/modular_pipelines/modular_pipeline.py

@@ -143,6 +143,7 @@ def _helios_pyramid_map_fn(config_dict=None):
         ("ideogram4", _create_default_map_fn("Ideogram4ModularPipeline")),
         ("krea2", _krea2_map_fn),
         ("qwenimage", _create_default_map_fn("QwenImageModularPipeline")),
+        ("qwenimage21", _create_default_map_fn("QwenImage21ModularPipeline")),
         ("qwenimage-edit", _create_default_map_fn("QwenImageEditModularPipeline")),
         ("qwenimage-edit-plus", _create_default_map_fn("QwenImageEditPlusModularPipeline")),
         ("qwenimage-layered", _create_default_map_fn("QwenImageLayeredModularPipeline")),

src/diffusers/modular_pipelines/qwenimage21/__init__.py

@@ -0,0 +1,47 @@
+from typing import TYPE_CHECKING
+
+from ...utils import (
+    DIFFUSERS_SLOW_IMPORT,
+    OptionalDependencyNotAvailable,
+    _LazyModule,
+    get_objects_from_module,
+    is_torch_available,
+    is_transformers_available,
+)
+
+
+_dummy_objects = {}
+_import_structure = {}
+
+try:
+    if not (is_transformers_available() and is_torch_available()):
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    from ...utils import dummy_torch_and_transformers_objects  # noqa F403
+
+    _dummy_objects.update(get_objects_from_module(dummy_torch_and_transformers_objects))
+else:
+    _import_structure["modular_blocks_qwenimage21"] = ["QwenImage21AutoBlocks"]
+    _import_structure["modular_pipeline"] = ["QwenImage21ModularPipeline"]
+
+if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
+    try:
+        if not (is_transformers_available() and is_torch_available()):
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        from ...utils.dummy_torch_and_transformers_objects import *  # noqa F403
+    else:
+        from .modular_blocks_qwenimage21 import QwenImage21AutoBlocks
+        from .modular_pipeline import QwenImage21ModularPipeline
+else:
+    import sys
+
+    sys.modules[__name__] = _LazyModule(
+        __name__,
+        globals()["__file__"],
+        _import_structure,
+        module_spec=__spec__,
+    )
+
+    for name, value in _dummy_objects.items():
+        setattr(sys.modules[__name__], name, value)