feat(diffusers/pipelines): add pipelines and required modules of QwenImage in Diffusers Master

[Dong1017]

What does this PR do?

Adds

1. QwenImage Pipelines and Required Modules

(Comparable with Diffusers Master)

a. Pipelines

• mindone.diffusers.QwenImagePipeline
• mindone.diffusers.QwenImageImg2ImgPipeline
• mindone.diffusers.QwenImageInpaintPipeline
• mindone.diffusers.QwenImageEditPipeline
• mindone.diffusers.QwenImageEditInpaintPipeline

b. Modules

• mindone.diffusers.models.AutoencoderQwenImage
• mindone.diffusers.models.QwenImageTransformer2DModel
• mindone.diffusers.loaders.QwenImageLoraLoaderMixin

2. add UTs of pipelines

• All UTs were setup according to Diffusers Master, accessed in Sep 17, 2025.
  • tests/diffusers_tests/pipelines/qwenimage/test_qwenimage.py
  • tests/diffusers_tests/pipelines/qwenimage/test_qwenimage_img2img.py
  • tests/diffusers_tests/pipelines/qwenimage/test_qwenimage_inpaint.py
  • tests/diffusers_tests/pipelines/qwenimage/test_qwenimage_edit.py
• Using MindSpore 2.7.0 can pass both fp32 and bf16 UTs.
• Using MindSpore 2.6.0 can pass bf16 UTs, while fp32 will happen to TypeError.

Fix

TypeError raised due to mindone/diffusers/loaders/lora_conversion_utils.py.

• converted_state_dict[diffusers_down_key] = down_weight * scale_down -> converted_state_dict[diffusers_down_key] = Parameter(down_weight * scale_down)
• converted_state_dict[diffusers_up_key] = up_weight * scale_up -> converted_state_dict[diffusers_up_key] = Parameter(up_weight * scale_up)

Usage

• QwenImagePipeline

import mindspore as ms 
from mindone.diffusers import QwenImagePipeline 

pipe = QwenImagePipeline.from_pretrained("Qwen/Qwen-Image", mindspore_dtype=ms.bfloat16) 
prompt = "A cat holding a sign that says hello world" 
# Depending on the variant being used, the pipeline call will slightly vary. 
# Refer to the pipeline documentation for more details. 
image = pipe(prompt, num_inference_steps=50)[0][0] 
image.save("qwenimage.png") 

• QwenImageImg2ImgPipeline

import mindspore as ms 
from mindone.diffusers import QwenImageImg2ImgPipeline
from mindone.diffusers.utils import load_image

pipe = QwenImageImg2ImgPipeline.from_pretrained("Qwen/Qwen-Image", mindspore_dtype=mindspore.bfloat16)
url = "https://raw.githubusercontent.com/CompVis/stable-diffusion/main/assets/stable-samples/img2img/sketch-mountains-input.jpg"
init_image = load_image(url).resize((1024, 1024))
prompt = "cat wizard, gandalf, lord of the rings, detailed, fantasy, cute, adorable, Pixar, Disney"
images = pipe(prompt=prompt, negative_prompt=" ", image=init_image, strength=0.95)[0][0]
images.save("qwenimage_img2img.png")

• QwenImageInpaintPipeline

import mindspore as ms 
from mindone.diffusers import QwenImageInpaintPipeline 
from mindone.diffusers.utils import load_image 

pipe = QwenImageInpaintPipeline.from_pretrained("Qwen/Qwen-Image", mindspore_dtype=ms.bfloat16) 
prompt = "Face of a yellow cat, high resolution, sitting on a park bench" 
img_url = "https://raw.githubusercontent.com/CompVis/latent-diffusion/main/data/inpainting_examples/overture-creations-5sI6fQgYIuo.png" 
mask_url = "https://raw.githubusercontent.com/CompVis/latent-diffusion/main/data/inpainting_examples/overture-creations-5sI6fQgYIuo_mask.png" 
source = load_image(img_url) 
mask = load_image(mask_url) 
image = pipe(prompt=prompt, negative_prompt=" ", image=source, mask_image=mask, strength=0.85)[0][0] 
image.save("qwenimage_inpainting.png") 

• QwenImageEditPipeline

import mindspore as ms 
from PIL import Image 
from mindone.diffusers import QwenImageEditPipeline 
from mindone.diffusers.utils import load_image 

pipe = QwenImageEditPipeline.from_pretrained("Qwen/Qwen-Image-Edit", mindspore_dtype=ms.bfloat16) 
image = load_image("https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/yarn-art-pikachu.png").convert("RGB") 
prompt = ("Make Pikachu hold a sign that says 'Qwen Edit is awesome', yarn art style, detailed, vibrant colors") 
# Depending on the variant being used, the pipeline call will slightly vary. 
# Refer to the pipeline documentation for more details. 
image = pipe(image, prompt, num_inference_steps=50)[0][0] 
image.save("qwenimage_edit.png") 

• QwenImageEditInpaintPipeline

import mindspore as ms 
from PIL import Image
from mindone.diffusers import QwenImageEditInpaintPipeline
from mindone.diffusers.utils import load_image

pipe = QwenImageEditInpaintPipeline.from_pretrained("Qwen/Qwen-Image-Edit", mindspore_dtype=mindspore.bfloat16)
prompt = "Face of a yellow cat, high resolution, sitting on a park bench"
img_url = "https://raw.githubusercontent.com/CompVis/latent-diffusion/main/data/inpainting_examples/overture-creations-5sI6fQgYIuo.png"
mask_url = "https://raw.githubusercontent.com/CompVis/latent-diffusion/main/data/inpainting_examples/overture-creations-5sI6fQgYIuo_mask.png"
source = load_image(img_url)
mask = load_image(mask_url)
image = pipe(prompt=prompt, negative_prompt=" ", image=source, mask_image=mask, strength=1.0, num_inference_steps=50)[0][0]
image.save("qwenimage_inpainting.png")

• LoRA Infer

from mindone.diffusers import DiffusionPipeline, FlowMatchEulerDiscreteScheduler
import mindspore
import math

scheduler_config = {
    "base_image_seq_len": 256,
    "base_shift": math.log(3),  # We use shift=3 in distillation
    "invert_sigmas": False,
    "max_image_seq_len": 8192,
    "max_shift": math.log(3),  # We use shift=3 in distillation
    "num_train_timesteps": 1000,
    "shift": 1.0,
    "shift_terminal": None,  # set shift_terminal to None
    "stochastic_sampling": False,
    "time_shift_type": "exponential",
    "use_beta_sigmas": False,
    "use_dynamic_shifting": True,
    "use_exponential_sigmas": False,
    "use_karras_sigmas": False,
}
scheduler = FlowMatchEulerDiscreteScheduler.from_config(scheduler_config)
pipe = DiffusionPipeline.from_pretrained(
    "Qwen/Qwen-Image", scheduler=scheduler, mindspore_dtype=mindspore.bfloat16
)
pipe.load_lora_weights(
    "Qwen/lightx2v/Qwen-Image-Lightning", 
    weight_name="Qwen-Image-Lightning-8steps-V1.0.safetensors", 
    adapter_name="qwenimage-lora"
)
pipe.fuse_lora()
pipe.unload_lora_weights()

prompt = "a tiny astronaut hatching from an egg on the moon, Ultra HD, 4K, cinematic composition."
negative_prompt = " "
image = pipe(
    prompt=prompt,
    negative_prompt=negative_prompt,
    width=1024,
    height=1024,
    num_inference_steps=8,
    true_cfg_scale=1.0,
    generator=None,
)[0][0]
image.save("lora_pic/qwen_fewsteps_lora.png")

Performance

Experiments are tested on Ascend Atlas 800T A2 machines with MindSpore 2.7.0

Pipeline	Weight Loading Time	Mode	Speed (s/it)
QwenImagePipeline	13m7s	Pynative / jit	9.93 / 8.63
QwenImageImg2ImgPipeline	12m53s	Pynative / jit	9.56 / 11.47
QwenImageInpaintPipeline	13m10s	Pynative / jit	4.80 / 4.35
QwenImageEditPipeline	14m22s	Pynative / jit	13.25 / 13.93
QwenImageEditInpaintPipeline	13m40s	Pynative / jit	13.98 / 13.77
QwenImagePipeline + LoRA	13m27s	Pynative / jit	2.96 / 4.47

Limitation

QwenImageEditPipeline and QwenImageEditInpaintPipeline will load modules from Qwen-Image-Edit. The use of these two pipes requires manually changing image_processor_type from Qwen2VLImageProcessorFast to Qwen2VLImageProcessor in Qwen-Image-Edit/processor/preprocessor_config.json

Notes

1. require transformers==4.52.1
2. The produced pictures are nearly identical to those by Torch, when setting consistent random seed and hidden states from the text encoder.
3. TODO: jit mode; LORA test; UTs of modules

Before submitting

• This PR fixes a typo or improves the docs (you can dismiss the other checks if that's the case).
• Did you read the contributor guideline?
• Did you make sure to update the documentation with your changes? E.g. record bug fixes or new features in What's New. Here are the
  documentation guidelines
• Did you build and run the code without any errors?
• Did you report the running environment (NPU type/MS version) and performance in the doc? (better record it for data loading, model inference, or training tasks)
• 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.
@xxx

[SamitHuang]

Can add an inference example and lora fine-tune example in examples folder, which helps introduce QwenImage

[Dong1017]

How to fix the requirement of transformers==4.52.1?

The main reason for using transformers==4.52.1 rather than transformers==4.50.0 is to avoid AttributeError and keep consistent with the requirements from Qwen-Image.
Using transformers==4.50.0 will raise the following AttributeError:

../../transformers/src/transformers/models/qwen2_5_vl/modeling_qwen2_5_vl.py:1517: in __init__
    super().__init__(config)
../../transformers/src/transformers/modeling_utils.py:1898: in __init__
    self.generation_config = GenerationConfig.from_model_config(config) if self.can_generate() else None
                             ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
        decoder_config = model_config.get_text_config(decoder=True)
        if decoder_config is not model_config:
            default_generation_config = GenerationConfig()
>           decoder_config_dict = decoder_config.to_dict()
                                  ^^^^^^^^^^^^^^^^^^^^^^
E           AttributeError: 'dict' object has no attribute 'to_dict'

../../transformers/src/transformers/generation/configuration_utils.py:1287: AttributeError

Upgrading transformers from 4.50.0 to 4.52.1 or highr version will solve this error.