diffsyn代码有bug？为什么Qwen-Image-Edit-2509用DiffSynth-Studio推理比diffusers效果差很多

[huxian0402]

@pengqun @Lupino @calmhawk @co63oc @wenmengzhou
用DiffSynth-Studio最新的推理代码，相同的配置和测试集，跑出来比diffusers推理效果差很多，这个是什么原因呢？

测试图片：

DiffSynth-Studio推理结果：

diffusers推理结果：

下面是我的测试代码：
def test_qwen_edit_diffusers():
import os
import torch
from PIL import Image
from diffusers import QwenImageEditPlusPipeline, QwenImageEditPipeline

model_path = "Qwen-Image-Edit-2509"
pipeline = QwenImageEditPlusPipeline.from_pretrained(model_path, torch_dtype=torch.bfloat16)
print("pipeline loaded from", model_path)
pipeline.to('cuda')
pipeline.set_progress_bar_config(disable=None)

image = Image.open("08_0.jpg").convert("RGB")
inputs = {
    "image": image,
    "prompt": "Remove the black area.",
    "width": 1024,
    "height": 1024,
    "generator": torch.manual_seed(42),
    "true_cfg_scale": 4.0,
    "negative_prompt": "",
    "num_inference_steps": 40,
    "guidance_scale": 1.0,
    "num_images_per_prompt": 1,
}
with torch.inference_mode():
    output = pipeline(**inputs)
    output_image = output.images[0]

output_image.save("qwen_edit_diffusers.jpg")
print(f"Image saved at qwen_edit_diffusers.jpg")

def test_qwen_edit_diffsyn():
import os, sys
sys.path.append(os.getcwd())
from diffsynth.pipelines.qwen_image import QwenImagePipeline, ModelConfig
from PIL import Image
import torch

base_model_path = "Qwen-Image-Edit-2509"
pipe = QwenImagePipeline.from_pretrained(
    torch_dtype=torch.bfloat16,
    device="cuda",
    model_configs=[
        ModelConfig(model_id=f"{base_model_path}", origin_file_pattern="transformer/diffusion_pytorch_model*.safetensors", skip_download=True),
        ModelConfig(model_id=f"{base_model_path}", origin_file_pattern="text_encoder/model*.safetensors", skip_download=True),
        ModelConfig(model_id=f"{base_model_path}", origin_file_pattern="vae/diffusion_pytorch_model.safetensors", skip_download=True),
    ],
    tokenizer_config=None,
    processor_config=ModelConfig(model_id=f"{base_model_path}", origin_file_pattern="processor/", skip_download=True),
)

image = Image.open("08_0.jpg").convert("RGB")
output_image = pipe(
        edit_image=image,
        prompt="Remove the black area.",
        height=1024,
        width=1024,
        seed=42,
        cfg_scale=4.0,
        num_inference_steps=40, 
        negative_prompt="",
    )
output_image.save("qwen_edit_diffsyn.jpg")
print(f"Image saved at qwen_edit_diffsyn.jpg")

[Artiprocher]

@huxian0402 2509 是个多图编辑模型，因此即使处理单图输入，也应使用列表

output_image = pipe(
        edit_image=[image], # Look at here!
        prompt="Remove the black area.",
        height=1024,
        width=1024,
        seed=42,
        cfg_scale=4.0,
        num_inference_steps=40, 
        negative_prompt="",
    )

这是我们的代码生成的图片：

[huxian0402]

@huxian0402 2509 是个多图编辑模型，因此即使处理单图输入，也应使用列表

output_image = pipe(
edit_image=[image], # Look at here!
prompt="Remove the black area.",
height=1024,
width=1024,
seed=42,
cfg_scale=4.0,
num_inference_steps=40,
negative_prompt="",
)
这是我们的代码生成的图片：

验证了下，是这个问题，必须传入list，才会走encode_prompt_edit_multi -> resize到384。修改后，结果和diffusers基本对齐了，谢谢大佬