bcmi / phdnet-painterly-image-harmonization Goto Github PK

1. py3.10 don't have torch 1.10.2+cu113
2. remove skimage=0.0 (fixed type)
3. numpy 1.19.5 is conflicted skimage
4. there is no pytorch 1.10.2.

i fixed as below and it works

$ pip install numpy
$ pip install torch==1.11.0+cu113 torchvision==0.12.0+cu113 torchaudio==0.11.0 --extra-index-url https://download.pytorch.org/whl/cu113
$ pip install -r requirements.txt

dominate==2.6.0
numpy
opencv_python_headless==4.5.5.64
Pillow==9.3.0
PyYAML==6.0
scikit_image==0.17.2
tensorboardX==2.5.1
tqdm==4.64.0
visdom==0.1.8.9
yacs==0.1.8

thanks for your work , but the pretrained model is not available

Hello, may I ask which part of the code DAT for Paintery Image Harmonization in Dual Domains in the paper comes from or inherits the downsample method in DAT's networks.py? I couldn't find the answer in the code. If you are willing to answer, I would greatly appreciate it.

作者您好，感谢您的工作，请问是否可以提供用于本地图片和谐的demo呢

Hi,

Thanks for the great work and the results look amazing!

For the pre-trained model here, is it possible to also share it on other platforms such as Google Drive? The Baidu Cloud keeps asking me to download its own app instead of the actual pre-trained model. Much appreaciated!

Also does the code strictly requires square images as input? Or does it work on any arbitrary input image resolutions? Thanks a lot!

作者你好，想请问下你们有尝试过增大和谐化图片的尺寸吗？我想在和谐化问题后面接识别和分割的应用，发现大多数的和谐化工作都在256×256的尺寸上进行，实验后resize或者简单超分会导致图片质量非常差，对下游应用来说还不太够用。想请教下作者，这块可以有什么处理思路呢？单纯改卷积尺寸可行吗？

用项目提供的预训练模型的到的图像，效果很差，带有蓝绿色的噪声纹理，例如：


不知道问题出现在哪里，能否给予指导

Thank you for your great work!

COCO selected mask and WikiArt dataset are shared via Baidu Netdisk, but I can't create a baidu account.
would you mind uploading these datasets on an alternative drive services?

Thank you.