我也在复现这篇文章

Dongshengjiang commented 4 years ago

你好，这篇利用HDRNET来做样式迁移的应用，我也在复现。目前有一些效果。文章细节缺失，代码也没有公开。复现挺难的。从文章看，要用6次AdaIN，然后如果正则损失过小，也会产生很多偏色。所以regW我设置为100000。最后的效果其实更像是整体颜色的迁移，因为hdrnet本身就是保边滤波，无法迁移样式。

mousecpn commented 4 years ago

你好，这篇利用HDRNET来做样式迁移的应用，我也在复现。目前有一些效果。文章细节缺失，代码也没有公开。复现挺难的。从文章看，要用6次AdaIN，然后如果正则损失过小，也会产生很多偏色。所以regW我设置为100000。最后的效果其实更像是整体颜色的迁移，因为hdrnet本身就是保边滤波，无法迁移样式。

可以分享一下您的结果吗，谢谢

Dongshengjiang commented 4 years ago

从结果看，更像是一个整体样式的颜色迁移。你是拿他用在哪里？滤镜还是其他？

TachibanaYoshino commented 3 years ago

A Closed-form Solution to Photorealistic Image Stylization ，This paper seems to have a better effect on color transfer. Do you understand? NVIDIA/FastPhotoStyle

Xingxiangrui commented 2 years ago

@Dongshengjiang 你好，我也复现了一下，效果确实不怎么好，确实更像是整体颜色的迁移。实际运用不推荐此篇论文。如果效果更好的话有另外两篇复现了效果很好。见下面两个，这两个我复现了效果很好：

[ICCV2019] Photorealistic Style Transfer via Wavelet Transforms [paper][code]
comment: Recommand to reimplement. Existing methods are limited by spatial distortions or unrealistic artifacts. This paper proposes a wavelet corrected transfer (WCT2) based on whitening and coloring transforms that allows features to preserve their structural information and statistical properties of VGG feature space during stylization.
[ACML2019] High-Resolution Network for Photorealistic Style Transfer [paper][code]
comment: This method uses high-resolution generation network for better performance.

mousecpn / Joint-Bilateral-Learning

我也在复现这篇文章 #4