search

NVIDIA / FastPhotoStyle

Style transfer, deep learning, feature transform
Other
11.11k stars 1.2k forks source link

Couldn't Reproduce the Results #67

Closed KangBK0120 closed 5 years ago

KangBK0120 commented 5 years ago

Hi, I found the same images(content and style) which you used in the paper to reproduce your results.

However, when I ran your codes, I couldn't get the result similar to yours.

Is there anything I missed?

image

Yijunmaverick commented 5 years ago

Thanks for your interests. This looks like a global transferring result, as shown in Example 1 of the tutorial. As mentioned in our paper, to achieve the best and reasonable result, we use masks to determine the correspondence (e.g., sky-to-sky, house-to-house) for most of times. Check Example 2 in our tutorial. For this example, we draw a mask with 3 labels for sky, house and ground. You can determine your own correspondence, as a kind of user control.

KangBK0120 commented 5 years ago

Thanks!