about
A Closed-Form Solution to Photorealistic Image Stylization (arxiv.org)
125 points by petercooper on Feb 21, 2018 | hide | past | pdf | 13 comments on HN

In plain words: It copies the colors and textures of one photo onto another in two direct steps: first stylize, then smooth so the result stays even across the picture. People preferred its results over other photo-style methods, and it ran much faster.

Abstract · A Closed-form Solution to Photorealistic Image Stylization

Photorealistic image stylization concerns transferring style of a reference photo to a content photo with the constraint that the stylized photo should remain photorealistic. While several photorealistic image stylization methods exist, they tend to generate spatially inconsistent stylizations with noticeable artifacts. In this paper, we propose a method to address these issues. The proposed method consists of a stylization step and a smoothing step. While the stylization step transfers the style of the reference photo to the content photo, the smoothing step ensures spatially consistent stylizations. Each of the steps has a closed-form solution and can be computed efficiently. We conduct extensive experimental validations. The results show that the proposed method generates photorealistic stylization outputs that are more preferred by human subjects as compared to those by the competing methods while running much faster. Source code and additional results are available at https://github.com/NVIDIA/FastPhotoStyle .

Yijun Li, Ming-Yu Liu, Xueting Li, Ming-Hsuan Yang, Jan Kautz
arXiv:1802.06474 · cs.CV · submitted Feb 19, 2018 · updated Jul 27, 2018
abstract · pdf · html · Accepted by ECCV 2018

add comment on HN

This is essentially a dupe of https://news.ycombinator.com/item?id=16424813 (which links to the corresponding github page, which then has a link to this paper)
Thanks. We merged most comments thither.
Do you ever find yourself using that word in real life too? You're the first person I've seen to use it non-ironically. I will use it from now on, I quite like it.
I do my bit.
For those clicking on the Github link and getting an error, remove the . at the end.

https://github.com/NVIDIA/FastPhotoStyle

Style: "sentence with no punctuation"

Content: "https://github.com/NVIDIA/FastPhotoStyle."

Stylized Content: "https//githubcom/NVIDIA/FastPhotoStyle"

So friggin' close to realism! Damn ...

I notice on almost all of these image transformation projects the example images provided are usually very low res. It's basically impossible to judge how well this really works with the thumbnail sized images on the github page.
Why are small images not good enough to demonstrate how well this really works? What kind of issues would you expect to see for larger images?
Visual artifacts are often visible only at higher resolutions. Take the first photo, as an example: wouldn't you be curious to see how the road looks at a higher resolution? From the thumbnail - it might as well be horrible/ unusable. Same thing about the clouds.
Seems to be a linked thumbnail. If you follow the link, github says: This file is too big to show. Sorry!
The above link is also good for seeing examples of the effect in action.
It's strange that they used a Creative Commons license for software.

https://creativecommons.org/faq/

"The only categories of works for which CC does not recommend its licenses are computer software and hardware."

(although they seem to have taken on the FAQ section talking about this in more detail)

Wow, I had no idea about this. I've used it on software before since it looks well-suited to software, and I suspect others are in the same boat. Thanks for mentioning this!