NVIDIA Develops Lightweight Algorithm "FastPhotoStyle" to Transfer the Characteristics of Photos to Another Photo

GPV's largest NVIDIA "transplant" the distinctive "style" of the image image to a completely different pictureFastPhotoStyle"We developed an algorithm called" FastPhotoStyle has already been released for free on GitHub.

[1802.06474] A Closed-form Solution to Photorealistic Image Stylization
https://arxiv.org/abs/1802.06474

Several techniques using deep learning have already been developed to transfer the visual features of the photo to another photo.

"Deep Photo Style Transfer" which transfers "appearance features of photographs" to another picture using deep learning - GIGAZINE


According to NVIDIA, traditional technology required high computing power for processing, and it took several minutes to process VGA (640 × 480 pixels) size photographs. However, it seems that FastPhotoStyle using the newly developed algorithm clears these computing cost limits.

FastPhotoStyle goes through two processes: "stylization" step and "smoothing" step. In the stylization step, I change the style of the reference photo source to the target photo and spatially arrange the style in the smoothing step. Unlike conventional algorithms, it is unnecessary to iterate over optimization, and it seems that both stylization and smoothing are closed form solutions.

Examples of using FastPhotoStyle to transfer features are as follows. The characteristic style of "Style Photo" on the left of the photo is "Stylized Content" on the right of the photo, which is replaced by FastPhotoStyle to "Content Photo" in the center of the photo.


NVIDIA publishes FastPhotoStyle with GitHub.

GitHub - NVIDIA / FastPhotoStyle


In addition, the technical idea of ​​FastPhotoStyle can be read in the following papers.

[1802.06474] A Closed-form Solution to Photorealistic Image Stylization
https://arxiv.org/abs/1802.06474