Luminous Landscape Forum

Raw & Post Processing, Printing => Digital Image Processing => Topic started by: Rajan Parrikar on August 24, 2019, 04:59:56 am

Title: AI Sky Replacement by Luminar
Post by: Rajan Parrikar on August 24, 2019, 04:59:56 am
Coming soon.

https://youtube.com/watch?v=0eA6dUECpAA
Title: Re: AI Sky Replacement by Luminar
Post by: Bart_van_der_Wolf on August 24, 2019, 06:31:49 am
Coming soon.

https://youtube.com/watch?v=0eA6dUECpAA

While an interesting idea, I can only hope that there is enough variation in sky images to prevent all images from different people showing the same sky. The ability to use/import one's own sky images can help, but requires a large database of sky images at different times of the day and directions to avoid repetition. It remains to be seen how well the AI can match the light direction in the image with a sky, and how it uses the reflections of the sky in other parts of the image.

In a number of years from now, I can see AI generating skies based on several image clues. Modeling realistic clouds is a very timeconsuming thing, but advances in CGI and raw processing power will open up new possibilities.

Cheers,
Bart
Title: Re: AI Sky Replacement by Luminar
Post by: Rajan Parrikar on August 24, 2019, 06:58:28 am
Bart, what is the magic sauce here? I mean, how it it different from the conventional mask and replace? Is the AI deployed in mask generation?
Title: Re: AI Sky Replacement by Luminar
Post by: Bart_van_der_Wolf on August 24, 2019, 07:11:29 am
Bart, what is the magic sauce here? I mean, how it it different from the conventional mask and replace? Is the AI deployed in mask generation?

Rajan, what I understand of it from reading bits and pieces here and there, the AI attempts to detect light-direction/angle in the (landscape) image, then looks up a sky image that roughly had the same light angle (presumably with horizontal flipping), then masks out the bland sky and replaces that with the more cloudy version it looked up from its database. I also assume it changes color temperature some to get a better match.

But I have no inside knowledge of how they actually implemented it, or how well it performs in this first implementation.

Cheers,
Bart