A New Magic Eraser

[Chris Kern]

Yet another interesting post-processing enhancement from Adobe based on machine-learning: a tool for removing window reflections from image files.  Publicly-available for testing in Photoshop/ACR and, according to the essay link in the previous sentence, apparently coming before too long to Lightroom.

Although it's apparently still in a fairly early stage of development, it's quite impressive.  I tried it out on an image that had required a lot of work in Lightroom with a round-trip to Photoshop, and it did at least as good a job of getting rid of the reflections in a single click than I previously had been able to achieve.  The first attachment is a lightly-edited rendered copy of the raw file before reflection removal; the second shows the same image after using the tool.  The new feature didn't remove the reflected red dots on the right side of the subject's skirt, but they could easily be excised with Lightroom's generative remove eraser.

The window reflection removal tool doesn't employ "generative AI" techniques, so it won't introduce the kind of artifacts that sometimes occur when tools based on machine-learning recreate parts or all of an image rather than simply modifying it.

I was also pleased to discover that the Adobe developers have published a journal article describing how they were able to achieve this latest magic.

(I have no connection to Adobe except as a customer.  I just happened to stumble upon the Adobe Blog post this afternoon.)

[Slobodan Blagojevic]

Thanks for sharing, Chris.

[Chris Kern]

Attached: an unsuccessful example.  The window reflection removal tool had essentially no effect.  It was unable to distinguish the reflections on the glass from the interior of the shop.

As I've discovered with other types of tools based on machine-learning, it's easy to determine when they work well and almost impossible to figure out what was going on inside the black box when they fail.  The details of how the neural network "learns" what its designers want it to accomplish during the training phase are typically opaque—often to the designers who created the software themselves and certainly to its end-users.  The machine has its own inscrutable internal logic: it works in mysterious ways.

[digitaldog]

I have been testing for a while, and indeed, some images work far better than others. It is one of those "try it" and see if you like it workflows; hopefully, the model will improve over time (?).

[francois]

Thanks for the article! I have not yet tested that tool but the technical read is interesting.

[KLaban]

Not one for RobC then.