[cprobertson1]

I was recently playing around with a neural net trained to remove parts of images.

Load a photo, then draw on it to highlight the bits you want removed. I've tried it with removing handles from doors, and streetlights that were ruining a skyline... I then used it to remove the horizon itself so it was just sky!

I thought I would share it here, as I found it really impressive... Its doing a better job than I can with my big human brain!

Cleanup.pictures - Remove objects, people, text and defects from any picture for free

[robgski]

That is quite impressive, I used it to remove street signs, powerlines and utility poles from a street scene, a fairly complex jumble of vertical, horizontal and arcing lines. This product removed the objects I did not want without affecting the objects I did want to remain in the scene.

I must consider this further.

[BigMackCam]

That is indeed very impressive - at least, given the first of my attempts. I'd love to see how it fares longer term...

[kiwi_jono]

Wow thats pretty impressive. As others have said its capable of doing a better job than myself (in many cases at least).
I might give it a go - although max size is 2k for pro thats good enough for some uses.

[texandrews]

Hmmmm...thanks for the tip.

[bertwert]

Thought I'd give it a challenge...

I decided this bear completely ruined the shot I was trying to take of the trees here.



So popped it into that photo cleanup thing to try and get rid of the pesky bugger.




Sure there's a few artifacts in there, but it's pretty impressive continuing most of the trees down and trying to emulate the patchwork of bushes. Squint a bit or take a couple steps back from the computer and you wouldn't even know anything is up...

[BigMackCam]

I decided this bear completely ruined the shot I was trying to take of the trees here.

LOL Best post of the month, Bert

[biz-engineer]

I decided this bear completely ruined the shot I was trying to take of the trees here.

Yeah, usually you want to remove branches in front of the bear. BTW, I can still see the shadow of the bear even after you removed the bear and the bushes behind the bear looks like digitally grown bushes.

[cprobertson1]

Yeah, usually you want to remove branches in front of the bear. BTW, I can still see the shadow of the bear even after you removed the bear and the bushes behind the bear looks like digitally grown bushes.

... I wonder what happens if you try to remove that dark patch and the digital bushes 🤔

--EDIT--
The AI doesnt quite know what to do with it and creates more artifacts. Interesting!

[BigMackCam]

... I wonder what happens if you try to remove that dark patch and the digital bushes 🤔

It replaces them with a bear?

[cprobertson1]

The thing to remember is that its a neural net that was trained on a certain dataset to handle certain things - what you're effectively seeing is how a computer imagines the image to look like without the removed area...

It has no way of knowing whether what it imagined is realistic or not - and it can only imagine within a certain scope (defined by, but not necessarily limited to the original dataset).

In other words, if you try to make it do something it wasn't trained on, your mileage may vary!

I still find it mighty impressive! 😄

[robgski]

In other words, if you try to make it do something it wasn't trained on, your mileage may vary!

Exactly so. I have the good fortune to work with a data science team and participate in some AI related conferences, and it is amazing to think of the advances that have been made in this field, especially related to images, but the amount of preliminary data gathering and data proofing necessary to make data sets to train AI over many iterations is astounding.

Several of the comments here touch on the benefit of such an AI application, the reduction of what I call "necessary drudgery", the repetitive tasks which all too often eat up a human being's time and prevent the type of improvements in the process or the efficiency of the human in the loop. Decreasing the drudgery allows the human to do more of the complex reasoning or fine detail work that humans still do better, often intuitively. This allows for improvements in operations at scale and velocity

In a "perfect" situation, the AI team and human worker interact with lots of feedback so the WRONG lessons are not being learned.

Personally, I'd be happy to have an AI assisted method of managing my teams inbox, as well as one that will create the spreadsheet formulas for me, as simple examples.

[texandrews]

Wonder if it works on reflections...?

[35mmfilmfan]

WOW !!!!! That is truly amazing ! I can see an afternoon of 'improvements' ahead of me.

And yes, it does seem to work on reflections too ! It even adds an unique number if you save the same image after making further adjustments !

[robgski]

Wonder if it works on reflections...?

Some cursory experiements indicate that it will work to eliminate reflections from a window only if whatever is behind the reflection is simple/nondescript.
If a reflection is obscuring the furnished interior of a room, or a person, this correction might eliminate some of the details behind the reflection as well. I imagine this will improve in future versions, since the use case is common enough to drive demand.