Strange Trouble with Generative Erase Tool

[Jordan26437343lemp]

I'm trying to get into using Lightroom's new AI Generative Erase tools, but odd (I think) behavior I've experienced was with an image in which I was trying to remove a car from the foreground. The good news is that in a way it did so, but the bad news is that it just replaced it with a different car! As if my problem wasn't the presence of the car itself but what kind of car it is! Obviously that's not the case -- what I was hoping for was a normal-looking street where the car was located. Is this is a known bug, or am I doing something wrong here? Thank you.

[hanomann]

I think the selection had the form of a car. Try to enlarge the selection a little, so that it gets another form.

[Jordan26437343lemp]

Thanks, but no, didn't work unfortuantely. I'm wondering what the rationale even is there, though. If it recognizes the object as a car and I'm using generative ERASE (not "replace!"), why would I need to do that? How would adding "non-car" subject matter help it erase the car? 

[johnrellis]

"I'm wondering what the rationale even is there,"

The Remove tool always replaces the selection with something else that it think matches, generated from its millions of training photos. If you don't select all of an object, its shadows, and its reflections and a generous boundary around it, then Remove will generate a replacement trying to match what remains, often generating a similar object.

[JohanElzenga]

To build upon what @johnrellis said: do not try to make a tight selection like you would do with traditional tools. Make a very generous selection and make sure you include everything that could be associated with the object. I have seen very funny results if you leave just a trace of the original object. For example: I tried to remove two big dustbins from a photo. The result was almost like Lightroom played a prank on me. Like it said: "If you remove the dustbins, then people will throw their rubbish on the floor", so it replaced the dustbins by a big pile of rubbish...

[johnrellis]

See this article for how to remove objects more reliably:
https://www.lightroomqueen.com/generative-remove-replace/

Most complaints about Remove are addressed in the article. But if it doesn't help, please attach the original photo here (not an export) so we and Adobe can see the issue in detail. If the forum won't let you attach it, upload it to Dropbox, Google Drive, or similar and post the sharing link here.

[Jordan26437343lemp]

I tried working with that a bit and will continue to try, but I feel I should attach the image (right out-of-camera as requested, not the exported jpeg) just in case. It's probably something Adobe should see! Here you go. 

PS. Thanks for the warning. No, the forum would not let me attach it directly.  https://www.dropbox.com/scl/fi/1cgvvh3g9stnjpr9tcpgc/DSC_3709.NEF?rlkey=u03ohy0j922n6yhwjkvypzb2o&st...

[johnrellis]

Which care are you trying to remove?

[johnrellis]

[This post contains formatting and embedded images that don't appear in email. View the post in your Web browser.]

I was able to quickly remove all the cars in the fore and middle-ground using Generative AI + Object Aware:

[Jordan26437343lemp]

That's incredible work! 

To answer your question, though, the one I was specifically trying to remove was that minivan -- the first (almost) full vehicle on the road. You wanted the original photo so that's what I provided, but in reality, I had already cropped the image a bit so that the first one you see just a little of was no longer even there without my having to erase anything. 

So, in the spirt of "teaching a man to fish," just to be clear, you clicked on the icon that looks like a little padlock, made sure the second little padlock was selected (not heal or clone), left "object aware" and "generative AI" checked, brushed over it, and just clicked "Apply?"

[johnrellis · Accepted solution]

@Jordan26437343lemp, I've attached a screen recording showing Remove mode removing that car from the uncropped original photo, using the options Generative AI and Object Aware.

"in reality, I had already cropped the image a bit so that the first one you see just a little of was no longer even there without my having to erase anything."

That crop is causing Remove to replace the car with another car. I've attached a second screen recording showing exactly what happens.

Remove always looks at all the pixels of a photo, even those that have been cropped out. When you crop out part of the car and then select it with Remove, not all the pixels of the car get selected. As it always does, Remove tries to replace the selected pixels with a new generated image that best matches the surrounding unselected pixels. Since there are bits and pieces of car surrounding the selection, Remove does a good job of generating a replacement that matches!  It can't simply replace the partially selected car with pavement since that would look very jarring against the remaining unselected parts of the car.

So the rule is: Remove first, then Crop. (Or remove an existing Crop, then Remove, then reapply the Crop.) The article above explains that there are other ways pixels can get cropped out: Lens Corrections, Transform, and with some cameras, the hidden cropping that occurs when you first import the photo.

This is tripping up many, many people, as seen in this thread. Adobe has designed all the Remove tools (Remove, Heal, Clone) to be applied first, before you use Crop or AI Masking. But that doesn't match the natural workflow of many people, who want to do the "big" edits first (Crop, Exposure, adjust with masks), decide if the photo is a keeper, and only then clean up instrusive objects and spots. In that workflow, it would be wasting time to clean up the photo first with Remove only to decide it's not a keeper.

[Jordan26437343lemp]

That's awesome -- thank you! Yes, that did work a lot better. Attached is the final product. 

[johnrellis]

It still boggles my mind what AI can do compared to just a few years ago.

[Jordan26437343lemp]

Very true! Makes me wonder what's still to come.