Microsoft has quietly shipped one of the most practical photo-editing models of the year, and it goes by the name Microsoft LensErase. Officially it's "LensErase: A Foundation Model for Erase Anything," and it does exactly what that label promises: you mark something in a picture, and the AI makes it vanish. Not just the object you circled. Its shadow, its reflection, and the watermark sitting on top of it go too.
What happened
Microsoft researchers published LensErase as a research release, and the workflow is almost insultingly simple. You give it two things — a photo and a mask — and it hands back a clean image. No text prompt to write, no sliders to fiddle with. The picture comes back with your selection gone and the background rebuilt underneath it.
The mask is where all the control lives. Think of it as a strip of painter's tape: it tells the model which pixels are off-limits and which ones it's allowed to rewrite. You can draw that mask by hand, or you can let a segmentation tool like SAM 2 auto-select the object for you. Either way, the model only touches what you've marked.
What the model can actually erase
The list is longer than you'd expect from a single tool. Watermarks and logos are covered, including the sneaky kinds: semi-transparent overlays, repeating marks, diagonal text running corner to corner, stripes, and small stamps tucked into the corners. It also removes objects — people, animals, furniture, whatever you point at — whether they're tiny details or big shapes eating up most of the frame.
It handles several victims at once, so you can tag a few different things scattered around the shot and clear them in one pass. It erases text while keeping the surfaces and neighboring elements intact. And it edits finished posters, pulling out individual headlines, logos, or badges and rebuilding the space they left behind.
The clever part: it can see what's hidden underneath
Here's the detail that separates LensErase from the erase brushes you've used before. Most tools black out the masked area and ask the AI to invent something to fill the hole. This one shows the model the full contents of the selection instead.
That's a big deal for watermarks. A semi-transparent overlay still lets the background show through, and those visible fragments give the model real information to rebuild from. Same logic with shadows and reflections: by looking at the object itself, the model can work out which dark smudges and mirrored shapes actually belong to it — even when they spill outside the mask you drew. A mechanism called TraceFusion then decides the true boundary of the change, so everything else in the photo stays untouched.
Trained on a mountain of before-and-after examples
To get there, the team assembled an enormous library of training examples, each one a matched set of original image, mask, and clean result. A big chunk of those were watermark cases, with separate batches for text, posters, and large-area removals. The scale is the story here: the model has seen just about every way a logo, a stamp, or a strip of text can be dropped onto a picture, so the odds are good it's met yours before.
What it means for you
You don't need to train anything or write code to benefit from this. Here's where it lands in real life.
At home: rescue the good photos
Family pictures are full of junk — a date stamp in the corner, a stranger in the background, a tourist sign ruining the shot. Today you either live with it or pay for a retouching app. Erase tools that understand shadows mean you'll clean up a photo in a minute on your phone instead of deleting it.
At work: slides and decks that look professional
Screenshots come with watermarks, old logos, and stray interface elements. If you build decks, write docs, or make tutorials, you'll strip those out without cropping half the image or blurring the spot into a grey smear.
For business: clean up your own product shots
Marketplaces and ad platforms have strict rules about how images look. Removing a competitor's watermark isn't okay — but removing your own outdated logo from a product photo, or clearing a prop out of a catalogue shot, is. For small brands shooting products on a kitchen table, that's a real saving on re-shoots.
For students: diagrams without the clutter
Textbook screenshots arrive covered in labels, page numbers, and publisher logos. Being able to delete a single annotation while keeping the diagram underneath makes study notes much cleaner.
For creators: posters that stay editable
Poster and thumbnail design usually means flattening everything and starting over when one headline changes. If the model can lift a single title off a finished poster and reconstruct the artwork behind it, you can version a design in minutes. If you want to see what's already available without waiting for a research release, the free AI tools at mykreatool.com cover a lot of the same quick jobs.
For extra income: retouching on demand
Photo cleanup is steady freelance work — real estate listings, old portraits, e-commerce catalogues. A tool that nails shadows and reflections cuts the time per image, which means more jobs per week at the same rate.
How to try it right now
Start with the free stuff, because that's genuinely all there is at the moment.
1. Read the paper — it's open and free, no account needed. The full write-up lives at LensErase on Hugging Face Papers.
2. Want a second copy with a different reader layout? There's a mirror at papers.cool.
3. Look for the GitHub link on the paper page. That's the "open source — well, almost" part: the project points to public code access rather than a finished consumer product.
4. Follow the team's releases on the Microsoft page on Hugging Face, and keep an eye on Hugging Face Papers — that's where models like this surface first.
5. Prepare your mask with SAM 2 if you want automatic selection, or draw it by hand in any editor. Then run image + mask through the model.
6. No patience for any of that? Use an existing object-removal tool today. Everything you've just read about is the direction the whole category is heading.
Upsides and what changes
No prompt-writing skills required
You don't have to describe what you want in words. Point at it, and it's gone. That's a much lower bar than prompt engineering, and it works the same for people who've never touched an AI tool.
Shadows and reflections finally count as part of the object
This is the thing that used to give away every retouched photo. A deleted lamp with its shadow still on the wall screams "edited." Teaching a model to treat the shadow as part of the object fixes the tell.
One pass for several objects
Marking five things at once instead of erasing them one by one is a small change that saves a lot of clicking on busy images.
Finished designs become editable again
Being able to lift a headline off a flattened poster and rebuild the art behind it is the difference between a two-minute tweak and a full redesign.
Limitations
Be clear-eyed about three things. First, what Microsoft published is a research release — there's a paper and a link to code, not a polished app with a friendly upload button, so you'll need some technical comfort or patience to use it directly. Second, the training data is enormous, but it can't cover every busy, high-detail background; complex textures behind a removed object may still look slightly off. And third, and most important, the ability to strip watermarks and logos cuts both ways — using it on images you don't own or haven't licensed is a copyright problem no matter how good the tool is. Treat it as a cleanup tool for your own material, not a way to get free stock photos.
Conclusion
Microsoft's LensErase goes after the one thing every other erase tool fumbled: the leftovers. Shadows, reflections, and see-through watermarks are what make retouching look fake, and a model trained on an enormous pile of before-and-after examples has seen enough of them to clean up properly. You can't buy it in an app store yet, but you can read exactly how it works today.
One action for today: open the LensErase paper, skim the example images, and pick one photo of your own that needs a cleanup. Bookmark it for the moment this kind of erasing lands in the tools you already use — because it's coming.
Comments 0