Remove Objects from Photos

Every photographer has a picture that would be perfect if one thing were not in it. A stranger walking through the frame at the exact wrong moment. A bin beside a beautiful doorway. A charging cable snaking across a product shot. The date stamp your old camera burned into the corner. For twenty years the fix was the clone stamp: copy a patch of background from next door, paste it over the offender, repeat a few hundred times, and hope nobody notices where the brick pattern stutters. It was slow, fiddly and unforgiving — a task for people who edited photos for a living.

Generative inpainting changed the nature of the job. Instead of copying pixels around, the software understands what the scene is made of and paints in what would have been there. It knows that paving continues in straight joints, that a hedge is made of leaves rather than a single texture, that a shadow falls in one direction across the whole frame. You mark the area that should change, and the model synthesises new content that continues the surroundings. The result is not a patch — it is a plausible piece of the original scene that was never actually photographed.

The practical consequence is that object removal has moved from the specialist’s toolbox into a browser tab. You drop in a photo, paint over what should disappear, wait a few seconds, and download a photo that looks as if the thing were never there. No layers, no selections, no tutorials. What follows is a guide to doing it well: the workflow, the settings that matter, the scenes that are easy and the ones that are not, and how to keep your photo private and at full resolution while you do it.

The tool is called a brush for a reason. You paint a mask — a set of pixels the AI is allowed to change — directly over the object, the person or the text you want gone. Everything outside that mask is off limits. The brush hint in the editor puts it plainly: paint the area the AI should change; everything outside stays exactly as it is. This single rule is what separates a controllable removal from a lottery. You decide the boundary; the model works inside it.

When you press Apply, the editor does not upload your whole photo. It cuts a working window around your brush strokes — the painted area plus enough surrounding context for the model to understand the textures, lighting and perspective it needs to continue — and sends only that window, together with the mask, to the image model. The faces at the other end of the frame, the licence plate in the corner, the paperwork on the desk behind your subject: none of it leaves your browser. The window and the mask are processed, returned, and discarded. Nothing is stored — no image, no mask, no prompt.

The returned fill is then blended back into your original. The editor resizes the generated window to the exact pixel dimensions it was cut from, feathers the mask edge so the seam disappears, and composites it over the untouched original. Because the model works on a window of at most 1536 pixels rather than the entire frame, a small removal happens at essentially full detail, and every pixel outside the mask is bit-for-bit what you loaded. A 24-megapixel photo comes back as a 24-megapixel photo with one region rewritten.

The whole process takes a minute or two once you know the rhythm. Here is the sequence that gives the cleanest result on the first pass, from loading the photo to checking the seams.

• Load the photo: drag it into the editor, choose it with the file picker, or paste it from the clipboard with Ctrl+V.
• Open the AI brush from the tool bar and set the Brush size a little larger than the finest detail you need to cover — you can shrink it later for edges.
• Paint over the object, person or text with a small margin: a few pixels of surrounding background inside the mask gives the model room to blend the fill.
• Include the shadow and any reflection the object casts. A removed person who still leaves a shadow on the sand looks stranger than the original.
• Use Erase to trim any strokes that strayed onto something you want to keep, or Clear to start the mask again.
• Choose Remove. Leave the description empty for a plain removal, or write something short and concrete like “empty pavement” if the background is ambiguous.
• Press Apply. The AI works for 10 to 30 seconds and the result appears in place; only the painted area has changed.
• Hold Compare to flick between before and after, zoom in on the edges of the filled region, and paint just the spots that need a second pass.
• Open Export, pick PNG, JPG or WebP, decide whether to keep the EXIF metadata, and download at the full resolution of your original.

Nine times out of ten, a disappointing removal is a mask problem, not a model problem. The most common mistake is painting too tightly. If your strokes trace the exact outline of a person, the model is asked to fill a person-shaped hole while a faint halo of their edge pixels remains around it — a ghostly outline that no amount of re-running will cure. Paint generously instead: a margin of a few pixels, more for soft or blurry edges, so the model sees clean background all the way around the region it fills.

Shadows and reflections are the second trap. Objects interact with the scene: a lamp post throws a stripe across the pavement, a glass on a table leaves a ring of light, a person standing on wet sand has a mirror image beneath them. Remove the object and leave its shadow, and the picture develops an uncanny quality that viewers notice before they can say why. Paint the shadow as part of the mask. The model understands that a shadow belongs to something and will lay down uninterrupted ground where it was.

Edges where the object meets something else — a person leaning on a railing, a bin against a wall corner — deserve a smaller brush and a zoomed-in canvas. Paint right up to and slightly over the boundary, so the model can decide how the railing continues behind the removed body rather than guessing from a half-covered edge. If the first pass leaves a hint of a sleeve or a smudge where two textures meet, do not repaint the whole mask; paint only that spot and apply again. Small, targeted second passes are faster and more precise than starting over.

The AI brush offers two closely related operations for this kind of work. Remove asks the model to continue the surrounding background into the painted area — the right choice when you simply want something gone. Replace asks it to put something specific there instead — the right choice when the background behind the object is unknown or you want to change it, for instance swapping a crowded pavement for empty flagstones, or a grey sky for a clear one. Remove can be used with no description at all; Replace needs one, and the editor will ask you to describe what should be there.

For plain removals, the mask does nearly all the work and the best description is often none. The model reads the surroundings and continues them. When the surroundings are ambiguous — a person standing where a wall meets a hedge, say — a short noun phrase settles it: “brick wall and hedge”, “wooden fence”, “grass”. Describe the result you want, not the action. “Empty wooden table” works better than “remove the cup and make it look like there was never a cup”. Nouns and materials help; long sentences, instructions and negations such as “no people” tend to confuse.

For replacements, be concrete and visual, and mention lighting when it matters: “clear blue sky with a few thin clouds” beats “nicer sky”, and a single word such as “sunlit” or “overcast” prevents the most common tell of an edit, a patch lit differently from the rest of the frame.

Some removals are trivial. A lamp post against a plain sky, a bird in front of a cloud, a cable across a wooden floor: the background is simple and continuous, and the model fills it perfectly on the first try. Grass, sand, water, walls, sky and tarmac are all forgiving. So are repeating textures — tiles, brick, fabric — as long as the mask is not so large that the model loses the pattern’s rhythm.

Hard removals are the ones where the background behind the object is both detailed and unknown. A person standing in front of a shop window with a display behind them, a dog sitting in front of a bookshelf, a car parked across a mural: the model has never seen what is behind the object and must invent it. The result will be plausible rather than accurate — books of the right size and colour, but not your books. This is not a flaw so much as a limit of the physics: information that was never captured cannot be recovered, only imagined.

For large or hard removals, work in stages so the model always has real context around each mask: remove the outer part of a crowd first, then the inner part. And for objects near the edge of the frame, a crop is often the cleanest removal of all — the Crop tool runs on your device, instantly, and costs nothing from your daily allowance.

A removal that comes back imperfect can almost always be fixed with a second, smaller pass. Before you re-run anything, hold Compare and zoom to 200 per cent on the edges of the filled region — that is where the problems live, and it tells you which of the following to try.

• A faint outline or halo of the object remains: the mask was too tight. Paint a wider margin around the whole shape and apply again.
• A shadow or reflection is still there: paint it as part of the mask. The model fills the ground where it was.
• The fill is blurry or the texture looks smeared: the mask is very large. Split it into two or three smaller passes so each works at full detail.
• A straight line — kerb, horizon, wall edge — is broken or offset: paint a narrow stroke across just the break and apply, or use Replace with “straight kerb line”.
• The patch is lighter, darker or a different colour from its surroundings: add a lighting word to the description (“in shade”, “sunlit”) or repaint including more of the surroundings.
• The wrong thing was filled in — a new object appeared where you wanted emptiness: use Remove without a description, or describe the background explicitly (“empty grass”).

It is worth being precise about what leaves your device, because most online editors are not. When you crop, straighten, adjust the light or blur the background, nothing is uploaded: those tools run in your browser using your own graphics hardware, and they are unlimited. When you press Apply on the AI brush, only the painted area — cropped with a margin of context — is sent to the server and forwarded to the image model. It is processed and returned in ten to thirty seconds, and nothing is kept: no image, no mask, no prompt. The button says so before you press it.

Resolution is preserved by design rather than by promise. The AI works on the painted window at up to 1536 pixels, and the editor blends the result back into your original at its native size. There is no downscaled export, no “HD” paid tier, no watermark in the corner. A 24-megapixel photo comes back at 24 megapixels, and you choose the format and quality yourself in the Export panel.

Because generative models run on expensive server hardware, AI operations are metered: twenty per day per visitor, shared with the image generator on the same site, with the count shown on the Apply button and a reset every twenty-four hours. That is enough for a serious afternoon of retouching if you use it deliberately — mask carefully, remove in small passes, and lean on the free on-device tools for everything that does not need intelligence. When the allowance is spent, cropping, light, colour, effects and background blur keep working without limit.

“Does it need an account?” No. There is no sign-up, no email address, and no watermark on the download. “Will it change anything I did not paint?” It cannot: pixels outside the mask are your original file, untouched. Use Compare to confirm this on any edit. “Can I remove several things at once?” Yes — paint all of them and apply once — but separate passes usually give cleaner fills, since each smaller mask works at full detail and you can judge each result on its own.

“Why does the fill look slightly soft on a very large mask?” The model works at up to about 1,500 pixels within the painted window. For small masks that is more than the original detail; for a mask covering a third of a large photo, the fill may be a touch softer than its surroundings. Smaller, targeted masks give the sharpest results. “Can I undo an AI removal?” Yes; the History panel keeps every step and Undo walks back through them. “Does it work on my phone?” Yes — paint with a finger on a zoomed-in canvas, pan with two fingers, and download to your photo library.

“What can’t I remove?” Watermarks and copyright notices from other people’s work, and anything that alters a real person’s identity or dignity — those requests are declined by the safety filter. A watermark you placed on your own photo is yours to remove.

The tourist in front of the cathedral, the cable across the product shot, the date stamp on the scan from 1998: each one is a minute’s work with a brush and a good mask. Paint generously, include the shadow, leave the description empty unless the background is ambiguous, work in small passes, and hold Compare before you download. The rest of the photo — the part you loved — stays exactly as you shot it.

PixGenia’s object remover is built on exactly the principles in this guide: paint what should disappear with the AI brush, and only that area is sent to a state-of-the-art image model, blended back into your original at full resolution. Twenty free AI operations a day, no account, no watermark, and unlimited on-device cropping, light and colour to finish the job. Drop in your photo and see for yourself.