Remove an Unwanted Object From a Photo With AI, Pass by Pass

The same technology that erases a watermark erases a stranger’s elbow from your holiday photo, and it is a much better use of it. Object removal differs from text removal in three ways that decide the outcome: the hole is bigger, the missing content is usually more structured, and there is nearly always a shadow you have to deal with as well.

What inpainting is good at here

WatermarkRemover is sold as a watermark tool, but the mechanism is general: you mask a region, and a LaMa-class model (Large Mask Inpainting; source) synthesizes that region from context elsewhere in the frame. Its wide receptive field is why it can extend a lawn, a pavement or a mountain ridge rather than smearing a blur.

Good targets, roughly in order of success rate: power cables and wires against sky; litter, bin bags and bollards on pavement; a distant stranger against a wall or sea; tripods and poles; date stamps and logos; blemishes and small skin details; a sign or poster on a plain wall.

Mask strategy: four rules

1. Include the shadow. The object is never just the object. A removed bollard with its shadow left behind looks worse than the bollard. Mask the contact shadow and the soft falloff, or the removal reads as a paste-over.
2. Give it margin. Paint a couple of pixels outside the object. A mask that sits just inside the edge leaves a permanent outline of the thing you removed.
3. Do not mask what you can keep. For a person standing in front of a wall, mask the person and not the wall behind them — extra area is context you threw away.
4. Split large regions into passes. One mask over 30% of the frame gives the model nothing to work from. Remove the torso, then the legs, then the shadow.

Where it visibly fails

Situation	Why it breaks	Better answer
Object in front of brick, fencing, tiles or a window grid	The pattern must continue exactly; drift is obvious	Manual clone/patch work with a reference line
Removed object should reveal a person or hand behind it	Anatomy is not extrapolatable	Generative fill, or accept the artefact
Object covers a face	Same reason	A face-aware model, not general inpainting
Very large object over complex scene	Little valid context remains	Re-shoot, composite from a bracketed set, or generative fill
Text in the scene (sign, licence plate)	Letterforms must be invented	Hand reconstruction; or crop
Reflections and water	Physically, removing the object should change the reflection too	Mask object and reflection as one region, expect softness

The tool is candid about the general version of this: small marks over fairly uniform backgrounds are the sweet spot, and large high-contrast subjects over detailed areas are hard for any inpainting model. Nothing about the browser implementation changes that — it is a property of the method.

A clean workflow

1. Load the photo: PNG, JPG or WebP up to 25 MB. No account, and nothing is uploaded — the model runs on your own CPU or GPU after a one-time download of roughly 200 MB.
2. Zoom to about 100% and mask the object plus its shadow, with margin.
3. Erase. Inspect at full zoom, then again zoomed out: a region can look fine at 100% and wrong at 20% because the structure does not continue.
4. Run a second, small pass on leftovers. Undo is per stroke, so a bad mask costs one action.
5. Export PNG. The untouched pixels stay identical to your original; only the painted region changes, so a lossless save preserves that.
6. Convert to JPG once at the end if the destination needs it.

On a phone the same tool falls back to WebAssembly and works, just more slowly — fine for one photo, painful for ten. Our image tools page lists the desktop editors worth having for the manual finishes.

Adjacent jobs, different tools

• Take the background away instead of filling a hole: that is matting — make an image background transparent.
• Clean a catalogue image: product photo background removal, where the shadow question is different because you want one.
• Small text or a handle: text watermark removal, same tool, much smaller mask.
• Prints and scans: scratch repair is an inpainting job, and the old photo guide puts it in the right order with scanning and upscaling.

One more thing worth saying: an image with people removed from it is still somebody’s image. Removing a tourist from your own holiday photograph is ordinary editing; removing a person from a photograph so it can be published as something it is not is a different act entirely. See the logo guide for the rights line.

Frequently asked questions

Can AI remove an object from a photo and fill the gap?

Yes, when the surrounding area is reasonably predictable. A LaMa-class inpainting model rebuilds the masked region using context from across the whole frame, so lawns, walls, roads and sky come back convincingly. Structured patterns, anatomy and legible text are where it visibly fails.

What is the most common mistake in object removal?

Forgetting the shadow. The object and its contact shadow are one visual event; mask both, including the soft falloff, or the result looks pasted.

Should I remove an object before or after upscaling?

Before. Repairing a 200-pixel hole is easier than repairing a 3,200-pixel one, and the extra detail from upscaling is texture the model then has to match. Order: clean, then correct colour, then upscale.

Does the photo get uploaded to a server for this?

With an on-device tool, no. WatermarkRemover decodes, inpaints and exports locally, with a one-time model download of roughly 200 MB. That matters more for photos of people you know than for anything else in this guide.

Will the file still look untouched?

Only the pixels you painted change; everything else is preserved at full resolution. If you export PNG you keep that exactly. Re-saving to JPEG is what alters the rest of the frame, by re-compressing it.