Four different machines with one apparent job
A mark on a wall, a cable across a floor, a spot on a face. Four tools remove them, and they work in completely different ways. Knowing which one to reach for is most of retouching.
Clone Stamp. You Alt/Option-click a source point, then paint. It copies pixels from source to destination exactly — same colour, same brightness, same texture, no blending. Total control, and it will leave an obvious repeated patch if you are careless.
Healing Brush and Spot Healing Brush. Same copying of texture, but then it discards the source's colour and brightness and mathematically blends the patch's edges into the surrounding tone. That is why healing a spot on a cheek works when cloning leaves a pale disc — the healing brush borrows the texture and inherits the local colour.
Content-Aware Fill / Patch. Select a region and the algorithm searches the *rest of the image* for patches that fit, then stitches them in. Everything it puts in was somewhere else in your photograph.
Generative Fill. A diffusion model generates new pixels from a text prompt and the surrounding context. It is not copying anything from your image. It is inventing.
The free landscape: GIMP has Heal and Clone tools that are genuinely fine, plus Resynthesizer, a plugin that does content-aware fill well and predates Adobe's. Krita has a *Smart Patch* tool. Photopea has both a healing brush and a clone stamp. None of the free tools has generative fill built in; you would generate separately and composite by hand, which is the workflow described in AI images.
Where each one fails, and why
The healing brush next to a strong edge. Heal a cable where a pale wall meets a dark table and you get a grey smear. The mechanism: the tool blends the patch's edges into what surrounds it, and what surrounds it here is both the wall and the table. It averages them. There is no setting that fixes this. Clone Stamp, sampling from along the same side of the edge, is the correct tool, and afterwards you can heal the middle of the patch.
The clone stamp on anything with visible structure. Clone a section of brickwork and the eye finds the repeated mortar joint instantly. Take several small samples from different places rather than one big one.
Content-aware fill on something unique. Ask it to remove a person standing in front of a doorway and it will fill the doorway with more wall, because it can only use what exists elsewhere in the frame. There is no other doorway to borrow.
Generative fill on a real product. It will invent a plausible bottle cap that is not your client's bottle cap. Where the object must be the actual object, generate nothing — clone, patch, or reshoot.
Do it non-destructively
The trick that costs nothing: create a new empty layer above the photo and, in the tool options bar, tick Sample All Layers (Photoshop, Photopea) or *Sample merged* (GIMP). Now heal and clone onto the empty layer. The repairs live above the photograph, you can erase any one of them, and you can toggle the whole retouch off to compare. Every professional retoucher works this way.
For skin, the standard technique is frequency separation — splitting the image into a blurred layer holding colour and tone and a second holding pores and hair, so you can even out blotchy colour without sanding away the texture that makes skin look like skin. GIMP does it with *Wavelet decompose* or G'MIC; Photoshop and Photopea with a duplicate, a Gaussian blur and Apply Image. Learn it, because the alternative — blurring skin — is why so much retouching looks like plastic.
The line where retouching becomes a claim
This part is not a style question.
Removing dust on the sensor, a lens flare, a bin in the corner of a wedding photo, a scratch on a scanned family picture — nobody is misled. This is the darkroom work photographers have always done.
The line is crossed when the edit changes what the picture asserts about the world.
- Documentary and news. Reuters, AP and the major photojournalism competitions prohibit adding or removing content. Photographers lose their careers over a cloned-out sensor spot in a war photograph. If the image is presented as a record, the only permitted work is what a darkroom could have done: exposure, contrast, colour, crop.
- Bodies. Several countries now legislate this. France has required commercial images to carry a "photographie retouchée" notice where body shape has been altered since 2017; Norway passed a similar labelling requirement in 2021; Israel's law goes back to 2012. The direction of travel is one way.
- Products. Showing a product that does not look like the product is a consumer protection matter in most jurisdictions long before it is a design matter.
- People's faces. Consent for someone's likeness is not something you can retouch your way past. If you would not have photographed them doing it, do not generate them doing it.
- Generated content. If generative fill has invented an object, a person or a background that was never there, you have made synthetic media, and several countries — India's IT Rules on synthetically generated information among them — require that to be visibly labelled. The rules are recent and still moving. A lawyer in your country answers your specific case; the shape of it is in making things with AI.
A working test that survives every future tool: would the person looking at this be surprised to learn what you did? Dust removal, no. A removed person, yes. Between those sits your judgement, and writing down what you did in the delivery note costs nothing.
Today: remove one mark on a plain wall twice — once with the clone stamp, once with the spot healing brush, both on an empty layer. Then try the same mark sitting on an edge.
Before you move on