Splitting texture from tone

The problem it solves

Skin has two kinds of imperfection and they need opposite treatments.

Tonal unevenness (a red patch on a cheek, a shadow under the eye, blotchy colour across a forehead) is large and soft. You fix it by smoothing.

Texture faults (a spot, a hair, a fleck of mascara) are small and sharp. You fix them by replacing.

Smooth the whole image and you destroy the pores, which is what makes retouched skin look like plastic. Heal every blotch and you get a lumpy patchwork, because the healing brush matches texture and averages tone across an area that legitimately varies.

Frequency separation splits the image into those two layers so each can be treated with its proper tool and neither affects the other.

Building it

The low frequency layer is a blurred copy: colour and tone with no detail. The high frequency layer is what the blur removed: detail with no colour or tone, held around a neutral mid-grey. Recombined, they reconstruct the original exactly.

In Photoshop or Photopea:

1. Duplicate the background layer twice. Name the lower copy low, the upper high.
2. Hide high. On low, apply Gaussian Blur. Choose the radius by eye: large enough that the blemishes disappear, small enough that the face’s structure (the nose, the eye sockets) is still recognisable. On a 20-megapixel portrait that is usually 5 to 10 pixels.
3. Show high and select it. Image > Apply Image. Set Layer to low, Blending to Subtract, Scale 2, Offset 128. (In a 16-bit document, use Blending Add, tick Invert, Scale 2, Offset 0.)
4. Set high to blend mode Linear Light.

The image now looks identical to the original, and it is: the arithmetic is exact, not approximate.

In GIMP the classic route needs no dialogue arithmetic. Duplicate twice, blur the lower copy, and set the upper copy’s mode to Grain extract with the blurred one beneath it; merge those two and set the result to Grain merge over the blurred layer. Same split, same reversibility. GIMP also ships Wavelet Decompose, which splits into several frequency bands at once, and the G’MIC plugin has Split Details, which is the most flexible of all.

Working on each layer

On high, use the Healing Brush and Clone Stamp. Spots, stray hairs, flecks. Because this layer holds no tone, healing here cannot leave a pale disc: there is no tone to get wrong. This is the single biggest practical gain of the technique.

On low, use the Lasso and blur, or a soft low-opacity brush. Select the blotchy region with a lasso feathered generously (20 to 40 pixels), then apply a Gaussian blur of roughly the same radius. The colour evens out and the texture above is untouched. Or sample a good colour nearby with the eyedropper and paint at 10 per cent opacity with a very soft brush.

The discipline that keeps it looking real: on low you are evening rather than flattening. A face has legitimate tonal structure: the shadow beside the nose, the warmth on the cheeks. Blur that away and the face goes flat and mask-like, which is the other way this technique fails.

Where it breaks, and why

Hard edges. At the jawline, where skin meets a dark background, the blur on low mixes the two. The high layer then holds a strong edge signature in compensation. Work on low near that edge and you get a visible halo: a pale or dark glow tracing the jaw.

The rule that follows: never work on the low layer across a hard edge. Stop the lasso short of the boundary, or mask the low-layer work away from it. If you must correct right up to an edge, do it on the original rather than in the separated stack.

Never clone across an edge on the high layer either. Cloning skin texture over the boundary between face and background copies the edge signature with it, and you get a ghost of the jaw in the middle of the cheek.

Wrong radius. Too small and the blemishes survive into low, so you are back to fighting them. Too large and the face’s structure lands in high, where healing it produces strange results. The test: look at low on its own. If you can still see the spots, the radius is too small. If the nose has disappeared, it is too large.

It does not fix shape. Frequency separation cannot move anything. A crooked nose, an awkward fold of fabric, a bulge in the wrong place is a warp problem, not a frequency problem.

Where it is actually used

Beyond skin: product packaging with printed texture, fabric with a weave, wood grain, scanned paper, painted walls, and anything where you need to correct colour without losing surface. A dented cardboard box responds very well: the dent’s shading is on low, the print on high, and you can flatten the shading while keeping the printing legible.

It is also the basis of texture transplantation: copy a patch of good texture from the high layer of one area and paste it over a destroyed area, then correct the tone separately on low. This is how a badly damaged region of a photograph is rebuilt.

The honest caution

Frequency separation is powerful enough to make bad retouching invisible to the retoucher. The high-frequency layer preserves pores, so the result reads as skin even after the low layer has been smoothed into something no face does.

Two checks. Toggle the whole retouch off and on at 100 per cent: if the person has become someone else’s idea of a person, reduce it. And view at the final output size, because at 400 per cent every pore is a decision and at 100 per cent almost none of them matter. Most over-retouching happens at zoom levels nobody will ever see the picture at.