{"course":{"slug":"image-editing","title":"Editing an Image","url":"https://addaly.com/learn/image-editing","level":"beginner","track":"create"},"lesson":{"slug":"sharpening-and-noise","title":"Sharpening adds no detail, and it goes last","url":"https://addaly.com/learn/image-editing/sharpening-and-noise","position":70,"of":78,"minutes":9,"answer":"Sharpening only exaggerates edges that already exist, so do it after resizing, at the output size, and after noise reduction rather than before.","keyIdea":"Sharpening only exaggerates edges that already exist, so do it after resizing, at the output size, and after noise reduction rather than before."},"body":{"format":"markdown","text":"## What Unsharp Mask actually does\n\nThe name is backwards on purpose, and it tells you the mechanism. The software makes a blurred (unsharp) copy of your image, subtracts it from the original, and what remains is a map of the edges. It then adds that edge map back at increased strength.\n\nThe result is that along every edge, the light side is made lighter and the dark side darker. That local contrast increase is what your eye reads as sharpness. No detail has been added. Nothing that was not in the file is now in the file. You have exaggerated the transitions that were already there.\n\nThree controls, and they mean exactly what the mechanism implies:\n\n- **Amount**: how much of the edge map is added back. 60 to 120 per cent for most photographs.\n- **Radius**: how wide a blur was used, and therefore how thick the resulting bright and dark rims are. Small, 0.5 to 1 pixel, for fine detail. Large radius over about 3 is not sharpening any more, it is local contrast, and it makes skies look dramatic and skin look grubby.\n- **Threshold**: how different two neighbouring pixels must be before the tool treats them as an edge. Set it to 0 and you sharpen the noise in a clear sky along with everything else. 2 to 4 protects flat areas.\n\nOver-sharpen and you get **halos**: a pale outline around dark branches against a sky, and a dark outline around bright ones. Once you have seen one you will see them everywhere, including on television.\n\n**High Pass** is the same idea with better control. Duplicate the layer, run *Filter > Other > High Pass* at 1 to 2 pixels, set the layer blend mode to Overlay or Soft Light, and adjust the layer opacity to taste. Because it is a layer, it takes a mask, so you can sharpen the eyes and not the skin. GIMP, Krita and Photopea all have both Unsharp Mask and High Pass.\n\n## Why it goes last\n\nEvery operation you do after sharpening changes the pixel scale that the sharpening was tuned for.\n\nSharpen a 4000-pixel-wide image, then resize it to 800 for a website, and the halos you created were sized for 4000 pixels. Shrinking the image does not shrink them proportionally in a clean way: the resampling stacks neighbouring halo pixels together, and the small version comes out crunchy with pale rims around every branch. Do it the other way round: resize to 800, *then* sharpen at 0.5 pixel radius, and the same photo looks clean.\n\n> Figure: The same photograph delivered at 800 pixels, two ways. Sharpened at 4,000, then resized: The halos were sized for a 4,000-pixel frame; Resampling stacks neighbouring halo pixels together; Pale rims round every branch, and flat areas that crawl; The noise was sharpened before anything reduced it. Resized to 800, then sharpened: Radius 0.5 to 1 pixel, tuned to the delivered size; Threshold 2 to 4, so a clear sky is left alone; Chroma noise reduced generously, luminance sparingly; High Pass on its own layer, so the skin can be masked out. Sharpening exaggerates edges that already exist, so everything done after it changes the scale it was tuned for. Resizing, noise reduction and compression all touch the pixels, which is why sharpening is last and at the size the picture will actually be seen.\n\nSame argument for noise reduction, which smooths pixels, and for JPEG compression, which quantises them. Sharpening amplifies whatever is there, including compression artefacts and noise. It goes at the end, after everything else, at the size the picture will actually be seen.\n\n## Three sharpenings, not one\n\nProfessionals sharpen up to three times for different reasons, and it helps to know which one you are doing:\n\n1. **Capture sharpening.** A small global correction for the softness introduced by the sensor's filter array and the demosaic process. Applied at full size, gently. Raw developers do this by default: darktable and RawTherapee both apply a modest amount unless you turn it off.\n2. **Creative sharpening.** Local and masked. Eyes, jewellery, the edge of a product, the texture of fabric. Never the whole frame.\n3. **Output sharpening.** At the final pixel dimensions, tuned to the medium. Print needs noticeably more than screen, because ink spreads slightly into paper: this is why an image that looks correctly sharpened on your monitor prints soft. Matte paper needs more than gloss.\n\n## Noise: two kinds, two different costs\n\nZoom into a photo shot at ISO 6400 and you see two things.\n\n**Chroma noise** is the coloured speckle: random red, green and blue blotches in what should be a smooth grey. Removing it is nearly free, because real colour information in a photograph changes slowly across the frame. You can push chroma noise reduction hard and lose almost nothing.\n\n**Luminance noise** is the grain-like variation in brightness. Removing it is expensive, because at the pixel level luminance noise and fine texture are the same thing: small local variations in brightness. The algorithm cannot tell the difference between grain in a sky and the weave in a shirt. Push the slider and skin turns to wax, grass turns to green paint, hair becomes a smooth mass. This is the plastic look, and it is far more damaging to a photograph than the noise was.\n\nSo: chroma reduction generously, luminance reduction sparingly, and mask it off the areas with real texture. Then sharpen. Denoise-then-sharpen is the fixed order; the reverse sharpens the noise and then smooths the detail.\n\nThe free route is strong here. darktable's profiled denoise uses measured noise profiles for specific camera and ISO combinations and outperforms most generic sliders. RawTherapee's is also good. Both cost nothing.\n\n## AI denoise, and the same warning as upscaling\n\nModel-based denoisers (Lightroom's Denoise, DxO DeepPRIME, several open-source options) do not smooth. They reconstruct. Trained on pairs of noisy and clean images, they predict what the clean version should look like, and on high-ISO photographs the results are startling: two or three stops of apparent improvement.\n\nThe warning is the same one from the first lesson. What comes back is plausible, not recovered. Fine repeated texture (fabric weave, distant foliage, small text on a sign) can come back as a convincing pattern that is not the pattern that was there. For a portrait or a wedding, nobody cares and the result is better. For anything documentary or evidential, you have altered the record.\n\nToday: take one photo, save two exports of it at 1000 pixels wide: one sharpened before the resize, one after. Look at them side by side at 100 per cent. You will not need to be told again."},"check":{"question":"A landscape is sharpened at full 4000-pixel width, then resized to 800 pixels wide for a website. The small version looks crunchy, with pale outlines around the branches. What is the mechanism?","options":[{"text":"The halos created by sharpening were sized for the large image, and the downsampling stacked them together, so sharpening should have been applied after the resize at the output size.","correct":true},{"text":"The resize used bicubic rather than Lanczos, and a better resampling filter would have avoided the outlines.","correct":false,"tempting":"Blames the resampling algorithm for artefacts that were written into the pixels before it ran."},{"text":"JPEG compression at the smaller size introduced the outlines, so exporting PNG would fix it.","correct":false,"tempting":"Attributes sharpening halos to compression blocking, which looks different and has a different cause."},{"text":"Noise reduction was applied before sharpening; reversing that order would have kept the edges clean.","correct":false,"tempting":"Inverts the correct order, and misidentifies a scale problem as an ordering problem between denoise and sharpen."}]},"prev":{"slug":"sizing-for-a-screen","title":"The width to export, and why the browser will not fix it for you","url":"https://addaly.com/learn/image-editing/sizing-for-a-screen"},"next":{"slug":"export-and-colour","title":"Why your image looks washed out on somebody else's phone","url":"https://addaly.com/learn/image-editing/export-and-colour"},"authorship":"Founder and series editor: Dr Mustafa Mun. Written by AI models to a written standard. Exam keys checked blind by a second model: 3,665 of 3,666 agreed. Academic review invited.","review":{"status":"Draft, open for academic review"},"cite":"Addaly (2026) 'Sharpening adds no detail, and it goes last', in Editing an Image. Edited by M. Mun. Addaly. Available at: https://addaly.com/learn/image-editing/sharpening-and-noise (Accessed: 31 October 2026).","licence":{"name":"CC BY-NC-SA 4.0","url":"https://creativecommons.org/licenses/by-nc-sa/4.0/"},"reader":{"schema":1,"title":"Sharpening adds no detail, and it goes last","courseTitle":"Editing an Image","module":{"index":8,"slug":"out-of-the-editor","title":"Out of the editor: size, sharpness and the delivered file"},"answer":null,"keyIdea":"Sharpening only exaggerates edges that already exist, so do it after resizing, at the output size, and after noise reduction rather than before.","hero":{"kind":"figure","figure":{"t":"figure","label":"Figure","figure":{"left":{"title":"Sharpened at 4,000, then resized","items":["The halos were sized for a 4,000-pixel frame","Resampling stacks neighbouring halo pixels together","Pale rims round every branch, and flat areas that crawl","The noise was sharpened before anything reduced it"]},"type":"compare","right":{"title":"Resized to 800, then sharpened","items":["Radius 0.5 to 1 pixel, tuned to the delivered size","Threshold 2 to 4, so a clear sky is left alone","Chroma noise reduced generously, luminance sparingly","High Pass on its own layer, so the skin can be masked out"]},"title":"The same photograph delivered at 800 pixels, two ways","caption":"Sharpening exaggerates edges that already exist, so everything done after it changes the scale it was tuned for. Resizing, noise reduction and compression all touch the pixels, which is why sharpening is last and at the size the picture will actually be seen."},"alt":"Figure: The same photograph delivered at 800 pixels, two ways. Sharpened at 4,000, then resized: The halos were sized for a 4,000-pixel frame; Resampling stacks neighbouring halo pixels together; Pale rims round every branch, and flat areas that crawl; The noise was sharpened before anything reduced it. Resized to 800, then sharpened: Radius 0.5 to 1 pixel, tuned to the delivered size; Threshold 2 to 4, so a clear sky is left alone; Chroma noise reduced generously, luminance sparingly; High Pass on its own layer, so the skin can be masked out. Sharpening exaggerates edges that already exist, so everything done after it changes the scale it was tuned for. Resizing, noise reduction and compression all touch the pixels, which is why sharpening is last and at the size the picture will actually be seen."}},"blocks":[{"t":"h2","id":"what-unsharp-mask-actually-does","runs":[{"text":"What Unsharp Mask actually does"}]},{"t":"p","runs":[{"text":"The name is backwards on purpose, and it tells you the mechanism. The software makes a blurred (unsharp) copy of your image, subtracts it from the original, and what remains is a map of the edges. It then adds that edge map back at increased strength."}]},{"t":"p","runs":[{"text":"The result is that along every edge, the light side is made lighter and the dark side darker. That local contrast increase is what your eye reads as sharpness. No detail has been added. Nothing that was not in the file is now in the file. You have exaggerated the transitions that were already there."}]},{"t":"p","runs":[{"text":"Three controls, and they mean exactly what the mechanism implies:"}]},{"t":"ul","items":[[{"text":"Amount","b":true},{"text":": how much of the edge map is added back. 60 to 120 per cent for most photographs."}],[{"text":"Radius","b":true},{"text":": how wide a blur was used, and therefore how thick the resulting bright and dark rims are. Small, 0.5 to 1 pixel, for fine detail. Large radius over about 3 is not sharpening any more, it is local contrast, and it makes skies look dramatic and skin look grubby."}],[{"text":"Threshold","b":true},{"text":": how different two neighbouring pixels must be before the tool treats them as an edge. Set it to 0 and you sharpen the noise in a clear sky along with everything else. 2 to 4 protects flat areas."}]]},{"t":"p","runs":[{"text":"Over-sharpen and you get "},{"text":"halos","b":true},{"text":": a pale outline around dark branches against a sky, and a dark outline around bright ones. Once you have seen one you will see them everywhere, including on television."}]},{"t":"p","runs":[{"text":"High Pass","b":true},{"text":" is the same idea with better control. Duplicate the layer, run "},{"text":"Filter > Other > High Pass","i":true},{"text":" at 1 to 2 pixels, set the layer blend mode to Overlay or Soft Light, and adjust the layer opacity to taste. Because it is a layer, it takes a mask, so you can sharpen the eyes and not the skin. GIMP, Krita and Photopea all have both Unsharp Mask and High Pass."}]},{"t":"h2","id":"why-it-goes-last","runs":[{"text":"Why it goes last"}]},{"t":"p","runs":[{"text":"Every operation you do after sharpening changes the pixel scale that the sharpening was tuned for."}]},{"t":"p","runs":[{"text":"Sharpen a 4000-pixel-wide image, then resize it to 800 for a website, and the halos you created were sized for 4000 pixels. Shrinking the image does not shrink them proportionally in a clean way: the resampling stacks neighbouring halo pixels together, and the small version comes out crunchy with pale rims around every branch. Do it the other way round: resize to 800, "},{"text":"then","i":true},{"text":" sharpen at 0.5 pixel radius, and the same photo looks clean."}]},{"t":"p","runs":[{"text":"Same argument for noise reduction, which smooths pixels, and for JPEG compression, which quantises them. Sharpening amplifies whatever is there, including compression artefacts and noise. It goes at the end, after everything else, at the size the picture will actually be seen."}]},{"t":"h2","id":"three-sharpenings-not-one","runs":[{"text":"Three sharpenings, not one"}]},{"t":"p","runs":[{"text":"Professionals sharpen up to three times for different reasons, and it helps to know which one you are doing:"}]},{"t":"ol","start":1,"items":[[{"text":"Capture sharpening.","b":true},{"text":" A small global correction for the softness introduced by the sensor’s filter array and the demosaic process. Applied at full size, gently. Raw developers do this by default: darktable and RawTherapee both apply a modest amount unless you turn it off."}],[{"text":"Creative sharpening.","b":true},{"text":" Local and masked. Eyes, jewellery, the edge of a product, the texture of fabric. Never the whole frame."}],[{"text":"Output sharpening.","b":true},{"text":" At the final pixel dimensions, tuned to the medium. Print needs noticeably more than screen, because ink spreads slightly into paper: this is why an image that looks correctly sharpened on your monitor prints soft. Matte paper needs more than gloss."}]]},{"t":"h2","id":"noise-two-kinds-two-different-costs","runs":[{"text":"Noise: two kinds, two different costs"}]},{"t":"p","runs":[{"text":"Zoom into a photo shot at ISO 6400 and you see two things."}]},{"t":"p","runs":[{"text":"Chroma noise","b":true},{"text":" is the coloured speckle: random red, green and blue blotches in what should be a smooth grey. Removing it is nearly free, because real colour information in a photograph changes slowly across the frame. You can push chroma noise reduction hard and lose almost nothing."}]},{"t":"p","runs":[{"text":"Luminance noise","b":true},{"text":" is the grain-like variation in brightness. Removing it is expensive, because at the pixel level luminance noise and fine texture are the same thing: small local variations in brightness. The algorithm cannot tell the difference between grain in a sky and the weave in a shirt. Push the slider and skin turns to wax, grass turns to green paint, hair becomes a smooth mass. This is the plastic look, and it is far more damaging to a photograph than the noise was."}]},{"t":"p","runs":[{"text":"So: chroma reduction generously, luminance reduction sparingly, and mask it off the areas with real texture. Then sharpen. Denoise-then-sharpen is the fixed order; the reverse sharpens the noise and then smooths the detail."}]},{"t":"p","runs":[{"text":"The free route is strong here. darktable’s profiled denoise uses measured noise profiles for specific camera and ISO combinations and outperforms most generic sliders. RawTherapee’s is also good. Both cost nothing."}]},{"t":"h2","id":"ai-denoise-and-the-same-warning-as-upscaling","runs":[{"text":"AI denoise, and the same warning as upscaling"}]},{"t":"p","runs":[{"text":"Model-based denoisers (Lightroom’s Denoise, DxO DeepPRIME, several open-source options) do not smooth. They reconstruct. Trained on pairs of noisy and clean images, they predict what the clean version should look like, and on high-ISO photographs the results are startling: two or three stops of apparent improvement."}]},{"t":"p","runs":[{"text":"The warning is the same one from the first lesson. What comes back is plausible, not recovered. Fine repeated texture (fabric weave, distant foliage, small text on a sign) can come back as a convincing pattern that is not the pattern that was there. For a portrait or a wedding, nobody cares and the result is better. For anything documentary or evidential, you have altered the record."}]},{"t":"p","runs":[{"text":"Today: take one photo, save two exports of it at 1000 pixels wide: one sharpened before the resize, one after. Look at them side by side at 100 per cent. You will not need to be told again."}]}],"check":{"question":[{"text":"A landscape is sharpened at full 4000-pixel width, then resized to 800 pixels wide for a website. The small version looks crunchy, with pale outlines around the branches. What is the mechanism?"}],"options":[{"text":[{"text":"The halos created by sharpening were sized for the large image, and the downsampling stacked them together, so sharpening should have been applied after the resize at the output size."}],"correct":true,"tempting":null},{"text":[{"text":"The resize used bicubic rather than Lanczos, and a better resampling filter would have avoided the outlines."}],"correct":false,"tempting":[{"text":"Blames the resampling algorithm for artefacts that were written into the pixels before it ran."}]},{"text":[{"text":"JPEG compression at the smaller size introduced the outlines, so exporting PNG would fix it."}],"correct":false,"tempting":[{"text":"Attributes sharpening halos to compression blocking, which looks different and has a different cause."}]},{"text":[{"text":"Noise reduction was applied before sharpening; reversing that order would have kept the edges clean."}],"correct":false,"tempting":[{"text":"Inverts the correct order, and misidentifies a scale problem as an ordering problem between denoise and sharpen."}]}],"order":[3,1,0,2],"why":null},"prev":{"slug":"sizing-for-a-screen","title":"The width to export, and why the browser will not fix it for you","minutes":9},"next":{"slug":"export-and-colour","title":"Why your image looks washed out on somebody else’s phone","minutes":10}}}