Clean a Noisy Product Reference Without Inventing Its Texture

By pitch.dogAdvanced15 min read20 September 2026

Advertising

Denoising does one job unusually well: it makes a photograph look as though it had been taken under better conditions than it was. That is precisely what makes it hazardous to a product reference, because the properties you are usually trying to communicate — a weave, a seam, an edge, a line of small print — live in the same part of the image as the noise you just removed.

So the useful question is not "is this version cleaner?" It is "does this version still carry the claims I need it to carry?" Those two questions can point in opposite directions, and the whole method below is about keeping them apart. Decide the claim first. Keep the original. Compare a restrained derivative against it at matching scale, in the regions that would change your answer. Then choose between cleaning, keeping the noise, and getting better information — and say which of those you did.

Write the caption before you touch the slider

The fastest way to find out what an image is supposed to demonstrate is to write the sentence it will appear under. A caption is a set of claims, and the claims have sizes.

"Textured navy swatch with a small woven label" asks for silhouette, colour, and an impression of texture. Every one of those is a low-frequency property. A fairly aggressive cleanup can serve them without much risk, because the information is large relative to the noise.

"The label is woven, not printed, and its edge is folded and stitched" is a different assignment. It asks about thread structure, about whether the edge is sharp or frayed, and about whether a line near the border is a seam. Those are small-scale claims, and they are exactly the ones a reconstruction step can damage or invent.

Keep going down: "the care line reads X," "the code is 30, not 38." Now you are asking an image to settle a question of fact about a physical object, at a scale where interference is likely. Some captures can answer that. Many cannot, and no amount of cleanup will change which kind you have.

Once you have the caption, mark the regions that would have to be good for it to be true. Choose them on the original, before you have processed anything — otherwise you will be drawn to whichever crops the derivative happens to flatter. Include at least one hostile region: the smallest type, the shadowed side, the edge where the material turns away from the light. Then write down, in the same note, the thing the caption needs, the source can't show. A capture that never resolved a fibre has not recorded it; a later treatment cannot recover what was never sampled.

Two useful distinctions fall out of this. "This helps a client see the overall look" is a legitimate use of a cleaned file. "This shows the actual fine texture" is a much stricter claim, and it needs its own evidence. Do not let the first slide into the second because the file looks convincing.

Keep the original, and change one thing

The comparison is only worth something if you can attribute the difference. So: one treatment variable per derivative, and a baseline that differs from it in nothing else.

Use the original raw file, not the in-camera JPEG. A JPEG has already been through the camera maker's own processing, including their noise handling, so it cannot serve as your untreated control — and if the capture exists only as JPEG, you have no baseline, which is itself an argument for recapturing rather than for cleaning. Render the raw with your ordinary settings for white balance, camera profile, and demosaicing, and turn the denoise control off for the baseline. Keep that baseline at native pixel dimensions. Then make the derivative from the same file with the same settings, plus denoising at a restrained amount — the least that removes the worst of the mottling, rather than the amount that produces the most satisfying screen.

No sharpening, no resizing, no new grade, no creative profile switch at the same time. Each of those is a second variable, and a second variable means you no longer know which change altered the letterforms you are staring at.

Record what you did while you are doing it, because defaults and paths move between releases. The help material for this kind of control is version-dependent — some published routes apply only to particular builds — so read the page that matches what you actually have installed, name the version in your note, and treat any path you have not walked yourself in that version as unverified. Your note needs the tool and version, the denoise amount, the base render settings, the output dimensions, the date, and the exact crop you are comparing. That note is not bureaucracy. It is the only thing standing between a defensible reference and a nice file with an unknown history.

It is worth knowing what the tool claims to be doing, as long as you keep the claim in its place. In an April 2023 post on Adobe's blog, a developer working on the feature describes denoising as a reconstruction step that runs alongside demosaicing, with stated goals of retaining fine texture while limiting artifacts (Denoise Demystified). That tells you the intent — and that intent is the reason a successful-looking result is not the same as a faithful one. Learned reconstruction implies inference. Inference is what you are checking for, not what vouches for the output. Vendor goals describe a design; they are not an accuracy test of your photograph.

Read the crops, not the picture

A full frame at the size the client will actually view it is a legitimate check — it is the only place you find out whether the cleanup mattered at all — but it will hide every failure mode you care about. Four signatures are worth looking for specifically, because they are the ways a reconstruction goes wrong in ways that still look plausible:

Letters that changed. Apertures closing, counters filling, strokes joining. A word that stays readable but is no longer the same word is the most dangerous outcome, because it looks like information.
Repetitions broken or invented. A knit, a weave, a row of perforations, a printed dot pattern. Check whether the period stays constant, and whether the direction of the repeat is still right.
Edges softened or hardened. A frayed boundary smoothed into a line, or a crisp boundary fuzzed out. This changes what a viewer can infer about how the material is finished.
Marks arriving or leaving. Small specks, faint registration marks, slight stains. Here you want to write down only what you can see, not what you conclude: "present in the source rendering, absent in the derivative" is an observation. "The mark is not on the material" is a claim you have not earned.

Judge at 1:1 for detail claims and at delivery size for overall claims, and alternate between the two images rather than studying them side by side for a long time. Your standard of "normal" drifts toward whichever version you look at longest, and after a few minutes the honest, grainy original will start to look like the damaged one.

And keep a third possibility open throughout: the noisy rendering is not ground truth either. It can obscure a thread that is there and it can produce a bright speck that suggests one that is not. Even with denoising off, a raw file has already been demosaiced — colour reconstructed from a pattern of single-channel samples — so the baseline is a consistent reference point, not a record of what was in front of the lens. If you need to get closer to reality, the fix is more sampling, not a different algorithm.

An invented comparison, worked through

Everything in this section is invented. No photograph was taken, no file was processed, and the numbers are stipulated so that the reasoning can be followed and checked — they are not measurements of anything. The product is a made-up one and the brand mark is a made-up mark.

Suppose a 24-megapixel capture, 6000 × 4000 pixels, of a 38 mm woven label sewn to a fine jersey swatch. Handheld indoors at ISO 6400; the label spans about 900 pixels across in the frame, or roughly 24 pixels per millimetre. A few consequences follow immediately.

The wordmark on the label has a cap height of about 2.6 mm, so it lands at roughly 62 pixels. The care block beneath it is much smaller: cap height around 1.2 mm, about 28 pixels, with strokes about 0.15 mm — call it 3.6 pixels. The jersey's columns repeat every 1.0 mm, about 24 pixels, and the individual yarn is around 0.25 mm, or 6 pixels. The label's cut edge frays by roughly 0.15 mm, which is that same 3.6 pixels. In this frame, luminance grain sits at a scale of a couple of pixels and the chroma mottling in patches of five to ten.

That arithmetic does not tell you what the algorithm did. It tells you where to look first, and why a whole-frame view can be reassuring. Against the luminance grain, the wordmark is roughly thirty times the noise scale, the columns about ten times, the yarn around three, and the care-block strokes essentially at it. Against the chroma mottling, whose patches run five to ten pixels, the same structures rank lower: the wordmark six to twelve times, the columns roughly three to five, the yarn at or below the patch scale, and the care block's strokes under it. Detail near the scale of the noise is where reconstruction has to guess, so that is where your attention belongs — and the yarn, comfortable against the grain, is not comfortable against the mottling. It also explains why the darker half of the frame is riskier than the lit half: shadow noise is stronger, so the effective ratio drops and previously safe structures move into the danger zone.

Now the stipulated comparison. At the size the client reviews — a long edge of, say, 1600 pixels — the derivative looks clean and the wordmark is legible, and if that were the only view, the job would appear done. At 1:1, four things differ.

The wordmark letters hold. Counters stay open; no letter gains or loses a stroke. The one change is a small ® mark: in the source rendering it is a faint blob of about six pixels, barely distinguishable from grain; in the derivative it is simply absent. The source was uncertain about it; the derivative is confident that it is not there. That is a change in kind, not just in degree.

The care block does not hold. Strokes merge, an aperture closes, and a two-digit code resolves into a shape the source never showed. The derivative's version is legible and plausible and, for the purpose of the caption, unusable.

The jersey columns survive in the lit central area and collapse near the left edge, where the light falls off. The result overstates the material's uniformity — it makes a claim ("this surface is even") that the capture does not support.

The label edge loses its fray, hardened into a thin clean line, and a faint periodic ripple appears a few pixels inside the boundary. At thumbnail size that ripple reads like a topstitched seam. Nothing in the source suggests one.

Set against that, a second capture was stipulated as available: tripod, ISO 200, a longer exposure, unchanged material and view, at roughly three times the original magnification — about 72 pixels per millimetre, which puts the care block's strokes near 11 pixels and the fray at a clearly visible width. In that capture, the letterforms resolve with the correct apertures, the yarn structure is unambiguous, and the edge is visibly frayed rather than cut clean. It also changes two things: the light is more directional, so the swatch reads warmer and lower-key than the flat-lit original, and the label sits at a slightly different angle. It is better evidence for construction and for the small type. It is not a replacement for the even-lit view.

No single image wins here, and that is the point. The derivative earns the broad look and the wordmark's letterforms — but not the ® mark, which it deleted rather than resolved; that mark has to come from the original rendering, where it is faint but present, or from the second capture, whose letterforms resolve at the larger scale. It does not earn the care block, the edge, the material's uniformity, or the seam-like ripple, which should simply be excluded from use. The second capture earns construction and the small type under its own conditions, and its conditions travel with it.

Keeping the noise, or getting better light

Before you push the denoise amount up to rescue a stubborn region, price the alternatives honestly.

Keeping the grain is often correct. If the caption asks for silhouette, colour, drape, or an impression of weight, and the capture's noise sits well below the scale of those properties, the derivative buys you nothing a smaller display size would not. That is worth saying plainly: the cheapest denoise is a smaller delivery size. If you never show the frame larger than a tile on a page, the mottling that bothers you at 1:1 may never exist for the audience you are writing for.

Recapturing is usually better information, under two conditions. The first is that you fix the variables that matter: same material, same view, same lighting intent, and a written note of whatever you deliberately changed. A better capture is evidence for the conditions it records and nothing else; a harder light makes texture legible but makes colour less comparable, a higher magnification makes small type readable but changes the perspective on the object's shape. The second is that you have something to compare it against. Recapture is not a way to avoid the baseline; it is a way to improve the source. Useful routes include more total light rather than more gain, a higher magnification specifically for the small type, a second frame with raking light for relief, and focus stacking when depth of field at close range is the limiting factor. Each one trades something.

If recapture is not available and the detail is still unresolved, narrow the claim. A reference that supports three claims is worth more than one that appears to support six. Write the delivery note so the reader knows which three. "Colour and overall texture only; care line not legible in source" is a complete and useful statement. Pushing the slider until the fabric looks decisive is not a fourth option, it is the failure mode.

What you ship, and what you say about it

The deliverable is not only an image. It is an image, a source, and a sentence about what that image can honestly show.

Ship the original alongside anything derived from it, with a link back to the source file and a short processing note on the derivative: tool and version, denoise amount, base render settings, no sharpening or resizing applied to the comparison copy, output dimensions, date. Where a reconstruction could reasonably be mistaken for observed material — a smooth surface standing in for a fine weave, a cleaned-up line standing in for a seam — say so near the image rather than in a footnote. A viewer who knows the surface has been reconstructed reads it differently from one who does not, and that difference is exactly the integrity of the reference.

Use each file only where the comparison supports it. If you need the broad look, the original or its restrained derivative will do. If you need construction, use the frame that resolves it, with its conditions attached. And do not repair unreadable wording into apparently photographic text: an invented letter that renders cleanly at 300 pixels is a fabricated product claim wearing a photograph's clothes.

That is the whole of it. The cleaned frame is optional. The honest sentence about what it can still be trusted to show is not.

Frequently asked questions

Why can a cleaner version still be less trustworthy for a product reference?

Denoising can remove or invent detail at the same scale as the properties a product reference often needs to communicate, such as a weave, seam, edge, or line of small print. The useful test is not whether the image looks cleaner, but whether it still carries the claims the caption needs. Decide the claim first, keep the original, and compare a restrained derivative against it in the regions that would change your answer.

What should the baseline be, and what should the derivative change?

Use the original raw file, not the in-camera JPEG, because a JPEG has already been processed by the camera and cannot serve as an untreated control. Render the raw with ordinary white balance, camera profile, and demosaicing settings, with denoise off, and keep it at native pixel dimensions. Make the derivative from the same file and settings with only restrained denoising added. Do not also sharpen, resize, regrade, or switch creative profiles, because that introduces a second variable.

Which signs should be checked at 1:1?

Look specifically for letters that changed, repetitions broken or invented, edges softened or hardened, and marks arriving or leaving. Judge detail claims at 1:1 and overall claims at delivery size, and alternate between the two images rather than studying them side by side for a long time. Also keep in mind that the noisy rendering is not ground truth either; even with denoising off, a raw file has already been demosaiced.

When is keeping the noise or recapturing better than pushing denoising further?

If the caption asks for silhouette, colour, drape, or an impression of weight and the noise sits well below those properties, keeping the grain or showing a smaller delivery size may cost nothing useful. Recapturing can be better information if the variables that matter are fixed and there is something to compare against. Useful routes include more total light rather than more gain, higher magnification for small type, raking light for relief, and focus stacking when depth of field limits close work. If recapture is unavailable, narrow the claim and state the limit in the delivery note.

What should ship with a cleaned derivative?

Ship the original alongside anything derived from it, with a link back to the source and a short processing note on the derivative covering tool and version, denoise amount, base render settings, output dimensions, date, and the exact crop being compared. Where a reconstruction could be mistaken for observed material, say so near the image rather than in a footnote. Do not repair unreadable wording into apparently photographic text, because an invented letter that renders cleanly can become a fabricated product claim.