Noise Deep Dive Part 3: The Denoising Mindset and RAW vs JPEG

This is Part 3 of the eight-part Noise Deep Dive supplement to Chapter 16: Noise Reduction. Part 1 (16D1_Physics_of_Digital_Noise.md) established where noise comes from at the sensor level. Part 2 (16D2_Old_Sensors_and_Visual_Diagnosis.md) taught you to read the noise before you touch a slider. This part is the hinge of the series: it fixes the attitude you bring to denoising, and it establishes where in the pipeline denoising should happen and why the file format you feed the denoiser determines the ceiling on the result.

Two topics, tightly linked:

The denoising mindset. You are not trying to make images clean. You are trying to make them convincing. That distinction changes every decision. It leads directly to content-dependent, local noise reduction — different strength for sky, skin, foliage, bark, and rock within a single frame.
RAW vs JPEG. The same denoising algorithm produces radically different results depending on whether it operates on linear raw mosaic data, on a 16-bit rendered TIFF, or on an 8-bit JPEG. Understanding why is the difference between fighting the file and working with it.

The masking mechanics that make local noise reduction possible live in Part 5 (16D5_Frequency_Masking_and_Sharpening.md). This part teaches the why and the decision framework. The specific Camera Raw and Photoshop tools are detailed in Part 4 (16D4_Camera_Raw_AI_Denoise_and_PS_Tools.md).

Learning Objectives

After completing this part, you will be able to:

State and defend the central denoising principle — that the goal is to remove disturbing noise while preserving real detail, not to produce a noise-free image — and explain why "clean" is the wrong target.
Distinguish between global noise reduction (one setting applied to the whole frame) and local (content-dependent, selective) noise reduction, and justify why appropriate noise reduction strength is a property of image content, not a single number.
Predict, for a given subject type (sky, fog, water, walls, skin, distant mountains, foliage, rocks, buildings, bark, fine textures), whether it tolerates aggressive or restrained noise reduction, and explain the perceptual and structural reasons for each.
Diagnose the "over-denoised" failure mode in smooth-but-textured subjects (foliage, bark, rock) and explain why destroying micro-texture destroys the character of the subject even when no obvious edge is lost.
Explain why noise reduction on RAW data is fundamentally superior to noise reduction on rendered pixels, citing linear-vs-gamma encoding, bit depth, mosaic/per-channel access, and the absence of baked-in demosaicing, sharpening, and noise reduction.
Rank RAW, 16-bit TIFF, and 8-bit JPEG as substrates for noise reduction and explain what each stage can and cannot recover, including why JPEG quantization and block artifacts compound noise into a harder problem.
Distinguish noise reduction before demosaicing (raw-domain / pre-demosaic) from noise reduction after demosaicing (post-demosaic), state what each can and cannot address, and locate each in the Adobe Camera Raw processing pipeline.
Choose the correct capture format and processing stage for a noise-critical shoot, and articulate the trade-off when only a JPEG or a delivered TIFF is available.
Recognize when a "noise problem" is actually a demosaicing artifact, a compression artifact, or an amplified-shadow problem, and route each to the correct treatment rather than reaching reflexively for a luminance slider.

Conceptual Foundation

Most photographers approach noise reduction with an unexamined assumption: that the objective is a clean image. They drag the Luminance slider until the grit disappears, congratulate themselves on the smoothness, and export. The result is technically less noisy and photographically worse — skies that look airbrushed, foliage that looks like a watercolor, rock that has turned to plastic, skin that belongs on a mannequin. The noise is gone. So is the photograph.

The error is in the goal, not the execution. Here is the principle that governs everything in this part, and it is worth stating in isolation because everything else follows from it:

You do not need to make the whole image noise-free. You remove the DISTURBING noise while PRESERVING real detail.

Read that twice. Two verbs, two objects. Remove the disturbing noise. Preserve the real detail. Neither verb says "everywhere," and neither says "completely." Noise reduction is not a cleaning operation applied to a whole surface. It is a targeted intervention against the specific noise that is bothering a specific viewer at a specific output size — nothing more.

This reframing has immediate, concrete consequences:

The target is perceptual, not numerical. The question is never "how much noise is left?" but "is the remaining noise disturbing at the intended output?" A print viewed at arm's length hides noise that screams at 100% on a monitor. The correct amount of noise reduction is the amount that makes the noise stop disturbing this output — and not one notch more, because every additional notch costs detail.
The target varies across the frame. Different content types have different noise tolerance and different detail value. A smooth sky shows every speckle and holds no detail worth protecting — denoise it hard. Weathered tree bark hides noise inside its own texture and is detail — barely touch it. The same slider value cannot be right for both. Therefore appropriate noise reduction is local, not global.
Detail is the thing you are protecting, not noise the thing you are chasing. Flip the frame of mind. Do not ask "where is the noise?" Ask "where is the detail I must not lose?" Then reduce noise everywhere except there, and reduce it only as far as the disturbance requires. This inversion is the mindset shift, and it changes which slider you reach for and how far you push it.

The second half of this part connects the mindset to the substrate. The most disciplined denoising decision-making in the world is capped by the data you feed it. Noise reduction on raw linear mosaic data has access to the original sensor signal at full bit depth, before any transformation has amplified, quantized, or blocked it. Noise reduction on an 8-bit JPEG is a salvage operation fighting the compression's own artifacts. Same operator, same intent, wildly different ceilings. The mindset tells you how hard and where to denoise; the format tells you how good the best possible result can be. You need both.

Terminology

Disturbing Noise

Definition: Disturbing noise is the subset of the total noise in an image that is actually perceptible and objectionable to a viewer at the intended output size and viewing distance. It is a perceptual category, not a physical measurement — the same noise floor can be disturbing in one output context (100% monitor view, large backlit print) and invisible in another (thumbnail, small matte print viewed at arm's length).

Meaning in photographic practice: Recognizing that only some noise is disturbing is the practical foundation of restraint. Every image contains noise; the question is whether this noise, at this size, in this content, bothers a viewer. Chrominance speckle in a smooth shadow is almost always disturbing. Fine luminance grain buried inside foliage texture is almost never disturbing. Treating all noise as equally in need of removal is the root cause of over-denoising. INTERPRETATION: A useful test question before every noise reduction decision — "Would a viewer, seeing this at the final size, be bothered by this noise, or only by the loss of detail if I remove it?" If the answer is the latter, do nothing.

Photoshop implementation: There is no "disturbing noise" control; the concept governs how far you push the controls in Chapter 16's Detail panel and the Reduce Noise filter. It also dictates evaluation discipline: judge noise at the actual output size (soft-proof, print test, or resized preview), not only at 100% zoom, because 100% shows noise the final output will never reveal.

Related concepts: Signal-to-noise ratio, output-size-dependent evaluation, viewing distance, local noise reduction, detail preservation.

Common misconception: "If I can see noise at 100%, I need to remove it." At 100% you are magnifying the image far beyond how it will ever be seen — often 3-6x the effective magnification of a large print viewed normally. Noise that is glaring at 100% may be entirely invisible in the delivered output. Evaluating and correcting at 100% systematically leads to over-denoising.

Local (Selective) Noise Reduction

Definition: Local noise reduction is the application of different noise reduction strength (or none) to different regions of a single image, chosen according to the noise level and detail content of each region. It is the opposite of global noise reduction, which applies one uniform setting to the entire frame.

Meaning in photographic practice: Almost every real photograph contains regions with radically different noise tolerance: a smooth sky next to detailed foliage, deep noisy shadows next to clean highlights, smooth skin next to sharp eyelashes. A single global setting must be a compromise that is too strong for the detailed regions and often still too weak for the worst shadows. Local noise reduction resolves the compromise by treating each region on its own terms — hard on the sky, gentle on the foliage, aggressive in the shadows, untouched in the highlights. INTERPRETATION: The recognition that "appropriate noise reduction is content-dependent, not a single global setting" is the single most important professional habit in this entire series. It is what separates a clean-looking, natural photograph from either a noisy one or a plastic one.

Photoshop implementation: Local noise reduction is achieved by applying noise reduction on a masked layer, or via luminosity/content masks, or via Camera Raw's local adjustment tools (masking / brushes / linear and radial gradients with per-mask noise sliders). The mechanics of building these masks — luminosity masks, blur-difference detail masks, frequency-based masks — are the subject of Part 5 (16D5_Frequency_Masking_and_Sharpening.md) and Chapter 13: Advanced Masking. This part establishes which regions need which treatment and why.

Related concepts: Global noise reduction, layer masks, luminosity masks (Chapter 13), content-dependent processing, detail preservation, per-content-type noise tolerance.

Common misconception: "Local noise reduction is only worth the effort on difficult images." The reverse is closer to the truth: nearly every image benefits, because nearly every image has both smooth regions (where noise shows and detail is scarce) and detailed regions (where noise hides and detail is precious). Global noise reduction is the special case that only works well when the frame is uniform in both noise and detail — which is rare.

Content-Dependent Noise Tolerance

Definition: Content-dependent noise tolerance is the principle that how much noise reduction a region can accept — and how much it needs — is determined by two properties of the depicted subject: how much the subject's own texture masks noise (visual masking), and how much real fine detail the subject contains that noise reduction would destroy.

Meaning in photographic practice: These two properties define a subject's position on a spectrum. A clear sky masks nothing (noise is fully exposed against smooth tone) and holds nothing (there is no fine detail to lose), so it tolerates and benefits from aggressive noise reduction. Tree bark masks heavily (its own chaotic micro-texture hides noise) and holds enormous detail (the texture is the subject), so it tolerates almost no noise reduction and needs almost none. Every subject sits somewhere between these poles. FACT: Human vision exhibits visual masking — the visibility of a small stimulus (noise) is reduced in the presence of a strong nearby stimulus of similar spatial frequency (busy texture). This is why identical noise is glaring in a smooth sky and invisible in foliage.

Photoshop implementation: There is no single control; the principle drives the entire local noise reduction strategy — which regions to mask for strong treatment, which to protect, and where to feather the transition. It is applied through the per-content-type decision table in the Theory section below.

Related concepts: Visual masking (perception), local noise reduction, spatial frequency, detail preservation, the noise reduction / detail trade-off (Chapter 16).

Common misconception: "Smooth-looking subjects can all be denoised hard." A subject can look smooth from a distance and still carry critical fine micro-texture — weathered rock, distant foliage, aged skin, tree bark. These "smooth-but-textured" subjects are the classic over-denoising trap: they invite aggressive treatment and punish it by collapsing into plastic featurelessness.

Micro-Texture (Subject Character)

Definition: Micro-texture is the fine, high-spatial-frequency surface structure that gives a subject its material identity — the grain of bark, the crystalline facets of rock, the roughness of stucco, the pores and fine lines of skin, the individual needles of a distant conifer. It occupies the same spatial frequency band as luminance noise, which is precisely why it is endangered by noise reduction.

Meaning in photographic practice: Micro-texture is often not "detail" in the sense of a resolvable edge you could point to, but it is the difference between a surface that reads as rock and one that reads as gray paint. Because it shares a frequency band with noise, and because it frequently has lower local contrast than a hard edge, edge-aware noise reduction algorithms fail to protect it — they see it as noise and smooth it away. INTERPRETATION: The loss of micro-texture is the most insidious noise reduction failure precisely because it is not obvious in isolation. No single feature is missing. The subject simply stops looking like itself. Photographers who evaluate only "is the noise gone?" never notice, because they are not asking "does this still look like bark?"

Photoshop implementation: Preserving micro-texture means keeping noise reduction off or minimal on the subjects that carry it, protecting them with masks (Part 5, Chapter 13), and — where some noise reduction is unavoidable — favoring detail-preserving controls (Luminance Detail high, moderate Luminance) or AI denoise, which is generally better at distinguishing texture from noise (see 16D4_Camera_Raw_AI_Denoise_and_PS_Tools.md).

Related concepts: Spatial frequency, visual masking, detail preservation, edge-aware filtering, over-denoising, the "watercolor"/"plastic" artifact.

Common misconception: "If no edges are lost, no detail is lost." Edges are only one kind of detail. Micro-texture is detail without discrete edges — a statistical roughness — and it can be entirely erased while every major edge in the frame survives intact. The image passes an edge-based inspection and still looks dead.

Linear Data (Raw / Scene-Referred)

Definition: Linear data is image data in which the recorded value is directly proportional to the number of photons collected — doubling the light doubles the value. Raw sensor data is (very nearly) linear before any tone curve or gamma encoding is applied. This is the state in which Camera Raw's raw-stage noise reduction operates.

Meaning in photographic practice: Linearity matters for noise reduction because no transformation has yet stretched the shadows apart. In linear data, the shadows occupy a tiny fraction of the numeric range and the noise within them has not been amplified by a tone curve. Noise reduction applied here operates on the noise in its original, un-stretched proportions — the most favorable possible condition. FACT: In a linear encoding, each photographic stop occupies a successively smaller share of the numeric range: the brightest stop uses half the values, the next a quarter, and so on, so the darkest stops are described by very few levels. Gamma/tone-curve encoding later redistributes these to match human perception, stretching the dark levels — and their noise — across a much wider output range.

Photoshop implementation: You never edit linear data by hand in Photoshop; Camera Raw processes it internally during raw development. The practical consequence is the recommendation, repeated throughout Chapter 16 and this series, to perform primary noise reduction in Camera Raw on the raw file rather than on rendered pixels.

Related concepts: Gamma encoding, tone curve, scene-referred vs output-referred, bit depth, shadow noise amplification, RAW development (Chapter 03).

Common misconception: "Raw data looks dark and flat, so it must have less information in the shadows." The flat, dark appearance of un-tone-mapped linear data is a display artifact — the data has not yet been mapped to perceptual brightness. The shadow information is present at full sensor bit depth; it simply has not been stretched into visibility yet. That un-stretched state is exactly what makes raw-stage noise reduction effective.

Demosaicing (and the Pre/Post Divide)

Definition: Demosaicing (also called debayering) is the interpolation process that converts the single-color-per-photosite mosaic captured by a color filter array (typically a Bayer array) into a full-color image with red, green, and blue values at every pixel. Noise reduction can be performed before demosaicing (on the raw mosaic, "pre-demosaic" or "raw-domain") or after demosaicing (on interpolated RGB pixels, "post-demosaic").

Meaning in photographic practice: The divide matters because the two stages address different noise. Pre-demosaic noise reduction operates on the actual per-photosite sensor readings before interpolation mixes them; it can suppress noise before the demosaicing algorithm amplifies or spreads it into color artifacts (maze patterns, colored speckle, zippering). Post-demosaic noise reduction cleans up what remains in the interpolated image but must work with noise that demosaicing has already correlated across neighboring pixels — noise that is now harder to distinguish from real detail. FACT: Demosaicing is an interpolation, and interpolating noisy mosaic data both spreads noise spatially and can convert single-channel noise into visible chrominance artifacts, because a noisy value in one color channel is blended into the RGB estimate of surrounding pixels.

Photoshop implementation: You do not choose the demosaicing stage manually in Camera Raw — Adobe's raw pipeline handles it internally, and ACR's Detail-panel and AI denoise operate within that pipeline where they have access to raw/near-raw data (see the pipeline discussion in Theory below). Photoshop's own filters (Reduce Noise, blur-on-channels) always operate post-demosaic on rendered pixels — they have no access to the mosaic at all.

Related concepts: Bayer array, color filter array, chrominance noise, RAW development (Chapter 03), linear data, AI denoise (16D4).

Common misconception: "Noise reduction always happens after the image is formed." For raw files, the most effective noise reduction happens within the raw conversion, entangled with demosaicing, before a conventional RGB image even exists. This is precisely why raw-stage noise reduction outperforms anything applied to a finished file.

8-bit JPEG as a Denoising Substrate

Definition: An 8-bit JPEG is a rendered, gamma-encoded, lossily-compressed image with 256 levels per channel, already subjected to in-camera (or in-software) demosaicing, sharpening, noise reduction, and tone mapping, and carrying JPEG's own quantization and 8x8 block compression artifacts. As a substrate for further noise reduction it is the worst common case.

Meaning in photographic practice: By the time you receive a JPEG, every advantage of raw-stage processing is gone and several new problems have been added. The tone curve has already amplified shadow noise; bit depth has been reduced to 256 levels, risking posterization when you push tones; and JPEG compression has introduced blocking and ringing that a noise reduction algorithm cannot distinguish from image content. You are not denoising a clean signal — you are denoising a signal plus baked-in noise reduction plus compression artifacts. INTERPRETATION: Noise reduction on JPEG is legitimate salvage work, but you should hold modest expectations and treat compression artifacts and quantization as part of the problem, not as separable from it.

Photoshop implementation: JPEGs are treated with the Reduce Noise filter, the Camera Raw Filter (not full raw development), or AI denoise where it accepts rendered files — always with the understanding that these operate on degraded, post-everything pixels. Aggressive tonal pushes on 8-bit JPEG risk posterization; convert to 16 bit before heavy editing to avoid adding banding, though this cannot restore the information the 8-bit quantization already discarded.

Related concepts: Quantization, posterization, JPEG block artifacts, gamma encoding, bit depth (Chapter 01), baked-in processing.

Common misconception: "Converting a JPEG to 16-bit TIFF before denoising gives me the benefits of a 16-bit file." Converting up to 16 bit prevents further degradation (banding from subsequent edits) but cannot recover the tonal information, the highlight/shadow headroom, or the pre-demosaic access that were lost when the JPEG was created. You cannot promote a JPEG back to raw quality; you can only stop making it worse.

Theory

Part A — The Denoising Mindset

"Clean" is the wrong target

The instinct to make an image clean comes from a category error: treating noise reduction as cleaning — the removal of dirt from an otherwise-perfect surface. Dirt is foreign to the surface; you can remove all of it and improve the surface. Noise is not foreign to the image. It is interleaved with the signal, occupying the same pixels and the same spatial frequencies as fine detail (Chapter 16 established this as the fundamental trade-off). You cannot remove all the noise without removing detail, because at the pixel level they are the same kind of thing — small, high-frequency variations. So "clean" is not merely hard to reach; it is the wrong place to aim, because the path toward it runs directly through your detail.

The correct target is convincing. A convincing photograph is one whose noise does not distract a viewer at the intended output and whose detail survives intact enough to read as real. Note what this target does not require: it does not require the absence of noise. A little residual grain in the right places is not a defect — it can be the very thing that keeps skin looking like skin and rock looking like rock. The mistake is not "leaving noise in." The mistake is "leaving disturbing noise in" or, far more commonly, "removing detail in pursuit of a cleanliness no viewer asked for."

You do not need to make the whole image noise-free. You remove the DISTURBING noise while PRESERVING real detail.

INTERPRETATION: The single most useful habit this principle produces is asking what you are protecting before deciding what to remove. Before touching a slider, identify the regions of real detail — the eyes, the foliage, the bark, the fabric weave, the frost crystals. Those are off-limits or nearly so. Then, and only then, look at the remaining smooth regions and ask whether their noise is actually disturbing at the output size. Usually only a fraction of the frame both contains disturbing noise and lacks detail worth protecting. That fraction is where noise reduction belongs. Everywhere else, restraint.

Why global noise reduction is structurally wrong

A single global noise reduction setting must satisfy contradictory demands simultaneously. Consider a common landscape: smooth sky, distant hazy hills, mid-ground foliage, foreground rock, and deep shadow under a tree. The sky wants strong smoothing (noise fully visible, no detail to lose). The foliage wants almost none (noise hidden, detail precious). The shadow wants very strong smoothing (severe noise) but has some texture to protect. No single number satisfies all of these. Whatever you choose:

Set it strong enough for the sky and shadow, and you have plasticized the foliage and rock.
Set it gentle enough for the foliage and rock, and the sky still shows speckle and the shadow is still a mess.
Split the difference, and you get the worst of both — a sky that is still slightly noisy and foliage that is slightly plastic.

FACT: This is not a tuning problem that a better global value would solve. It is structural. The optimal noise reduction strength is genuinely different for different content, so any single value is necessarily wrong somewhere. The only correct response is to make the strength local — to vary it across the frame according to content. This is why "appropriate noise reduction is content-dependent, not a single global setting" is stated as a principle rather than a preference.

The two axes: masking and detail value

Every subject's noise tolerance is set by two independent properties:

How much the subject masks noise (visual masking). Busy, high-contrast, high-frequency texture hides noise. Human vision literally cannot resolve fine noise fluctuations against a background already full of fine fluctuations. Smooth, uniform tone hides nothing — every speckle stands out against the flat field.
How much real fine detail the subject holds. Some subjects are structurally smooth and carry no fine detail (clear sky, out-of-focus background). Some are made of fine detail (foliage, bark, fabric, hair). Removing noise from the former costs nothing; removing noise from the latter costs the subject.

Cross these two axes and you get the tolerance of any subject:

Masks noise?	Holds detail?	Noise reduction strategy	Example subjects
No (smooth)	No (featureless)	Aggressive — nothing to lose, everything to gain	Clear sky, fog, out-of-focus background
No (smooth)	Yes (subtle detail)	Careful/moderate — noise shows but detail is fragile	Skin, calm water, gradients
Yes (busy)	No (chaotic but shallow)	Light — noise is already hidden	Coarse gravel at distance, heavy texture where detail is not the point
Yes (busy)	Yes (texture is the subject)	Minimal or none — treatment destroys character	Foliage, bark, rock, fine fabric

The bottom-right cell is the danger zone: subjects that both hide noise and consist of detail. Because they hide noise, they rarely need noise reduction. Because they consist of detail, they are catastrophically damaged by it. The only correct action for these subjects is usually to leave them alone.

The per-content-type decision framework

The following table is the practical core of the mindset. It ranks common subjects by how much noise reduction they tolerate and want, with the reasoning. Treat the strength labels as relative posture, not slider numbers — the actual numbers depend on the image, the ISO, the sensor, and the output, and are always a starting point, not a universal recipe.

Content type	Noise visibility	Real detail at risk	Recommended posture	Rationale
Clear sky	Very high — flat tone exposes every speckle	None	Aggressive luminance + color NR	Smooth gradient holds no detail; residual noise reads as a defect. The one place you can push hard without cost.
Fog / haze	Very high — low-contrast, smooth	Minimal (soft gradients only)	Aggressive, but watch for banding	Like sky, but low local contrast makes it prone to posterization/banding; denoise strongly yet check 16-bit smoothness.
Water (calm / reflective)	High in smooth areas	Subtle — reflection gradients, ripple micro-structure	Moderate; protect ripple detail	Broad smooth areas want smoothing, but ripples and reflection texture are real detail — do not flatten them.
Homogeneous walls (painted, plaster)	High — large uniform fields	Low (surface texture is often shallow)	Moderate to aggressive	Uniform fields show noise; most wall texture is not precious. Preserve just enough surface tooth to avoid a plastic look.
Skin	High in smooth cheeks/forehead	High — pores, fine lines are what make skin read as skin	Careful, moderate; favor detail preservation	Classic over-denoise trap. Some grain keeps skin alive; the "wax figure" effect is worse than the noise. Prefer detail-preserving or frequency-separated approaches.
Distant mountains	Moderate to high (atmospheric smoothing)	Moderate — ridgelines, faint texture	Moderate	Atmosphere already smooths them, so heavy NR looks natural, but faint ridge/vegetation texture gives scale and depth — keep some.
Foliage	Low — texture hides noise	Very high — leaves, needles ARE the subject	Minimal or none	Do not treat globally. Over-denoising turns foliage into green mush; the individual-leaf micro-texture is the entire character.
Rocks	Low — surface hides noise	High — crystalline/weathered micro-texture	Minimal	Rock reads as rock because of fine surface structure. Smoothing it produces "plastic rock" — instantly artificial.
Buildings	Mixed — smooth walls vs detailed edges	Mixed — flat facades tolerate NR, ornament/brick does not	Selective within the subject	Treat flat facades and glass moderately; protect brick courses, mortar lines, ornament, and edge crispness.
Tree bark	Very low — chaotic texture hides noise completely	Very high — bark texture is pure micro-detail	None or near-none	The purest example of the danger zone. Any meaningful NR visibly degrades bark into a smeared surface.
Fine textures (fabric weave, feathers, hair)	Low — self-masking	Very high — the weave/strand structure is the point	None or near-none; protect with masks	Same logic as bark: the fine structure and the noise share a frequency band; you cannot smooth one without the other.

RECOMMENDATION: Internalize the shape of this table rather than memorizing rows. It has one diagonal: as a subject moves from smooth-and-featureless (sky) toward busy-and-detailed (bark), the correct posture moves from aggressive toward hands-off. When you meet a subject not in the table, place it on that diagonal by asking the two questions — does it hide noise, and does it hold detail — and the posture follows.

Why over-denoising smooth-but-textured subjects destroys character

The subjects that get ruined most often are not the obviously detailed ones — photographers instinctively protect an eyelash. The victims are the smooth-but-textured subjects: foliage seen from a distance, weathered rock, bark, stucco, aged skin. From a normal viewing distance they read as relatively smooth, which invites aggressive treatment. But their smoothness is an illusion of scale — up close they are dense fields of low-contrast micro-texture. That micro-texture is exactly what noise reduction destroys, for three compounding reasons:

It shares a frequency band with noise. Micro-texture and luminance noise are both fine, high-frequency variation. An algorithm smoothing one smooths the other.
It often has lower local contrast than a hard edge. Edge-aware noise reduction protects high-contrast edges (it can detect them) but treats low-contrast micro-texture as noise, because statistically it looks like noise. So the very feature that defines these subjects is the one edge-aware filters fail to protect.
Its loss is not locally obvious. No single missing feature signals the damage. The surface just quietly stops looking like a material. INTERPRETATION: This is why over-denoising survives casual review — the reviewer checks "is it clean?" (yes) and "are the big edges sharp?" (yes) and never checks "does this still look like bark/rock/skin?" (no). The character has been removed without any alarm being raised.

The defense is the mindset itself: identify these subjects as detail-bearing before denoising, and either exclude them from noise reduction entirely or treat them with detail-preserving controls or AI denoise (which is meaningfully better at telling texture from noise — see 16D4). The how of excluding them — the masks — is Part 5. The why is this: their texture is not decoration on the subject. It is the subject.

Part B — RAW vs JPEG for Noise Reduction

The mindset tells you where and how hard to denoise. The substrate determines how good your best result can be. The same algorithm, the same intent, and the same care produce dramatically different ceilings depending on whether it operates on raw data, a 16-bit TIFF, or an 8-bit JPEG. This section explains why, from most to least favorable.

Why raw-stage noise reduction is superior — four reasons

1. Linear data: the tone curve has not yet amplified shadow noise. Raw data is (nearly) linear — value is proportional to photons. In this state the shadows occupy a tiny slice of the numeric range and their noise is in its original, un-stretched proportion. Rendering the image applies a tone curve / gamma encoding that stretches the dark values apart to match human perception — and stretches the shadow noise apart with them. FACT: Noise reduction applied before this stretch (raw stage) works on the noise in its compact, original form; noise reduction applied after (on a rendered TIFF or JPEG) must fight noise that the tone curve has already magnified. Chapter 16 quantifies this shadow-stretching effect; the practical upshot is that the shadows — where noise is worst — are exactly where raw-stage processing has its biggest advantage.

2. Full bit depth. Raw files carry the sensor's full precision — commonly 12 or 14 bits per channel, i.e. 4096 or 16384 levels. A noise reduction algorithm working at this precision can make fine distinctions between signal and noise and can smooth without introducing banding, because it has abundant levels to describe subtle gradients. A rendered 8-bit file has only 256 levels; smoothing near-uniform regions there risks creating posterization even as it removes noise.

3. Access to per-channel / mosaic data. On raw data, noise reduction (and demosaicing) can exploit the actual per-photosite readings and the known geometry of the color filter array. It can treat the noisier channels appropriately (the blue channel typically has the worst SNR — see Chapter 16 and Part 1) and suppress noise before interpolation spreads it into color artifacts. Once the image is a rendered RGB file, the mosaic is gone forever; every pixel is an interpolated blend, and per-channel structure has been mixed together.

4. No baked-in demosaicing, sharpening, or noise reduction to fight. A raw file is unrendered — no sharpening halos, no prior noise reduction smearing, no demosaicing artifacts frozen into the pixels. You get to make all of those decisions yourself, in the right order (noise reduction before sharpening; see Chapter 15 and Part 5). A JPEG has already had demosaicing, capture sharpening, and noise reduction applied by the camera — often crudely — and those decisions are now permanent. Any noise reduction you apply is a second pass on top of the camera's first, compounding detail loss (Chapter 16 warns against stacking noise reduction passes; the JPEG is the first pass).

The three substrates ranked

Substrate	Encoding	Bit depth	Mosaic access	Baked-in processing	Denoising ceiling
RAW (ORF, PEF, DNG, NEF, CR2/CR3, ARW, RAF, etc.)	Linear (scene-referred)	Full sensor precision (commonly 12-14 bit)	Yes — pre-demosaic	None	Highest. Full latitude, best shadow results, mosaic-aware.
16-bit TIFF	Gamma/output-referred	16 bit (no quantization/banding risk from editing)	No — post-demosaic	Demosaicing done; NR/sharpening depend on how it was rendered	Good intermediate. No mosaic or linear advantage, but no quantization or compression artifacts to fight.
8-bit JPEG	Gamma/output-referred	8 bit (256 levels)	No — post-demosaic	Demosaicing + camera NR + sharpening + tone curve all baked in	Lowest. Quantization + block artifacts compound the noise problem.

RAW formats. ORF (Olympus), PEF (Pentax), DNG (Adobe's open raw container), NEF (Nikon), CR2/CR3 (Canon), ARW (Sony), RAF (Fujifilm), and others differ in container and metadata but share the essential property: they carry unrendered, near-linear, full-bit-depth sensor data, generally still in mosaic form. UNCERTAINTY: Specific formats vary in details such as compression scheme (lossless vs lossy vs "compressed raw" options), whether they store the true mosaic or a partially processed variant, and bit depth per model. Some cameras offer reduced-bit-depth or lossy-compressed raw modes that erode the raw advantage. Verify the characteristics of a specific camera's raw format against its documentation rather than assuming; the general ranking above holds, but the size of the raw advantage depends on these specifics. Fujifilm's X-Trans arrays use a non-Bayer color filter pattern, which changes the demosaicing details but not the fundamental raw-stage advantages.

16-bit TIFF as the good intermediate. A 16-bit TIFF is what you get when you render a raw file out to a pixel format without throwing away precision. It has lost the linear encoding, the mosaic, and the pre-demosaic opportunity — so it is not equal to working on raw — but it retains enough precision that editing (including noise reduction and tonal pushes) does not introduce banding or quantization. RECOMMENDATION: When a raw file is unavailable but a 16-bit TIFF is (for example, a file delivered by a raw processor or a scan), it is a perfectly serviceable substrate for noise reduction — far better than a JPEG. When you must render before further work, render to 16-bit, never 8-bit, to preserve editing latitude. Note that a 16-bit TIFF's quality depends on how it was rendered — a TIFF exported with heavy in-processor noise reduction and sharpening already baked in carries those decisions, so a TIFF is only as clean a starting point as the render that produced it.

8-bit JPEG as the worst case. Everything works against you. The tone curve has amplified shadow noise. Only 256 levels remain, so pushing tones risks posterization and heavy noise reduction on near-uniform areas can create banding. And JPEG's lossy compression has added its own artifacts:

Quantization discards high-frequency information in 8x8 blocks, coarsening exactly the fine detail you are trying to protect and, in flat areas, throwing away the subtle gradations that would have hidden banding.
Block artifacts (blocking) create faint 8x8 grid discontinuities, most visible in smooth regions like sky — the same regions where you want smoothness. A noise reduction algorithm cannot tell a block boundary from image content.
Ringing / mosquito noise appears near high-contrast edges as faint oscillations, which noise reduction may smear rather than remove.

FACT: These compression artifacts are not noise in the sensor sense — they are deterministic, structured errors introduced by the codec — but they present to a noise reduction algorithm as content it must either preserve (wrongly) or smooth (also wrongly, taking real detail with them). This is why JPEG noise reduction compounds: you are simultaneously fighting shot/read noise, the camera's baked-in noise reduction, and the codec's artifacts, all entangled in 8-bit precision. INTERPRETATION: Denoising a JPEG is legitimate and sometimes the only option, but it is salvage. Set expectations accordingly, address obvious block/ringing artifacts as part of the problem, convert to 16-bit before any heavy tonal work (to avoid adding banding), and do not expect raw-quality results — the information required for them was discarded when the JPEG was written.

Noise reduction before vs after demosaicing

This is the most technically important distinction in the RAW-vs-rendered story, and it is usually invisible to the photographer because it happens inside the raw converter. It deserves explicit treatment.

Pre-demosaic (raw-domain) noise reduction operates on the mosaic — the array of single-color-per-photosite readings — before interpolation has built a full-color image. What it can do:

Suppress noise in the original per-photosite signal before demosaicing spreads it. Because demosaicing interpolates each pixel from its neighbors, noise present at that stage gets smeared across neighbors and, worse, converted into chrominance artifacts (a noisy single-channel value bleeds into the color estimate of surrounding pixels). Cleaning the mosaic first prevents this.
Reduce the color-speckle and maze/labyrinth artifacts that noisy demosaicing produces, because the demosaicing algorithm is now working on cleaner input.
Exploit the known color-filter-array geometry to treat channels according to their true noise characteristics.

What it cannot do: it cannot fix artifacts that only exist after interpolation, and it works best entangled with the demosaicing step itself rather than as a fully separate pass.

Post-demosaic noise reduction operates on the interpolated RGB image. What it can do:

Clean up residual luminance and chrominance noise that survives into the rendered image.
Apply the perceptually-oriented luminance/chrominance split that Chapter 16 describes (the Detail panel's separate Luminance and Color controls act on the near-rendered image).
Be applied selectively/locally with masks — the local noise reduction this part advocates is a post-demosaic (or post-render) operation.

What it cannot do: it cannot undo the correlation that demosaicing introduced. After interpolation, noise has been spread across neighboring pixels and mixed with real detail, so distinguishing the two is harder than it was on the mosaic. It also cannot recover from color artifacts that demosaicing already baked into the pixels; it can only try to smooth over them.

FACT: The most effective raw converters do the heavy lifting around the demosaicing step, where they still have mosaic and near-linear data, and then offer post-demosaic refinement. This is a large part of why raw-stage noise reduction beats anything applied to a finished file: the finished file has already passed the demosaicing stage, and the pre-demosaic opportunity is gone permanently. Photoshop's own filters (Reduce Noise, per-channel Gaussian blur — see Chapter 16) are always post-demosaic and post-render; they never see the mosaic.

Where each sits in the Adobe Camera Raw pipeline

Camera Raw processes a raw file through an ordered internal pipeline. You do not control the ordering directly, but understanding it explains why raw-stage noise reduction is effective and why its results interact with sharpening the way they do.

Pipeline stage (raw file)	What happens	Relation to noise reduction
Raw decode	Read the mosaic sensor data, near-linear	The state where mosaic/linear advantages exist
Raw-domain / demosaic-adjacent processing	Demosaicing plus the noise reduction and detail processing entangled with it	Where the Detail panel's noise reduction and (especially) AI Denoise do their most effective work — with access to near-raw data
Color / tone rendering	White balance, tone curve, color space, gamma encoding	Applies the shadow-stretching tone curve; noise reduction before this is the advantage
Local adjustments	Masking, brushes, gradients (with per-mask noise sliders)	Where local/selective noise reduction is applied within ACR
Output	Deliver to Photoshop (TIFF/PSD) or export (JPEG/TIFF)	After this, only post-render noise reduction is possible

VERSION NOTE: The exact internal ordering, the entanglement of AI Denoise with demosaicing, and the availability of per-mask noise controls depend on the Camera Raw version. AI Denoise in particular is implemented as a raw-stage, near-demosaic operation that outputs an enhanced DNG (see Chapter 16 and Part 4). Verify specifics against your version; the principle — that ACR's primary noise reduction operates early, on near-raw data, before the tone curve stretches the shadows — is stable across versions. UNCERTAINTY: Adobe does not fully publish the internal stage ordering of the raw pipeline; the table above reflects the well-established conceptual model and the observable behavior of the controls, not a documented internal specification.

The key inference from the pipeline: when you use Filter > Camera Raw Filter on an already-rendered layer in Photoshop, you are not getting raw-stage noise reduction. The Camera Raw Filter operates on the rendered pixels of that layer — post-demosaic, post-tone-curve, at whatever bit depth the document is. It offers the same controls as raw development but not the same data. This is a frequent and costly misunderstanding: photographers assume "Camera Raw noise reduction" is equally powerful wherever the Camera Raw interface appears. It is not — its power comes from the raw data it normally operates on, not from the sliders.

Photoshop Implementation

This part is deliberately concept-first; the tool-by-tool procedures live in Part 4 (16D4_Camera_Raw_AI_Denoise_and_PS_Tools.md) and the masking mechanics in Part 5 (16D5_Frequency_Masking_and_Sharpening.md). What follows is how to translate the mindset and the substrate knowledge into concrete Photoshop/ACR decisions.

Translating the mindset into a working procedure

Diagnose before dragging. With the file open (raw in Camera Raw, ideally), zoom to 100% and survey the frame region by region — but do it to build a map of content, not to start correcting. Note where the detail lives (foliage, bark, eyes, texture) and where the smooth, noise-exposed, detail-free regions are (sky, walls, out-of-focus areas). This map, not a slider, is your plan.
Set a global baseline for the universally-objectionable component only. Chrominance noise has no aesthetic value anywhere (Chapter 16), so a global Color noise reduction pass is safe and appropriate as a baseline. Luminance noise reduction is where content-dependence bites, so keep the global luminance value conservative — enough to tame the worst general grit, not enough to plasticize your detailed regions.
Add local strength where the map says smooth-and-noisy. For the sky, fog, plain walls, and deep shadows, add stronger noise reduction locally — in Camera Raw via masks/gradients with per-mask noise sliders, or in Photoshop on a masked, more-aggressively-denoised layer. This is where the sky finally goes clean without dragging the foliage down with it.
Protect the detail regions explicitly. Ensure foliage, bark, rock, fine textures, and key subject detail receive no more than the conservative global baseline — mask them out of any strong treatment entirely. This is the step most photographers skip, and it is the one that preserves character.
Evaluate at output size, not only at 100%. Before committing, judge the result at the intended output — soft-proof/resize for print, view at delivery pixel dimensions for web. Noise that is disturbing at 100% may be invisible at output; detail loss that is invisible at 100% (micro-texture) may still cost you at output. The mindset's target is perceptual, so evaluate perceptually.

Choosing the substrate deliberately

If you have the raw file, use it. Perform primary noise reduction in Camera Raw on the raw file — this is the whole point of Part B. Do not render to TIFF and then denoise unless you have a specific reason; you would be discarding the raw advantage.
If you only have a 16-bit TIFF, treat it as a solid intermediate. Denoise with the Camera Raw Filter (accepting it is post-demosaic) or the Reduce Noise filter, and remember its quality is capped by how it was rendered.
If you only have an 8-bit JPEG, convert to 16-bit before heavy tonal or noise work (Image > Mode > 16 Bits/Channel) to avoid adding banding — while understanding this recovers nothing already lost. Address JPEG block/ringing artifacts as part of the noise problem. Keep expectations modest; this is salvage.

The Camera Raw Filter trap

FACT: Filter > Camera Raw Filter applied to a rendered layer operates on rendered pixels, not raw data. It is a legitimate tool for local noise reduction on a masked layer (it gives you the familiar controls plus AI options), but it does not confer the raw-stage advantages of linear data, full bit depth, or pre-demosaic access — those are gone once the image is rendered. Use it knowingly, as a post-render tool, not as a substitute for having processed the raw file correctly in the first place.

Professional Workflow

Capture-stage decisions (the cheapest noise reduction)

The most consequential noise reduction decision is made before post-processing: shoot raw whenever noise is a foreseeable concern. RECOMMENDATION: For any low-light, high-ISO, high-dynamic-range, or shadow-recovery-likely situation, shoot raw — the raw-stage advantages in Part B are only available if you captured raw. Reserve JPEG-only capture for situations where noise is not a concern and workflow speed dominates (bright light, low ISO, high-volume delivery of small outputs). If a camera offers a reduced-bit-depth or lossy "compressed raw" mode, understand that it erodes the raw advantage; use full/lossless raw when shadow latitude matters. UNCERTAINTY: The magnitude of quality loss from compressed/reduced raw modes is model-specific; verify against the camera's documentation and, ideally, your own tests.

The mindset-driven local noise reduction workflow

Develop the raw file in Camera Raw. Set a conservative global luminance value and an adequate global color value (chrominance NR is safe globally).
Build your content map: identify detail regions (protect) and smooth-noisy regions (treat).
Apply local stronger noise reduction to the smooth-noisy regions — ACR local masks/gradients, or a masked layer in Photoshop after opening as a Smart Object. Deep shadows and sky are the usual targets.
Confirm the detail regions received only the conservative baseline (mask them out of strong treatment).
Sharpen after noise reduction (Chapter 15, Chapter 16, Part 5), and consider local sharpening that targets the same detail regions you protected from noise reduction — the logic is symmetric.
Evaluate at output size. Adjust. Add grain deliberately if a smooth region looks too clean relative to the rest of the frame (Chapter 16 on adding grain as a finishing step).

The masking that makes steps 3-5 possible is Part 5's subject; here the workflow point is the sequence of decisions — map, baseline, local strength, protect, evaluate — which is identical regardless of the specific masking technique used.

Working with delivered files you did not capture

When a client or archive delivers rendered files, your latitude is set by what you receive. Establish the substrate first: is it a raw, a 16-bit TIFF, or an 8-bit JPEG? RECOMMENDATION: Ask for the raw file if it exists and noise is a concern — the quality difference is not marginal. If only a JPEG is available, set expectations explicitly (with the client, and with yourself) that noise reduction will be salvage-grade, and price/plan the work accordingly. Never render a client's raw to 8-bit JPEG as an intermediate in your own workflow — you would be discarding latitude you will want later.

Common Mistakes

Aiming for "clean" instead of "convincing." The single most common and most damaging error. Dragging noise reduction until the grit is gone produces plastic skies, mushy foliage, and waxy skin. The target is a photograph whose noise does not disturb at output and whose detail survives — not a noise-free image. Some residual noise in the right places is correct, not a failure.

Applying one global setting to a mixed-content frame. A single value cannot be right for both a smooth sky and detailed bark; it will over-treat the detail or under-treat the smooth areas, or split the difference and do both badly. Vary the strength by content.

Over-denoising smooth-but-textured subjects (foliage, bark, rock, aged skin). These read as smooth from a distance and invite aggressive treatment, but their micro-texture is their character and shares a frequency band with noise. Smoothing them produces subjects that no longer look like their material — and the damage is invisible to an edge-based inspection, so it survives casual review.

Evaluating and correcting only at 100% zoom. 100% magnifies far beyond any real viewing condition, showing noise the output will never reveal and hiding micro-texture loss that the output will. Evaluate at the intended output size before committing.

Denoising a rendered file when the raw was available. Rendering to TIFF/JPEG and then denoising discards the linear-data, full-bit-depth, and pre-demosaic advantages. If the raw exists, denoise the raw.

Assuming Filter > Camera Raw Filter gives raw-stage noise reduction. The Camera Raw Filter on a rendered layer operates on rendered pixels — post-demosaic, post-tone-curve, at the document's bit depth. Same controls, not the same data, not the same result.

Pushing tones hard on an 8-bit JPEG without converting to 16-bit. Heavy tonal or noise work on 256 levels produces posterization and banding. Convert to 16-bit first to avoid adding damage — while remembering this recovers nothing the JPEG already discarded.

Treating JPEG compression artifacts as ordinary noise. Blocking and ringing are structured codec errors, not sensor noise. A luminance slider smears them rather than removing them and takes real detail along. Recognize them as a separate, compounding problem inherent to the substrate.

Stacking noise reduction passes on a JPEG. The JPEG already carries the camera's baked-in noise reduction. Your pass is the second, and each pass costs detail. One deliberate, correctly-scoped pass beats several reflexive ones (Chapter 16).

Removing noise everywhere and then wondering why the image looks dead. If you denoise the detailed regions along with the smooth ones, you strip the frame of the micro-texture that made it read as real. Protect detail first; treat noise second.

Summary

The mindset comes first, because it governs every subsequent decision. You do not need to make the whole image noise-free. You remove the DISTURBING noise while preserving real detail. "Clean" is the wrong target — it aims the operation straight through your detail, since noise and fine detail occupy the same pixels and frequencies. "Convincing" is the right target: noise that does not disturb at the intended output, and detail that survives enough to read as real. Only a fraction of any frame both contains disturbing noise and lacks detail worth protecting, and that fraction is where noise reduction belongs.

Because the right strength differs by content, appropriate noise reduction is local, not global. A subject's noise tolerance is set by two axes: how much its texture masks noise, and how much real detail it holds. Smooth-and-featureless subjects (sky, fog) tolerate aggressive treatment; busy-and-detailed subjects (foliage, bark, rock, fine textures) tolerate almost none — and the smooth-but-textured subjects are the classic over-denoising trap, because their character-defining micro-texture reads as noise to edge-aware algorithms and is destroyed without any obvious feature going missing. The per-content-type table encodes this as a single diagonal: from aggressive (sky) to hands-off (bark).

The substrate sets the ceiling. Raw-stage noise reduction is superior for four reasons: it works on linear data before the tone curve amplifies shadow noise; it has full sensor bit depth; it has per-channel/mosaic access; and it faces no baked-in demosaicing, sharpening, or noise reduction to fight. A 16-bit TIFF is a good intermediate — no linear or mosaic advantage, but no quantization or compression artifacts either — while an 8-bit JPEG is the worst case, where amplified shadow noise, 256-level quantization, and lossy block/ringing artifacts compound the problem into salvage work. Noise reduction before demosaicing (raw-domain) can suppress noise before interpolation spreads it into color artifacts; after demosaicing, noise has been correlated with detail and is harder to separate. Camera Raw does its most effective noise reduction early in the pipeline, on near-raw data, before the tone curve stretches the shadows — which is precisely why the Camera Raw Filter on a rendered layer, offering the same controls on already-rendered pixels, cannot match it.

Put together: the mindset tells you where and how hard to denoise; the substrate tells you how good the best possible result can be. Shoot raw when noise is foreseeable, denoise the raw locally with restraint, protect your detail explicitly, and judge at output size. Part 4 gives you the tools; Part 5 gives you the masks; this part gave you the judgment that directs both.

Exercises

Exercise 1: Building a Content Map

Open a landscape or urban image containing at least four distinct content types (for example: sky, foliage, a smooth wall or water surface, and a detailed textured subject such as rock or bark), shot at a moderate-to-high ISO.

Zoom to 100% and survey the frame without touching any slider. On a printout or a notes layer, mark each major region.
For each region, answer the two axis questions: (a) Does the subject's texture mask noise, or is the noise fully exposed? (b) Does the region hold real fine detail that noise reduction would destroy?
From your answers, assign each region a posture: aggressive, moderate, light, or none.
Compare your assignments to the per-content-type table in the Theory section. Where do they differ, and can you justify the difference from the specifics of this image?

Success criterion: You can produce a content map that assigns a defensible noise reduction posture to each region before making any correction, and you can state the reasoning in terms of masking and detail value.

Exercise 2: The Global Compromise Fails

Using the same image:

Apply a global luminance noise reduction strong enough to fully clean the smooth region (sky or wall). Note the value.
At 100%, examine your most detailed region (foliage, bark, rock). Describe what has happened to its micro-texture.
Now reduce the global luminance value until the detailed region looks natural again. Examine the smooth region. Is it still noisy?
Confirm that no single global value cleans the smooth region and preserves the detailed region.

Success criterion: You have demonstrated to yourself that the global setting is structurally, not just practically, unable to satisfy both regions — the motivation for local noise reduction.

Exercise 3: Evaluate at Output, Not at 100%

Take a moderately noisy image and a noise reduction setting you consider correct at 100% zoom.

Note the noise level and detail at 100%.
Resize a copy to the intended web delivery size (or soft-proof/print-preview at the intended print size and viewing distance).
Compare. Is the noise that bothered you at 100% still disturbing at output? Is any micro-texture loss now visible that was not at 100%?
Re-set your noise reduction based on the output evaluation. Did your correct setting change?

Success criterion: You can articulate the difference between 100%-view noise and output-disturbing noise, and you adjust noise reduction to the latter.

Exercise 4: RAW vs Rendered, Same Algorithm

Take a single high-ISO raw file.

Version A: In Camera Raw, apply luminance and color noise reduction on the raw file. Open into Photoshop.
Version B: In Camera Raw, set all noise reduction to 0. Open into Photoshop as a rendered 16-bit file. Now apply Filter > Camera Raw Filter with the same luminance and color values as Version A.
Compare at 100%, paying attention to the deep shadows specifically.

Success criterion: You can observe that the same control values produce a cleaner shadow result on the raw file than on the rendered file, and you can explain the difference in terms of linear data and the tone curve.

Advanced Exercises

Advanced Exercise 1: Local Noise Reduction Across Content Types

Using the content map from Exercise 1 and a raw file:

Set a conservative global luminance value and an adequate global color value in Camera Raw.
Using Camera Raw local masks (or, after opening into Photoshop as a Smart Object, masked layers — masking mechanics per Chapter 13 and Part 5), apply stronger noise reduction to the smooth-noisy regions only (sky, wall, deep shadow).
Explicitly confirm that the detailed regions (foliage, bark, rock) receive only the conservative baseline — mask them out of the strong treatment.
Evaluate at output size. If a smooth region now looks too clean relative to the textured regions, add a small amount of grain to it.

Success criterion: You produce an image where the sky and shadows are clean, the foliage/bark/rock retain full character, and no single region betrays either noise or plasticity at output size.

Advanced Exercise 2: JPEG Salvage

Take an 8-bit JPEG shot at high ISO (or export one from a high-ISO raw at moderate JPEG quality to guarantee visible compression artifacts).

At 100%, identify three distinct problems: sensor noise, the camera's baked-in noise reduction (look for smeared low-contrast areas), and JPEG artifacts (8x8 blocking in the sky, ringing near high-contrast edges).
Convert to 16-bit (Image > Mode > 16 Bits/Channel).
Apply noise reduction, and separately assess how it handles each of the three problems — which does it improve, which does it smear, which does it ignore?
Compare your best JPEG result to a raw-processed version of the same scene if you have one.

Success criterion: You can distinguish sensor noise from baked-in noise reduction from compression artifacts on a JPEG, and you can articulate why the JPEG's denoising ceiling is lower than the raw's, referencing quantization and block artifacts.

Advanced Exercise 3: The Smooth-but-Textured Trap

Find or shoot an image with a prominent smooth-but-textured subject filling much of the frame — weathered rock, tree bark, aged skin, or a stucco wall — at an ISO high enough to show noise.

Apply aggressive global luminance noise reduction. At 100%, confirm the noise is gone.
Now ask the character question: does the subject still look like its material (rock/bark/skin), or does it look like plastic/paint?
Undo. Apply minimal luminance noise reduction (or none), with detail preservation favored, and add color noise reduction as needed.
Compare the two versions at output size. Which reads as the real material? Which is "cleaner" but dead?

Success criterion: You can demonstrate and explain why over-denoising destroys subject character even when no discrete edge is lost, and you can produce a version that keeps the character by protecting micro-texture.

Blackbelt Challenge

You are handed three files of the same scene — a forest edge at dusk, ISO 6400 — and must design a complete noise reduction strategy for each, then reconcile the differences. The scene contains: a smooth twilight sky (upper third), distant hazy hills (behind the trees), a mid-ground of dense foliage and several tree trunks with detailed bark, foreground rock with weathered micro-texture, and deep noisy shadow beneath the trees. A person stands at the forest edge; their face is in the moderately-lit zone. The intended outputs are a 24x36 inch gallery print on baryta paper and a 1600 px social post.

File 1 is the camera raw (assume a full-bit-depth, lossless raw).
File 2 is a 16-bit TIFF rendered from that raw by a third party, with unknown in-processor noise reduction and sharpening already applied.
File 3 is an 8-bit in-camera JPEG of the same frame, with the camera's default noise reduction and sharpening.

Without opening Photoshop, produce a written strategy addressing all of the following:

Content map and posture. For each of the seven content zones (sky, hazy hills, foliage, bark, foreground rock, deep shadow, and the person's face), state the noise reduction posture (aggressive / moderate / light / none) and justify it using the two-axis framework (masking and detail value). Identify which zones are the over-denoising traps and explain what would be lost if you treated them like the sky.
Substrate ceilings. Rank the three files as denoising substrates and, for each, state specifically what you can and cannot achieve. For File 1, explain what the raw stage gives you that the others cannot (linear data, bit depth, mosaic/pre-demosaic access, no baked-in processing). For File 2, explain what the 16-bit TIFF retains and what it has already lost, and why its quality is capped by the unknown render. For File 3, enumerate the compounding problems (amplified shadow noise, quantization/posterization risk, block and ringing artifacts, baked-in NR) and explain why your face and foliage zones are especially compromised on this file.
Pre- vs post-demosaic reasoning. Explain, for File 1, why the deep-shadow and sky zones benefit specifically from noise reduction that occurs before or during demosaicing, and what would go wrong (in terms of chrominance artifacts) if the same scene were denoised only after demosaicing. Then explain why Files 2 and 3 can only receive post-demosaic treatment and what that costs.
Local strategy for File 1. Design the sequence of decisions — global baseline (which component is safe to treat globally, and why), then local strong treatment (which zones, and why), then explicit detail protection (which zones, and why). State how you would handle the person's face, in the moderately-lit zone, differently from both the smooth sky and the detailed bark. Reference (do not detail) the masking approach, deferring mechanics to Part 5.
The Camera Raw Filter question. Suppose a colleague proposes to skip raw development on File 1, render it flat to 16-bit, and then do all noise reduction with Filter > Camera Raw Filter "because it has the same sliders." Explain precisely why this is inferior, in terms of the pipeline, even though the controls are identical.
Output reconciliation. The 24x36 baryta print and the 1600 px social post will reveal noise and detail very differently. Explain how the disturbing-noise target differs between them, whether your per-zone postures change between outputs, and why evaluating only at 100% would mislead you for both. State where you might add grain, and to which output.
Failure-mode prediction. For each file, describe what the image would look like if you (a) aimed for "clean" and applied aggressive global luminance noise reduction, and (b) protected only obvious edges while denoising everything else uniformly. Be specific about which zones would collapse first, what the bark, foliage, rock, and face would look like, and — for File 3 specifically — how the JPEG artifacts would interact with the noise reduction.

This challenge requires you to integrate the entire part: the convincing-not-clean mindset, the two-axis content framework, the per-content-type postures, the RAW/TIFF/JPEG substrate ranking, the pre/post-demosaic distinction, the Camera Raw pipeline, and output-dependent evaluation — into three coherent, defensible strategies for one scene, predicting both the results you want and the failures you are avoiding, without touching the software.