Noise vs. Grain

Noise and grain are one texture wearing two verdicts. Both are a fine, random pattern laid across the whole frame that has nothing to do with the subject: the same fizz, pixel for pixel. Call it noise when it arrived uninvited and competes with the picture; call it grain when it was chosen and serves the picture. Nothing in the texture itself tells you which word applies. The difference is authorship, and authorship is what this category judges.

Draw the boundary with sharpness first, because that border is where the confusion lives. Sharpness and focus are about the real detail: whether the fine structure is crisp, and whether the sharp plane landed where it belonged. Noise-grain is about the false texture sitting on top of that detail and fighting it: a static the eye must see past to reach the eyelash, the leaf-edge, the weave. The two categories are neighbours with opposite jobs. Sharpness tries to resolve what the scene actually contained; unmanaged noise, or a clumsy attempt to remove it, buries that same detail from above.

The texture comes in two families, and telling them apart is half the craft. Luminance noise is brightness-only: a grain-like grit, tonal fluctuation with no color in it. Restrained, or added on purpose, it can read as film and even flatter a frame. Chroma noise is color: random red, green, and purple blotches mottling the flat, dark regions. Chroma noise has no aesthetic use, ever: it is always damage. One family can become a choice; the other never graduates. So the category does not ask whether texture is present. It asks whether the texture serves the photograph or happened to it.

A 6 wears a texture it did not choose, and it fails in one of two opposite directions. The first is raw noise left standing: chroma blotches mottling the shadows and the sky, a luminance grit fizzing across the fine structure so the detail arrives through a layer of static. The second is the over-correction, noise reduction pushed until the file goes plastic: waxy surfaces, foliage melted to watercolor, the fine detail the picture depended on smeared into a smooth dead skin. The second failure is the cruel one, because it is a sharpness loss manufactured in the name of looking clean. Both directions share one signature: the texture layer, not the subject, is the first thing a careful eye meets. The file records what happened to it, and a judge can feel that nobody was at the controls.

An 8 resolves the texture into intent, and it can do so in either direction. Sometimes the texture is simply gone: chroma killed completely, luminance restrained just enough that the grit disappears and the detail survives, so the eye reaches the subject and never notices the surface it crossed. Sometimes it is the opposite: a clean, even grain laid over the whole frame on purpose, a chosen surface that unifies the image and gives a digital file the tactile weight of film. The judge's tell is one question asked at 100%: is this texture a decision or an accident? A color blotch is always an accident. A uniform luminance grain can be a signature. The whole distance between the two scores is the distance between a surface that was decided and a surface the file was merely left with: control over the one layer most frames either ignore or bulldoze.

Start with what noise actually is, because the diagnosis happens before the file exists. Light itself is grainy: photons arrive at the sensor randomly, like rain on pavement, and when plenty arrive the randomness averages out invisibly. When few arrive, the randomness is a visible fraction of the signal: that is shot noise, the noise of missing light, and it appears on any sensor, however perfect, whenever the light runs thin. Beneath it the electronics add a second, much smaller hiss of their own in the act of reading the sensor out: read noise, the hardware's baseline hum, which only matters in the deepest, most starved tones, where the signal is thin enough to sink to its level. So most of the noise in a photograph is not the camera malfunctioning. It is the frame reporting how little light it was given.

That is why high ISO is the symptom and not the cause. ISO is raised precisely when the light is thin, and amplifying a thin signal amplifies the randomness riding on it just as hard: a high-ISO frame is an underexposure, amplified, which is exactly how it should be read. And the report always files from the shadows first: the darkest tones hold the fewest photons and sit nearest the read-noise floor, which is why an underexposed shadow earmarked to "lift later" is the exact place noise breeds. The pre-shutter read is therefore two questions: how dark are the regions the picture depends on, and how high will the ISO have to climb to hold them?

At the desk, punch to 100% and read the texture (do not just note that it is there) because the two families take different fixes and confusing them wastes the whole edit. Random purple-and-green speckle, worst in flat dark areas and clear sky, is chroma noise, and it cleans away almost for free. A brightness-only grit with no color in it is luminance noise, and removing it costs detail. Then ask the question the whole lesson turns on: is this texture competing with the detail (a flaw to remove) or is it a cohesive surface the frame could commit to and call grain? The answer decides whether the next move is a denoise slider or a grain slider.

The fight is won at exposure and light, long before any denoise slider exists. Buy light wherever it is cheap so the ISO can stay low: open the aperture a stop, put the camera on a tripod so a static scene can drop to base ISO on a slow shutter, add a reflector or a burst of flash. One caution the sharpness lesson owns: do not trade noise for shake by dragging the shutter past the hand-held floor. A grainy sharp frame beats a clean blurred one, and a clean sharp one beats both; the tripod and the flash exist to buy the third option.

The single largest lever is an exposure decision, not an editing one: expose to the right. A brighter capture collects more photons, and more photons mean quieter light: the shot-noise randomness shrinks relative to the signal, and the whole file rides further above the read-noise floor. Pull that brighter capture back down in post and the shadows come back clean. The exposure lesson carries the full mechanics; the noise consequence is the point here, because the inverse move is the classic self-inflicted wound: underexpose now, lift the shadows later, and the lift amplifies the thinnest, floor-hugging part of the signal into a mess no slider fully undoes. That is not revealing noise. It is manufacturing it.

Know the gear's honest ceiling, because it is physical, not a firmware setting. A larger sensor gathers more total light from the same scene, so at the same ISO it sits further above both noise sources: an ISO 3200 a full-frame body shrugs off can be a grainy wall on a small phone sensor, and no menu option negotiates with that. Long exposures bring their own texture: minutes of accumulated heat build thermal noise and hot pixels that ISO never touched, which is what in-camera long-exposure noise reduction, a dark-frame subtraction, exists to remove. And if the picture wants grain, decide it here or decide it in post, but decide it: a chosen texture is authorship, an inherited one is just weather.

Post holds enormous latitude on one half of this problem and almost none on the other, so separate the two jobs: conflating them is the classic mistake. Chroma noise reduction is nearly free: the color speckle carries almost no real image information, because true color detail is coarse next to the fine mottle, so chroma NR can be pushed hard and lose essentially nothing. Do it first and do it fully, and the ugliest, least defensible part of the texture is simply gone. (The only real limit sits at extreme strengths, where genuinely fine color detail, a tiny distant light, a thin colored thread, can drain toward grey; a 100% check catches it.) Luminance noise reduction is the expensive job, and the reason is structural: luminance noise lives on the same fine frequencies as real detail, so the tool cannot tell the grit from the eyelash (every step of smoothing trades genuine texture for cleanliness). Push it too far and the file goes plastic, watercolor, waxy: a sharpness failure built by hand in the name of removing noise. Restraint, plus a mask that spares the detailed areas and treats the smooth skies and shadows where noise shows most, is the entire discipline. And keep the order the sharpness lesson set: noise reduction before sharpening, because sharpening amplifies whatever grit is left behind.

Modern AI denoise has genuinely moved this wall: the good tools rescue far more luminance noise with far less detail loss than the old sliders, and that is a real advance, not a marketing one. But verify the result at 100%, because a tool that clever can also invent texture (synthetic hair, fabricated leaf-edges) that looks like detail and was never in the file. And the finishing move most editors never learn: adding grain is a legitimate technique. A uniform luminance grain laid over the whole frame unifies a file, hides the seams of heavy noise reduction, and hands a clinical digital image the honest, tactile feel of film. It is the exact inverse of noise (chosen, even, cohesive) and it is how the texture layer turns from a defect into a decision.

Now the honesty clause, and it splits cleanly along the two families. Chroma noise is nearly always recoverable: a color blotch, even on a critical tone, cleans up. Severe luminance noise does not: when a shadow was starved by stops at capture and then lifted, its detail is buried in randomness that no denoise un-buries without smearing what little survived. That frame is capture-limited (the exposure decision coming back to bill the file) and the honest fix was more light or a brighter exposure at the shutter, never a slider today. A subject built entirely of high-ISO mush cannot be made crisp; the only remaining choice is whether it reads as honest grain or as sanded-down plastic. Post cleans texture. It cannot mint the detail the light never gave.