All posts
By Pravit Gandhi··8 min read

Why log footage gets noisy when you grade it

Why is my S-Log3 footage grainy? Lifting log shadows amplifies chroma noise, and the cause is usually exposure rather than the grade you applied.

The grade did not add the noise. It was already in the file, sitting in the deep shadows where a log curve spends very few code values, and normalizing to a viewable image stretches that region until you can see what was there all along. Which is why the complaint always arrives in the same shape: the clip looked acceptable while it was flat, and the moment it got lifted the shadows went blotchy and green. Blackmagic states the connection in its own manual, noting that "increasing the contrast of underexposed clips often increases whatever noise is within an image." Underexposure is the cause; the grade is what reveals it.

The cause is exposure, and log makes it more expensive

Every log format is a deal: give up contrast in the recorded file, get more of the scene's range inside it. Nothing in that deal improves the sensor. The noise floor sits where the sensor puts it, and a log curve encodes the region near that floor using very little signal.

Sony's Cine EI guidance is direct about the cost of getting exposure wrong. Its help guide states that "recording at the base ISO enables you to take full advantage of the image sensor performance," and that "in general, shooting at an EI equal to or lower than the base ISO value is recommended." On rating higher, Sony does not hedge: "you can emphasize gradation in highlights when shooting at a higher EI, far from the same value as the base ISO, but this will increase noise and graininess, and color grading will become more difficult."

Note the last clause. Sony is not only saying the picture will be noisier, it is saying the grade will be harder, which is the part people discover in post rather than on set. The S-Log3 version is in how to expose S-Log3 correctly.

What lifting a log shadow actually does

Here is the mechanical part, with our own numbers rather than a hand-wave.

We computed how each published log curve distributes 10-bit code values across scene exposure, validating each implementation against that manufacturer's own published anchors first, and published the full ladder. For Sony S-Log3, the one-stop step of scene light from six stops under middle grey to five stops under consumes 19 code values. The step from middle grey to one stop over consumes 76, and in the straight-line part of the curve the slope settles at about 79 per stop.

So the deepest stops are recorded roughly four times more coarsely than the mid-range. That is not a defect, it is what lets the format hold a wide scene in an ordinary file. But it means normalizing has to take that 19-code-value region and open it out across a much larger share of a display's range, and whatever variation was inside those 19 values gets expanded by the same factor as the signal, because the transform cannot tell the two apart. Sensor noise comes up, quantization steps come up, and neither was created by the grade.

It also explains why the problem worsens faster than the underexposure does: two stops under costs more than twice what one stop costs, because you are sliding into a part of the curve where the code values per stop are still falling. When the same arithmetic produces stepping in smooth areas rather than speckle, that is banding after grading and export.

Why the colour noise is the ugly part

Noise arrives in two flavours that do not look alike. Luma noise is variation in brightness: fine, film-like, often tolerable. Chroma noise is variation in colour: blotches, usually green and magenta, drifting through the shadows. Almost nobody minds the first and almost everybody minds the second.

Part of that is taste and part is structural, coming from how the file stores colour. Apple's ProRes white paper defines video, in its glossary, as image sequences that "typically use the Y'CBCR color space and subsampled chroma channels, usually with one of the following patterns: 4:2:2, 4:2:0, or 4:1:1." Subsampled means colour is recorded at lower spatial resolution than brightness. Fujifilm's X-H2 manual gives the per-codec version: H.264 records "at a depth of 8 bits" with 4:2:0 chroma subsampling, while H.265 records at 10 bits with either 4:2:0 or 4:2:2.

A single noisy chroma sample in 4:2:0 therefore does not colour one pixel, it colours a block. That is why chroma noise reads as patches rather than grain, and why it draws the eye in the flat shadow areas where there is nothing else to look at. Blackmagic's controls carry the corollary: the manual notes you "may find you can raise the Chroma Threshold higher than the Luma Threshold with less noticeable artifacting."

What noise reduction gets back, and what it does not

It removes noise. It does not recover detail, because detail that never separated itself from the noise floor was not recorded, and nothing can distinguish a shadow texture that was never there from one averaged away.

Resolve's two approaches make the trade explicit. Temporal noise reduction analyzes "images across multiple frames in order to isolate noise from detail," with motion estimation to exclude moving subjects. Spatial noise reduction instead smooths "regions of high-frequency noise throughout the image, while attempting to avoid softening by preserving detail." The warning attached to both is the same sentence, repeated for luma and chroma: "too high a setting may eliminate fine detail from the image."

Blackmagic's recommended sequence is temporal first, because it "tends to be better at preserving detail accurately in regions of the image where there's little motion," then spatial for what remains. The manual also calls out a trap that costs people afternoons: "zooming really far into a clip while applying noise reduction may encourage you to use higher quality settings than are necessary," because you see changes nobody notices at actual size.

Where noise reduction goes in the chain

The common advice is to denoise before you lift, on the reasoning that you should not amplify noise you could have removed first. That advice is half right, and the manual shipping with the software most people use says so.

Blackmagic gives both orders and refuses to pick. Noise reduction at the start of a node tree "lets you pre-emptively eliminate any noise before it becomes a problem," and "the result can be smoother, but you may also notice that the edge detail within the image is a bit softer." At the end, "the noise reduced regions of the image aren't quite as smooth, however the edge detail within the image may be visibly sharper as a result." The verdict is one sentence: "Neither result is universally better or worse than the other."

So the real rule is narrower than "always denoise first." Put it early when the shot is badly underexposed and the grade will be a big lift, because there amplification dominates. Put it late when the shot is close to correctly exposed and you care about edge detail.

One ordering point is not optional. Noise reduction belongs after the log conversion, for the same reason the noise became visible at all: judging it on the un-normalized picture means judging it where the shadows are compressed into almost nothing, so you will consistently under-apply. Normalize, balance exposure and white balance, decide about noise, then build the look. Why that sequence is fixed is in why CST, ACES and LUTs look different.

What to do with footage that is already shot

Accept a lower ceiling and stop lifting so far. The most effective move on an underexposed log clip is usually to not fully rescue it: let the shadows sit lower than you wanted, and keep the deepest stops out of the part of the curve you stretch hardest.

Then treat chroma and luma separately rather than reaching for one slider, taking chroma out aggressively and luma sparingly. Where noise is confined to particular regions, the manual points out that noise reduction can be limited with qualifiers and windows like any other correction, "for example, in shadows and background regions," while sparing faces. And if some clips are clean and some are not, resist matching the good ones down to the bad ones. When a grade feels wrong and noise is not the reason, the diagnosis is in why does my grade look off.

Leumos AI, our browser-based grading studio, does not fix underexposure, and no tool honestly can. What it does is apply the input transform consistently across every shot in an upload. It is in closed beta with a waitlist at leumos.ai.

Frequently asked questions

Why is my S-Log3 footage grainy?

It was underexposed, and S-Log3 encodes the deepest stops very coarsely. Our computation from Sony's published curve puts 19 code values on the stop of scene light from six stops under middle grey to five stops under, against about 79 per stop in the straight-line region. Normalizing expands that thin region, and the noise in it expands too.

Should I apply noise reduction before or after my grade?

Blackmagic's own manual declines to give a universal answer, stating that neither order is universally better. Early gives a smoother result with slightly softer edges; late keeps edges sharper but leaves noise-reduced areas less smooth. Choose early for badly underexposed shots needing a large lift, late when the shot is close to correct.

Does noise reduction bring back shadow detail?

No. It removes noise; it cannot recover detail that was never separated from the noise floor. Resolve's manual warns for both luma and chroma that too high a setting may eliminate fine detail from the image.

Why is the colour noise worse than the grain?

Because colour is stored at lower spatial resolution than brightness. Apple's ProRes glossary describes video as typically using subsampled chroma channels in 4:2:2, 4:2:0 or 4:1:1 patterns, and Fujifilm's manual confirms H.264 records 8-bit 4:2:0. One bad chroma sample therefore tints a block of pixels rather than one.

Sources