CONCEPT Cited by 1 source

Grain intensity scaling function¶

Definition¶

A grain intensity scaling function is a piecewise-linear function from pixel value → grain intensity transmitted alongside AR grain coefficients in the AV1 Film Grain Synthesis bitstream. During decode, it controls how strongly synthesized grain is added at each point of the image — brighter regions get one intensity, darker regions another, mid-tones yet another (Source: sources/2025-07-03-netflix-av1scale-film-grain-synthesis-the-awakening).

The scaling function is "estimated during the encoding process" and "models the relationship between pixel value and noise intensity using a piecewise linear function. This allows for precise adjustments to the grain strength based on video brightness and color."

Why a scaling function is necessary¶

Film grain is not uniformly visible across an image. In real film stock and in digital sensor noise:

Dark regions saturate grain into invisible territory or swamp it with shadow detail — grain is weak here.
Highlights / bright whites — grain tends to be overwhelmed by the signal or clipped.
Mid-tones — this is where grain is most visible, and where viewers perceive the "filmic look".

A single grain intensity applied uniformly to every pixel would either under-grain the mid-tones or over-grain the shadows. The scaling function avoids both failure modes by making intensity a function of local pixel brightness — so the reconstructed grain layer has the same "patchy, intensity-dependent" character as real film grain.

Why piecewise-linear¶

Piecewise-linear is the simplest parametric family rich enough to capture the intensity-vs-brightness relationship:

Easy to estimate on the encoder — fit segments to the residual's local variance-vs-mean relationship.
Compact to transmit — a small number of (pixel_value, intensity) break points per frame or per GOP.
Trivial to evaluate on the decoder — linear interpolation between the nearest two break points, per pixel. This is a first-class design constraint: the decoder must run on commodity consumer devices (TVs, phones, set-top boxes).

Non-linear alternatives (cubic splines, learned networks) would be more expressive but none of them are justified given the grain-intensity relationship is empirically well-approximated by a small piecewise-linear curve.

Illustration (Netflix Fig. 2)¶

  Grain intensity
      │                    ╱───╮
      │                 ╱        ╲
      │              ╱             ╲____
      │           ╱
      │      ╱
      │ ╱
      └──────────────────────────────────► pixel value
        dark      mid-tones       bright

Netflix's Fig. 2 shows this curve next to a grayscale SMPTE bars frame with the resulting grain applied: the mid-grey bars have visibly heavier grain than the darkest / brightest bars, matching the curve's peak.

Interaction with the AR model¶

The AR model — see concepts/auto-regressive-grain-model — describes the shape of the grain (coarseness, spatial correlation). The scaling function describes the amplitude of the grain as a function of pixel brightness. The two are orthogonal:

AR coefficients answer "what does a grain speck look like?"
The scaling function answers "how strongly does grain appear at this brightness?"

Together they specify a per-pixel grain value at the decoder:

  grain(x, y) = scale(decoded_pixel(x, y)) · template_patch(x, y)

where template_patch is the 32×32 random-position patch extracted from the AR-generated 64×64 noise template, and scale(·) is the piecewise-linear function interpolating between transmitted break points.

Bitrate cost of the scaling function¶

The scaling function is a handful of (pixel_value, intensity) break points plus colour-channel metadata — a few tens of bytes per transmitted unit (frame, GOP, or per- sequence). Vanishingly cheap compared to the bitrate saved by denoising the source before compression. Together with the AR coefficients it is the compact "side channel" that makes decoder- side synthesis for compression a net bitrate win.

Seen in¶

sources/2025-07-03-netflix-av1scale-film-grain-synthesis-the-awakening — canonical wiki source; Fig. 2 illustrates the scaling function applied to a SMPTE bars frame.