Skip to content

CONCEPT Cited by 1 source

Film Grain Synthesis (FGS)

Definition

Film Grain Synthesis (FGS) is an AV1 codec tool that removes film grain from the source video before compression, transmits a compact parameter set describing the grain, and reconstructs the grain layer on the decoder after decompression. The compressed bitstream carries a denoised video plus grain metadata; the decoder re-synthesizes the grain from the metadata and adds it to the decoded frame.

FGS has been part of the AV1 standard since its inception, but was only enabled at scale by Netflix in 2025 — four years after the codec itself shipped (Source: sources/2025-07-03-netflix-av1scale-film-grain-synthesis-the-awakening).

The compression problem FGS solves

Film grain — whether from real film stock, sensor noise, or post-production-added grain for artistic mood — is "notoriously difficult to compress". Block-transform codecs (AV1, HEVC, VP9, H.264) rely on spatial and temporal correlation to compress efficiently. Grain, by construction, has no exploitable correlation: it is near-random noise at per-pixel scale. That makes it the adversarial input to a block-transform codec. Conventional codecs therefore face a forced choice:

  • Preserve the grain and blow bitrate by a large factor, or
  • Crush the grain with heavy quantization and lose the artistic intent.

FGS ejects the grain from the compressed bitstream entirely, neither preserving it as texture nor crushing it. The bitstream carries a denoised (low-entropy) signal plus a handful of grain parameters; the decoder re-adds the grain from those parameters, producing output that looks grainy to the viewer while costing almost nothing to transmit.

The two grain components

FGS models grain as two orthogonal parameter sets:

Grain pattern — auto-regressive (AR) model

The pattern component — shape, coarseness — is captured by an auto-regressive model whose AR coefficients {a₀, a₁, a₂, …} define how a noise pixel depends on previously-synthesized neighbouring samples. The coefficients are estimated from the residual between source and denoised video ("essentially capturing the noise"), and drive generation of a 64×64 noise template at decode time. At playback, random 32×32 patches are extracted from the 64×64 template and tiled onto the decoded frame. See concepts/auto-regressive-grain-model.

Grain intensity — piecewise-linear scaling function

The intensity component — how strong the grain appears in dark vs bright regions — is captured by a piecewise-linear scaling function mapping pixel value to grain intensity, estimated during encoding. This models the empirical fact that film grain is "adapted to the areas of the picture" — typically more visible in mid-tones than in blacks or highlights. See concepts/grain-intensity-scaling-function.

Together the two components let the decoder reconstruct grain that "closely recreates the look of the original video" without the original grain ever being in the compressed bitstream.

The encoding pipeline shape

FGS turns the encoder into a three-stage transformation — see concepts/denoise-encode-synthesize:

  1. Denoise the source video (encoder-vendor choice — AV1 standard does not mandate a denoiser).
  2. Encode the denoised signal (standard AV1 toolchain).
  3. Estimate AR coefficients + scaling function and transmit them as grain metadata alongside the compressed video.

The decoder side is fully specified by the AV1 standard: generate a 64×64 noise template from the AR coefficients, tile random 32×32 patches across each decoded frame, scale per-pixel by the scaling function, and add to the decoded video. The decoder-side procedure is explicitly designed to be cheap on commodity consumer devices — TVs, phones, set-top boxes — so the grain layer can be reconstructed in real time without specialised hardware.

Why this is an architectural move, not a compression trick

FGS is an instance of the more general decoder-side synthesis pattern: don't transmit a signal, transmit the parameters of a generator that can reconstruct it. This works whenever the signal is expensive to compress as texture but cheap to describe as parameters. Grain is the canonical example: 100 KB of grain parameters vs megabytes of encoded grain texture, for a perceptually equivalent result.

The architectural cost is:

  • Encoder complexity — grain parameter estimation + denoiser are real pipeline stages.
  • Decoder cost — a small, bounded per-frame cost.
  • Evaluation cost — reference metrics like VMAF / PSNR / SSIM (see concepts/visual-quality-metric) break down because FGS output is sample-wise different from the source even when perceptually equivalent. A synthesized grain pattern is not the same noise instance as the source grain; reference metrics will score it as heavily distorted even when viewers cannot tell.

The Netflix post hedges its quality claim accordingly: it speaks of "high-quality video" and "preserving the artistic integrity of film grain" rather than "higher VMAF".

Scope on the wiki

  • Only documented AV1 use case so far. Film-grain synthesis is AV1-specific; HEVC / VP9 / H.264 do not have an in-standard analogue.
  • Not the same as post-processing grain overlays. Player-side post-processing (e.g. dithering filters) is cosmetic; FGS is a bitstream-level tool with grain parameters transmitted from the encoder.
  • Deployment constraint is the long tail of AV1 decoders. FGS requires the decoder to implement the grain-synthesis procedure, which means device-compatibility testing is a real cost of the rollout. See patterns/codec-feature-gradual-rollout.

Seen in

Last updated · 319 distilled / 1,201 read