AVIF Film Grain Synthesis at Low Bitrates

AVIF manages film grain synthesis by stripping natural high-frequency noise during encoding, analyzing its characteristics, and reconstructing visually identical grain during playback using lightweight metadata. Traditional compression algorithms waste significant data trying to encode random grain pixels, often resulting in heavy blurring, blockiness, or severe color banding at low bitrates. By decoupling texture from the underlying image structure, AVIF preserves the authentic, organic feel of film and fine textures without consuming precious bandwidth.

The Problem with Film Grain in Traditional Compression

Film grain and digital sensor noise consist of pseudo-random, high-frequency signals. Traditional transform-based codecs like JPEG, and even modern codecs without synthesis, treat these chaotic variations as essential image details. At low bitrates, encoders cannot afford the data budget required to represent these rapidly shifting pixels. The compression algorithm typically discards these high frequencies, leading to two common visual flaws:

• Waxiness and Smearing: Faces and smooth surfaces lose micro-texture and look artificially smooth or "plastic."
• Banding and Blocking: Gradual tonal shifts (such as skies or dark shadows) turn into stepped, flat-colored blocks rather than smooth gradients.

How AVIF Film Grain Synthesis Works

AVIF inherits its film grain synthesis (FGS) mechanism directly from the AV1 video standard. The process relies on parametric modeling and is divided into distinct stages:

1. Grain Removal (Denoising)

Before compression, the encoder applies an integrated or pre-processing denoising filter to the source image. This step separates the structural image data from the high-frequency noise. The underlying picture becomes significantly easier to compress, allowing the encoder to achieve high compression ratios without creating motion-like distortion or block boundaries.

2. Parameter Estimation

Rather than discarding the removed noise, the encoder analyzes its mathematical properties. It measures several statistical attributes:

• Grain scale and size: The resolution and frequency of the grain particles.
• Intensity correlation: How grain visibility changes across different luminance (brightness) levels.
• Chrominance relationship: How noise in color channels correlates with brightness.

3. Metadata Storage

Instead of saving the grain as raw pixel data, the encoder translates the grain characteristics into a compact set of numerical parameters. These values are stored as lightweight metadata within the AVIF container (typically using only a few dozen to a few hundred bytes).

4. Decoder-Side Synthesis

When an AVIF-compatible viewer or browser renders the image, it performs the synthesis step:

• The decoder reconstructs the smooth, base image.
• It reads the film grain metadata parameters.
• Using a deterministic pseudo-random number generator (PRNG) combined with an autoregressive filter, the decoder generates a synthetic grain texture matching the original profile.
• The generated grain is blended directly onto the rendered image just before display.

Why This Technique Excels at Low Bitrates

Film grain synthesis fundamentally changes the rate-distortion tradeoff for photographic and cinematic imagery:

• Bandwidth Efficiency: Encoding mathematical formulas requires a negligible fraction of the data needed to encode actual pixel noise, freeing up the bitrate budget for edge sharpness and color accuracy.
• Banding Concealment: The reconstructed grain breaks up quantization steps across gradients, naturally masking color banding in dark areas and skies.
• Perceptual Realism: Human vision interprets micro-texture as sharpness. Even though the synthesized grain is not an exact pixel-for-pixel replica of the original noise, the human eye perceives it as equivalent, maintaining artistic intent and photographic quality at extremely low file sizes.