Why PSNR Fails for AV1 Film Grain Synthesis

AV1 film grain synthesis drastically improves perceived video quality at lower bitrates by removing natural grain before encoding and synthetically regenerating it during playback. While this technique creates a visually pleasing, cinematic experience that masks compression artifacts, traditional metrics like Peak Signal-to-Noise Ratio (PSNR) register it as severe signal degradation. This article explains why the mathematical nature of PSNR makes it fundamentally incapable of measuring the subjective benefits of AV1's film grain synthesis.

How AV1 Film Grain Synthesis Works

Natural film grain and high-frequency sensor noise are notoriously difficult for video codecs to compress. Because noise is mostly random, motion compensation and spatial prediction cannot effectively predict it, forcing encoders to spend a large portion of the bitrate budget simply preserving static-like patterns.

AV1 solves this with Film Grain Synthesis (FGS). During encoding, the video is analyzed to estimate the statistical characteristics of the grain, such as its pattern, intensity, and frequency. The encoder then denoises the video, encodes the clean image at a much lower bitrate, and sends the grain characteristics as lightweight metadata. Upon decoding, the player generates pseudo-random synthetic grain according to these parameters and blends it back onto the reconstructed frame.

The Mathematical Flaw of PSNR

Peak Signal-to-Noise Ratio relies entirely on Mean Squared Error (MSE), which computes the squared difference between the pixel values of the original frame and the decoded frame at identical spatial coordinates \((x, y)\).

Because the grain generated by AV1 is stochastic (pseudo-randomly simulated rather than an exact copy of the source grain), the synthetic grain particles almost never align perfectly with the original grain on a pixel-by-pixel basis. A pixel that was bright in the original noise pattern might be dark in the synthesized pattern.

From an MSE standpoint, this misalignment registers as a massive error. When calculating PSNR, this discrepancy significantly lowers the score, falsely indicating that the video quality has degraded, even though the image may appear sharper and more detailed to human viewers.

Human Perception vs. Deterministic Matching

The human visual system does not evaluate grain by comparing individual pixel positions. Instead, human eyes evaluate grain statistically and contextually:

• Texture Consistency: The brain perceives grain as an overarching texture that gives the image an organic, filmic look.
• Artifact Masking: Grain acts as natural dither, effectively masking compression artifacts such as color banding in flat areas and macroblocking in dark scenes.
• Perceived Sharpness: The presence of grain increases the high-frequency content of the display, making soft edges appear visually crisper.

Because PSNR evaluates deterministic, pixel-for-pixel accuracy rather than structural or statistical fidelity, it treats deliberate, perceptually beneficial noise as corruption. Turning film grain synthesis on will almost universally lower a stream's PSNR score while simultaneously improving its subjective viewing quality.

Better Alternatives for Assessing AV1 Video

Because standard PSNR fails to capture the benefits of FGS, relying on it for AV1 rate-distortion optimization or quality verification leads to misleading conclusions. Modern workflows rely on alternative approaches:

• Denoised-Reference PSNR/SSIM: Evaluating the reconstructed image before film grain is added against a pre-denoised version of the source to measure actual underlying compression fidelity.
• Tuned Perceptual Metrics: Using metrics like VMAF (Video Multi-Method Assessment Fusion) with grain-aware models or combined with spatial metrics like CAMBI (to measure banding reduction).
• Subjective Human Testing: Utilizing Mean Opinion Score (MOS) evaluations, which remain the ultimate ground truth for validating synthetic grain performance.