Optimize AVIF

AVIF optimization through Krunkit uses libavif, the reference encoder maintained by the Alliance for Open Media. The encoder exposes a speed parameter from 0 (slowest, most efficient) to 10 (fastest, least efficient). Each speed step adjusts dozens of internal parameters including transform block sizes, RDO (rate-distortion optimization) aggressiveness, and intra-prediction mode evaluation depth. Moving from speed 6 (Krunkit's default) to speed 4 typically reduces file size by 8-15% at the cost of 3-5x longer encoding time.

AVIF's tile-based encoding offers optimization opportunities unique to this format. Large images can be divided into independent tiles that encode and decode in parallel. For a 4000x3000 image, encoding with 2x2 tiles enables multi-threaded WASM processing and reduces the browser's peak memory usage from ~180 MB to ~50 MB. However, tile boundaries introduce redundancy because prediction cannot cross tiles — using fewer tiles produces smaller files when memory is not a constraint.

Chroma subsampling provides significant optimization potential in AVIF. The default 4:2:0 subsampling reduces color resolution to one-quarter of luminance resolution, which is imperceptible in most photographs because the human visual system is far less sensitive to color detail than brightness detail. However, for content with saturated color edges — red text on blue backgrounds, color calibration targets, fashion photography with vivid fabrics — switching to 4:4:4 subsampling preserves color fidelity at a 15-25% file size increase.

AVIF uniquely supports film grain synthesis, a technique borrowed from AV1 video encoding. Instead of compressing grainy textures pixel by pixel, the encoder analyzes the grain pattern, encodes a smooth underlying image, and stores grain parameters that the decoder uses to regenerate synthetic grain during display. For photographs shot at high ISO or images with intentional film grain aesthetic, this feature reduces file size by 20-40% while maintaining the visual texture that gives the image its character.