How librav1e Encodes Water and Foliage Textures

This article examines how the Rust-based AV1 encoder, librav1e, processes and adapts to highly detailed, complex textures such as moving water, grass, and foliage. It covers the specific encoding tools, adaptive quantization methods, and partition decisions librav1e utilizes to preserve high-frequency details while maintaining optimal bitrate efficiency.

The Challenge of High-Detail Textures

Textures like foliage, grass, and water are notoriously difficult for video encoders to compress. These scenes contain high-frequency spatial details and chaotic, non-uniform motion. Standard compression algorithms tend to smooth out these areas, resulting in a blurry, smeared effect, or they over-allocate bits, causing massive bitrate spikes. To solve this, librav1e employs several sophisticated adaptation strategies.

Adaptive Quantization (AQ)

The primary tool librav1e uses to handle complex textures is Adaptive Quantization. AQ dynamically adjusts the quantization parameter (QP) across different regions of a single frame.

In highly detailed areas like foliage, the human eye is less sensitive to minor compression artifacts because the visual field is already busy. However, the eye is highly sensitive to blockiness or sudden loss of detail (blurring). librav1e’s AQ algorithms analyze the spatial variance of each block. It then redistributes bitrate away from flat, low-contrast areas (where compression artifacts would be highly visible) and allocates it to the complex, textured zones to preserve the sharp edges of leaves and water ripples.

Dynamic Block Partitioning

AV1 allows for extremely flexible block partitioning, ranging from 128x128 down to 4x4 pixels. librav1e utilizes this flexibility to adapt to texture complexity:

• Large Blocks: Used for flat or homogeneous areas, such as a clear sky behind a tree.
• Small Blocks: Used for highly detailed areas, such as the leaves of the tree.

For water and foliage, librav1e performs recursive rate-distortion optimization (RDO) to determine the best partition layout. By breaking down chaotic texture areas into smaller blocks, the encoder can more accurately predict motion and spatial patterns, preventing block boundaries from tearing or pixelating.

Directional Intra Prediction and Loop Filters

When encoding complex static textures or slow-moving foliage, librav1e relies heavily on AV1’s advanced intra prediction. With 56 directional intra prediction modes, the encoder can precisely match the angles of branches, grass blades, or water waves.

Once predicted and quantized, librav1e applies in-loop filtering tools to clean up the output:

• Constrained Directional Enhancement Filter (CDEF): CDEF identifies the direction of edges in the texture and applies a directional smoothing filter. This removes ringing artifacts around fine details, like individual pine needles, without blurring the actual edges.
• Loop Restoration Filter (LR): This filter uses Wiener and Self-Guided Restoration to restore fine-scale details lost during the quantization process, acting as a final polish for complex textures.

Film Grain Synthesis Support

While not a direct compression of the texture itself, librav1e supports AV1’s film grain synthesis. Highly detailed textures often contain high-frequency noise that behaves similarly to film grain. Instead of spending precious bits trying to encode this random noise frame-by-frame, librav1e can analyze the noise, strip it out during encoding, and send parameters for the decoder to regenerate the texture noise on the playback side. This keeps the visual representation of water spray or rustling leaves looking natural and sharp without bloating the file size.