AVELUNE GEN 1 / UNFROZEN
historicalhistorical · non-normative

Avelune Video Candidate 0.0

Historical Avelune video architecture and coding experiments, retained for provenance and superseded by Draft Generation 1.

View authoritative source
Warning

HISTORICAL / NON-NORMATIVE: superseded by the Draft Generation 1 001-* specification.

Status: EXPERIMENTAL / NOT FROZEN

This document separates architectural requirements that are already intentional from disposable coding experiments.

Architectural invariants under evaluation

  1. Frames have immutable frame IDs.

  2. Every inter frame explicitly identifies its dependency frame IDs; no mutable semantic reference-slot array exists.

  3. Dependencies never cross epoch boundaries.

  4. Decoder resource requirements are bounded from syntax before large allocations.

  5. Entropy state is restartable at independently decodable spatial regions.

  6. Encoder search strategy is never part of the decoder contract.

  7. Lossless coding must be a first-class valid mode, not a separately wrapped format.

Intended source representations

Candidate profiles are expected to cover integer 8/10/12-bit planes, monochrome, 4:2:0, 4:2:2, 4:4:4, RGB/GBR and optional alpha. Exact profile signaling is not frozen.

Candidate coding pipeline

The serious candidate architecture is:

partition -> prediction -> residual transform -> quantization -> entropy coding
          -> inverse operations -> one reconstruction-filter stage

Planned tools are described in docs/architecture/codec-direction.adoc; they are not normative until experiments justify them.

Experiment 0 syntax

The first executable experiment intentionally uses a much smaller subset to test state and container assumptions:

  • 8-bit planar YUV 4:2:0;

  • intra frames are linearly predicted sample-by-sample and encode prediction errors;

  • inter frames predict from the same-position sample in one explicitly named reference frame;

  • residuals are signed integers zig-zag mapped to unsigned values and encoded with bounded LEB128;

  • a quantizer step of one is mathematically lossless;

  • no spatial transform, motion search, filtering, chroma tools or entropy model is implied by this experiment.

A decoder implementing Experiment 0 is not evidence that the above mechanism belongs in Draft Generation 1. It exists to validate exact arithmetic, explicit reference dependencies, packetization and independent epochs.

Exact experiment reconstruction

For every sample:

prediction = intra_predictor(...) OR reference_sample
error = source - prediction
reconstructed = clamp(prediction + dequantized_error, 0, 255)

For lossless Experiment 0, dequantized_error == error and clipping MUST therefore recover the original sample exactly.