Adds format version 2 alongside version 1 (still written for plain 8-bit
images, so existing readers keep working). Full specification in
docs/FORMAT.md.
- GPU block payloads: BC1-BC7, BC6H, ETC2/EAC, ASTC 4x4-12x12, in tiles
of 16x16 blocks. Built-in BC1/BC3/BC4/BC5/BC7/BC6H encoders and
BC1-BC7 decoders (BC6H/BC7 bit-exact with a reference decoder).
- Lossless 16-bit, half and float pixels as byte planes over the 8-bit
coder; order-preserving float mapping, transforms at sample width.
- Cubemaps, cube arrays, texture arrays, volumes, animation frames.
- Channel meanings, normal convention, linear mask, primaries/transfer,
mip filter and coverage-preserving mips (yapf_tex_build_mips).
- Small mips stored first, file size in the header, surface/tile tables:
decode from a file prefix and single tiles (virtual texturing).
- yapf-sum checksums on header, tables and every tile; needs bits so
decoders refuse cleanly what they cannot read.
- Optional entropy stage: per-surface Huffman code books with a
neighbourhood context (~10% smaller photos) and per-lane packing of
GPU blocks (~3x smaller on disk); planes decoded interleaved.
- Notes: author, licence, source, tool, ICC and application tags.
- CLI: --format, --entropy, --coverage, --normal, --map, colour options,
16-bit/HDR input, pack cube/array/volume/anim, info, verify, export.
- Tests for all of the above, including fuzzing; v1 API reads 8-bit v2.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01DVrVthWN8GsmgHGUhhsCYm
