An experimental, vibecoded video codec built from scratch using Claude Code agent teams. Explores lapped transforms, perceptual vector quantization (PVQ), and rANS entropy coding as a patent-free alternative to the H.264/H.265 lineage.
This was a learning experiment, not a production codec. The goal was to test one-shot agent team workflows on a domain I had zero prior experience in — and to see how far a simulated expert team could push a novel architecture. Write-up: One-Shot Wonder | Claude Agent Teams
At comparable luma quality (~49 dB): 18.6x larger than H.264. The architecture produces competitive perceptual quality but cannot match H.264's compression efficiency without adopting the same tools (B-frames, sub-pel MC, CABAC-level entropy coding).
| QP | Sinter PSNR | Sinter Size | H.264 PSNR | H.264 Size |
|---|---|---|---|---|
| 4 | 49.10 dB | 136 KB | 74.59 dB | 33 KB |
| 20 | 33.70 dB | 37 KB | 55.12 dB | 11 KB |
| 28 | 28.41 dB | 13 KB | 48.37 dB | 7 KB |
(256x256 testsrc, 30 frames. Full BD-rate data in SCOREBOARD.md.)
~5,000 lines of C across 12 improvement loops:
- Lapped transforms (TDLT, 20% Malvar lifting) — eliminates blocking artifacts structurally
- Hybrid PVQ/scalar quantization — preserves texture where H.264 smooths it away
- Dual-interleaved rANS with 10 PARA-adaptive CDFs and cross-frame carry
- Inter prediction (P-frames, integer-pel full search, median MV prediction, skip mode)
- Chroma inter (luma MV at half-resolution for YUV 4:2:0)
- CLI tools with Y4M I/O (
sinterenc/sinterdec) - 32 passing tests, entirely patent-free
make && make test
# Build CLI
gcc -O2 -I libsinter tools/sinterenc.c -Lbuild -lsinter -lm -o build/sinterenc
gcc -O2 -I libsinter tools/sinterdec.c -Lbuild -lsinter -lm -o build/sinterdec
# Encode/decode
ffmpeg -f lavfi -i "testsrc=duration=1:size=64x64:rate=10" -pix_fmt yuv420p -f yuv4mpegpipe input.y4m -y
./build/sinterenc -i input.y4m -o encoded.sntr -q 12
./build/sinterdec -i encoded.sntr -o decoded.y4mThe expert team's consensus: the 18.6x gap is the product of missing standard tools, each multiplicative:
| Missing Feature | Cost | Notes |
|---|---|---|
| No sub-pel MC | 1.5-2x | Integer-pel can't track sub-pixel motion |
| No B-frames | 1.5-2x | Patent risk on bidirectional prediction |
| PVQ overhead | 1.5-2x | Gain-shape is more verbose than run-level |
| Fewer entropy contexts | 1.2-1.5x | 10 CDFs vs CABAC's hundreds |
Realistic patent-safe ceiling: 4-6x H.264 with half-pel + better contexts. Matching H.264 would require rebuilding most of its toolset, defeating the purpose.
Built with Claude Code. The expert team was assembled by asking Claude to name five experts for building an open-source codec competing with H.264/H.265:
- Jim Bankoski — Google architect behind VP8 and VP9; open-codec veteran
- Timothy Terriberry — Mozilla/Xiph engineer; drove Daala and AV1 research
- Monty Montgomery — Xiph.org founder; built Ogg, Vorbis, and Theora ecosystems
- Gary Sullivan — Microsoft; co-chaired the H.264 and HEVC standardization efforts
- Jens-Rainer Ohm — HEVC co-chair; deep expertise in perceptual coding theory
Plus a principal engineer (facilitator) and patent reviewer. All personas simulated by Claude.
- C11 compiler (GCC or Clang)
- make
- ffmpeg (for Y4M generation)
- No external dependencies