Rust rewrite of FFmpeg, verified against FFmpeg's output bit-for-bit.

This codebase is AI-generated. Written by Claude (Anthropic) via Claude Code, directed and reviewed by a human. The AI reads FFmpeg's C source and reimplements it in Rust. Every conformance claim below is verified by automated CI on every commit, comparing to FFmpeg's output.

The intention of this project is to push the boundaries of what is possible with AI rewriting codebases in Rust. It provides no additional features compared to FFmpeg, and despite incorporating FFmpeg's assembly code, is significantly slower.

The H.264 decoder is ~30K lines of Rust across 25 modules. It implements:

• CAVLC and CABAC entropy coding
• All intra prediction modes (4x4, 8x8, 16x16, chroma)
• Quarter-pel motion compensation (6-tap FIR luma, bilinear chroma)
• 4x4 and 8x8 IDCT with Hadamard DC transforms
• In-loop deblocking filter
• MMCO and sliding-window reference management
• Weighted prediction (uni/bi)
• B-frames with direct prediction (spatial + temporal)
• High profile: 8x8 transforms, custom scaling matrices
• MBAFF interlaced (partial — field/frame MB switching, CABAC context adaptation)
• Frame-level threading with wavefront deblocking
• aarch64 NEON assembly for MC, IDCT, and deblocking (feature-gated)

Architecture details: H264.md. Known FFmpeg behavioral differences: DIVERGENCES.md.

FFmpeg has hundreds of codecs and formats. wedeo currently covers a small subset. Major gaps for parity:

• Video codecs — VP9, HEVC/H.265, MPEG-2, MPEG-4 Part 2, VP8, Theora. H.264 is missing interlaced (MBAFF/PAFF), 10-bit, and 4:2:2/4:4:4.
• Video encoding — no encoders exist yet (H.264, H.265, AV1 via rav1e)
• Muxers — only WAV. No MP4/MOV, MKV/WebM, or MPEG-TS muxer.
• Demuxers — no MKV/WebM, MPEG-TS, FLV, or AVI demuxer (MP4 and WAV only, plus symphonia-backed formats)
• Filters — trait skeleton exists but no functional filter graph (no scale, crop, overlay, fps, etc.)
• Player — no seek, no subtitle rendering, no hardware-accelerated decode
• Infrastructure — no interruptible I/O (network streams), no chapter/program support, no avformat_find_stream_info equivalent

# From source (includes AV1 support via rav1d)
cargo install --git https://github.com/sharifhsn/wedeo wedeo-cli

# Also from source — AV1, H.264, and audio all work
cargo install --git https://github.com/sharifhsn/wedeo wedeo-play
wedeo-play video.mp4

# Cargo.toml — core crates are on crates.io
[dependencies]
wedeo = "0.1.2"
wedeo-codec-h264 = "0.1.2"    # H.264 decoder
wedeo-format-mp4 = "0.1.2"    # MP4 demuxer
wedeo-symphonia = "0.1.2"     # audio codecs (AAC, MP3, FLAC, etc.)

# AV1 requires a git dependency (rav1d is not yet on crates.io)
wedeo-rav1d = { git = "https://github.com/sharifhsn/wedeo" }

cargo build
cargo nextest run            # or cargo test
cargo clippy

Decode a file and compare against FFmpeg:

cargo run --release --bin wedeo-framecrc -- input.264
ffmpeg -bitexact -i input.264 -f framecrc -

FFmpeg's native test suite.

./scripts/fetch-fate-suite.sh                    # downloads full suite (~1.2 GB)
FATE_SUITE=./fate-suite cargo nextest run -p wedeo-fate

204 test vectors from the ITU JVT conformance suite, with MD5 ground truth from the Fluster project. No FFmpeg required — comparison is against ITU-provided checksums.

python3 scripts/fetch_jvt.py                     # download vectors (~50 MB)
python3 scripts/suite_runner.py --suite jvt-avc-v1,jvt-fr-ext --format yuv420p

The unified suite runner also wraps the FATE suites:

python3 scripts/suite_runner.py --suite all --format yuv420p
python3 scripts/suite_runner.py --suite fate-cavlc --save-snapshot
python3 scripts/suite_runner.py --suite fate-cavlc --check-snapshot   # regression check

wedeo/
  crates/
    wedeo-core/          libavutil   — Rational, Buffer, Frame, Packet, errors
    wedeo-codec/         libavcodec  — Decoder/Encoder traits, codec registry
    wedeo-format/        libavformat — Demuxer/Muxer traits, I/O, InputContext
    wedeo-filter/        libavfilter (stub)
    wedeo-resample/      libswresample (rubato)
    wedeo-scale/         libswscale (dcv-color-primitives)
  codecs/
    wedeo-codec-h264/    H.264 decoder — 30K lines, NEON assembly, 55 benchmarks
    wedeo-codec-pcm/     PCM codec — 17 formats
  formats/
    wedeo-format-h264/   H.264 Annex B demuxer
    wedeo-format-wav/    WAV demuxer + muxer
    wedeo-format-mp4/    MP4/MOV demuxer
  adapters/
    wedeo-symphonia/     Wraps symphonia (FLAC, Vorbis, MP3, AAC, WavPack)
    wedeo-rav1d/         Wraps rav1d (AV1)
  bins/
    wedeo-cli/           CLI tool
    wedeo-play/          Video player — wgpu+winit, ffplay-style A/V sync
  tests/
    fate/                FATE cross-validation harness
  scripts/
    suite_runner.py      Unified conformance runner (FATE + JVT)
    fetch_jvt.py         JVT vector downloader
    conformance_full.py  FATE conformance report
    regression_check.py  Quick regression check
  test_suites/
    h264/                JVT manifest JSONs (tracked)

Each FFmpeg library maps to one Rust crate. Codecs and formats register themselves via inventory at link time — no central enum.

Five parallel jobs run on every PR (all required to merge):

• Lint — clippy + rustfmt
• Test — 462 unit and integration tests via nextest
• FATE Regression — no previously-passing FATE test may regress (framecrc vs FFmpeg)
• JVT Regression — no previously-passing JVT test may regress (MD5 vs ITU checksums)
• Deny — license allow-list and advisory audit via cargo-deny

Conformance baselines are committed in test_suites/baselines/. To update after expanding coverage:

python3 scripts/suite_runner.py --suite fate-cavlc,fate-cabac \
  --save-snapshot --snapshot-dir test_suites/baselines

Pre-commit hook available: pip install pre-commit && pre-commit install

The FFmpeg/ submodule is pinned to n8.1. Optional — only needed to read the C source or build a debug FFmpeg for development:

git submodule update --init
cd FFmpeg && ./configure --disable-optimizations --enable-debug=3 \
  --disable-stripping --disable-asm && make -j$(nproc) ffmpeg

See CONTRIBUTING.md.

This repo is designed for AI-assisted development. CLAUDE.md contains the full project context: architecture, conventions, debugging procedures, and technical requirements. It is the canonical reference for any AI agent working on this codebase — read it before writing code.

The llms.txt file provides a machine-readable project summary following the llms.txt convention.

LGPL-2.1-or-later. See LICENSE and COPYING.LGPLv2.1.