BadAppleStein

Reconstructs any video as a mosaic of pages from a source library of PDFs and images. A pure-C pipeline with SIMD-accelerated matching and hardware encoding support.

Installation

Pre-built binaries

curl -fsSL https://raw.githubusercontent.com/TheFirstIstari/BadApplestein/master/install.sh | sh

Build from source

Requires: ffmpeg (libavformat, libavcodec, libavutil, libswscale), optionally mupdf for PDF support, and libomp for parallel matching.

git clone https://github.com/TheFirstIstari/BadApplestein.git
cd BadApplestein
make              # builds src/badapplestein
make install      # installs to /usr/local/bin (override with PREFIX=)

Usage

# Build a source library from images/PDFs
badapplestein build ~/Documents/pdfs/

# Encode a video using the library
badapplestein encode input.mp4 output.mov

# Shorthand (auto-detects encode mode)
badapplestein input.mp4 output.mov

# Use a preset
badapplestein input.mp4 output.mov --preset 4k

Subcommands

Command	Description
badapplestein build <sources_dir>	Build a source library from images/PDFs
badapplestein encode <input> <output>	Encode a video using the library

Options

Flag	Description
--library <dir>	Library directory (default: ./ or ~/.badapplestein/library/)
--preset <name>	Output preset: 8k, 4k, 1080p, 720p
--width <n>	Output width (overrides preset)
--height <n>	Output height
--fps <n>	Output fps
--codec <name>	Output codec (default: auto-detect best available)
--no-hw	Disable hardware encoding
--max-frames <n>	Process only N frames (0 = all)
--threads <n>	Thread count (0 = auto)
--keep-manifests	Keep temp manifests after encoding
--verbose, -v	Verbose output
--json	JSON output mode

Pipeline

input.mp4 → arrange → manifests/*.bin → render → output.mov
              ↓            ↑
         source lib    (pdf_path, page_idx) per tile

Building

make              # Build unified binary
make install      # Install to $PREFIX/bin (default: /usr/local)
make clean        # Remove build artifacts

Development (legacy standalone binaries)

make arrange      # Build arrange-only binary
make render       # Build render-only binary
make build-library

Source Library

The arrange stage matches against a pre-built feature library:

• features.bin — Quantized feature vectors (multi-resolution, 1-8 bits/cell, grayscale or color)
• registry.bin — Maps library entries to source file paths and page indices

Generate a test library:

python3 gen_test_lib.py --pages 2000 --seed 42 --output test_lib

Performance

BadAppleStein outputs JSON progress that can be captured for benchmarking:

# Run with JSON output
./src/badapplestein encode input.mp4 output.mov --json 2> progress.json

# Or use hyperfine for timing
mise run benchmark-arrange
mise run benchmark-render

Performance characteristics (v1.0.0):

• Arrange: ~1-5 fps depending on resolution and match quality threshold
• Render: ~10-30 fps with atlas cache, hardware encoding
• Cache hit rate: 60-90% for typical videos (reduces source image loads)

The --max-frames option is useful for quick performance testing.

Note: The unified binary handles both arrange and render stages automatically in a single encode command; standalone arrange and render binaries are available for development and debugging.

Key Files

File	Description
src/main.c	Unified CLI entry point
src/arrange.c	Video decode + greedy block solver + feature extraction + matching
src/render.c	Manifest loading + source page rendering + frame assembly
src/match.c	L1 feature matcher (SIMD: AVX2/SSE2/NEON, OpenMP-parallel)
src/video.c	FFmpeg libav decode/encode + image loader
src/imgops.c	Grayscale conversion, resize, feature extraction
src/pdf.c	mupdf PDF rasterization + libav image fallback
src/cli.c	CLI framework (option parsing, progress, JSON output)
src/build_library.c	Build a source library from images/PDFs
src/vt_prores.m	macOS VideoToolbox ProRes hardware encoding

License

MIT — see LICENSE.