qrstream 0.8.3

Encode and decode files via QR code video streams using LT fountain codes

QRStream

中文文档

Transfer arbitrary files through QR code video streams. Built on LT Fountain Codes (Luby Transform) for reliable, feedback-free data transmission — the original file can be fully recovered even if some frames are lost.

How It Works

Encoder                                     Decoder
┌──────────┐   LT Fountain    ┌──────────┐   Screen cap   ┌──────────┐   QR detect    ┌──────────┐
│   File    │ ────────────── → │ QR Video │ ──────────── → │  Video   │ ────────────→ │ Recovered│
└──────────┘   zlib + base45  └──────────┘                └──────────┘   LT decode    │   File   │
                                                                                       └──────────┘

1. Encode: Split the file (optionally zlib-compressed) into blocks, generate redundant coded blocks via LT fountain codes, serialize each into a V3 protocol frame, base45-encode into QR alphanumeric-mode symbols, and output an MP4 video.
2. Decode: Extract QR codes from video using WeChatQRCode (highly robust), base45-decode, CRC32-validate to discard corrupted frames, feed into the LT decoder for belief propagation (peeling), and reconstruct the original file. Legacy base64 and COBS videos (pre-v0.6) continue to decode via a fallback path.

Key Features:

• LT Fountain Codes: Rateless erasure codes — naturally tolerant of frame loss, blur, and occlusion
• Base45 + QR Alphanumeric Mode: RFC 9285 base45 packs data into QR alphanumeric mode (5.5 bits/char vs 8 for byte mode); smaller and faster than base64 at the same QR version
• WeChatQRCode Detector: Far more robust than standard QR detectors for phone-captured screens (perspective, moire, lighting)
• Adaptive Sample Rate: Automatically selects optimal sampling strategy based on detection rate and frame repetition
• Targeted Recovery: After initial scan, precisely re-scans video segments where missing blocks are expected
• Low-Memory Paths: mmap-backed encoding and streaming decode-to-file for large inputs

Installation

From PyPI with pip

pip install qrstream

Use either command after installation:

qrstream <command> [options]
# or
qrs <command> [options]

You can also run it as a module:

python -m qrstream <command> [options]

From PyPI with uv

uv tool install qrstream

Then run:

qrstream <command> [options]

For one-off execution without a persistent install:

uvx qrstream <command> [options]

Development Install

git clone https://github.com/ddddavid-he/qrstream-enhanced.git && cd qrstream-enhanced
uv sync --dev

Requirements

• Python >= 3.10 (3.10 – 3.14 tested)
• Dependencies: opencv-contrib-python, numpy, rich, segno

Usage

qrstream <command> [options]
qrstream -V
qrstream --version

qrs is kept as a short alias, and python -m qrstream works as well.

Encode (File → QR Video)

qrstream encode <file> -o output.mp4 [options]

Option	Default	Description
<file>	-	Input file path
-o, --output	required	Output video path
--overhead	2.0	Encoding redundancy ratio (multiple of source block count)
--fps	10	Output video frame rate
--ec-level	1	QR error correction: 0=L(7%), 1=M(15%), 2=Q(25%), 3=H(30%)
--qr-version	25	QR code version 1-40 (higher = denser)
--border	standard 4-module quiet zone	Quiet-zone width as a percentage of QR content width (--border 10 = 10%, --border 0 disables it)
--lead-in-seconds	0.0	Insert white lead-in frames before the first QR frame
--no-compress	-	Disable zlib compression
--force-compress	-	Force compression for large V3 inputs (higher memory usage)
--qr-mode	alphanumeric	QR payload encoding: alphanumeric (base45, default, denser) or base64 (byte mode, fallback)
--legacy-qr	-	Accepted but ignored (kept for CLI backward compatibility)
--codec	mp4v	Video codec: mp4v or mjpeg (faster but larger files)
-w, --workers	min(CPU count, 4)	Parallel worker threads for QR generation. The auto-picked default is capped at 4 because QR matrix generation (segno.make()) is pure-Python (GIL-bound), so more than ~4 worker threads mostly contend on the GIL without adding real parallelism. Pass a larger value explicitly to override the cap on CPU-rich machines if profiling shows benefit.
-v, --verbose	-	Print extra detail (progress bars always shown)

Decode (QR Video → File)

qrstream decode <video> -o output_file [options]

Option	Default	Description
<video>	-	Input video path (MP4, MOV, etc.)
-o, --output	required	Output file path
-s, --sample-rate	0 (auto)	Sample every Nth frame (0 = adaptive probing)
-w, --workers	All CPU cores	Parallel worker threads for QR detection. WeChatQRCode is implemented in C++ and releases the GIL during detection, so more threads scale close to linearly on multi-core machines.
-v, --verbose	-	Print detailed progress

Examples

# Encode a PDF (default: base45 alphanumeric mode, 2x redundancy)
qrstream encode report.pdf -o report.mp4 --overhead 2.0 -v

# Decode video (adaptive sample rate + targeted recovery)
qrstream decode report.mp4 -o report_recovered.pdf -v

# Encode with high error correction (for phone screen capture)
qrstream encode data.bin -o data.mp4 --ec-level 3 --qr-version 15

# Add a larger quiet zone and white lead-in before recording
qrstream encode slides.zip -o slides.mp4 --border 10 --lead-in-seconds 1.5

Python API

from qrstream.encoder import encode_to_video
from qrstream.decoder import extract_qr_from_video, decode_blocks, decode_blocks_to_file

# Encode (default: base45 alphanumeric mode)
encode_to_video("input.bin", "output.mp4", overhead=2.0, verbose=True)

# Add recording-friendly quiet zone and white lead-in
encode_to_video("input.bin", "output.mp4", border=10.0, lead_in_seconds=1.5)

# Decode to memory
blocks = extract_qr_from_video("output.mp4", verbose=True)
result = decode_blocks(blocks, verbose=True)

# Better for large files: stream directly to file with incremental decompression
written = decode_blocks_to_file(blocks, "recovered.bin", verbose=True)
print(f"wrote {written} bytes")

Project Structure

project-root/
├── pyproject.toml             # Project config & dependencies
├── src/qrstream/
│   ├── cli.py                 # CLI entry (encode/decode subcommands)
│   ├── encoder.py             # LT encode → QR frame generation → MP4 video
│   ├── decoder.py             # Video frame extraction → QR detect → LT decode → file rebuild
│   ├── lt_codec.py            # LT fountain code primitives (PRNG, RSD, BlockGraph)
│   ├── protocol.py            # V3 protocol serialization + base45 codec (legacy base64/COBS decode supported)
│   └── qr_utils.py            # QR generation (segno) + detection (WeChatQRCode)
├── tests/
│   ├── test_lt_codec.py       # LT codec unit tests
│   ├── test_protocol.py       # V3 protocol + base45 tests
│   ├── test_decoder.py        # Decoder validation + probe strategy tests
│   ├── test_roundtrip.py      # Pure LT codec roundtrip tests (no video I/O)
│   ├── test_qr_generate.py    # QR generation correctness + glog(0) regression
│   ├── test_e2e_encode_decode.py  # End-to-end encode→video→decode SHA256 tests
│   └── test_optimizations.py  # Perf optimizations + WeChatQR + legacy-fallback tests
└── dev/
    ├── benchmark.py           # Performance benchmarks
    ├── perf-profile/          # cProfile hotspot analysis scripts
    ├── test-container/        # Podman test container
    └── wechatqrcode-mnn-poc/  # WeChatQRCode MNN acceleration POC

Technical Details

V3 Protocol Format (24-byte header + 4-byte trailing CRC)

Offset  Size  Field
  0      1    version      0x03
  1      1    flags        bit0=zlib compressed, bit1=high-density mode (base45 alphanumeric)
  2      8    filesize     uint64 BE (encoded payload size; compressed size when zlib is on)
 10      2    blocksize    uint16 BE
 12      4    block_count  uint32 BE  K = ceil(filesize / blocksize)
 16      4    seed         uint32 BE  PRNG seed
 20      2    block_seq    uint16 BE  monotonically increasing sequence number
 22      2    reserved     reserved (currently 0)
 24      ...  data         blocksize bytes of encoded data
 ...     4    crc32        CRC32(header[0:24] + data)

• Default encoding uses V3 + base45 alphanumeric QR.
• The decoder tries base45 → base64 → COBS in order, preserving compatibility with pre-v0.6 videos.
• V3 extends filesize to uint64 and block_count to uint32, supporting larger files and block counts.

Encoding Modes

Mode	QR Content	QR Mode	Overhead	Default
Base45 alphanumeric	raw bytes → base45 → 0-9A-Z $%*+-./:	Alphanumeric (5.5 bits/char)	~67% (but uses denser QR mode → net denser than byte mode)	Yes
Base64	raw bytes → base64 string	Byte (8 bits/char)	~33%	No (--qr-mode base64)
COBS (legacy)	raw bytes → COBS → latin-1 string	Byte	~0.4%	Removed in v0.6; decode-only fallback for old videos

Base45 (RFC 9285) is the default because QR's alphanumeric mode is denser per character than byte mode — at V25/M the base45 payload per frame is ~30% larger than base64, and in practice produces 20–25% smaller videos and 10–20% faster encode/decode.

Large Files & Low-Memory Paths

• For large V3 inputs, the encoder uses mmap for random access, avoiding loading the entire file into memory.
• When the input is large enough, V3 encoding automatically disables zlib compression to preserve the low-memory path; use --force-compress to override.
• The decoder supports streaming writes with incremental decompression, reducing memory overhead.
• Large file decoding shows LT block decoding progress and output write progress bars.

Decoding Pipeline

1. Probe phase: Sample 3 spread-out windows in the video (120 frames each by default), measure detection rate and repetition per window, pick the most conservative sample_rate
2. Main scan: Detect QR codes in parallel at the adaptive sample rate, feeding into the LT decoder in real time
3. Targeted recovery: If the first pass didn't recover all blocks, use linear regression on observed (seed, frame) pairs to locate missing seeds and re-scan those segments precisely
4. LT decode: Belief propagation (peeling) to recover all source blocks
5. Output writeback: Write recovered blocks sequentially; incremental decompression in compressed mode

LT Fountain Code Parameters

Parameter	Value	Notes
Degree distribution	Robust Soliton Distribution	c=0.1, delta=0.5
PRNG	SplitMix64 mixer + LCG (a=16807, m=2^31-1)	Non-linear seed mixing eliminates sequential-seed correlation
XOR	numpy vectorized + in-place	10-50x faster than pure Python
Decoding	Belief Propagation (Peeling)	Iterative elimination on bipartite graph

Testing

# Unit tests (default — fast, no video I/O)
uv run pytest tests/ -v

# End-to-end encode→video→decode tests (10 KB, 100 KB, 500 KB + glog regression)
uv run pytest -m e2e -v

# Real-world phone-recording tests (requires fixture videos)
uv run pytest -m slow -v

Decoder native crashes

If qrs decode exits with trace trap or a SIGSEGV/SIGTRAP message, you are hitting a known unfixed crash in the WeChat QR detector bundled with opencv_contrib (upstream issue opencv_contrib#3570). Since v0.7.7, qrs decode runs detection in subprocess helpers by default, so a single crashing frame is caught and treated as a dropped frame — the decode continues and completes as long as LT overhead (see --overhead) is ≥ 1.5.

To see whether the subprocess sandbox caught any crashes, look for a summary line like [sandbox] detector crashed N time(s) during decode. If you believe the sandbox overhead is unnecessary on your input, you can disable it at your own risk with --detect-isolation off.

License

MIT