Fast lossless image codec that compresses better than PNG.
Griffin automatically adapts its compression strategy to the image content — using Huffman entropy coding for photographic images and dictionary compression for text/document images. No manual tuning needed.
All speeds are raw pixel throughput (width x height x channels / time), single core, best-of-3 iterations. Ratio is encoded size / uncompressed size using the original channel count.
Methodology: Single core (Intel Core i7-13700H P-core), sequential execution, pre-allocated buffers and reusable contexts. WebP lossless uses method=0/quality=0 (fastest). JPEG-XL lossless uses effort=3.
- Tecnick — 182 photographic images at 1200x1200 from the Tecnick SAMPLING dataset. The standard benchmark for lossless image codec evaluation.
- DocBank — 200 document page images from the DocBank dataset. Academic papers rendered to PNG.
# Tecnick (182 images, ~370 MB)
curl -L "https://sourceforge.net/projects/testimages/files/SAMPLING_8BIT_RGB_1200x1200.tar.bz2/download" -o tecnick.tar.bz2
mkdir -p images/tecnick
tar xjf tecnick.tar.bz2 -C images/tecnick
rm tecnick.tar.bz2Pre-built binaries are available so the community can independently reproduce and verify these results on their own hardware and datasets.
Single-file executable, no dependencies. Available in the bin/ directory. Supports single-file and batch (directory) mode with per-file CSV reporting.
chmod +x bin/dvid3-x86_64.AppImage
# Single file
./bin/dvid3-x86_64.AppImage encode --in photo.png --out photo.grif
./bin/dvid3-x86_64.AppImage decode --in photo.grif --out photo.png
# Compression levels: 0=AUTO (default), 1=FAST, 2=BEST
./bin/dvid3-x86_64.AppImage encode --level 2 --in photo.png --out photo.grif
# Batch encode a directory (recursive), with per-file CSV report
./bin/dvid3-x86_64.AppImage encode --in images/ --out encoded/ --force --report encode.csv
# Batch decode
./bin/dvid3-x86_64.AppImage decode --in encoded/ --out decoded/ --force --report decode.csvBatch mode measures codec time only (excludes PNG load/save I/O), reuses contexts and pre-allocated buffers, and prints aggregate statistics with avg/median/p5/p95 speeds.
Pre-built static library and header in the bin/ directory:
| Library | Header |
|---|---|
libgriffin.a |
griffin.h |
Pure C API, caller owns all memory:
#include "griffin.h"
uint8_t out[griffin_encode_max_size(width, height)];
int encoded_size = griffin_encode(pixels, width, height, out, sizeof(out));
// With explicit compression level
int encoded_size = griffin_encode_level(pixels, width, height, out, sizeof(out), GRIFFIN_BEST);
// Timed variant excludes caller overhead
double seconds;
int encoded_size = griffin_encode_timed(pixels, w, h, out, sizeof(out), &seconds);Compile and link:
gcc -I bin/ my_benchmark.c bin/libgriffin.a -lstdc++ -lpthread -o my_benchmarkPre-built native module in python/. To use:
pip install numpy pillow # dependencies
PYTHONPATH=python python bench_codecs.py images/tecnick/The module API:
import griffin
import numpy as np
from PIL import Image
img = np.array(Image.open("photo.png").convert("RGBA"))
encoded = griffin.encode(img) # AUTO (default)
encoded = griffin.encode(img, level=2) # BEST
decoded = griffin.decode(encoded)
# Timed variants for benchmarking (measures codec time only)
encoded, enc_seconds = griffin.encode_timed(img)
decoded, dec_seconds = griffin.decode_timed(encoded)Griffin requires AVX2 (Intel Haswell 2013+ / AMD Excavator 2015+). The Python module checks at import time and raises a clear error on unsupported CPUs.
Released for non-commercial evaluation only. No warranty. See LICENSE for details.
This software uses third-party libraries under their respective open-source licenses. See THIRD_PARTY_NOTICES for full license texts.
For commercial licensing: dfaconti@aurynrobotics.com