ZipPlus Model Card

A pre-trained 4-layer GRU model for neural file compression. Each compressed file contains its own adapted model — no external model needed to decompress.

This is a pre-trained ByteGRU model for Zip+.

What is this

Zip+ compresses any file into a PNG image using a neural network (GRU + range coding). Each compressed file embeds its own adapted model:

file.txt → [ByteGRU + Range Coding] → file.txt.zpng.png → [embedded model] → file.txt

Every PNG is self-contained — decompress even if you lose the original model file!

Model Details

• Architecture: 4-layer GRU over byte embeddings
• Embedding dim: 64 → Hidden dim: 512
• Trained on: FineWeb-Edu (10BT of educational web text) + adaptive per-file training
• Entropy coding: Range coding via Constriction
• Output format: PNG where payload + model live in RGB pixel bytes
• Magic header: ZPNG (first 4 bytes)

Requirements

• Python 3.10+
• PyTorch (CUDA recommended)
• Constriction (pip install constriction)
• Pillow
• numpy
• huggingface_hub

pip install torch constriction pillow numpy huggingface_hub

Quick Start

Compress a file (with auto-adaptation)

python inference.py compress myfile.txt -o myfile.zpng.png

• Automatically adapts model to your file (50 steps)
• Embeds adapted model in PNG for self-contained decoding

Decompress

python inference.py decompress myfile.zpng.png -o restored.txt

Loads the model embedded in the PNG — no external files needed!

Training (optional)

python train.py --grid 128 --steps 10000

Auto-downloads FineWeb-Edu if no corpus specified.

Performance

• Text files: ~5-20% of original size
• Works best on files > 10KB
• Smaller files: embedding overhead (~21MB) may exceed compression gains

Warnings

• Embedding adds ~21MB to output — worth it for large files
• GPU recommended for training and compression
• Lossless — verified via SHA256 checksums

License

MIT. I'm not liable if this eats your thesis/pixels/anything.