vaepack 0.1.1

VAE-based image compression using latent space quantization and JPEG XL

vaepack

VAE-based image compression using latent space quantization and JPEG XL.

vaepack encodes images into the latent space of a Variational Autoencoder (SDXL), quantizes the latents to configurable bit-depth, and compresses the result with JPEG XL inside a compact .zvae container.

Alpha software. This project is experimental and not intended for production use. APIs and behavior may change without notice.

Installation

Recommended

Install the package with VAE support:

pip install "vaepack[diffusers]"

Minimal

Install the core package only:

pip install vaepack

This is enough for container inspection and non-VAE utilities, but encoding and decoding with the SDXL VAE requires the diffusers extra.

Note: if you need GPU acceleration, install the appropriate PyTorch build first. See pytorch.org.

From source

git clone https://github.com/DVA305-VT26-Grupp4/Projektarbete.git
cd Projektarbete
pip install -e ".[diffusers]"

Quick start

Python API

from vaepack import compress, decompress

state = compress("photo.png", "photo.zvae", quant_bits=10, metrics=True)
print(state.metrics)

state = decompress("photo.zvae", "recon.png")

Command line

zipvae compress photo.png -o photo.zvae --quant-bits 10 --metrics
zipvae decompress photo.zvae -o recon.png --reference photo.png --metrics

What the package provides

• High-level compress() and decompress() helpers for notebooks and scripts
• A stage-based pipeline API for step-by-step experimentation
• A .zvae container format for compressed latents
• JPEG XL latent compression via imagecodecs
• PSNR and SSIM metrics
• A zipvae command-line interface

Current limitations

• The project is still alpha quality.
• The packaged VAE backend is currently sdxl.
• The packaged codec backend is currently jxl.
• The first VAE load downloads weights from Hugging Face.
• Kodak benchmark assets and notebooks live in the source repository, not in the PyPI package.

Examples and benchmark assets

The repository contains the batch benchmark script and the Colab notebook used for Kodak experiments:

• Repository
• Kodak batch script
• Colab notebook

For the full source, module-level docstrings, and issue tracker, use the project links on PyPI or visit the repository directly.

License

MIT