OpenViSTAI · Agents · Blockchain · Crypto
Submit a project
Menu
AI Models

Generative Compression

Generative Compression is an MIT-licensed Python research implementation of learned image compression with generative adversarial networks, based on Agustsson and colleagues' extreme learned image compression method. It includes scripts for training and compressing a single image with a checkpoint, plus an experimental conditional GAN variant that uses semantic label maps.

Explore the practical guide ↓

Independent OpenViST project profile, based on linked sources.

Generative Compression — OpenViST project profile in AI Models.
OpenViST project illustration. See the linked sources for product details.
Understand the project

Practical overview

When it fits

Use it when you want to inspect or experiment with the documented GAN-based compression method in Python, and can work with its legacy TensorFlow environment and dataset configuration.

Who it is for

  • Computer-vision researchers investigating learned image compression.
  • Developers prototyping GAN-based compression workflows or testing the method on image datasets.

Documented uses

  • Train a learned image-compression model from image paths organized in an HDF5-backed Pandas dataframe.
  • Compress a single image with a model checkpoint and compare the reconstruction with the original.
  • Explore conditional reconstruction using semantic label maps.
Before you choose

Requirements & limits

The repository is public and identifies its license as MIT. It is code intended to run in a Python environment, not a hosted compression service.

  • The README lists Python 3.6, Pandas and TensorFlow 1.8 as dependencies.
  • Training requires a Pandas dataframe saved as HDF5, with image paths and a configured dataset location.
  • The conditional GAN variant needs semantic-label-map paths and is explicitly described as experimental.
  • The README reports hallucinated content in some reconstructions, including greenery in buildings.
  • The README says pretrained models trained with TensorFlow 1.3 appeared to load with TensorFlow 1.8, but does not guarantee compatibility.

Check current requirements in the official documentation. Source-based guidance does not mean OpenViST has run or independently tested this project.

Source-based practical guide

Compress one image with a trained model

Run the documented single-image compression command and save a side-by-side reconstruction comparison.

Before you start

  • Python 3.6, Pandas and TensorFlow 1.8, as listed by the README.
  • A model checkpoint and input image.
  • For the supplied noise-sampling model, retain the hyperparameters under config_test in config.py; otherwise match test-time parameters to training.

Steps

  1. Clone the repository and change into its directory.

  2. Run python3 compress.py -r /path/to/model/checkpoint -i /path/to/image -o path/to/output/image, substituting the checkpoint and image paths.

  3. Find the side-by-side comparison under the path specified by directories.samples in config.py.

Expected outcome

A side-by-side comparison of the input and its reconstruction is written in the configured samples directory.

Guide limitations

  • The README's command includes an output path argument but says the comparison is saved under directories.samples in config.py.
  • The README notes that reconstructions may hallucinate content.
  • The listed environment is TensorFlow 1.8; pretrained models were trained with TensorFlow 1.3, and loading is not guaranteed.
Documentation 1 ↗
Source-based practical guide

Prepare image paths and start model training

Set up the documented image-path dataset configuration and invoke the training script.

Before you start

  • Python 3.6, Pandas and TensorFlow 1.8.
  • Training images and a Pandas dataframe with a path column, saved as HDF5.
  • A configured HDF5 path under the directories class in config.py.

Steps

  1. Prepare image paths in a Pandas dataframe column named path and save the dataframe as an HDF5 file.

  2. Set the HDF5 file path under the directories class in config.py.

  3. Clone the repository and change into its directory.

  4. Inspect available training arguments with python3 train.py -h, then run the documented example: python3 train.py -opt momentum --name my_network.

Expected outcome

The training process is invoked; the README says reconstructed samples and TensorBoard summaries are written periodically and checkpoints are saved every 10 epochs.

Guide limitations

  • The README documents ADE20K and Cityscapes preparation examples, with resizing requirements, rather than a universal dataset workflow.
  • The documented example does not specify all training arguments or a guaranteed successful run.
Documentation 1 ↗
Evidence and transparency

Sources & review

Official repository ↗
Source record details
Source availability
Active
Sources checked
17 October 2026
Original website title
Justin-Tan/generative-compression · Official homepage title · retained in its source language
How we review sourcesSuggest a correction

Related options

High-Fidelity Generative Image Compression — The README points to this separate implementation as an option for higher-bitrate compression and says it supports arbitrary image sizes and resolutions.