Practical overview
When it fits
Use it for research or experimentation where compact image representations and perceptual reconstruction are acceptable, and where you can run a PyTorch codebase locally.
Who it is for
- Researchers and developers exploring learned image compression
- Practitioners who want to train or use HIFIC checkpoints for image compression experiments
Documented uses
- Compress image directories using a trained model checkpoint
- Reconstruct images from the model's compressed representation
- Train a base model and warm-start the GAN-based compression stage
- Use lossless entropy coding utilities for compressed latent representations
Requirements & limits
The repository is public and licensed under Apache-2.0. It requires a local Python and PyTorch environment.
- The compression workflow requires a model checkpoint.
- Training calls for a diverse dataset with more than 100,000 images and a two-stage workflow.
- The model is designed for realistic, perceptual reconstructions, not exact recovery; sensitive images and important documents are discouraged.
- Image dimensions are described as flexible, but processing remains subject to memory constraints.
Check current requirements in the official documentation. Source-based guidance does not mean OpenViST has run or independently tested this project.
Compress images with a trained HIFIC model
Encode images with a model checkpoint and request reconstructed output.
Before you start
- Python environment with PyTorch nightly and the repository's listed dependencies
- Repository requirements installed with pip
- A trained model checkpoint and an image directory
Steps
Clone the repository and install the requirements listed in requirements.txt.
Run compress.py with the image directory, checkpoint path, and --reconstruct option.
Expected outcome
The script saves a binary compressed representation and performs reconstruction; the resulting format can be decoded with routines in compress.py.
Guide limitations
- Processing is subject to available memory.
- Reconstructions are perceptual rather than exact; the authors caution against sensitive image contents.
Train a HIFIC model from an image dataset
Train a base compression model and use its checkpoint to warm-start the GAN model.
Before you start
- PyTorch nightly and repository dependencies
- A diverse color-image dataset of more than 100,000 images
- Dataset path configured in DatasetPaths in default_config.py
Steps
Configure the dataset path in default_config.py.
Train the base model with the compression model type and a selected regime.
Use the base checkpoint to warm-start training with the compression_gan model type.
Expected outcome
The documented workflow produces a trained GAN-stage model checkpoint for use with the repository's compression routine.
Guide limitations
- The README points to the user guide for fuller training instructions.
- Training details and outcomes depend on the dataset and chosen configuration.
Sources & review
Source record details
- Source availability
- Active
- Sources checked
- 16 October 2026
- Original website title
- Justin-Tan/high-fidelity-generative-compression · Official homepage title · retained in its source language
Related options
TensorFlow Compression HIFIC implementation — The README identifies this as the official TensorFlow code release for HIFIC; this repository provides a PyTorch implementation.