[Paper] [Project Page] [Full results] [Poster]
This repository contains the source code used in experimental evaluations for the paper "NIF: A Fast Implicit Image Compression with Bottleneck Layers and Modulated Sinusoidal Activations".
04/2025: Updated with reproducibility improvements and novel environment setup
The Python environment can be handled using uv. Interpreter and requirements will be automatically installed when necessary.
To encode a single image run the encode.py scripts:
uv run -m encode.py <CONFIGURATION_PATH> <INPUT_IMAGE> <COMPRESSED_IMAGE>
For example, to encode the image #3 from the Kodak dataset located at "test_images/kodak/3.png" and save it at "kodim03.nif" run:
uv run -m encode configurations/nif/kodak/120.yaml test_images/kodak/3.png kodim03.nif
Analogously, you can decode a compressed image using the decode.py script:
uv run -m decode <CONFIGURATION_PATH> <COMPRESSED_IMAGE> <DECOMPRESSED_IMAGE>
To decompress the image encoded in the example above, use:
uv run -m decode configurations/nif/kodak/120.yaml kodim03.nif kodim03_decoded.png
Input files should be encoded in lossless PNG format. Output file names end with ".nif" by convention, but this is not forced by code.
To reproduce the experiments discussed in the paper, run the corresponding script in the folder "scripts/datasets/". These scripts will generate a bash file named "schedule.sh", filled with the sequence of commands that will encode a whole dataset. Logs will be saved under "logs/" and compressed files, along decompressed images and statistics will be saved under "results/".
For instance, to generate a "schedule.sh" to reproduce experiments on Kodak use:
./scripts/datasets/kodak.sh
Then execute it with ./schedule.sh.
If you want to perform a full encode-decode experiment on a single image and to export stats, on Linux/MacOS systems you can use the "experiment.sh" script as follows:
uv run bash experiment.sh <CONFIGURATION_PATH> <INPUT_IMAGE> <OUTPUT_FOLDER>
The compressed and then decoded images, along with stats saved in JSON, will be exported in the indicated output folder.
The kodak dataset is included in this repository, along with compressed ".nif" files that can be decoded using this software.
The sample from CelebA is the same used by Strumpler et al [2022]. To download it, follow the instructions at https://github.com/YannickStruempler/inr_based_compression/#datasets. Alternatively, after downloading the .zip archives from the official Google Drive, run the following bash script:
./scripts/datasets/celeba.sh <CELEBA_DOWNLOAD_FOLDER>
Where CELEBA_DOWNLOAD_FOLDER contains the .zip archives.
The ICB dataset can be downloaded at http://imagecompression.info/test_images/rgb8bit.zip. It is necessary to convert it to PNG, for this conversion we have used ImageMagick:
magick <input_file>.ppm <output_file>.png
A bash script to automate the process is provided in ./scripts/datasets/icb.sh