Towards Lossless Implicit Neural Representation via Bit Plane Decomposition

This repository contains the official implementation of the following paper:

Towards Lossless Implicit Neural Representation via Bit Plane Decomposition (CVPR 2025) (Accepted)

 

 

Overall Structure

Installation

Our code is based on Ubuntu 20.04, pytorch, CUDA (NVIDIA RTX 3090 24GB, sm86) and python.

For enviorment setup, we recommend to use CONDA for the installation:

conda env create --file env.yaml
conda activate linr

We include 8-bit and 16-bit image fitting training code and 1.58-bit weight INR of our main paper.

Experiment

The basic train code are following:

Lossless Image Fitting :

python train.py --img_path YOUR_IMAGE_PATH/FANCY_IMAGE.png --num_bits bit_depth --save --img_size 256

Image Fitting with 1.58-bit INR :

python train_ternary.py --img_path YOUR_IMAGE_PATH/FANCY_IMAGE.png --num_bits bit_depth --save --img_size 128

Command-line Arguments

• --img_path: (Required) Path to the input image.
• --num_bits: Bit depth for decomposition (choose 8 or 16).
• --save: Flag to save output images during training.
• --img_size: Size (width/height) to which the input image will be resized (default is 256).

Additional optional arguments:

• --gpu: GPU id(s) to use (e.g., "0" or "0,1").
• --tag: Experiment tag for organizing logs and saved models.
• --flag: Additional custom flag for directory naming.
• --total_steps: Total number of training iterations (default: 10000).
• --lr: Learning rate for the optimizer (default: 1e-4).

They will prodice save file, Tensorflow Summary, text record of iteration.

Note that PIL do not support 16-bit image, so opencv (cv2) is used for saving and loading 16-bit image.

Acknowledgements

Our backbone models of the code are built on SIREN and NeRF. We thank the authors for sharing their codes.