GaussianSplatting2D

A Python library for image approximation via adaptive 2D Gaussian splatting. Iteratively optimizes a set of splats to reconstruct a target image with minimal memory footprint and allows quantization for shader use.

• 500 Splats after 1000 Epochs

Features

• Adaptive sample splitting and scaling based on gradient magnitude
• Quantized export for ShaderToy as uvec4 arrays
• Checkpointing and multi-bit quantization of splat parameters

Example Shader

https://www.shadertoy.com/view/Wc33W4

Installation

Install PyTorch

pip install -r requirements.txt

Usage

1. Place your target image in the project root.

2. Adjust parameters in the __main__ section of main.py if needed (epochs, sample count, thresholds).

3. Run training:

   python main.py

4. Interrupt with Ctrl+C to save the latest checkpoint.

5. Automatically copies splat parameters to clipboard, for use in ShaderGraph

Configuration Parameters

• num_samples – Initial splat count
• num_max_samples – Maximum allowed splats
• sigma_thre – Minimum standard deviation for splitting
• grad_thre – Gradient threshold for adaptive refinement
• num_epoch – Total training epochs
• num_iter_per_epoch – Iterations per epoch

Output

• Checkpoints saved under training/<image_basename>/<image_basename>.pt
• Quantized variants:
  • 16-bit-quantized-<name>.pt
  • 8-bit-quantized-<name>.pt
• Reconstruction and target side-by-side images in training/<name>/images/
• Clipboard export of uvec4 arrays for ShaderToy

License

This project is released under the MIT License.