gaussian-painters

Gaussian Painters

Sponsored by LingoSub: Learn languages by watching videos with AI-powered translations.

and ThumbnailsPro: Instant AI-generated Thumbnails, for videos that get clicks.

This is a fork of 3D Gaussian Splatting. Refer to the original repo for instructions on how to run the code.

How to create a Gaussian Painter dataset

After having installed the 3D Gaussian Splatting code, run the following command:

python create_dataset.py --img_path /path/to/image --output_dir /path/to/output_dir

You can disable the opacity_reset_interval argument by setting it to 30_000.

You can also set sh_degree to 0 to disable viewdependent effects.

This will create a dataset ready to be trained with the Gaussian Splatting code.

Experiments

• Orthogonal images (using create_dataset2.py)

https://github.com/ReshotAI/gaussian-painters/assets/16474636/4799f0b6-ed29-412e-9875-4a790ecbbaaf

• Steganography (using create_dataset3.py)

https://github.com/ReshotAI/gaussian-painters/assets/16474636/9a391361-7d5b-40cc-ab67-97e15e53a913

• Lenticular effect (using create_dataset5.py)

This code requires to install kornia using pip install kornia

https://github.com/ReshotAI/gaussian-painters/assets/16474636/356ad0f6-3bcb-46fe-a6f8-421138e54222

Visualize the "painting" process

Using the SIBR visualizer, you can visualize the "painting" process during the Gaussian Splatting optimization.

https://github.com/ReshotAI/gaussian-painters/assets/16474636/b29731b6-5fcc-43f5-a169-bfed2b109ce0

How it works?

The create_dataset script simply creates a COLMAP output directory with a single camera pointing at a plane. 100 points are sampled from the image and used as initial point cloud for the Gaussian Splatting optimization. A second perpendicular image is also created with a black image as target.