Neural Cellular Automata for Diverse Tree Growing

Our Project aims to enhance Minecraft landscapes with AI-generated 3D trees. Utilizing neural cellular automata, each tree evolves independently, creating diverse and realistic forests. Instead of manually designing a number of tree templates to use again and again, the Automata creates random trees that can repair themselves/regrow naturally.
The page is heavily inspired by Growing 3D Artefacts and Functional Machines with Neural Cellular Automata.

---

Requirements

• python3.8

Installation

For general installation

Depending on your setup, you might have to run the setup.py script after every change in code (because of hydra!).

If you want CUDA, first run this line otherwise skip

pip install torch==2.1.0 --index-url https://download.pytorch.org/whl/cu118

Run following to setup the project

pip install -r requirements.txt
python setup.py install

Usage

Configs

Each nca is trained on a specific structure w/ hyperparams and configurations defined in yaml config, which we use with hydra to create the NCA trainer class.

Example Config :

trainer:
    name: AcaciaTrees_8tree
    min_steps: 48
    max_steps: 65
    visualize_output: true
    device_id: 0
    use_cuda: true
    wandb: true
    checkpoint_interval: 200
    epochs: 3001
    batch_size: 8
    clip_gradients: true
    num_hidden_channels: 12
    embedding_dim: 4
    var_lr : .002
    var_loss_weight: 0.1
    random: false
    variational: true
    use_index: false
    model_config:
        normal_std: 0.1
        update_net_channel_dims: [64, 32, 32, 16]
    optimizer_config:
        lr: 0.002
    dataset_config:
        load_embeddings: false
        pool_size: 32
        spawn_at_bottom: false
        verbose: false
        sample_specific_pool: true
        equal_sized_samples: false
        load_entity_config: {'verbose': false}

defaults:
    - voxel

Additionally introduced parameters:

• embedding_dim - dimension of the embedding channel given to the input
• load_embeddings - true if we load embeddings for the prestored CSV file
• embedding_path - for pretrained embeddings loading
• clip_gradients - true if gradients clipped
• random - true if embeddings are sampled from the standard Gaussian
• variational - true if variational autoencoder is used
• var_lr - learning rate of the variational autoencoder
• var_loss_weight - weight of the variational loss in the total loss
• sample_specific_pool - true if we use specific sampling pool as starting points
• equal_sized_samples - true if all loaded sample trees have equal sizes
• use_index - embeddings dimension is filled with normalized id of the tree

Generation and Training

See training Python file for ways to train your models and inference Python file to produce inference results using pretrained models.

Paper

The paper can be found here.

Authors

Cedric Caspar ccaspar@student.ethz.ch, https://github.com/CedricCaspar

Nicolas Blumer nblume@student.ethz.ch, https://github.com/Thahit

Oleh Kuzyk okuzyk@student.ethz.ch, https://github.com/Olezhko2001

Piyushi Goyal pgoyal@student.ethz.ch, https://github.com/piyushigoyal