This is the official codebase for CityCraft: A Diffusion and LLM Powered Framework for Automatic Comprehensive Virtual City Generation.

Development environment:

• OS: Ubuntu 22.04
• CUDA: 11.7
• PyTorch: 2.0.1

Create a new environment and install the requirements:

conda create -n citycraft python=3.10
conda activate citycraft
pip install -r requirements.txt

Step 1: Download osm data from "https://huggingface.co/datasets/GearlessJ333/CityCraft/tree/main" and put it in the folder "CityCraft/dataset/CityCraft_OSM/"

cd dataset
unzip layouts.zip

After this step, you will the the following structure:

CityCraft/
    dataset/
        CityCraft_OSM/
            layouts/
                1.png
                2.png
                3.png
                ...

Step 2: prepare training features

cd dataset/extract_features
python extract_ratio_features.py

This will exract latent image features and ratio features for training. After this step, you will have the following structure:

CityCraft_OSM/
    layouts/
    features/
        image_ratio_features/

Step 1: Download assets data from "https://huggingface.co/datasets/GearlessJ333/CityCraft/tree/main" and put it in the folder "CityCraft/dataset/city_render/"

Step 2: Unzip data and features

cd dataset/CityCraft_Buildings
unzip assets.zip

This will extract the assets data and features. After this step, you will have the following structure:

CityCraft_Buildings/
    assets/
        assets_features/
        assets_filtered/
        assets_renders/

Our layout generation model is a diffusion transformer DiT. Specifically, we used DiT-B/2 and our training script is developed upon FastDiT.

cd layout_generation

accelerate launch --multi_gpu --num_processes 4 --mixed_precision bf16 train.py --model DiT-B/2 --ckpt-every-epoch 10 --global-batch-size 256  --lr 1e-5 

accelerate launch --multi_gpu --num_processes 4 --mixed_precision no train_class_ratio.py --model DiT-B/2 --ckpt-every-epoch 10 --global-batch-size 256 --lr 1e-5

You can adjust all training parameters on your need. You also need to config Accelerator before training.

The training results will be stored in your designated 'results-dir'.

You can login to Weight and Biases to monitor the training process.

Step 1: Download pretrained model from "https://huggingface.co/GearlessJ333/CityCraft/tree/main" and put it in the folder "CityCraft/pretrained_models/"

Step 2: Sample from pretrained models

cd layout_generation
python sample_uncondition.py

Step 3: Post-process the generated layouts

cd layout_generation
python utils/post_process.py

step 1 (optional): put your selected layout images in the folder "urban_planning/layouts/", and modify llm_planning.sh to set planning target and requirements.

step 2: set API key in llm_planning.sh

api_key = 'your_api_key'

You need an API key for OpenAI models, you can get one from here.

step 3: start planning

cd urban_planning
sh llm_planning.sh

After step, the plan will be generated in the folder "urban_planning/planning_result/" at the output directory you set in llm_planning.sh.

Example Plan Output:

{
    "0": {
        "location": [
            23,
            0,
            17,
            15
        ],
        "x": 8.5,
        "y": 7.5,
        "neighbors": [
            1,
            9,
            18,
            21,
            26
        ],
        "description": "A mid-rise residential building with children's playground ",
        "function_primary": " residential ",
        "function_secondary": " children's playground ",
        "scale": " mid-rise ",
        "style": " modern ",
        "num_floor": 5,
        "reason": " Positioned away from traffic roads and complements nearby residential areas while fulfilling user requirements for housing and recreational spaces.",
        "distance_to_roads": [
            9.848857801796104
        ]
    },
    "1": {
        "location": [
            23,
            40,
            20,
            50
        ],
        "x": 25.0,
        "y": 10.0,
        "neighbors": [
            0,
            9,
            3,
            18,
            21
        ],
        "description": "A three-story residential building with modern amenities ",
        "function_primary": " residential ",
        "function_secondary": " community space ",
        "scale": " mid-rise ",
        "style": " modern ",
        "num_floor": 3,
        "reason": " Satisfies user requirements for residential buildings, maintains a safe distance from traffic roads, and reflects the desired architectural style.",
        "distance_to_roads": [
            1.0
        ],
        "required_type": "residential buildings"
    },

    ...

step 4: select assets

In "select_buildings.sh", set the path to the planning result you want to select assets from, then run:

sh select_buildings.sh

After this step, .pkl file containing the selected assets will be generated in the folder, for example: CityCraft/urban_planning/planning_result/layout_1/selected_buildings.pkl.

step 1: Download Blender Installer

Go back to the root directory CityCraft/:

mkdir Blender

Download Blender from here, please select the version that matches your operating system, and put it in the folder "CityCraft/Blender/".

(For example, we used here to install Blender 4.0.0 for our Linux x86_64 system; You can put it anywhere but just remember to set the path later when you run the rendering script.)

step 2: Install Blender

cd Blender
tar -xJvf blender-4.0.2-linux-x64.tar.xz

step 3: Download extra assets and textures

Download extra assets and textures from "https://huggingface.co/GearlessJ333/CityCraft/tree/main" and put it in the folder "CityCraft/render_scene/"

cd render_scene
unzip extra_assets.zip
unzip textures.zip

After this step, you will have the following structure:

render_scene/
    extra_assets/
    textures/

step 4: Start Rendering

Note: You need to set the path to the layout image and the selected assets pkl file in the render_scene.sh file. You also need to set the path to the Blender executable file to where you installed it.

cd render_scene
sh render_scene.sh

After this step, you will have the rendered scenes in the folder "render_scene/rendered_scenes/".

Sample Scenes:

If you feel our work is helpful, please consider citing:

@article{deng2024citycraft,
  title={CityCraft: A Real Crafter for 3D City Generation},
  author={Deng, Jie and Chai, Wenhao and Huang, Junsheng and Zhao, Zhonghan and Huang, Qixuan and Gao, Mingyan and Guo, Jianshu and Hao, Shengyu and Hu, Wenhao and Hwang, Jenq-Neng and others},
  journal={arXiv preprint arXiv:2406.04983},
  year={2024}
}