Frankenstein: Generating Semantic-Compositional 3D Scenes in One Tri-Plane

SIGGRAPH Asia 2024 (Conference Track)

Project page | arXiv | Video

News

• [2024.11] Inference code released.

Installation

conda create -n frankenstein python=3.9 -y
conda activate frankenstein
pip install torch==2.0.1 torchvision==0.15.2
pip install -r requirements.txt

We slightly modify the diffusers:

• add "num_layers_mid" to UNet2DModel in diffusers/models/unet_2d.py
• add pipeline_ddpm_triplane.py in diffusers/pipelines/ddpm/

Pretrained Model

For bedroom:

• Download bedroom vae (2.3G) to vae_training/configs/vae_vroid/last.ckpt
• Downlaod bedroom diffusion (383.2M) to diff_denoising/configs/diff_vroid/unet/diffusion_pytorch_model.safetensors

For livingroom:

• Download livingroom vae (2.3G) to vae_training/configs/vae_vroid/last.ckpt
• Downlaod livingroom diffusion (383.2M) to diff_denoising/configs/diff_vroid/unet/diffusion_pytorch_model.safetensors

For vroid:

• Download vroid vae (4.4G) to vae_training/configs/vae_vroid/last.ckpt
• Downlaod vroid diffusion (1.1G) to diff_denoising/configs/diff_vroid/unet/diffusion_pytorch_model.safetensors

Inference

For bedroom (layout condition):

# generate layout
# move 1-th (-o 1) cabinet (-i 2) along z-axis (-d 2)
python paint_layout.py -b 62d0964d-a9c3-4f54-a1d5-4709a289193f_MasterBedroom-11117 -t bedroom -d 2 -i 2 -o 1

# generate latent tri-plane conditioned by layout
python test_diff.py -e diff_denoising/configs/diff_bedroom/ -u painting_bedroom/62d0964d-a9c3-4f54-a1d5-4709a289193f_MasterBedroom-11117/

# decode to 3D mesh
python test_latents.py -e vae_training/configs/vae_bedroom/ -s -l diff_denoising/configs/diff_bedroom/triplane_xxxxxx/

Additional examples can be selected from diff_denoising/semimg_bedroom.json.

For livingroom (layout condition + wall remeshing):

# generate layout
# move 0-th (-o 0) cabinet (-i 1) along x-axis (-d 0)
python paint_layout.py -b de513c79-61f9-4689-b671-fad361e605a3_LivingDiningRoom-16466 -t livingroom -d 0 -i 1 -o 0

# generate latent tri-plane conditioned by layout
python test_diff.py -e diff_denoising/configs/diff_livingroom/ -u painting_livingroom/de513c79-61f9-4689-b671-fad361e605a3_LivingDiningRoom-16466/

# decode to 3D mesh ('-w' activate the wall remeshing algorithm)
python test_latents.py -e vae_training/configs/vae_livingroom/ -w -l diff_denoising/configs/diff_livingroom/triplane_xxxxxx/

Additional examples can be selected from diff_denoising/semimg_livingroom.json.

For avatar (uncondition):

# generate latent tri-plane from random noise (unconditional)
python test_diff.py -e diff_denoising/configs/diff_vroid/

# decode to 3D mesh
python test_latents.py -e vae_training/configs/vae_vroid/ -s -l diff_denoising/configs/diff_vroid/triplane_xxxxxx/

Citation

If you find our code or paper helps, please consider citing:

@article{yan2024frankenstein,
    author    = {Han, Yan and Yang, Li and Zhennan, Wu and Shenzhou, Chen and Weixuan, Sun and Taizhang, Shang and Weizhe, Liu and Tian, Chen and Xiaqiang, Dai and Chao, Ma and Hongdong, Li and Pan, Ji},
    title     = {Frankenstein: Generating Semantic-Compositional 3D Scenes in One Tri-Plane},
    journal   = {ACM SIGGRAPH Asia Conference Proceedings},
    year      = {2024},
}

Acknowledgments

Some source codes are borrowed from DiffusionSDF, Diffusers, IDR.