bevy_zeroverse ♾️

arXiv | project page

bevy zeroverse synthetic reconstruction dataset generator. view the live demo.

capabilities

• depth/normal rendering modes
• plücker camera labels
• generate parameteric zeroverse primitives
• primitive deformation
• procedural zeroverse composite environments
• online torch dataloader
• safetensor chunking
• hypersim semantic labels
• primitive boolean operations
• primitive pbr wireframe
• primitive 4d augmentation

dataloader

from bevy_zeroverse_dataloader import BevyZeroverseDataset
from torch.utils.data import DataLoader

dataset = BevyZeroverseDataset(
    editor=False, headless=True, num_cameras=6,
    width=640, height=480, num_samples=1e6,
)
dataloader = DataLoader(
    dataset, batch_size=4, shuffle=True, num_workers=1,
)

for batch in dataloader:
    visualize(batch)

chunked dataloader

requires nvjpeg: https://developer.nvidia.com/nvjpeg

mat-synth

• download the mat-synth dataset here
• resize the mat-synth dataset (4k is heavy) using python mat-synth/resize.py --source_dir <path-to-mat-synth> --dest_dir assets/materials
• material basecolor grid view (cargo run -- --material-grid or live demo)

compatible bevy versions

credits

• lgm
• mat-synth
• zeroverse