Project page | Paper | Video
This repo contains the official implementation for the paper "BlockFusion: Expandable 3D Scene Generation using Latent Tri-plane Extrapolation".
Pytorch3d is required for postprocess. Come to offical page for installation instructions.
We modified diffusers to adapt triplane structure. Users should build based on the ./src
# clone this repo
python setup.py install
Pretrained weights of MLP, VAE, and Diffuser(Condition/Unconditioned). Download the model here [10GB]
and extract to ./checkpoints.
To do uncondtional inference, run
python unconditioned_prediction.py --batch 4 --save output/uncond
To do single block conditional inference, run
python conditioned_prediction.py --layout samplelayouts/exp1_32-56-24/0_0.npy --save output/cond
We provide some sample layouts for demo. If you would like to draw your own layout maps, you can use drawtkinter.py to create your own layout. Here is a tutorial:
The unit of the graduations are in meters, furnitures need to be in reasonable scale. Walls should surround the floor and furniture should be placed on the floor.
The output layouts directory named after expname_xscale-zscale-stride, xscale-zscale-stride are given in decimeter.
Usage:
python drawtkinter.py --h 56 --w 56 --expname draw
Full pipeline(conditional prediction + extrapolation + resample) is contained in fullpipeline.py. Before running, specifying the layout directory.
python fullpipeline.py --layout samplelayouts/exp2_32-56-24 --resample 15 --save output
If you find our code or paper helps, please consider citing:
@article{blockfusion,
title={BlockFusion: Expandable 3D Scene Generation using Latent Tri-plane Extrapolation},
author={Wu, Zhennan and Li, Yang and Yan, Han and Shang, Taizhang and Sun, Weixuan and Wang, Senbo and Cui, Ruikai and Liu, Weizhe and Sato, Hiroyuki and Li, Hongdong and Ji, Pan},
journal={ACM Transactions on Graphics},
volume={43},
number={4},
year={2024},
doi={10.1145/3658188}
}