Easy to install & run

Install

source scripts/install.sh

Download

Instant3D network weights

Download and put it in ckpts folder.

Huggingface login

huggingface-cli login

Enter your token. Huggingface token

You need to login to download sd2.1 in ckpts folder.

Run

soruce scripts/run.sh

Result

Text prompt: a white man wearing a steampunk hat is singing

RGB

fakes01_0.mp4

Depth

fakes01_depth_0.mp4

Normal

fakes01_norm_0.mp4

Official Instant3D Codes

This is the official implementation of our paper Instant3D: Instant Text-to-3D Generation.

The project page is here.

1. Building the required Conda environment by requirements.txt

2. Downloading the pre-trained ckpts

2.1 Our Instant3D network weights

(revise args.py --evaluate accordingly)

2.2 Pretrained CLIP and Stable Diffusion weights (revise args.py ----pretrained_path)

2.2.1 The CLIP weight for ViT-L/14

2.2.2 downloading SD 2.1, sign in hugging face account through

from huggingface_hub import login login('')

3. Install tiny-cuda-nn

4. DDP running

python -m torch.distributed.launch --nproc_per_node=1 --master_port=2349 ./main.py --batch-size 1 --random_seed 666 --gpu-ids 0

Citation

@article{li2023instant3d,
title     = {Instant3D: Instant Text-to-3D Generation},
author    = {Li, Ming and Zhou, Pan and Liu, Jia-Wei, and Keppo, Jussi and Lin, Min and Yan, Shuicheng and Xu, Xiangyu},
journal   = {International Journal of Computer Vision},
year      = {2024},
}