Project Page | Paper | Weights | Live Demo 🤗
Zero-1-to-3: Zero-shot One Image to 3D Object
Ruoshi Liu1, Rundi Wu1, Basile Van Hoorick1, Pavel Tokmakov2, Sergey Zakharov2, Carl Vondrick1
1Columbia University, 2Toyota Research Institute
- Zero123-XL and Objaverse-XL released!
- Checkout the amazing image-to-3D results at Threestudio developed by Stability AI!
- Threestudio has recently implemented single-view 3D reconstruction with zero123!
- Stable-Dreamfusion has recently implemented 3D reconstruction with zero123 using Instant-NGP and SDS loss from DreamFusion!
- We have released training script and objaverse renderings.
- Live demo released 🤗: https://huggingface.co/spaces/cvlab/zero123-live. Shout out to Huggingface for funding this demo!!
- We've optimized our code base with some simple tricks and the current demo runs at around 22GB VRAM so it's runnable on a RTX 3090/4090(Ti)!
conda create -n zero123 python=3.9
conda activate zero123
cd zero123
pip install -r requirements.txt
git clone https://github.com/CompVis/taming-transformers.git
pip install -e taming-transformers/
git clone https://github.com/openai/CLIP.git
pip install -e CLIP/
Download checkpoint under zero123 through one of the following sources:
https://huggingface.co/cvlab/zero123-weights/tree/main
wget https://cv.cs.columbia.edu/zero123/assets/$iteration.ckpt # iteration = [105000, 165000, 230000, 300000]
Note that we have released 4 model weights: 105000.ckpt, 165000.ckpt, 230000.ckpt, 300000.ckpt. By default, we use 105000.ckpt which is the checkpoint after finetuning 105000 iterations on objaverse. Naturally, checkpoints trained longer tend to overfit to training data and suffer in zero-shot generalization, though we didn't empirically verify this. 300000.ckpt is trained for around 6000 A100 hours.
Run our gradio demo for novel view synthesis:
python gradio_new.py
Note that this app uses around 22 GB of VRAM, so it may not be possible to run it on any GPU.
Download image-conditioned stable diffusion checkpoint released by Lambda Labs:
wget https://cv.cs.columbia.edu/zero123/assets/sd-image-conditioned-v2.ckpt
Download and unzip valid_paths.json.zip and move the valid_paths.json file under the view_release folder.
Run training command:
python main.py \
-t \
--base configs/sd-objaverse-finetune-c_concat-256.yaml \
--gpus 0,1,2,3,4,5,6,7 \
--scale_lr False \
--num_nodes 1 \
--seed 42 \
--check_val_every_n_epoch 10 \
--finetune_from sd-image-conditioned-v2.ckpt
Note that this training script is set for an 8-GPU system, each with 80GB of VRAM. As discussed in the paper, empirically the large batch size is very important for "stably" training stable diffusion. If you have smaller GPUs, consider using smaller batch size and gradient accumulation to obtain a similar effective batch size. Please check this thread for the train/val split we used in the paper.
Download our objaverse renderings with:
wget https://tri-ml-public.s3.amazonaws.com/datasets/views_release.tar.gz
Disclaimer: note that the renderings are generated with Objaverse. The renderings as a whole are released under the ODC-By 1.0 license. The licenses for the renderings of individual objects are released under the same license creative commons that they are in Objaverse.
Check out Stable-Dreamfusion
Note that we haven't extensively tuned the hyperparameters for 3D recosntruction. Feel free to explore and play around!
cd 3drec
pip install -r requirements.txt
python run_zero123.py \
--scene pikachu \
--index 0 \
--n_steps 10000 \
--lr 0.05 \
--sd.scale 100.0 \
--emptiness_weight 0 \
--depth_smooth_weight 10000. \
--near_view_weight 10000. \
--train_view True \
--prefix "experiments/exp_wild" \
--vox.blend_bg_texture False \
--nerf_path "data/nerf_wild"
-
You can see results under:
3drec/experiments/exp_wild/$EXP_NAME. -
To export a mesh from the trained Voxel NeRF with marching cube, use the
export_meshfunction. For example, add a line:vox.export_mesh($PATH_TO_EXPORT)under the
evaluatefunction. -
The dataset is formatted in the same way as NeRF for the convenience of dataloading. In reality, the recommended input in addition to the input image is an estimate of the elevation angle of the image (e.g. if the image is taken from top, the angle is 0, front is 90, bottom is 180). This is hard-coded now to the extrinsics matrix in
transforms_train.json -
We tested the installation processes on a system with Ubuntu 20.04 with an NVIDIA GPU with Ampere architecture.
The design of our method fundamentally alleviates the Janus problem as shown in the 3D reconstruction results above and many results in the Stable-Dreamfusion repo. By modeling camera perspective in an explicit way and training on a large-scale high-quality synthetic dataset where we can obtain ground truth for everything, the ambiguity and bias of viewpoint existing in text-to-image model is significantly alleviated.
This is also related to the prompting tricks used in DreamFusion where prompts like "a back view of" is inserted at the beginning of the text prompt. Zero-1-to-3 models such change of viewpoint explicitly and finetune on Objaverse to ensure both consistency after viewpoint change and accuracy of queried viewpoint.
This repository is based on Stable Diffusion, Objaverse, and SJC. We would like to thank the authors of these work for publicly releasing their code. We would like to thank the authors of NeRDi and SJC for their helpful feedback.
We would like to thank Changxi Zheng and Chengzhi Mao for many helpful discussions. This research is based on work partially supported by the Toyota Research Institute, the DARPA MCS program under Federal Agreement No. N660011924032, and the NSF NRI Award #1925157.
@misc{liu2023zero1to3,
title={Zero-1-to-3: Zero-shot One Image to 3D Object},
author={Ruoshi Liu and Rundi Wu and Basile Van Hoorick and Pavel Tokmakov and Sergey Zakharov and Carl Vondrick},
year={2023},
eprint={2303.11328},
archivePrefix={arXiv},
primaryClass={cs.CV}
}