Xin Kong · Shikun Liu · Xiaoyang Lyu · Marwan Taher · Xiaojuan Qi · Andrew J. Davison

EscherNet is a multi-view conditioned diffusion model for view synthesis. EscherNet learns implicit and generative 3D representations coupled with the camera positional encoding (CaPE), allowing precise and continuous relative control of the camera transformation between an arbitrary number of reference and target views.

conda env create -f environment.yml -n eschernet
conda activate eschernet

Run demo to generate randomly sampled 25 novel views from (1,2,3,5,10) reference views:

bash eval_eschernet.sh

CaPE is applied in self/cross-attention for encoding camera pose info into transformers. The main modification is in diffusers/models/attention_processor.py.

To quickly check the implementation of CaPE (6DoF and 4DoF), run:

python CaPE.py

Download Zero123's Objaverse Rendering data:

wget https://tri-ml-public.s3.amazonaws.com/datasets/views_release.tar.gz

Filter Zero-1-to-3 rendered views (empty images):

cd scripts
python objaverse_filter.py --path /data/objaverse/views_release

Configure accelerator (8 A100 GPUs, bf16):

accelerate config

Choose 4DoF or 6DoF CaPE (Camera Positional Encoding):

cd 4DoF or 6DoF

Launch training:

accelerate launch train_eschernet.py --train_data_dir /data/objectverse/views_release --pretrained_model_name_or_path runwayml/stable-diffusion-v1-5 --train_batch_size 256 --dataloader_num_workers 16 --mixed_precision bf16 --gradient_checkpointing --T_in 3 --T_out 3 --T_in_val 10 --output_dir logs_N3M3B256_SD1.5 --push_to_hub --hub_model_id ***** --hub_token hf_******************* --tracker_project_name eschernet

For monitoring training progress, we recommand wandb for its simplicity and powerful features.

wandb login

Offline mode:

WANDB_MODE=offline python xxx.py

We provide raw results and two checkpoints 4DoF and 6DoF for easier comparison.

GSO30: We select 30 objects from GSO dataset and render 25 randomly sampled novel views for each object for both NVS and 3D reconstruction evaluation.

We use the 10 scenes from google_scanned.tar under folder 40_scenes for NVS evaluation.

We use the all 8 NeRF objects for 2D NVS evaluation.

We collected 16 real world object-centric recordings using a Franka Emika Panda robot arm with RealSense D435i Camera for real world NVS evaluation.

We collected Text2Img generation results from internet, Stable Diffusion XL (1 view) and MVDream (4 views: front, right, back, left) for NVS evaluation.

To get 2D Novel View Synthesis (NVS) results, set cape_type, checkpoint, data_type, data_dir and run:

bash ./eval_eschernet.sh

Evaluate 2D metrics (PSNR, SSIM, LPIPS):

cd metrics
python eval_2D_NVS.py

We firstly generate 36 novel views with data_type=GSO3D by:

bash ./eval_eschernet.sh

Then we adopt NeuS for 3D reconstruction:

export CUDA_HOME=/usr/local/cuda-11.8
pip install git+https://github.com/NVlabs/tiny-cuda-nn/#subdirectory=bindings/torch
cd 3drecon
python run_NeuS.py

Evaluate 3D metrics (Chamfer Distance, IoU):

cd metrics
python eval_3D_GSO.py

TODO.

To build locally:

python gradio_eschernet.py

We have intensively borrow codes from the following repositories. Many thanks to the authors for sharing their codes.

• Zero-1-to-3
• SyncDreamer
• MVDream
• NeuS

If you find this work useful, a citation will be appreciated via:

@article{kong2024eschernet,
    title={EscherNet: A Generative Model for Scalable View Synthesis},
  author={Kong, Xin and Liu, Shikun and Lyu, Xiaoyang and Taher, Marwan and Qi, Xiaojuan and Davison, Andrew J},
  journal={arXiv preprint arXiv:2402.03908},
  year={2024}
}