This repository is for the new neral light field (NeLF) method introduced in the following ECCV'22 paper:

R2L: Distilling Neural Radiance Field to Neural Light Field for Efficient Novel View Synthesis
Huan Wang ^1,2, Jian Ren ¹, Zeng Huang ¹, Kyle Olszewski ¹, Menglei Chai ¹, Yun Fu ², and Sergey Tulyakov ^1
1 Snap Inc. ² Northeastern University
Work done when Huan was an intern at Snap Inc.

[TL;DR] We present R2L, a deep (88-layer) residual MLP network that can represent the neural light field (NeLF) of complex synthetic and real-world scenes. It is featured by compact representation size (~20MB storage size), faster rendering speed (~30x speedup than NeRF), significantly improved visual quality (1.4dB boost than NeRF), with no whistles and bells (no special data structure or parallelism required).

Below we only show the example of scene lego. You may test on other scenes simply by changing all the lego word segment to other scene names. Scripts have been doubled-checked. You should be able to run them simply by copy-paste.

git clone [email protected]:snap-research/R2L.git && cd R2L

sh scripts/download_example_data.sh

• conda create --name R2L python=3.9.6
• conda activate R2L
• pip install -r requirements.txt (We use torch 1.9.0, torchvision 0.10.0)

• Download models:

sh scripts/download_R2L_models.sh

• Run

CUDA_VISIBLE_DEVICES=0 python main.py --model_name R2L --config configs/lego_noview.txt --n_sample_per_ray 16 --netwidth 256 --netdepth 88 --use_residual --trial.ON --trial.body_arch resmlp --pretrained_ckpt R2L_Blender_Models/lego.tar --render_only --render_test --testskip 1 --experiment_name Test__R2L_W256D88__blender_lego

There are two major steps in R2L training. (1) Use pretrained NeRF model to generate synthetic data and train R2L network on the synthetic data -- this step can make our R2L model perform comparably to the NeRF teacher; (2) Finetune the R2L model in (1) with the real data -- this step will further boost the performance and make our R2L model outperform the NeRF teacher.

The detailed step-by-step training pipeline is as follows.

Train a NeRF model:

CUDA_VISIBLE_DEVICES=0 python main.py --model_name nerf --config configs/lego.txt --experiment_name NeRF__blender_lego

You can also download the teachers we trained to continue first:

sh scripts/download_NeRF_models.sh

To test the download teachers, you can use

CUDA_VISIBLE_DEVICES=0 python main.py --model_name nerf --config configs/lego.txt --pretrained_ckpt NeRF_Blender_Models/lego.tar --render_only --render_test --testskip 1 --experiment_name Test__NeRF__blender_lego

Use the pretrained NeRF model to generate synthetic data (saved in .npy format):

CUDA_VISIBLE_DEVICES=0 python utils/create_data.py --create_data rand --config configs/lego.txt --teacher_ckpt Experiments/NeRF__blender_lego*/weights/200000.tar --n_pose_kd 10000 --datadir_kd data/nerf_synthetic/lego:data/nerf_synthetic/lego_pseudo_images10k --experiment_name NeRF__blender_lego__create_pseudo

If you are using the downloaded teachers, please use this snippet:

CUDA_VISIBLE_DEVICES=0 python utils/create_data.py --create_data rand --config configs/lego.txt --teacher_ckpt NeRF_Blender_Models/lego.tar --n_pose_kd 10000 --datadir_kd data/nerf_synthetic/lego:data/nerf_synthetic/lego_pseudo_images10k --experiment_name NeRF__blender_lego__create_pseudo

The pseudo data will be saved in data/nerf_synthetic/lego_pseudo_images10k. Every 4096 rays are saved in one .npy file. For 10k images (400x400 resoltuion), there will be 309600 .npy files. On our RTX 2080Ti GPU, rendering 1 image with NeRF takes around 8.5s, so 10k images would take around 24hrs. If you want to try our method quicker, you may download the lego data we synthesized (500 images, 2.8GB) and go to Step 3:

sh scripts/download_lego_pseudo_images500.sh

The data will be extracted under data/nerf_synthetic/lego_pseudo_images500. Using only 500 pseudo images for training would lead to degraded quality, but based on our ablation study (see Fig. 6 in our paper), it works farily good.

Train R2L model on the synthetic data:

CUDA_VISIBLE_DEVICES=0 python main.py --model_name R2L --config configs/lego_noview.txt --n_sample_per_ray 16 --netwidth 256 --netdepth 88 --datadir_kd data/nerf_synthetic/lego_pseudo_images10k --n_pose_video 20,1,1 --N_iters 1200000 --N_rand 20 --data_mode rays --hard_ratio 0.2 --hard_mul 20 --use_residual --trial.ON --trial.body_arch resmlp --num_worker 8 --warmup_lr 0.0001,200 --experiment_name R2L__blender_lego

If you are using the downloaded lego_pseudo_images500 data, please use this snippet:

CUDA_VISIBLE_DEVICES=0 python main.py --model_name R2L --config configs/lego_noview.txt --n_sample_per_ray 16 --netwidth 256 --netdepth 88 --datadir_kd data/nerf_synthetic/lego_pseudo_images500 --n_pose_video 20,1,1 --N_iters 1200000 --N_rand 20 --data_mode rays --hard_ratio 0.2 --hard_mul 20 --use_residual --trial.ON --trial.body_arch resmlp --num_worker 8 --warmup_lr 0.0001,200 --experiment_name R2L__blender_lego

Convert original real data (images) to our .npy format:

• For blender data:

python utils/convert_original_data_to_rays_blender.py --splits train --datadir data/nerf_synthetic/lego

The converted data will be saved in data/nerf_synthetic/lego_real_train.

• For llff data:

python utils/convert_original_data_to_rays_llff.py --splits train --datadir data/nerf_llff_data/flower

The converted data will be saved in data/nerf_llff_data/room_real_train.

Finetune the R2L model in Step 3 on the data in Step 4:

CUDA_VISIBLE_DEVICES=0 python main.py --model_name R2L --config configs/lego_noview.txt --n_sample_per_ray 16 --netwidth 256 --netdepth 88 --datadir_kd data/nerf_synthetic/lego_real_train --n_pose_video 20,1,1 --N_iters 1600000 --N_rand 20 --data_mode rays --hard_ratio 0.2 --hard_mul 20 --use_residual --trial.ON --trial.body_arch resmlp --num_worker 8 --warmup_lr 0.0001,200 --save_intermediate_models --pretrained_ckpt Experiments/R2L__blender_lego_SERVER*/weights/ckpt_1200000.tar --resume --experiment_name R2L__blender_lego__ft

Note, this step is pretty fast and prone to overfitting, so do not finetune it too much. We simply set the finetuning steps based on our validation.

The quantitative and qualitative comparison are shown below. See more results and videos on our webpage.

In this code we refer to the following implementations: nerf-pytorch and smilelogging. Great thanks to them! We especially thank nerf-pytorch. Our code is largely built upon their wonderful implementation. We also greatly thank the anounymous ECCV'22 reviewers for the constructive comments to help us improve the paper.

If our work or code helps you, please consider to cite our paper. Thank you!

@inproceedings{wang2022r2l,
  author = {Huan Wang and Jian Ren and Zeng Huang and Kyle Olszewski and Menglei Chai and Yun Fu and Sergey Tulyakov},
  title = {R2L: Distilling Neural Radiance Field to Neural Light Field for Efficient Novel View Synthesis},
  booktitle = {ECCV},
  year = {2022}
}