This repository contains the code needed to train Mega-NeRF models and generate the sparse voxel octrees used by the Mega-NeRF-Dynamic viewer.

The codebase for the Mega-NeRF-Dynamic viewer can be found here.

Note: This is a preliminary release and there may still be outstanding bugs.


      title={Mega-NeRF: Scalable Construction of Large-Scale NeRFs for Virtual Fly-Throughs}, 
      author={Haithem Turki and Deva Ramanan and Mahadev Satyanarayanan},



conda env create -f environment.yml
conda activate mega-nerf

The codebase has been mainly tested against CUDA >= 11.1 and V100/2080 Ti/3090 Ti GPUs. 1080 Ti GPUs should work as well although training will be much slower.


Mill 19

  • The Building scene can be downloaded here.
  • The Rubble scene can be downloaded here.

UrbanScene 3D

  1. Download the raw photo collections from the UrbanScene3D dataset
  2. Download the refined camera poses for one of the scenes below:
  1. Run python scripts/ --image_path $RAW_PHOTO_PATH --dataset_path $CAMERA_POSE_PATH

Quad 6k Dataset

  1. Download the raw photo collections from here.
  2. Download the refined camera poses
  3. Run python scripts/ --image_path $RAW_PHOTO_PATH --dataset_path $CAMERA_POSE_PATH

Custom Data

The expected directory structure is:

  • / Torch file that should contain the following keys:
    • ‘origin_drb’: Origin of scene in real-world units
    • ‘pose_scale_factor’: Scale factor mapping from real-world unit (ie: meters) to [-1, 1] range
  • ‘/{val|train}/rgbs/’: JPEG or PNG images
  • ‘/{val|train}/metadata/’: Image-specific image metadata saved as a torch file. Each image should have a corresponding metadata file with the following file format: {rgb_stem}.pt. Each metadata file should contain the following keys:
    • ‘W’: Image width
    • ‘H’: Image height
    • ‘intrinsics’: Image intrinsics in the following form: [fx, fy, cx, cy]
    • ‘c2w’: Camera pose. 3×3 camera matrix with the convention used in the original NeRF repo, ie: x: down, y: right, z: backwards, followed by the following transformation:[camera_in_drb[:, 1:2], -camera_in_drb[:, :1], camera_in_drb[:, 2:4]], -1)


  1. Generate the training partitions for each submodule: python scripts/ --config configs/mega-nerf/${DATASET_NAME}.yml --dataset_path $DATASET_PATH --output $MASK_PATH --grid_dim $GRID_X $GRID_Y
    • Note: this can be run across multiple GPUs by instead running python -m --standalone --nnodes=1 --nproc_per_node $NUM_GPUS --max_restarts 0 scripts/ <args>
  2. Train each submodule: python mega_nerf/ --config_file configs/mega-nerf/${DATASET_NAME}.yml --exp_name $EXP_PATH --dataset_path $DATASET_PATH --chunk_paths $SCRATCH_PATH --cluster_mask_path ${MASK_PATH}/${SUBMODULE_INDEX}
    • Note: training with against full scale data will write hundreds of GBs / several TBs of shuffled data to disk. You can downsample the training data using train_scale_factor option.
    • Note: we provide a utility script based on parscript to start multiple training jobs in parallel. It can run through the following command: CONFIG_FILE=configs/mega-nerf/${DATASET_NAME}.yaml EXP_PREFIX=$EXP_PATH DATASET_PATH=$DATASET_PATH CHUNK_PREFIX=$SCRATCH_PATH MASK_PATH=$MASK_PATH python -m parscript.dispatcher parscripts/run_8.txt -g $NUM_GPUS
  3. Merge the trained submodules into a unified Mega-NeRF model: python scripts/ --config_file configs/mega-nerf/${DATASET_NAME}.yaml --ckpt_prefix ${EXP_PREFIX}- --centroid_path ${MASK_PATH}/ --output $MERGED_OUTPUT


Single-GPU evaluation: python mega_nerf/ --config_file configs/nerf/${DATASET_NAME}.yaml --exp_name $EXP_NAME --dataset_path $DATASET_PATH --container_path $MERGED_OUTPUT

Multi-GPU evaluation: python -m --standalone --nnodes=1 --nproc_per_node $NUM_GPUS mega_nerf/ --config_file configs/nerf/${DATASET_NAME}.yaml --exp_name $EXP_NAME --dataset_path $DATASET_PATH --container_path $MERGED_OUTPUT

Octree Extraction (for use by Mega-NeRF-Dynamic viewer)

python scripts/ --config configs/mega-nerf/${DATASET_NAME}.yaml --dataset_path $DATASET_PATH --container_path $MERGED_OUTPUT --output $OCTREE_PATH


Large parts of this codebase are based on existing work in the nerf_pl, NeRF++, and Plenoctree repositories. We use svox to serialize our sparse voxel octrees and the generated structures should be largely compatible with that codebase.


View Github