TLDR: NeRFiller completes missing regions in 3D scenes via an off-the-shelf 2D generative inpainter.

Project page at https://ethanweber.me/nerfiller/. Paper at https://arxiv.org/abs/2312.04560.

git clone [email protected]:ethanweber/nerfiller.git

pip install -e .

Download the NeRF synthetic data

python nerfiller/scripts/download.py blender

Download the meshes that we will complete with NeRFiller

python nerfiller/scripts/download.py meshes

Download the backpack capture from SPIn-NeRF, formatted as a Nerfstudio dataset

python nerfiller/scripts/download.py nerfstudio

Download the novel-view camera paths.

python nerfiller/scripts/download.py camera-paths

Download the LaMa inpainter weights

python nerfiller/scripts/download.py models

Convert the blender datasets to Nerfstudio format.

Use blender_to_nerfstudio.py for converting blender data to nerfstudio datasets. Run ns-launch blender-to-nerfstudio --dry-run to to see all the commands you need to run, one per dataset, for a total of 8. Running it without --dry-run will run all the commands. I prefer to run all commands in a tmux session in case they take a while to run.

ns-launch blender-to-nerfstudio

Use create_nerfstudio_dataset.ipynb to render the meshes into nerfstudio datasets. This will require going into the notebook and running it from start to completition, where you modify the variable dataset_name each time. The dataset variables should be the name of the meshes in the folder data/meshes. This is where you'd want to add your own custom data. For example, let's look at the dataset cat with an explanation for each file.

cat/
  |- jerrypillow.jpg    # texture file
  |- model.mtl          # model material file
  |- model.obj          # model obj (name is important)
  |- occlusion.obj      # occlusion obj (name is important)

The filenames for model.obj and occlusion.obj are important, and should be the same for each dataset so the notebook can properly process it based on the specified dataset_name. The model will reference it's material and texture files, and the occlusion file will be used to mark regions of the mesh in 3D which we want to inpaint. We call the texture file jerrypillow.jpg because the cat plushie is named Jerry. 🙂

Cat mesh	Occlusion mesh (manually made)	Cat mesh & occlusion

The cat mesh was created by scanning a scene with Polycam, exporting the mesh and loading it into Blender, and then deleting vertices and faces. The occluder wsa manually created in Blender too. Both the mesh (left) and occluder (middle) were explored to the files model.obj and occlusion.obj, respectively, where we keep the material and texture files for the model. The composited result is on the right.

If you want to use your own data, you can make a folder like specified above. Blender is useful is you want to export to OBJ files, which is what we require. In the notebook, you'll need to add another entry for your dataset in the dataset_configs dictionary.

Run ns-launch --help for replicating our experiments. Run commands with --dry-run to see what the individual jobs will be. Edit nerfiller/experiments/data.py if you'd like to run on your own data! You'll need to modify the beginning lines of the file.

Next, we will show how to run experiments on all of our data for completing large unknown 3D regions. We refer to these experiments as the "occluder" experiments. This will involve editing the file nerfiller/experiments/occluder.py. You'd first uncomment as follows and run ns-launch train-occluder. This will train models on the visible regions.

methods = [
    ("nerfacto-nerfiller-visualize", "none"),
    ("nerfacto-nerfiller", "none"), # this must be trained first, as the following depend on it
    # ("individual-inpaint-once", "individual-sd-image"),
    # ("individual-inpaint-once", "individual-lama"),
    # ("grid-prior-du-no-depth", "none"),
    # ("grid-prior-du", "none"),
    # ("individual-inpaint-du", "none"),
    # ("grid-prior-du-random-noise", "none"), # this is extra for an ablation
]

Next, you'd uncomment as follows, omitting the first two methods. These methods will start from the nerfacto-nerfiller method checkpoint, which was trained in the previous step. grid-prior-du is our main method, and the rest are baselines.

If you want to run the individual-inpaint-once baselines, first run ns-launch inpaint-occluder. This will inpaint the missing regions once, and before we train the NeRF. The du methods will inpaint iteratively with the dataset update strategy. The inpaints will be in the folders data/nerfstudio/<dataset_name>-inpaint/<inpaint_method>/<timestamp>.

methods = [
    # ("nerfacto-nerfiller-visualize", "none"),
    # ("nerfacto-nerfiller", "none"), # this must be trained first, as the following depend on it
    ("individual-inpaint-once", "individual-sd-image"),
    ("individual-inpaint-once", "individual-lama"),
    ("grid-prior-du-no-depth", "none"),
    ("grid-prior-du", "none"),
    ("individual-inpaint-du", "none"),
    ("grid-prior-du-random-noise", "none"), # this is extra for an ablation
]

Finally, you can run ns-launch render-occluder --help and ns-launch metrics-occluder --help (in order, without --help or --dry-run). In general, we following a convention of training, rendering out the result, and finally, computing metrics. We use the notebook notebooks/inpainting_consistency_table.ipynb for aggregating results and generating LaTeX tables.

If you want to create side-by-side comparison videos, use the notebook notebooks/video_comparison.ipynb.

Use notebooks/inpaint.ipynb to play around with inpainting a dataset in a notebook.

For the experiments on the Blender dataset, the series of commands will be ns-launch inpaint-blender, ns-launch train-blender, ns-launch render-blender, and ns-launch metrics-blender. In this case, "blender" is used instead of "occluder" (as used previously). These are our one-word nicknames we use in the code to distinguish the different types of experiments performed in the paper.

In order to do reference inpainting, you can place a reference image named reference.png inside a Nerfstudio dataset folder. E.g., if you look at data/nerfstudio/cat, you'll see the file data/nerfstudio/cat/reference.png. You can inpaint images with notebooks/inpaint.ipynb until finding an example image that you like and then place it inside the Nerfstudio dataset folder. Then, you'll need to run the method for inpainting with a reference. Edit nerfiller/experiments/data.py as follows. Comment out lines of the occluder_dataset_names set such that only the datasets you care about are present. Mark if they are synthetic datasets or not in the variable synthetic_dataset_names as done with other datasets. This will change settings of Nerfacto, such as turning off scene contraction or setting a white or black background during training. Modify dataset_name_method_modifications as done for other examples if you want to use a text prompt with the reference inpaint. Finally, call ns-launch train-reference --dry-run to see the commands that would run. If you're happy with it, run without --dry-run or simply copy and paste the printed command and run on your own.

If you find this code or data useful for your research, please consider citing the following paper:

@inproceedings{weber2023nerfiller,
    title = {NeRFiller: Completing Scenes via Generative 3D Inpainting},
    author = {Ethan Weber and Aleksander Holynski and Varun Jampani and Saurabh Saxena and
        Noah Snavely and Abhishek Kar and Angjoo Kanazawa},
    booktitle = {CVPR},
    year = {2024},
}