Customizable configs#

Our dataclass configs allow you to easily plug in different permutations of models, dataloaders, modules, etc. and modify all parameters from a typed CLI supported by tyro.

Base components#

All basic, reusable config components can be found in nerfstudio/configs/ The Config class at the bottom of the file is the upper-most config level and stores all of the sub-configs needed to get started with training.

You can browse this file and read the attribute annotations to see what configs are available and what each specifies.

Creating new configs#

If you are interested in creating a brand new model or data format, you will need to create a corresponding config with associated parameters you want to expose as configurable.

Let’s say you want to create a new model called Nerfacto. You can create a new Model class that extends the base class as described here. Before the model definition, you define the actual NerfactoModelConfig which points to the NerfactoModel class (make sure to wrap the _target classes in a field as shown below).


You can then enable type/auto complete on the config passed into the NerfactoModel by specifying the config type below the class definition.


class NerfactoModelConfig(ModelConfig):
    """Nerfacto Model Config"""

    _target: Type = field(default_factory=lambda: NerfactoModel)

class NerfactoModel(Model):
     """Nerfacto model

        config: Nerfacto configuration to instantiate model

    config: NerfactoModelConfig

The same logic applies to all other custom configs you want to create. For more examples, you can see nerfstudio/data/dataparsers/, nerfstudio/data/

See Also

For how to create the actual data and model classes that follow the configs, please refer to pipeline overview.

Updating method configs#

If you are interested in creating a new model config, you will have to modify the nerfstudio/configs/ file. This is where all of the configs for implemented models are housed. You can browse this file to see how we construct various existing models by modifying the Config class and specifying new or modified default components.

For instance, say we created a brand new model called Nerfacto that has an associated NerfactoModelConfig, we can specify the following new Config by overriding the pipeline and optimizers attributes appropriately.


method_configs["nerfacto"] = Config(
        model=NerfactoModelConfig(eval_num_rays_per_chunk=1 << 14),
        "proposal_networks": {
            "optimizer": AdamOptimizerConfig(lr=1e-2, eps=1e-15),
            "scheduler": None,
        "fields": {
            "optimizer": AdamOptimizerConfig(lr=1e-2, eps=1e-15),
            "scheduler": None,

After placing your new Config class into the method_configs dictionary, you can provide a description for the model by updating the descriptions dictionary at the top of the file.

Modifying from CLI#

Often times, you just want to play with the parameters of an existing model without having to specify a new one. You can easily do so via CLI. Below, we showcase some useful CLI commands.

  • List out all existing models

    ns-train --help
  • List out all exist configurable parameters for {METHOD_NAME}

    ns-train {METHOD_NAME} --help
  • Change the train/eval dataset

    ns-train {METHOD_NAME} --data DATA_PATH
  • Enable the viewer

    ns-train {METHOD_NAME} --vis viewer
  • See what options are available for the specified dataparser (e.g. blender-data)

    ns-train {METHOD_NAME} {DATA_PARSER} --help
  • Run with changed dataparser attributes and viewer on

    # NOTE: the dataparser and associated configurations go at the end of the command
    ns-train {METHOD_NAME} --vis viewer {DATA_PARSER} --scale-factor 0.5