> ## Documentation Index
> Fetch the complete documentation index at: https://docs.cyberun.cloud/llms.txt
> Use this file to discover all available pages before exploring further.

# Run a Nerfstudio workflow

> Build a Gaussian splatting workflow with the wizard, then run it on an agent.

The Nerfstudio wizard in Cloud builds a workflow for a 3D Gaussian
splatting pipeline (`splatfacto` or `splatfacto-big`) without
hand-editing JSON. You provide input captures and parameters; the
wizard outputs a workflow you can save and run.

The wizard must appear on the workflow create flow. If not, the
Nerfstudio isn't enabled on this deployment (contact your team admin), or no agent
in your team carries the `nerfstudio` label.

## Before you start

* A capture you want to reconstruct. Three input kinds are
  supported:

  | Input kind              | What it is                                     |
  | ----------------------- | ---------------------------------------------- |
  | `images_zip`            | A ZIP of still images of the subject.          |
  | `video`                 | A video file that will be sampled into frames. |
  | `processed_dataset_zip` | A pre-processed Nerfstudio dataset ZIP.        |

* A team agent carrying the `nerfstudio` label (and the `gpu`
  label, of course).

## Build the workflow

1. Open **Workflows** in the sidebar.

2. Click **Create workflow** and pick **Nerfstudio wizard**.

3. Fill in the wizard:

   | Field                 | Notes                                                                               |
   | --------------------- | ----------------------------------------------------------------------------------- |
   | Input kind            | `images_zip`, `video`, or `processed_dataset_zip`.                                  |
   | Training method       | `splatfacto` (faster, lower quality) or `splatfacto-big` (slower, higher quality).  |
   | File parameter key    | The name of the file parameter callers will supply. Defaults to something sensible. |
   | File parameter label  | The human-readable label shown on the run form.                                     |
   | Max iterations        | Training iterations, 1000–100000.                                                   |
   | Camera type           | Free text. Leave blank to let the runtime pick.                                     |
   | Run process\_data     | Toggle. Enables Nerfstudio's pre-processing step.                                   |
   | Export Gaussian splat | Toggle. Adds the `.ply` export step at the end.                                     |

4. Click **Save** to create the workflow.

The wizard shows the generated workflow spec inline (in a
collapsible preview) before saving so you can review it.

## Run

The new workflow appears in the **Workflows** list. From the detail
page:

1. Click **Run**.
2. Upload the input (images ZIP, video, or processed dataset).
3. Submit.

The task moves through `pending → waiting → queued → running → completed`.
Long-running training tasks emit progress events with step counts
that the task detail page renders inline.

## Output

When the task succeeds, the output panel shows:

* A primary artifact (the splat `.ply` or the video, depending on
  toggles).
* A storage key referencing the rest of the output (config, file
  list, manifest).

You can re-run the workflow with new inputs or different parameters
by submitting another task.

## Related

* [Workflows](/cloud/concepts/workflows) — concept.
* [Tasks](/cloud/concepts/tasks) — lifecycle and progress events.
* [Agents](/cloud/concepts/agents) — what carries the `nerfstudio`
  label.