Docs

Recipes

Recipes create a workbook with a tested starting structure, setup instructions, files, and one or more runnable cells or controls. They are starting points, not a claim that every dependency, model, or workload is already present.

Choose a starter#

Creating any recipe creates a workbook only. It does not start compute. Start the selected CPU or GPU runtime when you are ready to run it.

Workbook starter What it prepares What to know
Blank workbook A clean place for your documents, source files, and code. Best when you already know the project structure.
Work with video FFmpeg actions to encode or resize video, extract eight frames, or export audio as WAV. Select a workbook media file. Imported audio is preserved when present; only the explicit test-pattern demo generates a tone. Each run saves its declared outputs and a manifest.
Render with Blender A GPU Blender scene with render controls and saved artifacts. Use GPU compute for the recipe's intended rendering path. Inspect the generated files after the run.
Train a model A small PyTorch GPU example with synthetic data, live loss metrics, and a checkpoint. This is a GPU health-check and starter example, not general dataset training or a production-trained model. Replace the script and data deliberately.
Generate images ComfyUI on a GPU with persistent model and output directories. The ComfyUI application and dependencies are prepared, but models are not bundled. Add models you are authorized to use.

Work with video#

Select a workbook media file and choose one action:

  • Encode clip makes a bounded MP4 excerpt from an image or video, plus a poster and metadata. Source audio is kept when present; silent input stays silent.
  • Resize video fits the video inside the selected output box without stretching, padding, or enlargement. It saves an MP4, poster, and metadata.
  • Extract 8 frames saves eight evenly spaced PNGs and a manifest with their requested timestamps.
  • Extract audio as WAV saves the first audio stream as a WAV and a manifest. It reports an error if the source has no audio.

Inputs are limited to 256 MiB. Processing is limited to 60 seconds; resize and extraction reject longer inputs rather than silently truncating them. Output files are limited to 15 MiB each. Manifests record the operation, settings, source hash, detected streams, output hashes and sizes, FFmpeg version, and audio origin. Each run has separate outputs and retained source inputs.

For a generated example, leave the file empty and explicitly select test-pattern under Encode clip demo source, then select Encode clip. This is the only path that generates a tone. This starter is a set of media-processing actions, not a non-destructive timeline editor. Quickstart walks through the demo. Existing workbooks keep their existing scripts; recipe updates apply when creating a new workbook.

Render with Blender#

The Blender recipe creates a procedural scene and saves render artifacts. Start the runtime before a render, wait for the run result, then use Files to inspect the image and other outputs. Keep the scene files and controls in the workbook when you expect to revise the render later.

Train a model#

The training starter runs a short synthetic PyTorch workload on CUDA, streams loss values, and saves a checkpoint plus metrics. It validates the GPU path and shows a repeatable pattern for saving outputs. Before using it with real data, choose the data source, checkpoint policy, runtime limits, and cost controls appropriate to the project.

Generate images#

The ComfyUI recipe starts the application on GPU compute and can expose an authenticated browser app link once the server is listening. Keep the server job running while you use the app. Store chosen models under the persistent model directory and generated assets under the persistent output directory. Copy small results into workbook-visible outputs when you want them in the Files explorer.