Graph Workflows¶
The graph is the visible record of a VIPP analysis.
Each node performs one operation. Each connection says which output feeds which input. Parameters are adjusted in the inspector, and node outputs can be viewed as thumbnails, napari layers, histograms, labels, or tables.
Core Graph Features¶
| Feature | Use |
|---|---|
| Typed ports | Prevent images, masks, labels, and tables from being mixed accidentally. |
| Multi-input nodes | Express operations such as labels plus intensity image, image plus PSF, or two-channel colocalization. |
| Multi-output nodes | Split channels, split axes, skeleton keypoints, or graph tables into separate outputs. |
| Named tunnels | Reuse a source such as a channel, mask, ROI, or reference image without drawing long wires. |
| Live source subtitles | See the layer, sample, file, or collection representative bound to each Image Source; hover for the complete value. |
| Workflow tabs | Keep independent graphs, results, caches, histories, inspectors, paths, compute requests, and Batch workspaces open together. |
| Graph notes | Record reasoning next to workflow steps. |
| Graph search | Find nodes, operation IDs, tunnel names, and batch output tags. |
| Workflow JSON | Save the graph, node parameters, connections, canvas positions, tunnels, portable schema-6 compute and Run/Bypass intent, stable SourceItems, an optional versioned Batch workspace, and selected UI state. |
| Actual-run provenance | Record the CPU or CuPy-family implementation and fallback that produced each accepted result; authored intent alone is not proof. |
| Python export | Generate a version-locked headless program that uses the same execution service, compute request, progress, cancellation, and optional output provenance. |
Good Graph Habits¶
- Keep the source image and key masks inspectable.
- Use tunnels when a channel, mask, ROI, or reference image feeds many nodes.
- Add graph notes at scientific decision points.
- Detach VIPP for longer workflows, then resize the floating window freely in width and height or maximize it. Reattaching restores napari's dock constraints.
- Hide the node library after building the graph when you need more canvas space.
- Zoom out for whole-workflow review, then zoom in for parameter tuning.
- Add explicit
Batch Outputnodes before running a workflow over a folder. - Read actual-run badges before assuming that Auto, Prefer GPU, or Custom used a GPU.
- Save workflow snapshots before major parameter sweeps.
Avoid¶
- Treating the final table as trustworthy without inspecting masks and labels.
- Reusing a workflow on structurally different data without retuning.
- Hiding scientific decisions in file names or external notes only.