Inspection Model

The point of VIPP is not only to run a pipeline. It is to make each processing decision inspectable.

Inspection Surfaces

Surface	Use
Node thumbnail	Quick scan of what each stage is doing.
Napari inspect layer	Full-resolution inspection of the selected node output.
Pinned layer	Keep important outputs visible while editing other nodes.
Histogram panel	Check intensity distributions and threshold choices.
Input histogram	Review cutoffs for threshold-like nodes.
Label-volume histogram	Tune label filtering against the incoming object-size distribution.
Table preview	Check measurement columns, row count, and units before export.
Graph notes	Record why a parameter or branch exists.

Calculation And Display Are Different Contracts

VIPP separates a node's declared scientific calculation from the compact view used to draw it. Where a complete-population summary is described below, every required value contributes even though the result is grouped for display.

Inspector item	What is exact	What is display-only
Node thumbnail contrast	Stack Percentile/Min-max summarizes the complete node output. Native uint8/uint16 Percentile uses an exact histogram with identical CPU/CuPy limits; Min-max uses an exact native reduction; float and other-dtype percentiles retain the exact NumPy-compatible CPU path. Slice contrast instead normalizes the spatially sampled current view at the selected detail.	Low/Standard/High/Very High backing detail, Slice versus Stack scope, contrast mode, colormap, and the CPU/GPU statistics backend affect the card image only. Slice limits can change slightly with detail; Stack limits cannot.
Intensity histogram	Every finite value in the selected slice or stack is counted.	Counts are grouped for drawing: two bars for Boolean data, one for a constant image, 256 for non-negative integer data within 0–255, and otherwise 128 display bins. These bins are not the automatic-threshold bin setting.
Automatic-threshold guide	The guide uses the node's saved scope and scientific method parameters.	Line colour, plot scale, and the chart's display bins do not change the mask.
Colocalization scatter	Every ROI voxel contributes to the configured density; ROI population and the number meeting both thresholds are exact. Threshold-independent density can be reused while an exact recount runs.	Interactive rendering is capped at 1,024 bins per axis; graph scatter nodes can request up to 4,096. Colormap, logarithmic display, guide styling, and rendered output size do not change thresholds or metric tables.
Inspect or pinned image contrast	The final display range spans every finite value and zero, using exact finite extrema.	Contrast limits affect only rendering in napari, never the stored node output or downstream calculations.

Progressive inspection during a run

As each node completes in a background run, its card becomes current and its sampled thumbnail updates immediately; later cards can still be calculating or waiting. If the active cache policy already retains that node, selecting or pinning it and opening its table or metadata uses the same newly completed run-scoped payload. An otherwise prunable Low-memory intermediate gets the card state and thumbnail without forcing VIPP to retain its full-volume payload.

This is a progress view, not by itself a partial scientific commit. Normal success publishes the new workflow cache only when the complete run is accepted against the same graph and source revisions. Cancellation, a newer edit, or a superseding run removes the temporary overlays and restores the last coherent inspection state. On failure, VIPP may accept a verified source boundary; it may additionally accept completed processing data from a cleanup-failed result, but only with matching actual-implementation provenance. Unreported processing overlays are discarded and their earlier coherent inspection state is restored.

Thumbnail detail, scope, and statistics provenance

VIPP renders cards at Low (90 × 55), Standard (180 × 110), High (360 × 220), or Very High (720 × 440) detail inside the same card viewport. VIPP retains that source image and scales from it, so High and Very High can improve HiDPI display, downsampling, or maximum graph zoom; they do not guarantee a larger on-screen card. Very High uses four times the backing pixels of High. This is a spatial display choice. Changing it rerenders the graph but neither reruns scientific nodes nor invalidates cached exact Stack limits.

Stack contrast provides a stable, resolution-independent brightness window across T/Z/C by summarizing the complete node output once. Slice normalizes the selected detail's spatially sampled current view on CPU and avoids that full-output scan. It is intentionally responsive, but Low/Standard/High/Very High can change its display limits slightly. Percentile Stack contrast uses the 0.5th/99.9th percentiles. Exact native-dtype histograms avoid a full float sort for uint8 and uint16; Min-max uses a faster exact native reduction. Float and other-dtype percentiles stay on the exact NumPy-compatible CPU path in this release.

Presentation statistics have their own local Auto / CPU / Prefer GPU preference. Auto uses the complete output dtype and byte count. Before the GPU histogram path is warm, its conservative crossover is 384 MiB for uint8 and 512 MiB for uint16; both become 32 MiB after a successful GPU calculation. These measured defaults are heuristics, not universal fastest guarantees, because distribution, hardware, CUDA startup, residency, and competing work can move the crossover. Render detail is not part of this decision. Prefer GPU is the explicit override. Main compute CPU always forces statistics CPU; main Prefer GPU biases statistics Auto to GPU; main Auto and Custom leave it adaptive.

Select a node to see that work in the compact Thumbnail contrast row near the top of its inspector: Calculating… / CPU · NumPy / GPU · CuPy / CPU fallback / Error. The separate scientific CPU/CuPy/cuCIM badge remains in the node title row and still identifies what produced the node output. Hover the inspector row or thumbnail for the detailed presentation decision; keyboard What's This help and screen readers receive the same text. Stack scans publish the active node/backend/phase through the toolbar without accepting partial limits. CPU integer histograms and min-max reductions advance and cancel between bounded chunks. An active GPU kernel/synchronization or exact float/other-dtype NumPy percentile may contain a non-interruptible inner pass; the UI identifies that phase and applies cancellation at the next cooperative boundary.

Reused and replaced Inspect layers

Recalculating the same logical node/output reuses a compatible active VIPP Inspect layer and replaces only its data reference with a non-writeable view of the exact output. VIPP preserves the working view—displayed dimensions and slice positions, camera zoom, translation, and rotation—and compatible user styling such as colormap, contrast, blending, opacity, visibility, gamma, interpolation, and compatible rendering settings. Physical layer scale continues to come from the output metadata; arbitrary napari transforms are not part of the saved display profile. This lets you zoom into a structure, tune a node in isolation, and compare the same region after each calculation without the viewer jumping back to a default. Pinned layers are separate napari viewer artifacts and do not receive this saved per-output profile behavior.

Display profiles are remembered independently for each node, output port, and RGB display surface and are saved as presentation state in the workflow. When you switch to another logical output, VIPP restores that output's own compatible profile or initializes safe defaults; styling from the previous output does not leak into it. Use the inspector header's reset action when the selected output should return deliberately to VIPP defaults.

VIPP creates a replacement Inspect layer when the presentation class or layout is genuinely incompatible, such as Image versus label-ID Labels, or an RGB layout change. A Boolean mask pinned as a Labels overlay requires a uint8 presentation copy. The original cached Boolean array remains the scientific output used by downstream nodes and saving. Any pending display calculation for a superseded output is invalidated.

Large generated layers

When an inspect, pinned, or RGB-channel layer is large, VIPP first supplies a provisional dtype-based contrast range so napari can show the layer without scanning it on the user-interface thread. VIPP then calculates the exact full finite range in the background and updates the layer.

This temporary range does not mean the processing result is approximate. It is only a display window. If you manually adjust the napari contrast limits while the exact calculation is pending, VIPP keeps your adjustment rather than overwriting it; the exact finite range remains available in the layer metadata.

For large inputs, exact histogram guides, colocalization summaries, Auto Contrast, and generated-layer ranges may briefly show a calculating state. Selecting another node or changing the input invalidates stale display results.

Dragging a manual input-histogram guide does not change the input population. VIPP therefore retains the exact displayed counts while moving Binary Threshold, Hysteresis, or explicit Rescale/Clip guides, and recalculates only the node output after the short interaction delay. Its output histogram then refreshes because that population really did change. Parameters that change an automatic guide refresh that guide separately. A different connected image, slice, or histogram scope still invalidates the input counts because it selects a different population.

For Rescale Intensity, both low and high input guides are draggable. Dragging a guide derived from exact percentiles changes the cutoff mode to explicit values, retains the other exact cutoff, and recalculates the node. A click that does not move the guide is not an edit. The same no-movement rule prevents accidental Binary Threshold changes. Save the workflow after accepting a drag; guide position is a presentation affordance for a persisted scientific value.

For colocalization, threshold scrubbing moves compatible guides immediately and keeps the threshold-independent density visible while the exact full-ROI count is recomputed. Rapid requests are coalesced. A density is reused only while channels, ROI, native intensity context, and histogram definition still match. The linked inspector and resizable pop-out colormap selectors redraw from that cached density without recalculating a scientific result.

Split Channels presentation output

Split Channels produces one scientific graph output per channel, but its node card and inspector need one channel to present at a time. If downstream connections use exactly one distinct output port, VIPP presents that channel in the split node's thumbnail, napari inspect or pinned layer, histogram, output metadata, dimension controls, and Save selected output... action. Multiple branches may consume the same port; that remains one distinct used output.

If no output is connected, or the graph uses two or more different channel ports, these presentation surfaces use the saved Thumbnail channel instead. The automatic choice never changes Thumbnail channel, rewires a connection, or alters the arrays supplied to downstream calculations. It only keeps all ways of inspecting the Split Channels node focused on the same unambiguous channel.

Recommended Inspection Sequence

For segmentation:

1. Inspect the raw channel.
2. Inspect background correction or denoising.
3. Inspect the threshold mask.
4. Inspect label IDs.
5. Inspect measurement tables.
6. Save the workflow only after the intermediate stages make sense.

For colocalization:

1. Inspect both channels separately.
2. Inspect the ROI mask if used.
3. Inspect scatter/threshold guides.
4. Inspect colocalized voxels.
5. Inspect metric tables.

For skeletons:

1. Inspect the binary mask.
2. Inspect the skeleton.
3. Inspect keypoints or graph overlay.
4. Inspect branch/component labels.
5. Inspect measurement tables.