Review Images

New in 0.16.0a4

Review Images requires a runtime containing this presentation-only node. It is not available in 0.16.0a3 despite the unchanged workflow schema.

Review Images is a presentation-only sink: required Image A, optional Image B, and no scientific output image or table. Its compact graph card shows the two inputs and Open review…, without an output port, output thumbnail or No output placeholder. Its separate window has an inspector sidebar and one overlay or two selectable linked napari panes. Each pane is a cohesive framed card containing its header/selector, native image view and footer title. Use the review guide for steps.

Input contract

Kind	Accepted representation	Controls
Scalar intensity	Real numeric image with explicit YX, ZYX, TYX or TZYX axes	Black/white levels, colour map, fixed colour scale, visibility cutoff, opacity, visibility and scalar-only 3D rendering
Binary mask	Boolean array with explicit spatial/time axes	Foreground colour, opacity and visibility; background 0 transparent
Label image	Explicit labelled image carrying non-negative integer IDs	Categorical colours, opacity and visibility; background 0 transparent
RGB/RGBA	Explicit encoded RGB/RGBA metadata with a trailing component axis of length 3/4	Original colours, opacity and visibility; no scalar contrast/colormap controls

RGB is never inferred merely because a dimension has length three or four. Encoded colour inputs use uint8 or finite float values in the explicit 0–1 display range; unsupported encodings must be converted explicitly upstream. Ordinary C stacks need Extract Channel or Composite to RGB upstream. Ambiguous/inferred axes need an explicit declaration upstream. Review requires resident current NumPy results; it is not a new large-file loading strategy.

Paired inputs must agree in spatial and time axes, shape, scale, translation and compatible units. There is no hidden registration, time broadcasting, axis reorder, interpolation, intensity normalization or scientific casting. The original scientific buffers are read-only.

Display and navigation

• Top toolbar: Side by side or Overlay; 2D slice or 3D volume where the spatial data supports it; Link viewpoint, Axes and Scale bar. Icons follow the active palette without replacing the controls' text labels.
• 3D → 2D: switching a volume to 2D resets the plane to XY and exposes Z navigation. Deliberately selecting XZ or YZ slices remains supported; reopening an already-2D saved recipe preserves its selected plane.
• Pane selectors: A, B or A + B above the left and right panes, where a second input is connected.
• Shared navigation below the panes: T and spatial-slice positions, where applicable, always stay synchronized. The spatial-slice label follows the currently undisplayed axis (Z, Y or X). The editable number is the one-based current position; / N is a separate total-count label immediately after the box. The calibrated world-coordinate caption remains visible beside it.
• Link viewpoint: optionally synchronizes pan, zoom and camera rotation. Disabling it allows different camera viewpoints, not different timepoints or slices. It is not an alignment transform or a registration guarantee.
• XY / XZ / YZ: orthogonal slice planes in 2D, or axis-aligned viewpoints in 3D. Oblique selects an angled 3D view where supported. These are display orientation presets, not resampling or scientific image transforms; they sit below the panes beside Fit views.
• Axes: show or hide the 3D orientation axes in both panes together. Enabled by default and effective only in 3D; this display setting is saved with the review recipe. Arrows and labels appear on a transparent background without an opaque box. This is presentation only; source arrays are unchanged.
• Scale bar: beside Axes, show or hide scale bars in both panes in 2D and 3D. Enabled by default, with a transparent background. A bar appears only where compatible spatial units make calibration meaningful; the switch does not supply or repair calibration. Visibility is saved with the review recipe.
• Link A/B contrast: optional for two scalar inputs; colours and opacity remain independent. Off by default.
• Fit views: fit the review data. Reset display: restore presentation defaults without recalculating the workflow.
• Read-only · Native analysis resolution: status text in the same bottom row as Oblique/XY/XZ/YZ and Fit views, rather than a separate row.

Scalar black/white levels must be finite and ordered. Opacity ranges from 0 to 1. The visibility cutoff is display alpha, not a numerical segmentation threshold. Ordinary graphics texture/colormap precision applies, so displayed boundaries are not an exact scientific acceptance test. The scalar colour bar retains its black/white mapping as low values are hidden. Lock colour scale prevents edits to the scalar black/white limits when a fixed range is meaningful.

Display recipes, including the orientation preset, axes/scale-bar visibility and per-input 3D rendering controls, are stored per review node in workflow UI metadata. Current camera, timepoint and slice positions are runtime-only navigation. They do not become processing parameters or scientific cache keys. Review nodes have no output for downstream analysis, batch image publication or numeric export. Use Compare Images for quantitative RMSE/correlation/SSIM/PSNR instead.

Live updates and current results

A visible open review retains its last complete inputs during upstream edits and calculation, clearly marked waiting/noncurrent. It does not initiate recalculation. After calculation finishes and the complete, current reviewed inputs are accepted, A/B are replaced together; partial previews do not replace one side independently. Same-grid replacement preserves camera/navigation, contrast and display styles.

Failure, cancellation, supersession or a missing required review source cannot publish a current pair. An unrelated branch may fail while the complete A/B results are accepted; current review status is not a claim that every node in the workflow succeeded. A one-input review applies the same acceptance rule to its connected Image A.

Grid changes (including axes, shape or calibration), image-kind changes and connected-input presence changes require reopening. There is no implicit resampling. Visible review inputs are presentation retention targets even in Low-memory cache mode, still subject to the normal memory/resource guards. Retention does not make arbitrary volumes fit memory or turn waiting data into current scientific results.

3D colour-rendering boundary

Scalar volumes offer 3D rendering controls per input, only in 3D:

Choice	Meaning and controls
Maximum intensity (MIP)	Default. Maximum displayed intensity along each viewing ray, without depth weighting or surface lighting.
Depth-weighted intensity	Nearer structures receive more prominence. Attenuation defaults to 0.05 and must be finite and non-negative. It is not opaque foreground occlusion.
Surface (isosurface)	A lit display surface at a source-intensity value, initially the black/white midpoint if no level is set. Interior structure can be hidden. It does not calculate a scientific segmentation or surface mesh.

Depth weighting and isosurface lighting change apparent brightness. They are not quantitative colour-scale/intensity review. In these modes the colour bar is labelled Base colour scale (before shading): its numeric limits describe the base mapping, not the shaded pixel brightness. Lock colour scale protects black/white limits only; rendering mode, attenuation and surface level remain independent presentation controls. Use 2D slices or unshaded MIP to inspect the base mapping; numerical analysis still uses scientific results, not renderer pixels. Two scalar inputs in A + B are combined additively, without physically correct inter-layer occlusion.

These controls are saved presentation metadata and do not alter scientific arrays, measurements or cache keys. RGB retains the component-MIP adapter below; the scalar choices do not change mask or categorical-label rendering. Categorical label volumes use napari's native volume presentation. Boolean masks use a shaded 3D surface rather than a flat foreground maximum-intensity silhouette. Interpolation and lighting approximate the displayed surface; they do not smooth or change the Boolean mask. Sampled voxel facets can remain visible, so this is not an exact geometric reconstruction. Mask 2D slices use nearest-neighbour display with a transparent zero background. Native RGB/RGBA 2D slices retain encoded colours.

Ordinary Boolean masks share their source buffer for 3D display. The rare case of Boolean True stored with noncanonical bytes requires a separate read-only display buffer, costing one byte per voxel. VIPP checks available host memory before allocating it and reports an actionable error if headroom is insufficient. The scientific mask and its stored bytes are never changed.

RGB 3D uses a clearly labelled display-only adapter: separate red, green and blue component maximum-intensity projections combined additively. Component maxima can originate at different depths. This is not voxelwise RGB volume ray-casting, not a single joint-channel projection and not parity with the custom napari-racc shader. RGB examples therefore open in 2D slices.

RGBA 3D is unavailable because this adapter cannot preserve per-voxel alpha. Use 2D slices; VIPP does not silently discard the alpha channel.

Two panes may use additional graphics memory even when they share read-only source arrays. Large-volume memory/performance, acquired-microscopy accuracy, minimum-dependency versions and native Linux/macOS remain unqualified for this feature. Local Windows checks and synthetic fixtures are not those release gates. Renderer output is for visual review, not numerical validation.