ScanRuler by CNC Kitchen

Use it live at https://scanruler.stefan-755.workers.dev/ — nothing to install.

Check a 3D scan against what it should have been — entirely in your browser. Three workspaces share the loaded scan:

• Elements — pick features and get automatically fitted spheres, cylinders and planes, construct points, lines and planes from them, and create the distance and angle dimensions you actually need between them, the way metrology software like GOM Inspect does it. Aimed at ball bars and other calibrated artefacts. The measured elements can serve as datums for a 3-2-1 alignment into the global coordinate system, and be exported as a STEP file of analytic geometry.
• Deviation — paint how far the scan strays from what it should have been as a colour map over the part, measured either against a nominal CAD part (loaded as a mesh or as a STEP file tessellated in the browser, and best-fitted onto the scan) or against one fitted element — is this face flat, is this bore round, does this surface sit where the datum says.
• Thickness — paint the wall thickness of the part itself over it, with no reference model and no alignment: load one file and measure.

Everything runs locally. Your scan files are processed in your browser and never leave your computer.

First time here? An amber ring pulses around the one control to press next — open the scan, fit an element, measure between them — and moves on by itself as each step is done. It is read off the state of the work rather than from a script, so doing the steps out of order or undoing one keeps it honest, and it never rings a control that cannot be pressed yet. A workspace stops hinting once you have carried it through on two separate visits, so a reload always gives them back to you the first time; ◉ HINTS in the status strip switches the guidance off outright, and switching it back on starts it over.

Elements: fitting and measuring features

How to use it

1. Open your scan (STL, PLY, or OBJ — meshes only, units assumed mm), or drag & drop it anywhere in the window. (CAD goes in the Deviation workspace's reference slot, which also takes STEP.)
2. Press Sphere, Cylinder or Plane, then click a point on that feature in the 3D view. The tool automatically selects the surface around your click — the spherical patch without leaking onto the connecting rod, the cylinder wall without climbing onto its end faces, the flat face without crossing an edge — and marks it in the colour the element will get, with the fit itself shown as a neutral grey ghost. If the scan of that feature is broken into unconnected patches, click a point on each one — every pick feeds the same fit. Press Create … when the preview looks right (Undo point drops the last pick, Cancel or Esc discards). The finished element stays on screen in its own colour.
3. Press New dimension, pick the measurement type, and select the two elements to measure between. The value previews live; Add dimension keeps it. Copy summary puts everything on your clipboard.
4. Nothing is final: the ✎ key on any element or dimension row re-opens it in the box it was made in — see Changing what you have made.

Each fitted element reports its size and its sigma — the RMS deviation of the scan from the ideal geometry, i.e. how round, how cylindrical or how flat the scanned surface actually is. Cylinders also report the length and the arc of wall the fit rests on, planes the size of the measured patch.

Marking the surface by hand

Automatic surface selection is right almost all of the time, and wrong exactly where a scan is worst: a rounded edge that lets the region creep onto the next face, a bore broken up by noise, a feature you deliberately want to measure on one clean band rather than on everything the tool can reach.

Set Surface to Marked by hand in the element being created and the fit takes what you mark, and nothing else. The tools are the same ones the local fine fit is marked with — Navigate, Window, Brush, Lasso, the Erase switch, Mark faces pointing away too and Clear marking — and they behave identically here: nothing is armed until you pick a gesture, left-drag marks and right-drag rubs out while one is, Shift-drag still orbits, and Navigate or Esc hands the plain drags back to the camera without touching what is already marked. A second Esc discards the element, the way it always has.

The brush Ø in millimetres sets how wide a brush stroke is — it starts sized to the part. A ring on the surface under the cursor shows the footprint before you commit to it, in the element's colour while marking and dark while rubbing out. The marked surface wears the colour the element will get, the fit re-runs each time you lift the button, and Clear marking starts over.

A gesture takes whole triangles — the ones it actually covers — so what lights up is exactly what the fit is given. It never reaches through a thin wall or around an edge either, unless you ask it to with Mark faces pointing away too.

Everything else is unchanged — the same Gaussian best fit, the same outlier cut-off, the same reported sigma — so a hand-marked element and an automatic one are the same measurement, differently aimed. Changing Used points re-fits a hand-marked element on exactly the surface it was marked with.

Whichever tool you pick stays picked from one element to the next; switching workspaces puts it back to Navigate, so a gesture is never holding the mouse because of something you did somewhere else.

Which way the surface faces

Backfaces in the status strip colours the far side of every triangle. A scan is a surface, not a solid: where it has a hole, you are looking at the inside of the wall behind it, and in plain grey that reads as part. Switched on, it reads as a hole — which is also how an inverted normal gives itself away.

How the part is shown

View in the status strip picks what the part is made of and how it is lit, and it is remembered per browser. The stage stays the same grey either way.

Studio grey is the default: a matt grey part, evenly lit — the quietest thing to lay a coloured map or an element tint over.

Scanner blue is the one every handheld scanner's own software puts on screen: a glossy blue part under a hard light. It is worth switching to whenever you are looking at the surface rather than at a measurement. A tight specular highlight travelling over the part is the best way there is to see the shape of it, and tool marks, print layers, the faceting of a coarse mesh and the ripple of a bad scan all show up in that highlight long before they show up in matt shading. Every viewport follows — the reference half of the split view and both halves of the point picker with it.

Nothing measured changes colour with the scheme. The element tints, the deviation and thickness ramps and the grey a map paints where nothing was measured are all exactly the same in both: a reading that shifted with the lighting would be worth nothing. What does change is the reference where it is overlaid on the scan — amber against a blue scan instead of blue against a grey one, since its whole job there is to not look like the part underneath it.

Given a frame of its own it needs no such contrast, and is better without it: in the split view and in both halves of the point picker the reference is the same material as the scan, in the scheme's own bare-surface colour, so the only thing that differs between the two pictures is the shape.

Elements: fitted, picked and constructed

Elements don't have to come from the scan surface. Every element type offers a choice of creation methods:

Element	Created by
Point	pick on the scan surface · typed-in coordinates · midpoint of two points · intersection of a line/axis with a plane
Line	through two points · a cylinder's axis · intersection of two planes
Plane	fit to the scan · through three points · offset from a plane · midplane of two planes · typed-in normal + point
Sphere, Cylinder	fit to the scan

Anywhere a point is asked for, a sphere stands in with its center; anywhere an axis is asked for, a cylinder stands in with its axis — the standard metrology reduction. Constructed elements re-evaluate automatically when a source element is re-fitted, and deleting an element removes everything built on it.

Dimensions

Measurements are created deliberately, not generated for every pair — with a handful of elements the all-pairs list explodes, and half of it (distances between planes that will never be parallel, say) means nothing. Following GOM Inspect / PC-DMIS conventions:

Dimension	Value
Point – Point	center distance, with signed ΔX/ΔY/ΔZ components; between two spheres also surface gap (− radii) or outer span (+ radii)
Point – Axis	perpendicular distance to the axis
Point – Plane	perpendicular distance, signed along the outward plane normal
Axis – Axis	offset of near-parallel axes, or the distance at closest approach of skew axes
Axis – Plane	distance from the middle of a near-parallel axis, signed
Plane – Plane	distance between near-parallel planes, from the center of the first
Angle: Axis – Axis	0–90°
Angle: Axis – Plane	angle to the surface, 0–90°
Angle: Plane – Plane	via outward normals, 0–180° (opposing faces read 180°)

Fitted geometry is treated as finite: a plane is only the patch that was measured, an axis only the section the fit rests on. A dimension that has to extrapolate past that — a projection landing off the measured patch, skew axes whose closest approach lies beyond the fitted sections — says so with a warning, or refuses the value outright when it would be meaningless (planes more than 3° from parallel have no distance; use an angle dimension instead).

Extending a plane or a cylinder past what was measured

A fit is only as big as the surface it found: a plane stops at the edge of the patch the scan covered, a cylinder at the ends of the band the fit rests on. That is the right answer for a measurement and the wrong one for a datum you are about to hand to CAD, where a face usually needs to reach past the part and a bore needs to run through it.

While a cylinder or a plane is being created or edited, an Extend block appears under the preview: two millimetre fields for a cylinder (one per end) and four for a plane (one per edge, ±U and ±V — the patch's own two axes). The same sides carry grips in the viewport — arrows on the ends of a cylinder, bars along the edges of a plane — and dragging one types itself straight into its field. Make square grows the shorter axis of a plane out to the longer one, evenly on both sides so the patch keeps its middle; Reset puts everything back on the measured surface. Negative values pull an edge back in, as far as leaving the element a size at all.

Grips are grabbed with a plain left-drag; the camera steps aside for them the way it does for the marking brush, and Shift-drag still orbits. While a marking tool is armed both plain drags belong to it, so choose Navigate to reach the grips.

Nothing measured changes. The extension is carried beside the fit, not in it: the sigma, the reported patch size or fitted length, and the warning a dimension gives when it leaves the measured surface all go on describing the scan. What changes is the shape on screen and the shape in the STEP file — and the summary notes what an extended element is drawn at, beside what it was measured as.

Changing what you have made

Every element and every dimension carries a ✎ key next to its hide and delete keys. It re-opens the thing in the same box it was created in, with everything it was built from already in place:

Re-opening a…	brings back
fitted element	the points that were clicked on it — the fit re-runs and previews at once, so more picks or Undo point change the surface it rests on
hand-marked element	the marked surface itself, back on the part under the marking tools, ready to be added to or rubbed out
picked point	the point, so a click on the scan moves it
constructed element	its source elements and typed-in numbers, in their fields
dimension	its type, its two references and the sphere anchor

The creation method can be changed on the way through — a plane fitted to the scan can be re-made from three points — and the Name field renames it. What comes out is the same element: same id, same colour, same place in the list. Everything measured against it — dimensions, and constructions built on it — re-reads the new geometry instead of being rebuilt, so correcting a bad fit costs one edit rather than a rebuild of everything downstream. A construction cannot be pointed at itself or at anything already built on it, so no loop can be created. Cancel or Esc leaves the original untouched.

A dimension that changes group in the process — a distance turned into an angle — takes the next name of the group it has become, unless you named it yourself.

Aligning the part (3-2-1)

A scan arrives in whatever coordinates the scanner happened to use. Align part (3-2-1) sets where X, Y, Z and the zero point sit on the part instead. Three steps, each after the first optional:

1. Level with — a flat face, a cylinder, a line, or 3 points picked straight on the scan. Its direction is turned exactly onto the chosen axis (+Z by default) and levels the part.
2. Rotate with (optional) — a second direction (element or 2 picked points) so the part cannot spin around the first axis.
3. Zero point (optional) — a point, a sphere center, or 1 picked point that becomes the origin.

The slots mix and match freely — a fitted plane for levelling, two picked points for the rotation, an existing point for zero all in one alignment. Whatever levels or rotates also sets its own zero: a levelling face ends up at height 0, a cylinder lands on the axis it points along, and the zero point covers whatever is left. Fill the slots by clicking elements in the 3D view or picking points on the scan (they stay marked as Point 1, 2, 3 while you work). The part swings into the pose it would take as soon as a slot has what it needs, and again whenever you change the axis it points along, so the alignment is judged by looking at the part rather than by applying it to find out — with the panel reading how far it would rotate and move. Measured elements make the most accurate references — each one averages the thousands of scan points behind its fit, where a picked point is a single spot of scan noise.

Move / rotate by numbers does the same thing by hand: type how far to move (mm) and turn (°) along the global axes. The part turns about the zero point — about X, then Y, then Z — and moves after that. Useful for nudging a part into a nicer pose, or for applying a known offset exactly.

Applying an alignment or a manual move transforms the scan and everything measured on it — every element, dimension and typed-in coordinate moves with the part, so a fitted sphere stays on its ball and every dimension keeps its value. Re-fits after the alignment measure in the new frame. Reset alignment undoes all of it at once and puts the part back exactly where the scanner delivered it. (A scan-to-reference best fit from the Deviation workspace was measured in the old frame and is cleared — realigning there takes one click.)

STEP export

Export STEP writes the created elements to a STEP file (ISO 10303-21, AP214) as analytic geometry, not tessellation, in either of two forms — STEP as beside the button picks which, and the choice is remembered.

Solids & faces (the default) hands CAD geometry it can build on:

Element	Written as
Plane	a bounded planar face — an ADVANCED_FACE with four real edges, in an open shell — so it can be sketched on, offset and referenced
Cylinder	a closed solid body: the fitted wall, capped at both ends by flat lids
Sphere	a closed solid ball, two hemispheres meeting at an equator
Line, Point	a trimmed line and a point, as they always were

Each body is its own named shape representation, tied to the part the way the bodies of a multi-body file are, so the element names arrive in the CAD tree.

Construction surfaces is the older form, and still the honest one for handing over datums: every element as a trimmed analytic surface or curve in a single GEOMETRIC_SET, with no topology at all — planes as planar patches at their extents, cylinders and spheres as trimmed surfaces at their fitted radii. It is unmistakably reference geometry rather than a part, and the safer choice for an importer that chokes on bodies.

Either way the size written is the size on screen, extensions included, and coordinates are millimetres in the current frame — so aligning the part first hands CAD the elements in the datum system.

The viewport uses a parallel (orthographic) projection so nothing is foreshortened, and rotates freely around the model's bounding-box center with no fixed up-axis — you can turn the part all the way over without hitting a pole. Left-drag to rotate, right-drag to pan, scroll to zoom; the XYZ gizmo in the bottom-right corner shows the current orientation. On a touch screen the tablet gestures do the same three things — one finger turns, two fingers pan and pinch to zoom, a tap picks — and with the marking brush armed the single finger paints while two fingers still move the part.

Fitting uses a Gaussian best-fit (orthogonal least squares) with GOM-style used points presets (all / 3σ / 2σ / 1σ, default 3σ). The initial estimate is made robust with LMedS/RANSAC, and the point selection is a model-guided region grow over the mesh surface with normal-direction checks — so a click anywhere on a feature finds exactly that surface, even when it's fused to the rest of the part. Spheres and planes are solved in closed form (algebraic fit refined orthogonally, and the total-least-squares plane through the point cloud); the cylinder's five degrees of freedom are solved by damped Gauss-Newton, from a starting axis taken from the scatter of the surface normals.

Validated against GOM Inspect on a real structured-light scan: center distance agrees within a micrometer (148.6398 mm vs 148.64 mm), with matching point selections and fit sigma. The cylinder and plane fits are covered by unit tests against synthetic geometry with known dimensions, since the included scan has no such feature.

Deviation

The Deviation workspace paints how far the scan strays from what it should have been over the part itself. Measure against at the top of the panel decides what "should have been" means:

• Reference model — the whole nominal part, best-fitted onto the scan. Answers is this the shape it was drawn as.
• Fitted element — one plane, cylinder or sphere measured on this same scan in the Elements workspace. Answers is this face flat, is this bore round, does this surface sit where the datum says.

Both paint the same map, read through the same colour scale, with the same statistics, pinned readings and report underneath. They differ only in what it takes to get there — and both maps are kept, so switching between them loses neither.

Deviation from a nominal part

With Reference model chosen, an empty stage asks for the two models it needs, each its own drop target — drag and drop works anywhere in the window, and whichever slot is still empty takes the file. Once a part is on the stage the prompt gets out of the way for good: whatever is still missing is asked for by its row in the panel and a line above the model.

The scan is a mesh, as always. The reference takes a mesh too, or a STEP file straight from CAD — see below.

Both parts are on the stage as soon as both are loaded, the reference drawn as a translucent ghost over the scan. Press Align automatically and the scan walks onto it pass by pass — the refinement streams its intermediate poses out of the worker, which costs one matrix write per pass against tens of milliseconds of closest-point queries.

The reference never moves. A nominal part is the datum a measurement is taken against, so the alignment is applied to the scan and the world ends up in the reference's coordinates. Everything measured on the scan — the deviation map, pinned readings, and any elements fitted in the other workspace — rides along with it, so a fitted sphere stays on the ball it was fitted to.

The map is measured as soon as the fit lands, and the reference stands down once there is a map on the scan. Separate Show reference and Show scan switches bring either back; with the scan switched off the reference turns solid, which is how you check it is the right part — an aligned reference otherwise lies inside a scan of nearly the same shape and loses the depth test almost everywhere.

Hover the part for the deviation under the cursor, interpolated across the triangle rather than snapped to a vertex, and click to pin a reading where you want a number to stay.

Both parts, side by side

◫ Split view in the status strip puts the scan and the reference in two viewports next to each other with one camera between them: turn, pan or zoom either half and the other follows, whichever half the pointer is in. Because the fit carries the scan into the reference's coordinates, both halves are showing the same world — so a feature on the left sits exactly where its counterpart sits on the right, and the only thing that differs between the two pictures is the part.

The overlaid ghost stands down while the split is open: the reference has a half of its own, and the ghost was only ever a way of getting two parts into one frame. Everything the scan carries stays on the left — the map, the reading under the cursor, the pinned readings, the marking tools. Nothing is picked or measured on the reference half; it is the shape being compared against.

Both halves are one material under one light — the scheme's bare-surface colour and finish on either side, rather than the ghost's contrasting blue. Two pictures of the same part in the same grey leave the shape as the only difference between them, which is the comparison you opened the view to make. Switch the colour plot off and it is exactly that; leave it on and the map is the one thing marking the scan out from its nominal. Under Scanner blue it is worth the look on its own: the highlight lies along the same edge in both halves, scalloped into mesh facets on the scan and dead straight on the CAD part beside it.

It does not wait for an alignment either. Opened on an unfitted pair it shows each part where it actually is, which is how you check the reference is the part you meant before spending a fit on it.

The colour plot, off

▩ Colour plot in the status strip stops painting the map onto the scan and leaves the bare surface. The scale goes with it — histogram, figures and all: it is the key to colours that are no longer on the part, and being left with the part is the whole point of switching them off. It is there for the times you want the shape rather than the reading — the form of a face, a hole in the scan, the marks a finish left — which is most of what the reference beside it is for.

Nothing measured is lost by not looking at it. The map is still measured underneath, the reading under the cursor still reports it, pinned readings stay pinned, and switching it back on brings the scale back reading exactly what it read before — nothing is re-measured.

The reference straight from CAD

The nominal part is whatever the CAD system says it is, and exporting it to STL first means choosing a tessellation in a dialog that has nothing to do with the measurement. So the reference slot also takes STEP (.step / .stp, ISO 10303-21, AP203 / AP214 / AP242): the file's exact surfaces are tessellated here, in the browser, by meshStep. Coordinates come out in millimetres whatever unit the file declares, so an inch-native export needs no conversion beforehand.

How finely it is tessellated matters, because chord error is a systematic term in every reading taken against a curved face. The tolerance is scaled to the part — 0.01 mm on a 100 mm one, tighter on smaller — which puts it about a tenth of what a good structured-light scanner resolves, so the conversion disappears under the scan rather than being measured by it. The figure is reported in the status strip and stays on the reference slot, since it is the floor under everything the map says.

What it deliberately does not do is subdivide by length. Triangle count is driven by curvature alone: a flat face is exact at two triangles however large it is, and the usual size-adaptive default spends 238 510 triangles on a 20 mm cube that 12 describe perfectly. A bracket arrives as a few thousand triangles of exactly the right shape instead of a million of the same shape.

The conversion is audited, and the result is not taken on trust. A STEP file whose faces come through with cracks or holes leaves the solid without a reliable inside — and inside is where the sign of every deviation comes from — so that is reported as an error against the reference, with the advice to export a mesh from CAD instead. A file that only needed heuristic repair says so more quietly, in the status strip.

A scan is never a B-rep, so the scan slot does not offer STEP and turns one away by name if it is pushed at it; a STEP file dropped anywhere in the Deviation workspace goes to the reference.

The best fit

The fit is rigid — rotation and translation only, no scale. That is deliberate for a scanner accuracy tool: a scale error is one of the things you are trying to see, and a 7-parameter fit would quietly absorb it.

Alignment is a point-to-plane ICP with adaptive outlier rejection, and the increment is taken about the centroid of the correspondences so the solve stays conditioned on a part sitting far from the origin. The starting pose is picked by trying candidates and keeping whichever actually fits: the identity, which is right whenever both files already share a frame, plus the 24 rotations that map the scan's principal axes onto the nominal's. All 24 are needed, not just the four sign flips, because a nearly cubic part has nearly equal principal moments and a partial scan of it can rank its axes differently from the whole part — no amount of sign flipping recovers that, the axes have to be permuted.

Candidates are scored on the mean distance from every sample to the surface, capped so outliers cannot dominate. Scoring only the pairs that survive outlier rejection — the obvious thing — is actively wrong: the cut-off is a multiple of the median, so a pose that slides until only a well-fitting patch still corresponds scores better on fewer pairs, and the fit walks off the part chasing it.

If the automatic match fails or reports itself ambiguous, press Align by picking points… for a split screen with the scan on one side and the reference on the other, each freely rotatable. Clicks alternate — a feature on the scan, the same feature on the reference — and three pairs are enough. The picks only fix a coarse pose, solved in closed form by Horn's absolute orientation, and ICP does the rest, so they only have to be roughly right. Points that land nearly in a line are rejected as you place them: the rotation about that line would be unconstrained.

Local fine fit

The global fit weighs every point of the scan alike, which is right until the scan contains surface that is not the part — developer spray, print supports, the riser it was scanned on, fixturing, geometry the reference simply does not have. Those are fitted on purpose, they pull the whole alignment off, and no number of passes shakes them loose.

Local fine fit is the second pass: mark the surface that really is the part, and fit on that alone. It starts from the alignment already in hand and only corrects it — it is not a way to find a pose from scratch, so run the global fit first.

Four modes, one of which is always live:

Navigate	Marking off. Orbit, pan and zoom exactly as everywhere else. This is what the tools open in and return to after a fit — what is already marked stays marked.
Window	Drag a rectangle. Everything inside it is marked — the fastest way to drop a riser or a whole scanned-in fixture.
Brush	Drag over the surface with a round brush of a set diameter, as in the Elements workspace. For working along an edge.
Lasso	Draw a free outline; everything it encloses is marked. For a patch of spray that follows no straight line.

The three gestures are additive and all undone by the same gesture with the right button (or with Erase switched on, or with Alt). A live gesture takes both plain drags for as long as it is on, so Navigate — or Esc, or clicking the live tool again — hands them straight back. Shift-drag orbits while a gesture is live, and the middle button is untouched throughout.

Esc backs out one step at a time: the first press stands the gesture down and returns the camera, the second closes the local fine fit and clears the marking. Never both at once — Esc is the key you reach for to get the mouse working again, and losing a marking to that would be a trap.

While the tools are out, the rest of the faceplate fades back: loading models, the global fit and reading the map all belong to another step. It is a fade, not a lock — anything there still works, and comes back to full strength under the pointer.

Mark faces pointing away too decides whether a gesture reaches through the part. Off — the default — only surface turned towards you is taken, so a window over a closed part cannot quietly mark the far wall along with the near one. On, the gesture goes straight through, which takes a whole rib or boss in one sweep and is also the escape hatch for a scan whose normals are inverted. There is no depth test behind this, only the facing test: that is how CAD selection works everywhere, and it is why the switch exists.

Max search distance (1 mm by default) is a hard gate on the fit: a marked point that finds no reference surface within it contributes nothing, and counts against the pose exactly as a miss does. It is what stops a marked patch from sliding onto a neighbouring feature that happens to fit it better. A global fit already has the part within a few tenths, so a millimetre is generous; raise it and the fine fit can find a different feature and settle there instead. If nothing at all is in reach, the fit refuses rather than answering.

A selection that faces essentially one way is flagged: a single flat patch fixes the distance across itself and leaves the part free to slide along it and to spin about its normal. Mark a second surface facing another way.

The marking excludes surface from the fit, never from the reading — the map that follows is still measured over the whole scan, so the supports and the spray are still coloured, just no longer voting on where the part sits. Back to the global fit puts the whole-scan alignment back, exactly as it was.

The map

Deviation is the signed distance from each scan vertex to the nearest point on the nominal surface, positive where the scan sits outside the reference. The scan is queried against the nominal and never the other way round, because only the nominal is watertight and so only that direction has a well-defined inside.

The sign comes from an angle-weighted pseudonormal of whichever feature the closest point actually landed on — face, edge or vertex — not from the nearest triangle's own normal. Against a CAD part full of sharp pockets and bores a scanned surface projects onto those seams constantly, and signing by the face normal speckles every edge of the map with false inside/outside flips.

Reading the map:

Control	Unit	What it does
Range ±	mm	Half-width of the colour scale. Defaults to the rounded 95th percentile of the absolute deviation, so a handful of outliers on a fixture edge cannot flatten the whole part to green — and never opens wider than ±1 mm on its own, so a great many of them cannot either. A part that really is further out than that says so in the dark end caps; widen the scale by hand to read it.
Bands	—	Continuous jet, or quantised into bands when you want iso-deviation contours.
Histogram	—	The distribution, drawn beside the scale and sharing its axis, plus min / max / mean / RMS / sigma.
Max search distance	mm	How far a scan point may look for reference surface. Beyond it there is nothing to deviate from, so the surface is left plain grey and kept out of the statistics. Display only — it never affects the alignment, and moving it re-colours instantly.
Tolerance ±	mm	The band the within ± x mm figure under the scale counts. It does not change the colours.
▩ Colour plot	—	Whether the map is painted onto the scan at all (status strip). Off leaves the bare surface and takes the scale with it; the map stays measured, and the reading under the cursor and the pinned readings go on reporting it.

The ramp is jet — blue through cyan, green, yellow to red — pinned so that zero is a saturated green, with dark caps beyond each end so a reading that is off-scale is never mistaken for one that is merely large.

Validated against the included test pair (side bracket left.stl as nominal, block-marius.stl as the scan, 1.43 M triangles): the fit converges to 0.072 mm RMS, and a scan displaced by a random rotation and translation comes back to within 0.9 µm of the fit found in place.

Deviation from a fitted element

Choose Fitted element under Measure against and every plane, cylinder and sphere you measured in the Elements workspace appears on the part, each in its own colour. Click one on the model — or pick it from the dropdown — and that is the whole setup: no reference file and no alignment, because the element was fitted on this scan and is already in its frame. There is no Measure button either: the distance to a plane, a cylinder or a sphere is a handful of flops per vertex, so the map is computed on the main thread and simply follows every change to it.

The element in use is reduced to its outline, for two reasons. It lies exactly on the surface being read, so a translucent body there would wash the colour the reading is made of — and on a map the colour is the measurement. And an outline is not something clicks resolve through, so a click on the map it covers still pins a reading rather than re-selecting the element under it. The elements on offer stay bodies, faded, because a body is what you can aim a click at. Show elements on the part takes them all off for a clean screenshot, and with them the clicking; an element hidden by its own eye in the Elements workspace stays hidden here too.

A point and a line are not offered. The distance to them is unsigned, so there is no zero for a scale that runs warm one way and cool the other.

Three things turn a raw closest distance into a measurement here.

The region is the element as drawn. A plane is infinite and a cylinder is an endless tube; taken literally, a plane would paint a slab clean through the part. So a vertex counts only where it falls within the element's own extent — which is the extent the grips set, so extending a plane in the Elements workspace grows the measured region with it. That is how a plane fitted on one pad becomes a flatness map of the whole face it belongs to, and the outline drawn over the map is exactly the boundary of what was measured.

The sign follows the material, not the fit. Positive deviation is always too much material, so the tool has to know which side of the element the part is on. It reads that off the scan's own normals around the element as you choose it, because a fitted plane's normal points whichever way the fit happened to choose, and inside a bore the material is on the inner side — where a raw radial distance runs backwards. Flip turns it round when the detection is wrong.

A surface facing the wrong way is not the surface being measured. A plane fitted on the top of a 10 mm plate reaches the underside of it, which lies squarely inside the footprint and would be reported as ten millimetres of missing material. Surface must face the element leaves it out, the same way the wall thickness search steps over a surface that does not face back.

Control	Unit	What it does
Measure to	—	The element the map is measured against. Choosing one measures it — as does clicking it on the part.
Show elements on the part	—	The candidates on the model, which is also what makes them clickable.
Material side	—	Which side of the element the material is on, detected from the scan. Flip inverts the whole map.
Max search distance	mm	How far off the element a point may be and still be measured. Display only, so it can be dialled either way with the map following immediately. How far the element reaches sideways is set by extending it.
Surface must face the element	°	Leave out scan surface whose own normal points away from the element's — the far side of a wall, the back of a rib.

Everything below that — the colour scale, the bands, the histogram, the tolerance tally, hover readings and pinned ones, Copy report — is the same instrument as for a reference part, because it is the same map.

Validated end-to-end against a generated 20 mm CAD cube (npm run e2e:element-deviation): two planes fitted on two of its faces, chosen and swapped by clicking them on the model, and the map on each reading the face as flat to 0.000 mm, leaving the underside out while the facing filter is on, and reporting it as exactly −20.000 mm with the filter off.

Wall thickness

Switch to the Thickness workspace, load a part, and press Measure wall thickness. There is no reference model and no alignment — the measurement is of the part against itself.

The two methods

Ray — from every vertex a ray is fired straight into the material along the inward surface normal, and how far it travels before it comes out the far side is the wall thickness there. Exact wherever the two faces of a wall are parallel; a little long where they are not, by 1/cos of the angle between them. It runs at roughly 150 000 vertices per second — under five seconds for a 1.4 M-triangle scan.

An opening angle spreads a cone of rays around the normal and takes the shortest, which finds the narrow way across a chamfer or a tapered rib.

Sphere — a sphere is placed halfway along that ray and grown until it touches. It can never read longer than the ray and usually reads shorter, because it is not tied to the ray's direction: it finds a wedge square across, and at the corner of a block it reports the block rather than the long diagonal the blended normal points down. One extra closest-point query per vertex.

It is deliberately not the largest sphere tangent to the surface at the point. On a mesh from CAD nearly every vertex sits on a sharp edge, and no sphere of any size touches an edge from inside without poking out of it, so that definition collapses to nothing over most of the part — measured, on the included bracket, at a mean of 3.3 mm against the ray's 16.3 mm. Centring the sphere on the crossing has no such degeneracy and gives 7.7 mm.

What counts as the far side

Both faces of every triangle count as an exit. A scan is not reliably wound, and a far wall whose winding disagrees with the near one would otherwise be invisible to the ray looking for it.

A surface nearly edge-on to the ray, though, is not the other side of a wall — it is a rib the ray is running alongside, or the rim of an open scan. Max. deviation of normals is how far from squarely facing the ray a surface may be and still count; anything flatter is stepped over and the search goes on behind it. At the default 60° it also stops a 30° cone escaping sideways through a convex edge, which is why those two defaults belong together.

A ray with nothing behind it inside the search limit is left unmeasured — bare grey, and out of the statistics. On a closed part that is almost nothing; on a scan of one side only, it is most of the part, and the readings you do get are the distance across the body rather than across a wall. Wall thickness is only meaningful on a surface that has a back to it.

Reading the map

Control	Unit	What it does
Method	—	Ray along the normal, or the sphere across what it crossed.
Max. thickness	mm	The search stops here, and a point with nothing behind it inside the distance is left unmeasured. Defaults to a fifth of the part's bounding-box diagonal, rounded; tighten it to just past the wall you care about.
Rays / Max. opening angle	— / °	The cone: how many rays, and how far off the normal they may look.
Max. deviation of normals	°	How far the surface a ray lands on may be from facing it, before it is stepped over as not-a-wall.
Thin end / Thick end	mm	The two ends of the colour scale. They default to the 2nd and 95th percentile of the part, rounded — a thickness distribution has a long tail to the right, and letting it set the scale flattens every wall onto one colour.
Bands	—	Continuous, or quantised into iso-thickness bands.
Histogram	—	The distribution beside the scale, sharing its axis, plus min / max / mean / sigma.
Thinner than	mm	The wall the under x mm figure under the scale counts. It does not change the colours — set the thin end of the scale to it if you want the map itself to call it out.

Everything down to Max. deviation of normals shapes the search, so changing it means measuring again; everything below it is display, and takes effect as you turn it.

The ramp is the same jet as the deviation map but reversed: red is thin, blue is thick. Thickness has no signed zero to sit in the middle, and the end that needs to shout is the thin one.

Development

npm install
npm run dev       # local dev server
npm test          # unit tests + ballbar.stl and deviation acceptance tests
npm run build     # type-check + production build to dist/

The repository includes ballbar.stl, a real 3D scan of a ball bar used by the acceptance test in tests/ballbar.test.ts, and the pair side bracket left.stl / block-marius.stl — a nominal part and a structured-light scan of it — used by tests/align.test.ts. Both files are already aligned in GOM, so that test displaces the scan by random rigid transforms first; otherwise the automatic match would never be asked a real question.

Nine end-to-end smoke tests drive the real app in headless Chrome against a running dev server:

node scripts/e2e-smoke.mjs      # element fitting on the ball bar
node scripts/e2e-paint.mjs      # hand-marked surface fitting + back-face tint
node scripts/e2e-deviation.mjs  # load, align, measure, split-screen picking
node scripts/e2e-local-fit.mjs  # window / brush / lasso marking + local fine fit
node scripts/e2e-align.mjs      # 3-2-1 datum alignment + STEP export round-trip
node scripts/e2e-thickness.mjs  # measure wall thickness, scale, hover and pin
node scripts/e2e-step.mjs       # STEP reference geometry, measured end to end
node scripts/e2e-split.mjs      # side-by-side compare + the colour plot off
node scripts/e2e-extend.mjs     # extending an element by field and by grip

e2e-step.mjs builds its own pair rather than shipping one, so the answer is known before the app is started: the reference is a STEP cube and the scan is a fine mesh of the same cube with one face raised 0.2 mm and another sunk 0.15 mm. A correct import has to read those two numbers back off the map — sign included — and leave the other four faces flat. It does, to 66.6 % of the scan inside ±0.1 mm, which is exactly four faces of six.

e2e-split.mjs builds the same pair, and checks the split view the way you would by eye: both halves are photographed and reduced to the share of the frame the part covers and where that silhouette sits in it. Two viewports in one pose have to agree on both numbers, before and after a drag in either half and after a zoom — which is a check no amount of asserting on camera matrices would give, since the claim being made is about what is on the screen.

Deploying to Cloudflare

The app is a static Vite build (dist/) served by a Cloudflare Worker with no server code — wrangler.jsonc holds the whole configuration. Two ways to ship it:

• From your machine: npx wrangler login once, then npm run deploy (builds, then uploads dist/).
• On every push: in the Cloudflare dashboard, create a Worker from this Git repository (Workers Builds). Set the build command to npm test && npm run build — the config file supplies the rest. Pushes to main then build and deploy automatically.

Either way the site lands on a *.workers.dev URL — the live deployment is at https://scanruler.stefan-755.workers.dev/ — and a custom domain can be attached in the Worker's settings if you have one. The GitHub Actions workflow (.github/workflows/ci.yml) still runs tests and a build on every push and pull request, so a red suite is visible before Cloudflare ships it.

Roadmap

• More fit methods: Chebyshev (min-zone), min-circumscribed, max-inscribed
• More element types: cones, circles, slots
• Point-cloud (faceless PLY) support
• Export the coloured scan, and section views through the deviation map

License

AGPL-3.0-only — see LICENSE. Free to use, modify, self-host, and redistribute; if you distribute it or offer it over a network — and a browser app on any web server is offering it over a network — your version's complete source must be available under the same terms.

Want it inside closed-source software, or hosted without publishing your changes? Commercial exceptions are available — see COMMERCIAL.md. Contributions require the CLA in CONTRIBUTING.md, which also documents the strict dependency license policy (no third-party copyleft in the app — it would break the dual-licensing model; enforced in CI by license-check.yml). The STEP importer meshStep is AGPL, which the policy otherwise forbids, and is in only because it is ours as well: a commercial exception for this app covers it too.

The bundled fonts are third-party under SIL OFL 1.1 (license files alongside them in public/fonts/). The project name and the CNC Kitchen name and logo are trademarks and not covered by the code license.

By CNC Kitchen.