KullGames . ImageTool . Milestone 4.1

Turn a flat photo into a spinning 3D model.

Click one pixel of a subject. The tool magic-wand seeds it, runs GrabCut to prop the whole object off its background, inflates that flat silhouette into a rounded, watertight solid with the classic Teddy method, orbits a camera around it to bake a turntable GIF, and exports the mesh as STL + OBJ/MTL/PNG. No neural net, no depth sensor, no multi-view rig - just classical geometry, written entirely in C# against nothing outside the .NET base class library.

Magic-wand + GrabCut Teddy inflation Software rasterizer STL / OBJ export Zero dependencies

Everything below is a fresh render from this run. The 39-photo unattended batch swept the whole folder in 120.2 s (3.1 s/image, single-threaded): 30 graded a compact OBJECT, 2 partial, 7 honest scenes - every GIF round-trip verified, every mesh exported to STL and OBJ.

A campfire flame, cut out of a photo and inflated into a puffed 3D solid, rotating on a turntable.

The pipeline, on screen

Watch a photo become a solid.

Each stage of the pipeline emits a real image, so the whole path is legible - not a black box. Left to right: the selection, the cut, and the spin.

For each photo: the input frame, then the magic-wand selection outlined in place, then the RGBA cutout composited on a checkerboard. — Step 1 - select: the magic wand

A CIELAB tolerance flood grows from one click. Per photo this loops input photo -> selection outlined in place -> the isolated RGBA cutout on a checkerboard, so you can see exactly what got grabbed.

A side-by-side GIF. Left: the single-seed Tier-1 flood fill grabbing only a sliver. Right: GrabCut recovering the whole multi-colour subject, with the seed rectangle drawn. — Step 2 - cut: GrabCut beats the single-seed flood

LEFT = the Tier-1 wand (one seed colour), RIGHT = GrabCut (models the whole foreground and background colour distributions, cuts on a min-cut). On the campfire the wand grabs 2.5% (a sliver of the white-hot core); GrabCut recovers the whole 11.1% flame.

A campfire flame isolated on a transparent background - the GrabCut cutout that feeds the inflator. — The cutout that feeds the inflator

GrabCut's RGBA foreground - the exact bitmap the Teddy step domes into a mesh, and the same pixels painted onto the 3D form as per-vertex colour.

A synthetic multi-colour disc inflated into a plump biconvex lens, rotating. — Step 3 - inflate + spin: a textbook dome

A clean synthetic cutout inflates to a biconvex lens - the Teddy step isolated from any segmentation noise. 6,300 tris, watertight (0.00% boundary edges), 24-frame turntable.

WATERTIGHT 6,300 tris synthetic

Renders

Cut out, inflate, spin - across a whole folder.

Every tile is a real turntable from the unattended batch: an ordinary JPEG, auto-seeded with no click, cut out, inflated, and spun. Each carries the honest grade the batch driver assigned it - green OBJECT (a compact solid), amber PARTIAL, grey SCENE (a landscape with no clear subject, inflated into an abstract colour relief and labelled so).

A cluster of clover leaves inflated as several separate puffed medallions, rotating. — Clovers, honestly scattered

A genuinely disconnected mask inflates into separate leaf medallions, not one unified solid - the honest result, not a hidden-geometry guess.

OBJECT13.2% mask9,944 tris

A leafy tree canopy inflated and rotating on a turntable. — Tree canopy

Dense organic foliage - a broad, textured green mass. Auto-seeded at (656,450).

OBJECT12.9% mask24,060 tris

A weathered stump with a sapling, inflated and rotating. — Stump and sapling

A high-detail organic subject - the inflator's densest object mesh in the batch.

OBJECT9.3% mask21,416 tris

Brightly coloured kayaks inflated into a solid, rotating. — Kayaks

High-contrast man-made objects - saturated hull colours the per-vertex shading carries cleanly.

OBJECT7.1% mask11,500 tris

Seaside stone ruins inflated into a solid, rotating. — Seaside ruins

Geometric masonry against sky - a compact, well-bounded subject.

OBJECT9.1% mask6,624 tris

A golden-hour city skyline inflated into a relief, rotating. — Skyline, golden hour

A geometric cityscape band - buildings read as a jagged inflated ridge.

OBJECT6.5% mask4,904 tris

An aerial view of a town inflated into a relief, rotating. — Town from above

A dense aerial that still resolves a compact central subject - graded OBJECT.

OBJECT10.9% mask13,908 tris

A pale glacier fjord inflated into a solid, rotating. — Glacier fjord

Low-contrast whites and greys - the border-prior saliency still found the ice mass against the water.

OBJECT16.1% mask14,880 tris

An alpine valley subject, auto-seeded and inflated, rotating. — Alpine valley, no click

Saliency picked the seed at (538,142); GrabCut grabbed the ridge; the inflator built a 14,780-tri solid.

OBJECT13.6% mask14,780 tris

A foggy farm scene inflated into a low relief, rotating. — Foggy farm

A deliberately low-contrast, hazy input - still a solid, coherent OBJECT grade.

OBJECT26.6% mask19,596 tris

A winter dusk scene inflated into a pale solid, rotating. — Winter dusk

Muted, low-contrast snow and sky - a broad pale form.

OBJECT15.0% mask14,728 tris

A fiery sunset band inflated into a thin bright sliver, rotating. — Fire sunset

A high-contrast, thin bright band - inflates flat, exactly as the square-root height field should for a sliver.

OBJECT5.4% mask3,944 tris

A golden grain field at sunset inflated into a solid, rotating. — Grain at sunset

Warm, high-contrast crop rows against sky.

OBJECT24.7% mask14,900 tris

A mackerel-patterned sky inflated into a rippled relief, rotating. — Mackerel sky

Rippled cloud banding inflated into a textured relief.

OBJECT25.9% mask18,892 tris

A waterfall cascade inflated into a solid, rotating. — Waterfall

The bright cascade and cliff face as one coherent bright mass.

OBJECT27.4% mask13,640 tris

A beach-palms subject, auto-seeded and inflated, rotating. — Beach palms

A compact tropical subject against sky - a clean OBJECT grade.

OBJECT11.3% mask7,396 tris

A mountain at dusk inflated into a solid, rotating. — Mountain dusk

A broad ridge silhouette at low light.

OBJECT20.2% mask18,828 tris

An aerial of cascades graded PARTIAL, inflated and rotating. — Cascades from the air

A real region got grabbed, but too small and ragged to call a clean object - graded PARTIAL, and it says so.

PARTIAL3.1% mask12,412 tris

A dusk skyline graded PARTIAL, inflated and rotating. — Skyline at dusk

A thin, low-contrast band at the horizon - a borderline grab, honestly flagged PARTIAL.

PARTIAL4.1% mask2,972 tris

An aerial coastline graded SCENE, inflated into an abstract relief, rotating. — Aerial coast (no subject)

A pure landscape with no distinct subject - the seed lands on the most contrasty patch and inflates a colour relief, not an object. Graded SCENE.

SCENE0.3% mask660 tris

An airplane against a plain sky graded SCENE, a tiny inflated speck rotating. — Airplane (tiny subject)

A speck against a plain sky falls below the 3% subject floor - honestly graded SCENE rather than pretending it is a solid.

SCENE0.6% mask1,748 tris

Batch, unattended

Thirty-nine photos, zero clicks.

Point model3d-batch at a folder and it auto-seeds each photo with a border-prior saliency cue, runs the same GrabCut -> Inflate -> Turntable pipeline, and exports a mesh per image. Then it bakes two "see everything at once" artifacts and prints an honest per-image verdict table.

A 7-by-6 contact sheet of all 39 inflated subjects, each tile tagged OBJECT, PARTIAL, or SCENE in green, amber, or grey. — All 39 subjects, one glance

One angled turntable frame per photo, each tile tagged with its name and colour-coded verdict, exactly as the batch driver bakes it. The final tally: 30 OBJECT, 2 PARTIAL, 7 SCENE.

30 OBJECT2 PARTIAL7 SCENE

A concatenated rotation tour of the strongest subjects, spun one after another. — Rotation tour

The strongest subjects (objects first) spun back to back - capped at 20 to keep the GIF small, with the dropped names logged, never silently truncated.

Exports

Every mesh is a real, printable file.

The turntable is only the preview. Each inflated solid is written as binary STL and as OBJ + MTL + texture PNG - watertight, per-vertex coloured, and openable in any slicer (Cura, PrusaSlicer, Bambu) or mesh viewer (MeshLab, Blender, Windows 3D Viewer). Download the actual files this run produced:

Binary STL has no magic number, so a file sniffer reports "data" - that is expected; the byte size is exactly 84 + triangles . 50. Keep the OBJ, MTL, and texture PNG together in one folder so the photo maps on via the MTL's map_Kd.

Also in the box

Straight lines and geometric shape.

The same classical toolkit powers a recognition layer: a from-scratch Hough transform finds straight lines over Canny edges, and a shape descriptor (convex hull, rotating calipers, PCA axis) sorts a mask into TallThin / Wide / Round / Irregular. It reports geometry, not identity - it says "tall and thin", never "a person".

Per photo: an input frame, a Hough frame with detected cyan lines and an angle label, and a shape frame with the cutout outline and a magenta PCA axis plus a geometric-class label. — Hough lines + geometric class

Per photo: INPUT, then a HOUGH panel (Canny + Sobel edges + detected straight lines with their angle), then a SHAPE panel (cutout outline + magenta PCA principal axis + the geometric class). The self-test recovers known 0/45/90-degree lines to within 0.0 degrees.

How it works

Four classical steps, no learned model.

Select. A click seeds a CIELAB magic-wand flood; its bounding box seeds GrabCut, which models the whole foreground and background colour distributions as Gaussian mixtures and cuts the boundary with a min-cut over the pixel grid.

Height field. A two-pass chamfer distance transform of the mask gives each pixel its distance to the nearest edge - high in the middle of a wide region, low along a thin sliver.

Inflate. Dome it with h = Thickness . s . sqrt(dt . Dmax) - the square root gives a sphere-like profile, so wide areas bulge tall and thin areas stay flat. Grid-mesh the front and back, then stitch a side wall into one closed, watertight solid.

Render + export. A Pineda software rasterizer orbits a turntable camera, shading per-vertex photo colour under a directional light, tonemaps Reinhard + sRGB, and bakes the frames into a GIF89a - then writes the mesh as STL and OBJ/MTL/PNG, all with the library's own codecs.

Honest scope

A relief, not a reconstruction.

This is a single-image inflation: the back is a mirror of the front and the sides the camera never saw are smoothly invented. It is convincing head-on and in a gentle spin, not a metrologically correct model - the honest nature of any single-snapshot method that does not hallucinate hidden geometry. Quality rides entirely on the cutout: a clean mask gives a clean solid, a landscape with no subject gives an abstract colour relief - and the batch grader says which is which, out loud, on every tile.

The technique is Teddy (Igarashi, Matsuoka & Tanaka, SIGGRAPH 1999); the segmenter is GrabCut (Rother, Kolmogorov & Blake, SIGGRAPH 2004). Both are reimplemented from scratch in pure C# over the OnlyCSharp library, against nothing outside the .NET base class library.