Mac Studio — 1:1 Blender Model

A dimensionally exact, procedurally generated model of the current-generation aluminium Mac Studio, built with Blender's Python API (bpy) and driven headlessly.

Live site: https://vshengbro.github.io/mac-studio-model/ Interactive viewer: https://vshengbro.github.io/mac-studio-model/viewer.html

Verified dimensions

Measured from the evaluated mesh (all modifiers applied), against Apple's published tech specs.

Axis	Official	Model	Δ
Width (X)	19.70 cm (7.7 in)	19.7000 cm	+0.000 mm
Depth (Y)	19.70 cm (7.7 in)	19.7000 cm	+0.000 mm
Height (Z)	9.50 cm (3.7 in)	9.5000 cm	+0.000 mm

Height is the full 3.7 in including the rubber feet — the enclosure shell is 9.30 cm and the feet add 0.20 cm.

Everything else on the machine is measured off Apple's product photography and checked by:

python3 tools/verify_spec.py        # 26 rear + 16 front + 7 band + 20 glyphs

All four groups measure the photograph and the model independently rather than restating a constant, and the run exits non-zero if anything drifts more than 0.9 mm. Three further gates cover what that one cannot see:

python3 tools/check_builder.py      # the Blender script reads the spec, not a copy
python3 tools/check_internals.py    # the box it builds is physically sensible
python3 tools/check_otsu.py         # the comparison metric is handled everywhere

See docs/VERIFICATION.md for the full table, the calibration rules, and the 27 errors the verification process found.

Source of truth: https://www.apple.com/mac-studio/specs/ — fetched 2026-09-27, "Size and Weight" section. (Apple's older support page ID 121559 now 404s.)

Reproduce the overall-size measurement:

/Applications/Blender.app/Contents/MacOS/Blender --background \
  --python blender/verify.py          # prints VERIFY_OK / VERIFY_FAIL

What the model contains

Every dimension below except the three overall ones is measured from Apple's own product photography and checked by tools/verify_spec.py, which exits non-zero if any feature drifts more than 0.9 mm from the photographs. The full record, including the errors that process found, is in docs/VERIFICATION.md.

• Enclosure — anodised silver aluminium, 1.05 cm vertical corner radius, 0.38 cm top/bottom edge break.
• Front — 2× vertical USB-C and a 27.0 × 2.7 mm horizontal SDXC slot, grouped on the left; a 2.7 mm white status LED on the right. Every opening is a real boolean recess with a socket built wall-by-wall (bright metal walls, shadowed back plate, contact tongue).
• Rear — 4× Thunderbolt 5 (USB-C), 10Gb Ethernet (RJ-45), a three-lobed mains inlet, 2× USB-A, HDMI 2.1, 3.5 mm headphone jack, Touch ID. All front and rear connectors share one centre height, 23.5 mm above the foot plane. Every x position is within 0.1 mm of the photograph.
• Perforated rear field — one large grille, 171.5 mm wide, from 4.7 to 52.8 mm below the top. 1.86 × 1.57 mm lattice, alternate rows offset by a fifth of a pitch, 1.42 mm holes, 54% open.
• Base band — a 7.4 mm perforated band wrapping the whole perimeter (front, both sides, rear). Its lattice is not a scaled copy of the rear field: 1.962 × 0.906 mm, unstaggered, with obround 1.28 × 0.53 mm holes at 41% open. Genuine geometry, not a texture.
• Rear icons — the five engraved decals (Thunderbolt, Ethernet, USB, HDMI, headphone) are measured glyph geometry, not a font and not a drawing: each is a set of polygons, round-capped strokes and arcs whose bounding boxes match the photograph to within 0.28 mm. A Blender text object could never be checked this way, because font metrics say nothing about what gets rendered.
• Top / sides — plain, with no logo, text, or vents. The current enclosure has no top marking; the 2014–2020 shell carried an Apple logo.
• Internals — two blower assemblies with shrouds, ducts, hubs and 11 swept blades each; a finned heatsink; the copper heat-pipe plane; the logic board with memory and two SSDs; the power supply with its copper coil; a row of electrolytics standing on the floor frame; speaker and front I/O; the internal frame. Heights, widths and centre offsets are measured off Apple's X-ray cutaways — the blowers are 59 mm wide at ±62 mm, which the previous build had wrong by 25 mm and 17 mm respectively. See docs/VERIFICATION.md for what that reference can and cannot settle.

Materials: Principled BSDF throughout — silver Metallic 1.0 / Roughness 0.19, dark anodised grille, matte black cavity, rubber feet, emissive status LED.

Layout

blender/
  build_mac_studio.py   geometry + materials + studio lighting, writes the .blend
  render_views.py       loads the .blend, frames and renders 8 views
  export_gltf.py        exports the .blend to Draco-compressed GLB for the viewer
  verify.py             dimension check -> VERIFY_OK / VERIFY_FAIL
  inspect_bbox.py       per-object bounds report
  bool_sweep.py         boolean-solver parameter probe (diagnostic)
  cut_trace.py          per-cut bbox trace (diagnostic)
  rear_trace.py         rear-bay cut trace (diagnostic)
  build_trace.py        full-build stage trace (diagnostic)
docs/                   GitHub Pages site (published from /docs)
  index.html            project page: dimensions, gallery, gotchas
  viewer.html           interactive WebGL viewer (three.js)
  viewer.js             orbit nav, clipping planes, live PBR controls
  viewer.css
  style.css
  mac-studio.glb        0.40 MB Draco GLB, 84 meshes / 51,840 tris, real scale
  images/               12 renders, web-optimised (1400px progressive JPEG)
  thumbs/               420px gallery thumbnails
blender/mac_studio.blend  generated; not committed (see .gitignore)
renders/                12 MB of full-resolution PNG; not committed
reference/              Apple photos used for the fidelity check; not committed

The repo is ~1.2 MB. The 12 MB of full-resolution Cycles PNGs, the 45 MB .blend, and the Apple reference photos are all gitignored — regenerate them with the commands below. There is exactly one .blend, and it lives next to the build script that writes it; tools/check_blend.py fails if a second one appears at the repo root.

Interactive viewer

https://vshengbro.github.io/mac-studio-model/viewer.html

A three.js viewer over the exported geometry, for inspecting the model rather than just looking at renders:

• Orbit / zoom / pan, with 8 named camera views (keys 1–8)
• Clipping planes on X, Y, or Z — slice the body open to inspect the port cavities and the underside grille from the inside
• Per-part visibility for the grille, connectors, and feet, plus a wireframe overlay
• Live PBR controls for roughness, metalness, environment intensity, and tone-mapping exposure

The GLB is authored in metres (Blender centimetres × 0.01), so the model is at true real-world scale in the viewer. Draco compression takes it from 9.1 MB to 0.40 MB; viewer.js loads the decoder from a CDN.

Rebuild the GLB after any geometry change:

blender --background --python blender/export_gltf.py
npx gltf-pipeline -i docs/mac-studio.glb -o docs/mac-studio.draco.glb -d
mv docs/mac-studio.draco.glb docs/mac-studio.glb

Reproducing

# geometry + .blend + bbox report
/Applications/Blender.app/Contents/MacOS/Blender --background \
  --python blender/build_mac_studio.py

# 8 views into renders/ (arg 1 = output dir, arg 2 = samples)
/Applications/Blender.app/Contents/MacOS/Blender --background \
  --python blender/render_views.py -- renders 96

Camera distance is derived from the model's bounding sphere and the lens FOV, so re-framing survives geometry edits — no hand-tuned distance constants.

Rendered views

File	View
01_front.png	front elevation
02_side.png	side profile
03_rear.png	rear
04_hero.png	3/4 hero
05_top.png	top-down
06_bottom.png	underside (grille, feet, power button)
07_front_closeup.png	front panel close-up
08_rear_closeup.png	rear connector bay close-up

Blender gotchas hit while building this

Recorded because each one produced a plausible-looking wrong result rather than an error.

1. Non-planar n-gon caps break the EXACT boolean solver. The lofted shell's top and bottom caps are large non-planar n-gons. Booleans against it "succeed" but silently merge the cutter's outer half into the shell, so the bounding box grows by the cutter's overhang. Triangulating the body first (cut_from_body → triangulate_caps) fixes it. Guard the triangulation on len(p.vertices) == 3 over polygons — data.loop_triangles is a lazily-filled cache and its length is not a valid "already triangulated" test.
2. A cutter must straddle the skin. The convention used here: outer face 0.10 cm proud of the panel (discarded), inner face 0.50 cm in. A cutter that does not clearly cross the surface does not produce a recess.
3. Keep the cutter's bevel small relative to its depth. A 0.16 cm bevel on a 0.60 cm-deep cutter rounds away the part overlapping the skin and reproduces the absorption in (1). 0.06 cm works.
4. Boolean result is direction-sensitive. A socket helper that hard-codes "body extends toward −Y" silently builds front-panel sockets outside the enclosure. The inward parameter makes the axis explicit at every call site.
5. Auto-smooth needs edge subdivision. A flat panel lofted as one huge quad shades with visible banding. rounded_rect(..., edge_sub=4) splits every straight edge.
6. Blender 4.x renamed smooth-shading APIs — shade_auto_smooth (4.2+) / shade_smooth_by_angle (4.1); mesh.use_auto_smooth is gone.
7. BSDF socket renames in 4.x — Clearcoat→Coat Weight, Transmission→Transmission Weight. Print bsdf.inputs.keys() first.
8. A bevel wider than half the shortest side self-intersects; clamp it.
9. At roughness 0.19 the shell is a mirror — bare area lights give it nothing to reflect, so the bounce cards are load-bearing, not decoration.
10. Bounce cards and the floor must be excluded from bbox measurement or they swamp the product bounds (a 40 cm card next to a 19.7 cm product).
11. A rounded-rect path's corner ARC CENTRES are inset by the radius. For a rectangle of half-extent hx with corner radius r, the corner circle is centred at (hx - r, 0) — not (hx, 0). Using the edge midpoint as the centre pushes the entire path r outward and inflates the bbox by exactly that amount, which is maddening to spot in a render.
12. Never Solidify an open per-hole ring to make a perforation. Solidify offsets along the surface normal, which points outward on a ring laid on a vertical face, and the mesh ends up ~1.5 cm proud of the skin. Build each hole as a closed tube instead.
13. Budget a perforated band's Z extent before placing holes. Tube depth plus row spacing must fit between the foot line and the band top, or the lowest holes drop below the feet and Z grows by the overshoot.
14. A boolean groove needs a cutter that straddles the skin — and two of the three ways to get that wrong are silent. A cutter tangent to the shell and a cutter buried inside it both produce no cut at all; only a cutter proud of the skin is loud, and it leaves the difference behind as new geometry that inflates the bounding box by the overshoot (measured: +1.16 mm at proud=0.04). Worse, the quiet failures still pass verify.py — the bounding box is unchanged and nothing else looks wrong. Model the recess into the loft itself (grille_inset) and the question never comes up.
15. EXACT silently no-ops on open and degenerate cutters. A ring prism with no end caps is an open surface; sample its path every 8th point and the faces are so thin the solver treats the solid as degenerate. Both leave the vertex count unchanged and the boolean "succeeding". Cap all four rings and sample densely.
16. Ray-cast to check a cut, don't look at the render. A 300×200 offscreen render of the band plus a scene.ray_cast sweep across its height is what finally located this class of bug: the band read as smooth metal for a dozen render iterations, and a ray sweep showed the front face at a constant y=-9.850 the whole way down. After the fix the same sweep alternates between the groove floor and the hole barrels, and the render goes from 0.00% dark pixels to 30%.

Known deviations

• The .blend and the Cycles renders on disk are stale. Retracted. This said Blender could not start in the authoring sandbox, so the build had never run and everything on disk was a pre-rewrite artefact. That was wrong: Blender 4.5.4 runs fine here. The real reason the model looked wrong was three sign errors in the build (see below), plus two .blend files — the build writes blender/mac_studio.blend while five tools resolved their path to a stale 97-object copy at the repo root, so the renders and the published GLB were built from the broken machine while the gates measured the correct one. tools/check_blend.py now fails on a second .blend. To rebuild everything:

  SKIP_RENDER=1 /Applications/Blender.app/Contents/MacOS/Blender --background \
    --factory-startup --python blender/build_mac_studio.py
  /Applications/Blender.app/Contents/MacOS/Blender --background \
    --factory-startup --python blender/render_views.py -- renders 64
  python3 tools/publish_assets.py
  /Applications/Blender.app/Contents/MacOS/Blender --background \
    --factory-startup --python blender/export_gltf.py

  tools/publish_assets.py --check reports whether docs/images and docs/thumbs are older than the renders they are made from.
• Port micro-detail. Connector internals (pin pitch, latch cutouts, SDXC contact rows) are represented, not reproduced. A production asset would use a manufacturer CAD import.
• Finish. Machining anisotropy, the fine bead-blast texture, and the exact anodised tone are approximated with a single isotropic roughness value.
• No Apple logo or silkscreen. The current enclosure has none; the earlier (2014–2020) shell did. The rear port icons (Thunderbolt, Ethernet, USB, HDMI, headphone) and the Touch ID power glyph are modelled as engraved decals.
• Colour management is left at Blender's default Filmic/AgX view transform rather than being matched against a calibrated reference.