How heavy should a 3D model be? GLB detail vs file size

A 3D model in a report has one job: let people find an object and see its status. Every triangle that does not serve that job costs download time, memory and frame rate. This page measures where the weight sits in a GLB, what you can remove for free, and what you pay when you remove more.

The test plant: two tanks, twelve pumps, a pipe rack with lines, elbows and valves on a slab — **The test plant.** Generated for this page: 2 tanks, 12 identical pumps, 24 identical valves, a pipe rack with 12 lines, 75 objects, 164,284 triangles. Written the way a dense CAD export writes: 128 segments around a tank, 48 around a pipe.

11.9 MBas exported

0.6 MBafter 3 lossless steps

57 KB1 cm cap + Draco

01Where detail is decided

CAD and IFC can store a pipe as an exact cylinder. A GLB only stores flat triangles, so somewhere an export cuts every curve into straight segments. The gap between the true curve and the flat segment (the chord error) is the detail you keep or lose.

8 segments gap 11.4 mm on a 300 mm pipe

16 segments gap 2.9 mm on a 300 mm pipe

48 segments gap 0.32 mm on a 300 mm pipe

Chord error = r × (1 − cos(π / n)). Dashed: the exact pipe. Solid: what the GLB stores. 48 segments keep the surface within a third of a millimetre, far below what a screen can show.

Radius	Segments for 1 cm	Segments for 1 mm
30 mm (2 in line)	4	13
150 mm (12 in line)	9	28
1 m (vessel)	23	71
8 m (tank)	63	199

One fixed segment count wastes triangles on small pipes and starves large tanks. A cap in millimetres treats every size the same way, which is why the steps below cap the error, not the count.

02Which format carries what

No format wins on everything. The richest formats are the hardest to draw; the format built for drawing carries the least data. The usual answer is to send the geometry as GLB and keep the data in a table beside it.

Format	Owner	Geometry	Properties	In a browser
NWD	Autodesk (Navisworks)	Triangles	Yes, rich	No: Autodesk tools or a converter
IFC	buildingSMART, ISO 16739-1	Exact solids, extrusions or triangles	Yes, the richest (property sets, types, relations)	Through a converter or an IFC web library
STEP	ISO 10303	Exact curved surfaces (B-rep)	Little	No: tessellate first
FBX	Autodesk	Triangles, materials, animation	Limited	Through a converter
OBJ	Open, plain text	Triangles, groups, materials	No	Yes, but large and slow
glTF / GLB	Khronos Group	Triangles, scene tree, names	Names + extras JSON, or a separate tag table	Yes: built for runtime delivery
3D Tiles	OGC community standard	Many GLB tiles, streamed by view	Per-feature metadata	Yes, at city or site scale

03Seven steps, measured

The same plant, each step applied on top of the previous one. Green steps change nothing you can see. Amber steps remove detail you choose.

0As exported

11.87 MB100.0 % of raw · gzip 1472 KB

Every triangle owns its three corners. This is what many CAD exports write.

1Weld

3.73 MB31.4 % of raw · gzip 765 KB

Corners shared by neighbouring triangles are stored once, plus an index list.

2Dedup + instance

606 KB5.1 % of raw · gzip 229 KB

12 identical pumps and 24 identical valves stored once, placed many times.

3Simplify, 1 cm cap

357 KB3.0 % of raw · gzip 128 KB

Triangles removed where no surface moves more than 1 cm.

4Quantize

275 KB2.3 % of raw · gzip 120 KB

Positions as 16-bit integers, normals as 8-bit, instead of 32-bit floats.

5aMeshopt compression

98 KB0.8 % of raw · gzip 45 KB

Buffers packed with EXT_meshopt_compression. Needs the meshopt decoder.

5bDraco (instead of 4 and 5a)

57 KB0.5 % of raw · gzip 48 KB

Smallest file on disk. Needs the Draco decoder.

After the web server gzips the file, meshopt (45 KB) and Draco (48 KB) end up close. Both need their decoder in the viewer; a viewer without it shows nothing.

04The centimetre

Simplification merges triangles until a surface would move more than the cap. The cap is the only lossy choice that matters, so pick it on purpose: the same close-up at four caps.

Close-up of elbows and a valve handwheel, full detail — **No cap** · 164k triangles

Same close-up at a 1 cm cap — **1 cm** · 55k triangles

Same close-up at a 2 cm cap, elbows polygonal — **2 cm** · 30k triangles

Same close-up at a 5 cm cap, the handwheel is gone — **5 cm** · 20k · handwheel gone

CapTriangles drawnFileWhat you see

none

164,284

606 KB

Reference after the lossless steps.

1 mm

150,336

701 KB

8 % fewer triangles, file 16 % larger: the rewrite split shared buffers (27 to 35). Not worth it.

5 mm

87,264

504 KB

Close to the original.

1 cm

54,588

357 KB

Shapes kept. Elbows show facets close up.

2 cm

29,904

173 KB

Elbows turn polygonal close up. The 6 cm handwheel still there.

5 cm

19,808

104 KB

Handwheel and stem gone, elbows become cones. Overview only.

10 cm

16,360

79 KB

Same as 5 cm with rougher curves. Gains have flattened.

Cap below the smallest thing that must stay visible.Here a ring 6 cm thick survived 2 cm and vanished at 5 cm. Instruments, handwheels and small bore lines set the cap, not the tanks.

Cap in centimetres, not in percent.A relative error is a fraction of each mesh's size (meshoptimizer's default): 1 % of a 100 m slab is 1 m, 1 % of a valve is 5 mm. An absolute cap treats both the same.

Stop where the curve flattens.From 1 cm to 2 cm the file halves; from 5 cm to 10 cm it drops 24 % for a worse picture. For site overviews 2 cm, for equipment walk-downs 1 cm, on this plant.

Drop on purpose, not by accident.If tiny parts are not needed, remove them by size before simplifying and list what was removed. A part that collapses into a sliver looks like a modelling error.

05Before it goes into Power BI or a web app

The tag survives every step. Count the distinct element IDs before and after optimising. Welding, dedup and simplification rewrite buffers; the ID attribute must ride along (see how the click finds its row).
The viewer reads what you wrote. Instancing (EXT_mesh_gpu_instancing), meshopt and Draco are extensions. three.js reads all three when given the decoders; check yours before you ship.
Quantize knowingly. 16-bit positions split each mesh's bounding box into 65,535 steps: 1.5 mm on a 100 m mesh. Fine for a plant, not for a micrometre part.
Keep coordinates near zero. GLB positions are 32-bit floats: at 500 km from the origin the step is 3 cm. Subtract a site origin at export and store it once.
Size the tag table too. A Power BI visual receives data in windows of 30,000 rows and stops at 1,048,576 rows. One row per tagged object, not per triangle.
Keep the raw export. Every step here is a script. When the cap turns out wrong, re-run from the raw file, never from an already simplified one.

Companion paper A 3D model inside Power BI

The four routes, the tag join, and how a click finds its row.

Read →

Related tool Layout Studio

Place the 3D visual and its tables on a report page, download a .pbip.

Open →

Method. The test plant was generated for this page (no real project data). Steps run with glTF-Transform 4.5.1 (weld, dedup, instance, quantize, meshopt, draco), simplification with meshoptimizer 1.2.0 using an absolute error cap, renders with three.js 0.170. Sizes are bytes on disk; gzip at level 9. Your ratios will differ with your model: run the same steps on it.
Sources. glTF 2.0, KHR_mesh_quantization, EXT_mesh_gpu_instancing, EXT_meshopt_compression, KHR_draco_mesh_compression, meshoptimizer, glTF-Transform, OGC 3D Tiles, IFC, Power BI fetchMoreData limits.