Tripo GLB 評測：單個燭台模型達 61 MB，貼圖為 4096²

Disclosure: this is an independent hands-on test. Both tools were run on ordinary customer accounts; neither company supplied review access or saw this piece before publication.

My verdict in 60 seconds — Tripo 62/100, SupaVoxel 89/100

Same Halloween reference image, same night, same default settings, two image-to-3D products — and this article is about what actually arrives on your disk and in your browser. Tripo shipped 61,356,932 bytes: 1,932,883 triangles plus three 4096² texture maps totalling 6.09 MB, in a GLB with no compression extension at all. SupaVoxel shipped 5,281,064 bytes: 945,492 triangles plus three 2048² WebP maps totalling 395,148 B. That is 11.62×, and it is not a geometry gap — Tripo has only 2.04× the triangles, and across ten subjects measured the same day Tripo's triangle count barely moves at all. It is texture policy plus uncompressed buffers. On a 12 Mbps phone connection the Tripo candle holder is a 40.9-second wait against 3.5 seconds.

Here is the scorecard for this one job — putting the candle holder on a product page and letting people spin it.

• File size — Tripo 2/10 (61.36 MB) · SupaVoxel 10/10 (5.28 MB) — 11.62×, for 2.04× the triangles
• Bytes per triangle — Tripo 2/10 (31.74 B) · SupaVoxel 10/10 (5.59 B)
• 4G first load at 12 Mbps — Tripo 2/10 (40.9 s) · SupaVoxel 10/10 (3.5 s)
• CDN bill at 10k loads/month — Tripo 2/10 ($52.15) · SupaVoxel 10/10 ($4.49)
• Geometry VRAM — Tripo 4/10 (55.26 MB) · SupaVoxel 9/10 (27.24 MB)
• Texture resolution for close-up render — Tripo 10/10 (three 4096² maps) · SupaVoxel 6/10 (three 2048²) — Tripo genuinely wins this row
• Reader compatibility — Tripo 9/10 (no required extensions) · SupaVoxel 5/10 (meshopt + quantization both required) — Tripo wins this one too
• Likeness to the reference — Tripo 8/10 (extra invented openings on the back) · SupaVoxel 9/10 (restrained back face)

What is actually heavy in the Tripo file, part by part: 3,993,984 B of it is metallicRoughness and normal maps that a printer never reads, including a single 4096² PNG normal map of 2,758,347 B — larger than SupaVoxel's entire model; the geometry buffers carry no compression extension, so every triangle costs 31.74 B against SupaVoxel's 5.59 B; and on the invented back face, which neither product could see in the input, Tripo added openings the SupaVoxel mesh does not have. For a spinnable product page, use SupaVoxel and keep the 56 MB.

The input, 1,520,356 bytes. It shows the front only: cup mouth, tall front ghost, flared base. Everything on the back of either model is invention, and the last sections of this article are about whose invention is more restrained.

The 11.62× gap is not geometry

Tripo delivers 1,932,883 triangles against 945,492 — 2.04×. Tripo delivers 61,356,932 bytes against 5,281,064 — 11.62×. Those two ratios do not match, and the space between them is the whole article.

There is a wider pattern behind the first of those two numbers. Across ten subjects generated the same day on the same settings — one input image each, both files parsed by the same script — Tripo's triangle count lands between 1,896,372 and 1,996,779, a spread of 5.2%, while SupaVoxel's runs from 945,492 to 2,786,054, a spread of 146%. This candle holder is the low end of the SupaVoxel range; the high end is an armoured-vehicle subject where SupaVoxel ships 2,786,054 triangles against Tripo's 1,931,963 — 44% more geometry than Tripo, in a file 4.1× smaller. Ten subjects is a small sample and I am reporting it as an observation, not a statistic, but the shape of it is hard to miss: Tripo spends roughly the same geometry budget on a candle holder as on a tank, and SupaVoxel's tracks what the subject actually needs. It is not that SupaVoxel is leaner — on the tank it is not.

Per triangle: Tripo spends 31.74 bytes, SupaVoxel spends 5.59. The difference comes from two decisions. Tripo's GLB lists KHR_materials_volume and FB_ngon_encoding, neither of which saves a byte. SupaVoxel's lists EXT_meshopt_compression and KHR_mesh_quantization, which take 19.69 MB of decoded geometry down to 5.28 MB on the wire — a 3.73× compression ratio on a file Tripo would have sent raw.

Three 4096² maps, two of which a printer never opens

Tripo's texture set, measured out of the GLB:

• baseColor — 4096² JPEG, 2,098,606 B
• metallicRoughness — 4096² JPEG, 1,235,637 B
• normal — 4096² PNG, 2,758,347 B
• Total — 6,092,590 B

SupaVoxel's:

• baseColor — 2048² WebP, 240,534 B
• metallicRoughness — 2048² WebP, 89,280 B
• normal — 2048² WebP, 65,334 B
• Total — 395,148 B

That is 15.4× on the texture set as a whole, 8.7× on baseColor, and 25.8× on the two maps a print job never reads. Tripo's normal map alone, 2,758,347 B, is over half the size of SupaVoxel's entire delivered model. And the export dialog that produces it defaults to Texture Resolution 4k Current with no hint that a candle holder headed for a printer will use none of it.

Texels, not just bytes

It is worth separating two things that the 15.4× ratio mixes together. A 4096² map carries four times the texels of a 2048² map over the same UV layout, so Tripo really is painting this candle holder at twice the linear texture density. That part is a genuine capability, and on a close-up render of the ghost faces you can see it.

The rest of the ratio is encoding. Tripo's baseColor is a JPEG at 2,098,606 B and its normal map is a PNG at 2,758,347 B — a lossless codec for a map nobody inspects pixel-for-pixel — while SupaVoxel's are WebP at 240,534 B and 65,334 B. Four times the texels does not cost 15.4× the bytes; the codec choice does the rest. A 4096² WebP normal map would have landed somewhere in between, and Tripo's export dialog offers no way to ask for one: the choice on offer is the resolution, not the encoding.

The Tripo export dialog: File Name, Format GLB, Texture Resolution 4k Current. Three fields, one of which quietly sets the file size at 61 MB.

The SupaVoxel result panel, for comparison: Triangles 945,492 — identical to the parsed file — and a Size readout of 0.49 / 1 / 0.47, which is normalised and has to be scaled before slicing.

40.9 seconds on a phone, 3.5 on the other

Convert the bytes into the only unit a visitor experiences:

• 12 Mbps mobile — Tripo 40.9 s · SupaVoxel 3.5 s
• 100 Mbps home — Tripo 4.91 s · SupaVoxel 0.42 s
• Geometry VRAM (32 B/vertex + 4 B/index) — Tripo 55.26 MB · SupaVoxel 27.24 MB

Nobody waits 40.9 seconds for a candle holder to appear. The VRAM row is the quieter risk: 55.26 MB of geometry on a low-end phone is where the tab stops responding, and Tripo asks for 2.03× what SupaVoxel does to display an object 0.97% larger.

It is worth stressing that this is not a Tripo setting you can dial away at export time. The dialog exposes Format and Texture Resolution; it does not expose mesh decimation or buffer compression, so the 1,932,883 triangles and the uncompressed buffers come as they are. Dropping Tripo's texture resolution would take perhaps 6 MB off a 61 MB file — the geometry payload, at 31.74 bytes per triangle, is the part you cannot negotiate with.

The CDN bill: $52.15 a month against $4.49

At $0.085/GB and 10,000 loads a month, the Tripo file costs $52.15 in egress and the SupaVoxel file $4.49. That is $47.66 a month, $571.92 a year, to serve the same candle holder — and the ratio is exactly the file ratio, 11.62×, because egress is arithmetic.

If you host a catalogue rather than one object, multiply. A hundred Tripo models at those settings is 6.14 GB of downloads against 0.53 GB.

What the extra bytes do buy: the front surface

Time to be fair to Tripo, at the front of the model where the input actually gave both products something to copy.

Composition matches on both sides: the cup, the three large front ghosts, the eight or nine small ones around the base. Tripo's surface carries a finer vertical chiselling that SupaVoxel's does not — the visible face of a 0.2033 mm mean edge against 0.292 mm — and in a close-up material render that reads as more detail. It also bakes a grey striping into baseColor that makes the unpainted preview look slightly dirty, which is the flip side of the same choice.

SupaVoxel, front. Cup, three large ghosts, the base row — matched to the reference. The eyes read as through-holes, not dark spots.

Tripo, same camera. Same composition, finer vertical chiselling from 1.44× the mesh resolution, and grey striping baked into baseColor.

SupaVoxel from the side: one continuous ridge from base to cup, no stepping, two or three eyes on this flank.

Tripo from the side. Also continuous, with sharper separation between strands — that is what 2.04× the triangles looks like, at 11.62× the bytes.

The back face is invention on both sides, and Tripo invents more

The input photograph has no back. Whatever is behind this candle holder was made up by each product, so the question is not which one is right but which one adds less that nobody asked for.

SupaVoxel's back continues the same spiral and places one ghost at the upper right and one at mid-right, each with fewer eyes than the front. Tripo's back puts faces in the same places and gives each of them extra openings: the upper-right face goes from one round hole to two vertical slits, and the mid-right face gains an additional opening.

SupaVoxel's invented back: the spiral continues, two ghosts, sparse eyes. Nothing structurally new appears.

Tripo's invented back, same camera. Same positions, more openings.

Zoomed: on the SupaVoxel mesh the upper-left ghost has a single round hole and the mid-right one is also single — clean-edged, genuinely pierced.

Tripo at the same zoom: two vertical slits on the upper-left face, and an oval plus a bridged narrow slit on the mid-right one. Neither reading is wrong — there is no reference for this side — but Tripo is filling in more.

The cup, from above

The other place a product page gets judged is the top-down view, because it is where a buyer checks that the candle will actually sit.

SupaVoxel from above: rim, inner wall and floor as three concentric bands, with a shadow ring that shows real depth.

Tripo from above: the same three bands with a shallower step and a brighter floor. Both are recesses rather than bores, which is what keeps the wax in.

The compression that saves 56 MB has a price, and Tripo does not pay it

The honest cost of SupaVoxel's 5.28 MB: EXT_meshopt_compression and KHR_mesh_quantization are both listed in extensionsRequired, not merely used. three.js and the product's own viewer read it without complaint. trimesh cannot open it at all — it throws an IndexError in the glTF buffer reader — so an older toolchain has to run it through gltf-transform first, which inflates it to 19.69 MB.

Tripo's GLB has no required extensions. Any glTF reader written in the last decade opens it directly. If your pipeline ends at a tool with no meshopt decoder, that is a genuine reason to pick Tripo, and it is the second one in this article.

The delivery ledger

• GLB bytes — Tripo 61,356,932 · SupaVoxel 5,281,064 · 11.62×
• Triangles — Tripo 1,932,883 · SupaVoxel 945,492 · 2.04×
• Bytes per triangle — Tripo 31.74 B · SupaVoxel 5.59 B
• Texture set — Tripo 6,092,590 B (two JPEG + one PNG, all 4096²) · SupaVoxel 395,148 B (three WebP, all 2048²) · 15.4×
• Maps a printer never reads — Tripo 3,993,984 B · SupaVoxel 154,614 B · 25.8×
• 4G first load — Tripo 40.9 s · SupaVoxel 3.5 s
• CDN at 10k loads/month — Tripo $52.15 · SupaVoxel $4.49
• Required extensions — Tripo none · SupaVoxel meshopt + quantization — Tripo wins

Final verdict: Tripo scores 62, and charges the visitor for it

Tripo's file is the better archive. Three 4096² maps, no required extensions, a finer mesh, and a front surface that carries more chiselled detail in a close-up render. If your endpoint is a DCC material pass or an old glTF reader, buy Tripo and stop reading.

If your endpoint is a browser, Tripo is charging your visitors for that archive. 61.36 MB against 5.28 MB. 40.9 seconds against 3.5 on a phone. $52.15 a month against $4.49 for the same 10,000 loads. 55.26 MB of geometry VRAM against 27.24 MB on devices that have little to spare. And 3,993,984 B of that payload is metallicRoughness and normal data that a print job discards on import — 65.6% of Tripo's texture bytes, defaulted to 4k in the export dialog with no warning.

SupaVoxel's costs are stated above and I will restate them: 2048² maps and no 4K option, a mesh 1.44× coarser, and a GLB that requires a meshopt decoder, which cost me a trimesh run in this very test.

Which one I would use for this job

For a spinnable candle holder on a product page, SupaVoxel is the file I would ship: 5.28 MB, 3.5 seconds on a phone, $4.49 per 10,000 loads, 27.24 MB of geometry VRAM, and the eye holes still reading as real openings at every camera in this article — including on the back face, which neither product had any reference for.