Field notes · 18 October 2026
We measured 28 genuine Tripo outputs, then ran the generator live from one original drawing. This page reports what it gets right, what repairs it leaves you every time, and which route fits the thing you are actually making.
Input · one front drawing
P2 · 6,976 quads
H3.1 · 1.9 M triangles
01
Tripo is a fast way to get a coherent, likeness-preserving candidate. In every earlier study we drew on, a model accepted for use first went through local repair: decimation or re-bake, material overrides, weight fixes, a deliberate scale.
The leverage is not in generating more. It is in choosing the right route for the intended use, and knowing in advance which repairs that route always needs. The two main routes do not produce more or less of the same thing. They produce structurally different objects, and that difference decides what you can do next.
02
We welded every file by bit-identical vertex position and counted connected pieces (“shells”), open edges and non-manifold edges. The method is written down as one spec, implemented in Python and JavaScript, and the two implementations give identical results on all 28 files. Measured
Many open pieces
A defect for printing and booleans, and a possible gift for editing: the pieces may already be hair locks, eyes and garments.
One sealed skin
A watertight 2-manifold: a strong start for printing or as a bake source. Everything is fused into one skin, so nothing can move separately.
Turn the two routes yourself
One character, both routes, two genuine provider files. 14–16 MB per pair.
Nothing is downloaded until you press the button.
Every file, by triangle count and number of pieces
Point at or focus a dot to read that file.
Measured All 28 files are normalized to about one unit tall (0.947 – 1.000) with their base at the origin, whether the subject is a head, a body or a cat. Real-world scale is always a deliberate step.
Measured Without a metal/rough map, material factors are the glTF defaults (roughness 0.5, metal 0). All six P2 rig exports carry roughness 0.9 where the matching textured export carried 0.5: rigging changes the material.
Measured The generator: "Tripo" stamp survives editing. At least 18 derivative files still carry it. It is a hint, never proof of an untouched original; keep the original’s hash.
03
P2’s open pieces could be noise from the generator or a gift for editing. To find out, we rendered every one of the 94 pieces in the two P2 files above, each highlighted on its own, and named what it is. The labels come from an AI reviewer working from those renders and have not yet been checked by a person. Observed
By piece count
By triangles
P2 gives you an accessory layer for free, not a body–garment split. Hair locks, eyes, buttons and jewellery already come apart, so they can be recolored, swapped or rigged on their own. But the clothes that matter for deformation are fused to the body: a skirt to both legs, a blouse or bodice to the torso. Separating those still needs segmentation or a manual cut. Whether Studio’s segmentation works on P2 the way it did on HD armor is Open.
04
To compare the routes fairly, we drew one original character for the test (a curious courier, front view only, A-pose) and generated it once with each route in Tripo Studio. This was a supervised run on a free account, so nothing could be exported. The counts below are the Studio viewer’s own readouts, not our measurements. Observed
Front, wireframe. The quad layout follows the jacket, satchel strap and scarf.
| P2.0 | H3.1 | |
|---|---|---|
| Wall time | ≈ 1 min | ≈ 5 min |
| Topology | Quads · 6,976 faces (target 6,000 → +16%) | Triangles · 1,899,257 |
| Texture | None; geometry only | 4K PBR, remove-lighting on by default |
| Front | Every garment piece present, readable face loops | Very close to the drawing; sculpted seams and folds |
| Invented back | Plausible; moved the scarf tails behind (the prompt said so, the drawing did not) | Plausible hair locks, a strap across the back |
| Credits | 0 (free trial; normally quoted 100) | 55 |
05
HD fuses every garment into one skin. That made HD a dead end for anything that moves, until we tried Studio’s segmentation on a fused 1.96-million-triangle demo armor.
The armor is a demo model that ships with the Studio account, shown here only to discuss the segmentation result.
06
About 20 credits on top of a trial generation produced a real-time-sized textured character that already walks. Whether it deforms well is a separate question. In earlier studies auto-rig weights pinned fingertips to legs and folded a skirt into the thighs. Observed
07
None of these is a reason to avoid the generator. Each is a step to budget for, together with a check that shows whether it has been done.
Relighting a color map changed the apparent eye sockets and smile; the Studio look does not travel with the file. Remove-lighting is now on by default for HD texturing, which targets exactly this. Whether it fully works is untested without an export.
Observed Open
HD Ultra is about 2 M triangles and 46–59 MB. Treat it as a reference, donor or bake source, never a runtime asset as is. One festival cart went to 11,580 triangles and 556 KB, and shipped.
Measured
Fingertips weighted to legs (5,590 hand vertices re-pinned), a tail weighted to legs, an 8.15% leg-length asymmetry, a skirt folding into the thighs.
Measured Observed
Preset clips travel. A game controller that expects in-place clips needs a conversion step.
Measured
0 of 28 files had morph targets. Blinks, visemes and smiles must be authored. No separate eyeballs or lids were verified.
Measured
Targets are approximate: 6,000 gave 6,976 (P2) and 6,396 (retopology); an older 5k quad target gave about 4,100, and a 10k target exported as 20,222 triangles.
Measured Observed
08
This is a judgment drawn from the evidence above, not a benchmark.
Two tested ways to a 6k-quad textured character
P2 native trial generation, plus 20 for a 4K texture. Cheapest and fastest; overshot the face target by 16%.
H3.1 + retopology 55 plus 40. Hit the target within 6.6% and keeps the HD texture. Whether its even grid deforms better is Open.
09
We set a ceiling of 200 credits before starting. Every charge matched the quote shown before the click, and no balance changed without a cause we could name. The quotes are what this account saw that day, not a price list. Observed
The free plan is a viewer, not a pipeline. Export is locked and every generation is public. Without a file, nothing can be measured, opened in Blender or tested in an engine. That is why the live results in sections 04 to 06 are graded as observed rather than measured.
10
The measurements come from a small local review desk built for this study. It opens two candidates side by side under one camera, switches between lit, flat-albedo, clay and silhouette views, and measures shells, open edges, maps, rig and root travel straight from the file. It records a per-check verdict with its evidence, and keeps “not tested” as a valid answer. It runs offline and uploads nothing. It is a private research tool and is not distributed.
11
