We tested 41 AI image-to-3D engines (August 2026)

PolyMedium is not built on one engine. It routes every upload across a bench of engines, because the single most important finding in this data is that no engine wins on every subject: the engine with the best dog face loses the character, the cheapest engine ties the field on vehicles and melts on animals, and the most expensive one never wins raw fidelity but is the only one that ships game-ready topology. This page is the evidence. On 2026-08-29 we ran every live image-to-3D endpoint on fal.ai, all 41 of them, against our draft engine (TRELLIS 1) as a fixed control. 18 endpoints were screened head-to-head; the other 23 are excluded below, each with its reason. Survivors ran across five subjects chosen for specific failure modes, and every output was rendered as an identical six-view Blender ring so nothing could hide behind a flattering angle.

The headline findings, including the ones that go against us: Rodin v2.5 beat our own TRELLIS 2 on the game character (facial likeness, engraved armour, cloak embroidery). SAM-3 Objects matched TRELLIS 1 at the identical price and was 3x faster, and it beat our control outright on the dog's face. Meshy v7 was the one genuinely additive engine: game-ready topology by default, real part separation, rigging in the same call. On the other side: Hunyuan 3.1 Rapid cannot emit a GLB at all, the Hunyuan v2 family hallucinated blue wheels, and three engines quietly bill up to 3x their advertised price for texture.

Where we lost

A vendor-run review is only worth reading if the vendor can lose in it. We did, three times:

• Rodin v2.5 beats our TRELLIS 2 on the character. The reference's actual likeness comes through (brow, cheekbones, hair strands), plus engraved armour and embroidery on the cloak hem. It also beats both TRELLIS 2 and Hunyuan 3.1 Pro on the dog's head. The trade: a higher price band, and roughly 2x slower.
• SAM-3 Objects matches TRELLIS 1 at the same price and is 3x faster (24s vs about 75s). On the lying dog, our control's worst failure, it produces a face with actual eyes, nose and ears where TRELLIS 1 produces a blob. It is clearly worse on characters and it silently dropped the dragon's flame, so it is a second source, not a replacement. But at the draft price point it is a real peer.
• Meshy v7 is the one genuinely additive engine in the catalog. It never wins on raw fidelity, but it is the only engine that reconstructed the character's sword as a separate blade and the dragon's flame as a real volumetric jet, it ships game-ready ~30k-triangle topology by default, and it offers rigging and animation in the same call. Nothing in our stack does that.

Method

• One reference image per subject, byte-identical for every engine. The five references are cleaned, restaged images from prior validated runs, re-uploaded once and reused by URL. Byte-identity was verified by content length, which is what makes the reused TRELLIS 1 control GLBs valid rather than approximate.
• Two stages. Stage A screens the whole credible catalog on one subject (the go-kart). Stage B advances the survivors to all five subjects.
• A six-view ring, not a hero angle. Every GLB gets a 6-view ring at 60 degree steps from one headless Blender scene: identical three-point studio rig, AgX view transform, bounding-sphere framing solved once over all ring angles so scale is constant across engines. Engines disagree about which way "front" is, and a fixed azimuth would silently score some of them on their back.
• One material normalisation, applied to everyone. Principled BSDF metallic forced to 0. glTF defaults metallicFactor to 1.0 and the TRELLIS family omits it, so without this the control renders black. Applying it to all engines equally is the only way the grid stays comparable.
• Single runs, judged by eye — which is why this page calls itself a qualitative review, not a benchmark. Every survivor got one run per subject under a hard spend cap. That is enough to rank who is good at what and to catch structural failures; it is not enough to publish failure rates or variance, and we don't pretend otherwise. If you need quantitative numbers, run your own — the method above is fully described.

The five subjects

All 41 endpoints, accounted for

fal.ai's image-to-3D category listed 41 live endpoints on 2026-08-29, none flagged deprecated. Notably absent from the platform entirely: InstantMesh and Stable Fast 3D, both searched for by name and gone. We screened 18 endpoints and excluded 23; every exclusion has a stated reason, so the whole catalog is accounted for.

Stage A screen: 18 endpoints, one go-kart each

Identical input image, best-practice default parameters, one run each. Price is a relative band within the field on the run date, from ↑ (budget) to ↑↑↑↑ (premium). We don't republish vendors' per-run prices; they change, and their pricing pages are authoritative.

* TRELLIS 1's time is from the original production run; its GLB was reused after verifying the reference was byte-identical, so no wall clock was re-measured.

The 23 excluded endpoints, with reasons

Three billing traps

We don't republish per-run prices — vendors change them, and their pricing pages are the authoritative source. But three structural traps are worth recording, because the vendors' own pages do not lead with them:

Head-to-head: nine engines, five subjects

Each cell scores the engine against the TRELLIS 1 control on that subject. ++ clearly better, + better, = equal, - worse, -- clearly worse, ! a defect that blocks integration regardless of quality. n/r = not run. Price is a relative band within the field on the run date, from ↑ (budget) to ↑↑↑↑ (premium). We don't republish vendors' per-run prices; they change, and their pricing pages are authoritative.

TRELLIS 2 controls were run only on the two character-type subjects (dog, game character), per the brief.

Against the TRELLIS 2 control: on the game character, Rodin v2.5 beats TRELLIS 2 (face likeness, engraved armour, cloak embroidery) and Meshy v7 matches it while adding a separated sword; Hunyuan 3.1 Pro loses to TRELLIS 2 on facial detail. On the dog, TRELLIS 2 and Hunyuan 3.1 Pro are close (TRELLIS 2 keeps more fur texture, Hunyuan builds the better hindquarters), and Rodin v2.5 beats both on the head.

The renders, subject by subject

Each strip is the same six-view Blender ring: same rig, same tone mapping, same framing solve, same material normalisation. A curated selection is shown here, chosen to show each engine at its most characteristic — wins and failures both.

Sneaker: hard matting

Hard matting: laces, thin eyelets, a dark sole against a dark upper.

Go-kart: mechanical detail

Mechanical detail: tubular frame, negative space, an exposed engine.

Dog on a couch: the stress test

A real animal in a lying pose (pose-locked redo). The hardest case in the set.

Game character: stylized figure

A stylized character: armoured ranger, cloak, sword, layered belts.

Crayon dragon: made-real doodle

A made-real doodle: a child's crayon dragon turned into a toy-style reference.

Integration defects and quirks

The things that would cost engineering time, independent of render quality. These do not show up in anyone's marketing pages, but they decide whether an engine can actually ship.

Side experiment: text-to-3D scenes

Outside the tournament, one run of Meshy v7's text-to-3D at ultra + textured (154s) produced a coherent multi-fixture strip mall: five glazed storefront bays, brick pilasters, a parapet roof with individually modelled HVAC units and vent stacks, awnings, and a parking lot with light poles. No reconstructor in this bench can produce anything like it, because a reconstructor is bounded by one photo of one object. Honest caveats: parking stalls came back as raised rails rather than painted lines, the signage is texture smear, and this is a single run with a single prompt, so it establishes a capability, not a failure rate.

Provenance

Run 2026-08-29 against fal.ai's live catalog on that date. Engine names and versions are as listed on the run date; vendors ship new versions frequently, and this page describes these endpoints on this date, not the vendors' products in general. Price arrows are relative bands within the field on that date, from our own per-call spend; we deliberately do not republish per-run vendor prices. Scores are our qualitative judgement of the six-view renders against our own control, made by the team that operates these engines in production; where our engines lost, the page says so.