Meshy 7 Scored Its Own AI Against Real Geometry — and Surface Detail Still Stops at 59.8%

Key takeaways

Meshy 7, the company’s new image-to-3D foundation model, went live on 10 August and is open to every Meshy subscription tier — but downloading a generated model still requires the Pro tier or above.
Meshy published the method behind a new geometry alignment benchmark alongside it: generated meshes are scored directly against reference 3D models held out of training, with no human raters and no vision-language model acting as judge.
From a single input image, Meshy reports 81.0% on overall proportion, 79.7% on spatial distribution and 59.8% on surface detail against four competing models, where 100% is a reference scored against itself.
The widest margin — 5.3 points — is on surface detail, the one metric where no model tested cleared 60%.
Alignment measures how closely a mesh matches your picture. It says nothing about whether that mesh is watertight, manifold or printable.

What shipped

Meshy released Meshy 7 on 10 August and announced it publicly two days later, pairing the model with the method behind a geometry alignment benchmark of its own design. The announcement is built around one idea: the 3D result should agree with the image the user brought.

That is a narrower goal than it sounds, and deliberately so. The company argues the first era of image-to-3D is over — broken surfaces, scrambled structure and missing limbs are now rare enough among leading models that not breaking separates nobody. What is left is subtler: a body slightly too wide, a hand shifted out of place, an engraved line from the source image simply gone. Meshy calls that gap alignment, and it is what Meshy 7 was trained and scored against.

Access follows the existing tiers: open across all subscription levels, downloads gated behind Pro or above. Ultra Mode is single-image only for now, with multi-view support described as arriving shortly.

The benchmark is the more interesting release

The scoring method deserves more attention than the model. Instead of asking humans which mesh looks better, or handing the job to a vision-language judge, Meshy renders reference 3D models held out of training into input images from known camera positions. Each system generates a mesh from that image, and the mesh is compared against the original geometry it came from — so every test image carries a known correct answer.

Two design choices carry the weight. Comparing geometry against geometry, rather than renders, avoids estimating the source camera first and folding camera error into the geometry error. And before scoring, each generation is fitted to its reference by translation, rotation and uniform scaling only — never stretched along one axis — so a model’s own proportion errors cannot be quietly corrected away by the harness.

Agreement is then measured at three levels — overall proportion, spatial distribution and surface detail — across a reference set of characters, vehicles, sculptures and thin-structure props, under a front view, a raised top-quarter view, and four views combined.

Where the numbers land

Given one input image, Meshy 7 posted 81.0% on overall proportion, 79.7% on spatial distribution and 59.8% on surface detail, ahead of four competing 3D foundation models on all three measures. Meshy says the pattern holds under front-view input, which it describes as the strictest condition.

The shape of the result matters more than the topline. Overall proportion is bunched — the leading models sit close together, so an edge there means little. Surface detail is not bunched: no model tested reached 60%, and Meshy 7’s 5.3-point margin is more than double its margin on either of the other two metrics. That is what a real lead looks like, on the dimension nobody has solved yet.

Feed the models four views instead of one and the field converges to within roughly two points of one another, with Meshy’s closest competitor catching up rather than overtaking. Single-view input, the company notes fairly, is both where systems separate and the condition most users are actually in.

Three changes are credited for the gain, as 3D Printing Industry reported: an image encoder that reads inputs at multiple scales and higher resolution, training data rebuilt so every sample corresponds exactly to its target geometry with style, lighting and background stripped out, and alignment scoring moved inside the training loop rather than applied at the end.

Why it matters

These are vendor-run numbers on a vendor-designed benchmark, and the usual discount applies. The four competitors are not named, and the harness has not shipped — Meshy says the geometry benchmark will be published separately after launch, with a texture alignment benchmark to follow. Until then, “5.3 points clear” is a claim rather than a result. As Unite.AI put it, publishing the harness is what would turn a vendor claim into a number other labs have to answer. In fairness, the structural choices — held-out references, direct geometric comparison, no per-axis rescaling — are the ones an honest measurement would make.

For anyone who prints, there is a second and larger caveat: alignment is not printability. Every metric here scores how faithfully a mesh reproduces a picture. None of them ask whether the mesh is watertight, whether normals are consistent, whether wall thicknesses survive a slicer, or whether those thin-structure props would stand up coming off a build plate. 3D Printing Industry made the same point, noting the benchmark measures how closely output matches a source image rather than whether it meets physical printing constraints.

That gap is where the rework still lives. A model that holds dense shallow relief together — Meshy’s own example is a jade medallion with a coiled dragon on a nearly flat disc — is genuinely useful, because that detail is exactly what fragments in weaker generators and exactly what a good resin machine can resolve. But between a 59.8% surface-detail score and a finished part sit mesh repair, orientation, supports and a slicer that will happily choke on a non-manifold edge. Meshy 7 improves the first half of that pipeline. It does not remove the second.

FAQ

What does “alignment” mean in image-to-3D generation?

How closely a generated 3D model reproduces the specific image it was built from — correct proportions, parts in the right places, surface detail present at the right depth. It is distinct from whether the model simply looks plausible on its own.

Is Meshy 7 free to use?

Meshy 7 is available across all Meshy subscription tiers, but downloading a generated model requires a Pro-tier subscription or higher.

Does a high alignment score mean the model is ready to print?

No. Alignment measures fidelity to the source image only. Watertightness, manifold geometry, wall thickness and support requirements are not part of the benchmark, so generated models still need checking and repair before slicing.

When will the benchmark be public?

Meshy says the geometry benchmark method will be released separately after launch, with a dedicated texture alignment benchmark to follow. No date has been announced.