Meshy 7 Launches With a New Benchmark for Image-to-3D Accuracy

Meshy 7 went live on August 10 and the headline feature is not just better geometry, but a measurable way to prove it. The new release ships with what Meshy calls an image-3D alignment benchmark: a standardised scoring method that quantifies how closely a generated 3D asset matches the image it was built from. Until now, that comparison was largely subjective. A number changes the conversation.

What Changed Under the Hood

The underlying foundation model produces higher-fidelity geometry than Meshy 6, with cleaner surface detail and improved PBR materials. Metallic, roughness, and normal maps are all sharper. The release also introduces an ultra_mode parameter for single-view generation that pushes surface detail further, with multi-view support promised in a near-term update. Multi-view generation now conditions on all input views simultaneously and drives texturing from the combined data, which should reduce the guesswork that plagues single-image generation.

The Benchmark Angle

Geometry alignment is only part of the picture. Meshy says colour, material, and pattern alignment each need their own benchmarks, and a dedicated texture alignment benchmark will follow. That sequencing makes sense: geometry is the foundation, and the texture benchmark is harder to define objectively. If Meshy can establish credible open metrics, it gives buyers a way to compare against competing image-to-3D services instead of relying on marketing renders.

Access and Pricing

Meshy 7 is available across all subscription tiers as of August 10. Downloading generated models still requires a Pro subscription or above, which is the same gate as the previous version. The API also accepts meshy-7 as the model identifier for image-to-3D, multi-image-to-3D, and retexture tasks, so developers can integrate the new model immediately.

Why It Matters for Makers

Image-to-3D has been useful for concepting but unreliable for printing. Generated models often have non-manifold geometry, flipped normals, or base meshes that do not reflect the source image accurately enough to print without heavy manual cleanup. A model that scores well on an alignment benchmark is more likely to print cleanly on the first attempt. That is the practical test this release is aiming for.