Your images stay on your machine.

What a hosted generator decides for you

Meshy, Tripo, Rodin and Kaedim run on their own servers. That is a real architecture with real advantages: you need no graphics card, and the work happens on hardware far beyond a laptop. It is not a worse choice, it is a different arrangement, and it settles three questions before you get to ask them.

• Where your images go They leave your machine. For a personal photograph that may be fine. For a client's unreleased product, a set of reference shots under an agreement, or anything you do not own outright, it is a question you have to answer to someone else.
• What an attempt costs Generation is metered, by credit or by month. That puts a price on trying something twice, which is exactly what this kind of work is made of.
• How long it stays available The tool works while the service does. Pricing changes, plans are retired, companies are acquired. What you learned stays with you, but what you built the workflow on is not yours.

What running it here changes

LocalMesh installs the models once, about 27 GB, and then works with the network switched off. There is no upload step, no credit counter and no queue behind other people's jobs. The only thing between a photograph and a file is your own graphics card.

• One payment $39 once, after a 14 day trial. The key is perpetual: the version you have keeps opening for as long as you like. A year of updates is included, and another year costs $39 if and when you want it.
• Your files, immediately The mesh is written to the folder you chose, in OBJ, GLB or STL. Nothing is held back behind an export tier.
• The engine is open The generation core is published under Apache 2.0. You can read exactly what it does to your photograph, and run it without the application if you would rather.

What it asks of you in return

This page would be worth less if it only listed advantages, so here is the other side.

You need Windows and an NVIDIA card, because everything goes through CUDA. LocalMesh is built for 8 GB of video memory; smaller cards run it with heavier subjects handled one level down. You need about 38 GB of free disk while the models download, and 27 GB afterwards. And a generation takes minutes on a laptop card rather than seconds on a datacentre one: 6 minutes 30 to 8 minutes 30 at the standard level on an RTX 4060 Laptop, four photographs in.

If none of that suits your machine, a hosted service is the right answer and you should use one.

If you are coming from a hosted tool

The trial is 14 days and it is the full program, not a reduced one. The honest way to judge it is to take a subject you already generated elsewhere and run it here, then compare the two files in the software you actually work in rather than in a preview window.

Objects come out with metalness written as zero, which is deliberate: the predicted metalness measures as noise and turns objects chrome in Unreal and Blender. If a file you made elsewhere looked wrong in your engine, that is often why.