← Blog
2026-07-18 Deep Dive Forge Text to 3D

Inside the Forge: How a Prompt Becomes a 3D Model

You type a weathered brass diving helmet at three.ws/forge, and about half a minute later you are orbiting a real, textured 3D model you can download, view in AR, or drop into a scene. No account, no API key, no cost. This post explains what actually happens in between.

Try the Forge →

Step 1: your words become a picture

For a text prompt, an image model first paints your description as a clean reference view of the object, the way a concept artist sketches before a sculptor starts. If you uploaded photos instead, this step is skipped: your photos are the reference. Sketches skip it too and drive the geometry directly.

Step 2: the picture becomes a shape

A 3D reconstruction model looks at the reference and infers the full object: the sides it can see and the sides it has to imagine. Out comes real geometry with color and texture baked on, delivered as a GLB, the open standard that loads in Blender, Unity, Unreal, three.js, and every major 3D tool.

Step 3: the best available engine does the work

Here is the part most tools get wrong. Every other text-to-3D product wires up one model and hopes it suits your prompt. The Forge runs a grid of engines: Microsoft's TRELLIS on NVIDIA's cloud, Tencent's Hunyuan3D and TRELLIS on our own GPUs, community models on Hugging Face, and premium engines like Meshy and Tripo if you bring your own key.

Before your job starts, the Forge checks which engines are healthy right now and picks the best free one for your kind of prompt. Machines and vehicles route to the engine that is best at hard surfaces; people and creatures route to the one that is best at organic shapes. Free lanes always come first: a paid engine never runs while a healthy free one can serve you.

Step 4: if an engine stumbles, your job quietly moves

GPU services go down, queues fill, clouds have bad days. When that happens mid-generation, the Forge re-submits your job to the next engine in line automatically. You keep watching the same progress bar; you never find out it happened. Only if every lane fails do you see an error, and even then it arrives with one-click retry options instead of a dead end.

Step 5: the model is saved somewhere permanent

Raw engine output lands on temporary storage that expires within an hour, so the Forge immediately copies your model to a permanent CDN. That link keeps working forever. The generation is also recorded in your gallery, so closing the tab mid-job loses nothing: the platform finishes it in the background and it is waiting when you come back.

What the quality tiers mean

The honest part: draft and standard usually cost us nothing to produce, because they run on free lanes and our own GPUs. We charge for the top tier because it is genuinely better, not to pass a vendor bill along.

After the model exists

A finished model is a starting point, not a dead end:

Agents use the exact same machinery

Everything above is also a tool call. AI agents generate models through our free MCP server or plain HTTP, and pay per call in USDC over x402 when they want more, no account needed. The page you use and the API an agent uses are the same pipeline underneath.

Go deeper

← All posts