Today we are launching Zetta N1, our neural 3D generation model, together with the studio built around it.
N1 turns a description, a photograph, or several reference views into an actual 3D mesh — a file you download and use, not a render or a turntable. And because a generated mesh is a starting point rather than a shipping asset, everything needed to finish one is in the same place: rebuilding the geometry, laying out UVs, repainting surfaces, splitting a model into parts, and rigging humanoid characters.
Three ways in
Text to 3D. Describe an object and N1 produces it. Under the hood a prompt becomes a reference image and the image becomes a mesh, which is worth knowing because it explains what prompts are good at: single objects with a clear silhouette, materials, stylised work, and forty variations of an idea when only one has to survive.
Image to 3D. One photograph, one mesh. The most predictable route, because you are not asking a model to invent the object first — only to work out its shape from a picture you chose.
Multi-view to 3D. Several angles of the same object, so the far side stops being guesswork. A back view is the single most valuable thing you can add; it converts the largest inferred region into an observed one.
What comes after the mesh
This is the half that usually gets left out of a launch, and it is the half that decides whether an asset ships.
Retopology. Reconstruction produces dense, irregular geometry — correct in shape, badly built. Retopology keeps the shape and rebuilds it at a density you choose, in quads or triangles, at low, medium or high detail. This is what makes a generated model animatable, subdividable and affordable.
UV unwrapping. Generated UV layouts are a by-product rather than a design: fragmented, unevenly scaled, seamed wherever the algorithm found convenient. Unwrapping gives you a layout you can actually paint on, with an interactive view that highlights an island on the model as you hover it.
AI texturing. Repaint a surface without touching the shape, driven either by a description or by a reference image. The flow previews before it spends: the prompt edits a picture you can see — the model's own reference, or a frame of the mesh itself — and only the image you approve is applied.
PBR channel generation. From a base colour map, derive normal, roughness, metalness, emissive and ambient occlusion, each generated at the source texture's own dimensions so it lines up pixel for pixel. A colour map alone reads as flat plastic under real lighting; these are what make a surface read as a material.
AI segmentation. Split one mesh into separate named parts — a chair into seat, back and legs — each downloadable on its own, or together as a single file with the pieces kept separate under one parent.
AI rigging. A skeleton and skin weights for humanoid characters, with a pre-flight check that looks at your model before anything is spent and tells you plainly if it is not riggable.
The pre-flight check
That last point is worth dwelling on, because it is the kind of thing that usually only shows up in a support ticket.
Automatic rigging works on one shape: a single humanoid figure in a T-pose. An A-pose — arms angled down toward the hips — looks perfectly riggable and is not; the joints land in the wrong places and the shoulders deform badly. Vehicles, quadrupeds and props have no standard skeleton to match against at all.
So before a rigging job runs, the studio photographs your model from the viewer and has it looked at. If it is not a humanoid, or not in a T-pose, you get a sentence explaining why in plain terms — "the arms are angled down in an A-pose rather than straight out to the sides" — before any credits are spent. You can override it, because vision misjudges silhouettes and a false negative should never block a model that would have rigged fine.
Bring your own models
The tools are not limited to what N1 made. Upload a GLB and it becomes a library item like any other, with every operation available. That matters if you already have a mesh that needs cleaner topology, a UV layout, or a skeleton — the generator is one way in, not the only one.
Formats and the chain
Models come out as GLB by default, with OBJ and FBX available. Several operations need FBX and every generated model is GLB, so those jobs include a format conversion — planned by the server, priced as its own line, and included in the total you see before you run anything.
What N1 is not good at
A launch post that only lists strengths is a brochure. The honest list:
- Text and lettering. The image step writes letter-shaped marks rather than letters, and reconstruction bakes them in. Generate the object without text and add it as a decal.
- Thin repeated structures. Chain links, wire mesh, railings, spokes — these come back fused. There is no reliable fix.
- Interiors. An unseen cavity is an inferred cavity. Mugs come back solid.
- Exact dimensions. Nothing in the pipeline measures anything. Anything that must fit something else is a CAD job.
- Transparent and reflective objects. What the camera recorded is not the surface.
We would rather you knew this before your first generation than after your fifth.
Where it fits
N1 occupies concept and high-poly — it produces a detailed shape, which is what a sculpt or a concept-plus-model process would have produced. Everything downstream is unchanged: retopology, UVs, baking, materials, rigging, LODs, engine integration.
The front of the pipeline compresses; the back does not. That makes it strongest on volume work with low individual stakes — set dressing, background props, kitbash pieces, prototyping — and weakest on assets with a fixed specification, where correcting an approximation toward an exact target can cost more than hitting it directly.
Zetta P1
N1 is the neural half. Zetta P1, our procedural model, is built on top of Claude and approaches the problem differently: rather than reconstructing a shape from pixels, it reasons about structure and generates geometry as a set of operations. That suits things with rules — architecture, modular kits, mechanical parts, anything where exact dimensions and repeatable variation matter more than organic form.
The short version: N1 is better at "make me this object", P1 is better at "make me forty of these, and make them fit together." More on P1 soon.
Getting started
The studio is where everything above lives. The knowledge base explains each step in plain terms — including the ones the studio does for you — and the FAQ answers what it costs and what it cannot do.