Today we are adding a second generation engine. Zetta N2.5 sits beside Zetta N1 rather than replacing it, and the difference between them is not "newer" — it is a different trade.
N1 is the one that accepts anything: a bare prompt, a sketch, a photograph, or four reference views. N2.5 wants a picture. What it gives back for that narrower door is control — more of it than any other engine we run.
Quads, on any base
Quadrilateral topology is usually something you get only on a low-poly rung, if at all. Here it is a top-level choice, which means you can ask for it on a textured, high-resolution mesh.
That matters because quads are what subdivision and deformation need. A character that will be rigged, a surface that will be sculpted further, a panel that will be bevelled — all of them want edge loops rather than a triangle soup, and asking at generation time is far cheaper than rebuilding afterwards.
A mesh that arrives in parts
Ask for parts and a chair comes back as seat, back and legs, rather than as one shell you then have to cut. Each piece is its own mesh with its own material slot.
This is not the same as segmenting a finished model. Segmentation divides a surface that already exists; this builds the divisions in. The difference shows the moment something has to move independently — a lid, a drawer, a wheel.
Parts are only available on the geometry-only base, so a parted model arrives untextured. That is a real constraint rather than an oversight: texture the pieces afterwards, or pick a different base.
An Ultra band, and low-poly that was built
Two controls at opposite ends of the same axis.
Ultra is a second geometry quality above the standard one, for a subject whose fine structure is the point. It buys resolution, not a different shape — worth paying for on a hero asset and wasted on a background prop.
Low-poly here is built rather than reduced. The engine designs geometry to be small, roughly between a thousand and twenty thousand faces, instead of building freely and decimating until the silhouette breaks. Anyone who has run a decimator over a generated mesh knows the difference.
Where this adds up: product catalogues
A catalogue is, structurally, exactly one good photograph per product — plain background, even lighting, front on. That is the input N2.5 is strongest on, and it is an asset you already paid a photographer for.
What the option set adds is that the output can be shaped to the job rather than accepted as it comes. A rotatable viewer on a listing page wants a modest polygon count and good textures. An AR placement wants a small file. A configurator wants the product in parts, so a customer can change the legs without regenerating the chair. Three jobs, one photograph, three sets of options.
Two limits apply here in particular. The back of the product is inferred unless you supply it — for a listing where customers rotate all the way round, use the four view slots. And a generated mesh is not a measurement: if somebody will judge whether an object fits a space, set the dimensions deliberately rather than trusting a reconstruction.
What it gives up
- No bare prompt. A description becomes a reference image first and that picture is submitted. Usually that is what you want — the image is kept and reusable — but the picture is what gets reconstructed.
- No sketch mode. A line drawing is read as a photograph of a line drawing. That is N1's job.
- Parts and textures are exclusive. Parts live on the geometry-only base.
- Ultra and HD cost more, and neither changes the shape the engine chose.
- It is still a reconstruction. Better geometry, not production topology — retopology still applies.
Which one to use
N2.5 when you have a photograph and want control over what comes back. N1 when you have words, a sketch, or you are on the free tier.
And when a subject reconstructs badly, try the other. Quality varies far more by subject than by engine, so a shape that defeats one is frequently handled cleanly by another. Switching is a dropdown, not a new workflow.
N2.5 is in the 3D Studio now. The feature page has the capability-by-capability version.