Frequently asked questions

It turns descriptions, photographs and reference views into 3D models, then gives you the tools to finish them — rebuilding the geometry, repainting surfaces, splitting a model into parts, and rigging characters and creatures.

The output is a real mesh you download and use, not a render or a preview. It can also make Gaussian splats — photoreal objects and whole worlds you can fly through, from a picture, several pictures or a video — and 3D Pipelines chain the tools into one recipe you run on one picture or many. See text to 3D and image to 3D for what each input is good at.

To generate something, no. To ship it into a game or a product, some — a generated mesh is a starting point, and the finishing work is skilled work.

The knowledge base exists for exactly this: it explains what each step is for in plain terms, including the ones the studio does for you.

Generate. A verified account gets a one-time credit grant and can generate from a prompt or a single image with Zetta N1, upload its own models, and generate 2D images. Nothing above is time-limited or watermarked.

Seven things need a plan:

• The finishing tools — retopology, segmentation, texturing and rigging. They are the half of the product a subscription buys, on your own uploads as much as on generations; Generate 3D works on every plan.
• Zetta N2.5, Np1 and Np1.5 — the paid engines. N2.5 carries the widest option set; Np1 and Np1.5 are low-poly with 8K textures, and Np1.5 can return quads.
• Multi-view input. Without a plan a generation takes one image; a plan adds reference views, two on the entry tier and all four from Basic up.
• Choosing the polygon count. Every generation without a plan uses the model's own default of 500,000 — you cannot raise or lower it. With a plan you name the number, and how high it can go is the model's own limit rather than your tier's — Ultra detail is the one rung that starts at Basic.
• Low-poly generation and quad meshes, as an output of generation itself.
• Image editing, and with it the AI-derived PBR channels.
• Gaussian splats — which are also in early access, so a paid plan plus access from the Generate Splat tab.

3D Pipelines are open to every account, but each step keeps its tool's own rule: a pipeline step that needs a plan in the Studio needs one in a pipeline too. The pricing page lists tier by tier which is which.

For meshes there are five. Four of them build meshes and one rebuilds them. Gaussian splats have three engines of their own — see the question on splats below.

Zetta N1 takes every input — a prompt on its own, a sketch, one image, or four reference views — and produces geometry-only, textured or low-poly output at a polygon count you choose. It is on every plan, including without one, and it is the only engine that will work from words alone.

Zetta N2.5 has the widest option set: quad topology on any base, a mesh that arrives already split into parts, an Ultra density band, and low-poly that is built rather than decimated. It takes a picture rather than a prompt. Paid plans.

Zetta Np1 is low-poly all the way down — every mesh within a budget you name, up to 20,000 faces — with 8K textures, which only Np1 and Np1.5 write. That combination suits a small mesh carrying its detail in the surface. Paid plans.

Zetta Np1.5 is everything Np1 does plus quad meshes, with a ceiling two and a half times higher — up to 50,000 faces as triangles or 25,000 as quads. Paid plans.

Zetta R1 does not generate at all. It rebuilds the topology of a mesh you already have, to an exact polygon count rather than a band.

The short version: N1 accepts anything including a bare prompt, N2.5 gives you the most control, Np1 is small meshes with large textures, Np1.5 is the same with quads, and R1 is the one you send a finished mesh to. See the feature pages for the capability-by-capability version.

Use N2.5 as the default on a paid plan. It has the widest option set of the four that generate and takes every output option, from one image or from several views.

Use N1 when you are generating from a prompt alone — N2.5 needs a picture, and the studio will make one from your prompt first — or when you have no paid plan. It is the one that accepts every kind of input.

If you are unsure, start with N2.5.

GLB by default. Export, in the Studio's Inspect panel, writes GLB, FBX, OBJ and STL from any model on every plan, for free — FBX for animation packages, STL for a slicer, OBJ for anything that reads it. GLB carries geometry, PBR materials and textures in one file and is what most engines and viewers want — see choosing a format.

Gaussian splats are different files: Zetta Splat delivers a .splat, and worlds come as SPZ, which Inspect can also export as PLY splats, a GLB mesh or a GLB collider. See splat file formats.

Yes. Models you generate are yours to use commercially, including in shipped games, applications and products.

Two caveats worth knowing. First, what you put in matters: generating from a reference image you do not have rights to is a question about your input, not about our output. Second, recognisable trademarks and likenesses carry their own restrictions regardless. See licensing and rights.

Because nothing showed it. Single-image reconstruction infers everything it cannot see — for a symmetrical object that guess is usually good, and for anything with a distinct rear it is invention.

The fix is multi-view: supply a back view and the guess becomes an observation. It needs a paid plan, and every engine that generates takes it.

Text-to-3D writes an image before it builds a mesh, and image models produce letter-shaped marks rather than actual letters. Reconstruction then bakes those marks into the texture.

Generate the object without the lettering and add text as a decal or as separate geometry.

Yes. Upload a GLB and it becomes a library item like any generated one, with every tool available — retopology, segmentation, texturing and rigging.

A splat stores a scene as millions of small, soft, coloured blobs instead of a surface. It looks photographic — foliage, fabric and sheen included — and opens in the browser, but it has no polygons: you cannot retopologise, rig, texture or 3D print it.

Make a splat when you want to look at a place or an object as it is: a walkthrough, a backdrop, a digital-twin view. Make a mesh when the asset has to do something — move, collide, change material, print. A splat world can be exported as a mesh or a collider mesh when you need a surface after all. See splats versus meshes.

There are three. Zetta Splat captures an object or a place from one picture, with its real light and fine detail — choose Object and it removes the background first. Zetta Splat 2 builds a whole world out from a picture, several pictures or a video, to look around and fly through. Zetta Splat 2.1 does the same for larger spaces, big outdoor and indoor areas.

See choosing a splat engine for the side-by-side.

Yes, with Zetta Splat 2 or 2.1 in Environment mode. Attach one slow, steady take walking through the space, up to 100 MB as MP4, MOV or WebM. The more of the space the video shows from more than one angle, the more complete the world — mirrors, glass and people walking through confuse it.

Filming a video for a splat has the details.

Two reasons it can be closed to you. Gaussian splats come with a paid plan, and they are in early access on top of that. Press Request access on the Generate Splat tab and the team switches it on for your account.

Reconstruction density is a by-product of the algorithm rather than a decision, so it bears no relationship to what your project needs.

Retopology is the fix — it keeps the shape and rebuilds the geometry at a density you choose.

3D credits pay for generation and post-processing jobs — each operation has a fixed price, shown before you run it.

Working in 2D costs no credits: generating, editing and composing images are not metered at all. A credit is only spent when you turn a look into 3D.

Not metered is not the same as unrestricted, though. Generating images is open to every account; editing them, the 4K size and the Pro model need a paid plan.

Usually a format conversion. Segmentation, texturing and rigging need FBX, and every generated model is GLB — so those jobs include a conversion step, priced as its own line and included in the quoted total.

The quote you see before running includes it. Nothing is charged that was not shown.

It depends on your plan, and it is counted separately for each kind of job:

• Free and Tester — one job of any kind at a time.
• Basic — two of each kind at once.
• Creator — five of each kind, and three for texturing and conversion.
• Creator Pro — the most we can run: ten each of Zetta N2.5 generations, rigs, and segmentation and retopology; five each of Smart 3D generations and texturing and conversion.

Gaussian splat worlds and their exports count as kinds of their own, at the same levels: two at a time on Basic, five on Creator, ten on Creator Pro.

Zetta N1 runs one job at a time on every plan. Kinds do not hold each other up — two rigs running do not stop you starting a texture.

When your plan is already running as many of a kind as it may, a new one is not taken or charged. The Studio tells you, and shows the plan that would run it.

No. A failed job returns its credits, so a failure costs you nothing. The same goes for a failed step in a 3D Pipeline run.

Step by step. Each step costs what the same tool costs in the Studio and is charged when it is submitted — there is no hold on the whole run. The run shows its estimated total before it starts, and in Auto mode it checks your balance covers the estimate first, so it does not stall halfway.

Building and saving pipelines costs nothing. See Auto or Manual.

Monthly allowances reset with your billing period. Purchased and bonus credits carry over rather than evaporating at the month boundary.

A way to chain the Studio's tools into one recipe — for example generate, retopologise, lay out UVs, texture and rig — and run it on one picture or model, or on up to twenty at a time. Every step is the full Studio tool, with its engines and every setting.

Runs go two ways: Manual waits for you before each step, so you can check a result before paying for the next; Auto runs every step to the end. Runs happen on our servers, so closing the tab does not stop one, and every model a step makes lands in your library grouped with its run. See 3D pipelines, explained.

Remove background, AI edit, Draw views and Isolate Parts for pictures; Generate 3D (Zetta N1 or N2.5) and Generate Smart 3D (Zetta Np1 or Np1.5) to make the model; then retopology, segmentation, AI texturing, Smart UV and auto rigging to refine it.

Not yet in the palette: Gaussian splats and export — download files from the library. There are no ready-made templates yet; every pipeline starts empty. Ordering a pipeline covers the order that works.

It takes one picture — usually concept art of a character — and pulls it apart: the character on its own, re-posed with empty hands (a humanoid goes into a T-pose for rigging), plus a clean picture of each part it carries or wears. Whatever follows it in the pipeline runs once per part, so one run can turn a single picture into a character and a kit of props.

It needs a plan with image editing. See Isolate Parts.

It rebuilds a mesh with cleaner, more even geometry at a density you choose, keeping the shape.

You need it if the model will be animated, subdivided, textured properly, or shipped into a performance-sensitive target. You can skip it for a static render or a quick visualisation.

Almost always because the body plan is wrong. Rigging asks which kind of body it is looking at — humanoid, four legs, bird, snake, fish, six legs or eight legs — and that choice also picks the engine underneath. A quadruped rigged as a humanoid produces joints where the animal does not bend.

The second cause is the pose. A humanoid wants a T-pose: arms straight out to the sides, legs straight, facing forward. An A-pose, with the arms angled down toward the hips, looks perfectly riggable and is not — the joints land in the wrong places. See T-pose versus A-pose.

The studio checks the model against the plan you chose before charging you, and tells you what it found rather than producing a bad skeleton.

Vehicles, props and machines have no body plan here. Those are a segmentation job — split the parts and move them — rather than a rigging one.

Splitting one mesh into separate named parts — a chair into seat, back and legs. Useful for separate materials, independent movement, and selective editing.

One thing to know: the pieces meet exactly where the surface was cut, so there is no hidden geometry behind a join. A segmented drawer has no back face where it met the cabinet.

Almost always missing channels. A colour map alone has no roughness and no normal, and under real lighting that reads as flat plastic.

The other common cause is lighting baked into the colour — if the reference image had strong shadows, those shadows are now painted onto the surface and will fight whatever lighting you put it under.

Yes, on a paid plan. From a base colour map, the studio derives normal, roughness, metalness, emissive and ambient occlusion, each generated at the source texture's own resolution so it lines up pixel for pixel.

Without a plan you can still upload maps you already have — it is generating them that a plan adds.

Worth being direct about:

• Exact specifications. Anything that must match concept art line for line, or hit a dimension.
• Thin repeated structures. Chain links, wire mesh, railings, spokes — these reconstruct as fused masses.
• Interiors. An unseen cavity is an inferred cavity, so mugs come back solid and boxes do not open.
• Text and logos, as above.
• Transparent and reflective objects. What the camera recorded is not the surface.
• Mechanical parts that must fit other parts. Approximation and tolerance do not mix.

See common generation failures for what each one is caused by and whether it is fixable.

It replaces a stage, not a role. Generation occupies concept and high-poly — producing a detailed shape. Everything downstream of that shape is unchanged: retopology, UVs, baking, materials, rigging, LODs, engine integration.

In practice it moves effort from producing geometry to judging and finishing it, which rewards technical-artist skills over modelling throughput. Where AI fits in a pipeline sets out the honest version.

For volume work with low individual stakes — set dressing, background props, kitbash pieces, forty crates that each need to be plausible — it is genuinely competitive, and much faster.

For a hero asset with a fixed specification, a modeller wins, and correcting an approximation toward an exact target can cost more than hitting the target directly.

Knowing which category an asset falls into before you start is most of the skill.

On a paid plan, yes by default: every new model is private to your account, and you publish the ones you want seen. Keeping a model private is a paid feature — models made on a free account are public, so they can appear in the community gallery and anyone with the link can open them.

Either way, files are served through signed, expiring URLs rather than permanent public links. On a paid plan you can change a model's visibility whenever you like.

The record and its files are removed from storage. It is not a soft flag — the files go, and the operation cannot be undone, which is why the studio asks before doing it.

Yes. The same generation and post-processing operations available in the studio are available programmatically, which is what makes catalogue-scale and pipeline-integrated use possible.