VixScan
VixScan
A model from a single photo — generated, not scanned
How to photograph the object
Where this stands today
You can attach a photo on the site now — in the generator, on the AI tab, next to the description field. One of the two is enough: either a photo or a description. Before it is even sent, your browser scales the photo down and tells you what it sees on it, so you get the verdict before you spend a token. AI generation itself is not switched on yet; when it is, your photo will be ready — re-shooting an object later is usually harder than shooting it properly the first time.
This is the most important thing on the whole site. The photo decides the result more than anything we do with it afterwards — and we do not remove the background for you, so whatever is on it turns up in the model.
A plain background
White paper, a single-coloured wall, a table top. The less there is behind the object, the less the AI invents into the model. A living room behind the object is the most common reason the result is off.
An object, not a scene
This is the difference between a good result and an unusable one. The AI does not recognise the object — it casts whatever is in the photo. When the object is separated from the background, the cast and the object are almost the same thing and it comes out well. When it stands in a scene, the whole scene gets cast: a fence by a road comes out as a flat slab with a photo on it, a chair in a lit hall as a piece of the room. Better to carry the object to a blank wall than to shoot it where it stands.
Soft, even light
What sits in deep shadow is not visible in the photo — and what is not visible, the model has nowhere to take from. Daylight by a window beats a flash; a hard shadow across the object hides the shape that is supposed to come out.
The whole object in frame
What is not in the photo, the model has nowhere to take from. A cropped edge is not reasoned out — it is patched over. Leave a little room around the object.
One side, not all around
For now we take one photo. The AI invents the back — so pick the side on which the object is easiest to understand, not the prettiest one.
Open the generator and attach a photo
What you get, and what you do not
A likeness, not a copy
The AI makes a model that resembles the photo. Do not take measurements from it — for exact dimensions there is the parametric generator, not this one.
A statue, not a character
The model comes without a skeleton. In an engine it will be a static object — it cannot be bent or animated. We do not offer a skeleton at all yet, so that you cannot pick it and be refused afterwards.
Meshes and thin things
Wire mesh, grilles and wires get scattered into dots — on the AI grid a wire thinner than one cell has nowhere to be recorded. We can repair it, but we need the pitch from you, and before generating, not after.
Not for a printer
The model is made of separate shells, not one closed body. An engine does not mind; a slicer does. For printing use the parametric generators — those build an exact, closed body.
Trust the colour, not the material
The package is a GLB with a colour texture — that much is certain. It usually carries a metallic/roughness map too, but do not trust it: on a shiny object it marks glass and rubber as metal as well, and the model then renders almost black in an engine with no reflections around it. If that happens, set metallic to zero; the colour texture is unharmed. A fine relief (normal map) is sometimes included and sometimes not — do not count on it.
What the pitch is and how to measure it
If the object has wire mesh or a grille, we will need one number from you. It cannot be read back from the finished model — we tried two independent methods and neither found it.
Centre to centre
The pitch is the distance from the centre of one wire to the centre of the next — the clear opening plus the wire. If you measure only the hole between the wires, the number comes out smaller and the mesh is repaired denser than it really is.
Measure ten openings at once
A single opening measures badly — a ruler has millimetres and an opening is a few of them. Measure ten openings and divide by ten; that shrinks the measuring error tenfold.
Why we cannot get it ourselves
The AI scatters the grid into dots, and the pitch simply is not in those dots — we looked for it through neighbour distances and through periodicity, and got numbers unrelated to the truth. Without your number we cannot repair the mesh, and we would rather say so than guess.
How long it takes and what it costs
This is written up front on purpose. Generating takes minutes, not seconds, and it is unpleasant to find that out only while waiting at the screen.
Seven to fifteen minutes
The difference is not in your model but in whether the machine was already working or has to start up first. The first model after a pause is the longer one. The AI itself is over ninety per cent of that — our part is under half a minute.
You pay per generation
Not per format or per size. The same model downloaded in three formats and three sizes costs what one costs — because that is what it costs us. Only one thing is expensive: time on the graphics card.
A failure is not charged
When a task fails, the token comes back automatically. You pay for a result, not for an attempt. And when the result did not come out well, repeating is free — how many times depends on your level.
We ask for the mesh pitch up front
If the object has wire mesh or a grille, we will need the pitch from you — how many millimetres one cell spans. It cannot be read back from the finished model, so it has to be given before generating, not after.
Where it fits
We have tried it in two programs and measured the result. About the others we claim nothing yet — until we test it, it would be a guess.
Unreal Engine 5
Drop the GLB straight into the content browser. It arrives as a static mesh with its texture, in the right unit and not lying on its side — we measured that. Set the size yourself, though: the AI returns the model squeezed into a unit cube, so an object 2.8 m tall reached us as 1.0 m. Unreal switches Nanite on by itself for a dense mesh; that part needs no setting.
Blender
GLB goes straight in, with texture and colour, and at the same size as in Unreal. We measured the same part in both and it came out the same.
What Nanite does not tell you
Alongside the dense model Unreal keeps a simplified fallback — and it is in the low thousands of triangles even when the model has hundreds of thousands. While Nanite runs you never see it. Where it does not (an older machine, mobile, Nanite off), that fallback is what gets drawn — and it will not hold a grille or thin bars. If you are targeting such hardware, plan for it; the import will not tell you.