Drop a sculpt.
Cast it lighter.

A solid silver idol prices itself before a single hour of labour: weight times today's rate. Studios know this, which is why paper cores and hand-built hollows exist at all.

What no hand method can promise is that the surface you sold is the surface you cast. Hollowing by boolean or by remesh rebuilds the skin. Detail softens. And nobody can tell you by how much, because nothing measured it.

SculptScale carves the cavity and leaves the skin alone. Then it proves it left the skin alone.

Ask for 0.8 mm and you get about 0.8 mm. The voxel carve overshoots by a known amount; that overshoot was measured empirically and is inverted before carving, so the wall lands where you set it rather than wherever the algorithm finished.

Then it is checked on the output. Up to 400,000 points on the cavity face are queried against up to 4,000,000 samples of the skin, and the report prints two numbers: the median wall, and the thinnest 5%. The thinnest 5% is what fails in the kiln. An average has never saved a casting.

It measures the shell it built. It does not inspect your sculpt. SculptScale will not tell you where your uploaded model is thin before hollowing, and above 400,000 faces it skips the thickness check entirely and says so on the report. Thin-spot inspection of your own mesh is your CAD package's job.

Where the piece has a base worth opening, the underside opens along the wall's own edge — wax and resin drain and vent through an opening the shell already has, with nothing drilled through the sculpt.

Where there is no such base, it closes the shell and bores drain holes instead. Each bore is ray-tested to confirm it actually broke through before the report is allowed to claim it. A hole that is drawn but not open is worse than no hole.

“Why can’t I just scale this in Rhino?”

You can. Every package can. That is not the hard part.

The hard part is that scaling a hollow shell scales the wall along with it. Take a 0.8 mm shell to 130% and you are casting 1.04 mm and paying for the extra metal on every piece. Take it to 70% and you are casting 0.56 mm and finding out at the kiln. A scale factor is a ratio. A wall thickness is an absolute manufacturing decision. They do not belong in the same operation, and no uniform transform can keep both.

So SculptScale refuses. In these words:

Upload a PNG, JPG or WEBP and get a bas-relief. Form, ornament and texture depth are three separate dials, each set in millimetres — 2.0, 0.6 and 0.2 mm by default — rather than one “strength” slider you learn by ruining files.

It is deterministic geometry on the CPU. The same drawing and the same three numbers produce the same relief every time, with nothing queued and no model in the loop.

Then it scores how much of your drawing survived and hands back an overlay: green edges made it into the relief, red edges did not. You can see what was lost before you print, which is the only useful time to see it.

The score and the overlay exist only for reliefs SculptScale generated from your image. It cannot score an STL you modelled yourself, and scaling-down detail warnings work the same way — they fire on tool-made relief, never on your own sculpt.

Most of SculptScale is not a learned model, and we would rather say so than let you work it out.

The hollowing, the wall measurement, the surface proof, the sectioning, the scaling and the relief are computational geometry: deterministic, inspectable, the same answer twice. Nothing was trained on your files, or on anyone else's.

One path does use a generative model — turning a single photograph into a rough 3D blockout. That path is beta. Every output it produces carries this stamp: “AI blockout: topology and unseen side are invented — requires sculptor round-trip.” Its size is not read from your photograph; output is normalised to 60 mm and you rescale it yourself. Treat it as a starting mass for a sculptor, never as a model to cast.

Photo detail transfer has the same shape of limit: it decorates the surface facing the camera. The back and the sides get nothing.

One more piece of ordinary arithmetic, since it looks like cleverness: on upload, SculptScale checks whether a different orientation is clearly a better print pose, and rotates only when it clearly is. That is geometry too.

Scaling, hollowing and ring-sizing each require a named human approval before they run. Every step becomes its own immutable version with its lineage recorded, so “which file did we actually cast” still has an answer six months later.

There is no per-seat licence to buy — a studio gets one shared access code, created by us. But approvals are still signed by the person who gave them.

There is no self-signup, and that is deliberate for now. We create each studio's access code by hand, usually the same working day, so that someone has actually looked at what you are trying to make before you spend money on it.

One code per studio — not per person. Your first production export is free, and reports are free forever.