First multi-light-state scan

ALIVE rests on one claim: if you scan the same miniature in several lighting states, the reconstructions will align well enough to treat them as layers of one scene. If that fails, there is no pipeline — just a stack of incompatible point clouds. Everything else I could engineer around. This one I had to test.

A test miniature — surface detail, self-lit elements, deliberately awkward materials — on the turntable in front of the rig. Three lighting states: key only, practicals only, engine glow only. One camera path covering the full angular range, defined once in world coordinates.

Camera locked manual everything: exposure, white balance, focus. The lighting states are the only variable. That's the experiment.

Three passes, one state per pass, the rig flying the identical path each time. Several hundred frames per pass. The LED controller holds one named state per pass and confirms it over serial before the rig starts — no hand-flipped switches, no 'which state was pass two again'.

Total scan time per state is dominated by exposure, not motion. The dark states want long shutters, and the rig holds still at each station like it's bolted down — which, effectively, it is.

Each pass becomes its own Gaussian splat. Because every pass shares the same camera poses, the splats land in the same coordinate frame by construction — no cross-registration step, no drift between states. The camera solve from pass one is reused for all three.

Render the key-only splat and the practicals-only splat from the same novel viewpoint and flip between them: the geometry doesn't move. Only the light changes. That's the whole trick, working.

Alignment held, fully — the rig's repeatability carries the concept. What needs work: the darkest state is noisy in the shadows, which the splat happily bakes into floaters. Next iteration scans dark states with longer exposures and a denoising pass before training.

But the riskiest assumption in the whole pipeline held. Everything after this is engineering.