nx_gsplat_density.nx

nx_gsplat_density.nx -- ADAPTIVE SPLAT DENSITY (split / clone / prune), rung 4 of the fitting ladder. The fitter (gs_pos_grad_step_aniso) already measures, per primitive, everything this decision needs: gradbuf[g*3], gradbuf[g*3+1] -- the screen-space residual gradient gradbuf[g*3+2], wnorm[g] -- the Gauss-Newton curvature denominators pa/pc[g] -- the conic, hence the screen footprint via gs_splat_rx/ry depth[g] >= 0 -- visible in this view So densification is a CLASSIFIER over state that already exists. It derives no new quantity and re-renders nothing; a second measurement here would be a second ruler. EVERY THRESHOLD IS DERIVED FROM THE POPULATION OR THE FORMAT. NONE IS PICKED. The published 3DGS rule densifies on a view-space gradient above a FIXED tau (0.0002 in the paper). That constant is calibrated to their units, their scene scale and their float pipeline; copying it into an integer fixed-point renderer would be a picked number wearing a citation. So the bar here is the population's OWN MEAN, recomputed every step: DENSIFY iff |gx| + |gy| > mean(|gx| + |gy|) over VISIBLE primitives A mean needs no parameter, adapts to scene scale and to fitting progress by construction, and cannot go stale. L1 is used rather than L2 because this is an ORDERING against a mean of the same measure -- no square root, exact in integers, and monotone in the same direction as L2. SPLIT vs CLONE, also against the population's own mean, and the reasoning is the paper's: footprint > mean footprint -> SPLIT (over-reconstruction: one large splat spanning detail it cannot represent; two smaller ones can) footprint <= mean footprint -> CLONE (under-reconstruction: the region is under-covered, so add a second primitive of the same size) Footprint is the screen half-extent gs_splat_rx(ca) + gs_splat_ry(cc) -- the renderer's OWN support radius, so a splat is judged by exactly the extent it actually paints. PRUNE, derived from the FIXED-POINT FORMAT rather than from the population: a splat contributes c*T*alpha/GFXA to a channel; peak over T<=GFXA and alpha<=op gives c*op/GFXA. With the framebuffer quantised to 8 bits (c <= 255), that peak is below one quantisation step when op * 255 < GFXA Such a primitive CANNOT change any pixel at any depth or colour. It is invisible by the arithmetic of the format, not by a judgement about smallness. A view-dependent prune was CONSIDERED AND REJECTED, and the reason is multi-view correctness. Pruning on zero accumulated footprint in this view, or on depth[g] < 0, would delete primitives that are culled HERE and visible from another camera -- and the fitter is now multi-view (it cycles yaw across steps). A prune rule that is correct per-view and wrong per-scene is the worst kind, because every single-view test passes. Opacity is view-independent, so the shipped rule is.