nx_ng_gen3d.nx

nx_ng_gen3d.nx -- CAP-GEN-3D, generation 12: generative-3D (DreamFusion-class, arxiv 2209.14988). The honest CORE mechanism of text/image-to-3D: a 3D representation is optimized so that its RENDERED 2D VIEWS satisfy a 2D objective (a "prior"), via the DIFFERENTIABLE renderer -- "lifting" 2D supervision into a 3D scene, with no direct 3D supervision. Reuses the splat color-fit on the autograd tape (CAP-GAUSSIAN-SPLAT / CAP-DIFF-RENDER). Honest gated proof of the DEFINING property (why generative-3D needs multiple views / a prior): a 3D scene = N gaussians with shared learnable colors; each VIEW renders as a weighted sum of those shared colors (the gaussian's projected footprint+transmittance in that view). We generate target views from a known scene c*. T1 MULTI-VIEW LIFT -- optimizing colors from TWO 2D views recovers the 3D scene (|c_g - c*_g| small): 2D supervision is distilled into a 3D-consistent representation. T2 SINGLE-VIEW INSUFFICIENT -- view A alone is ill-posed (gaussians 0,1 overlap into one observed pixel -> only their SUM is seen) so c0,c1 are NOT recovered: this is WHY a generative prior / extra views are needed = the load-bearing reason the method exists. T3 BIT-EXACT -- the 2-view solve is deterministic (train twice -> identical integer colors). HONEST: the "prior" here is explicit target views (full DreamFusion = a learned text-to-image diffusion model driving the views via score-distillation = the follow-on); 1D pixels, fixed gaussian geometry. Q16, no float. Sovereign: nx_nofloat_autograd + syscalls. license_tier: ORIGINAL expect_exit: 0